Checkpoint-restart (c/r): fixes a couple of bugs and a DoS issue
(tested against v2.6.28-rc3).
We'd like these to make it into -mm. This version addresses the
last of the known bugs. Please pull at least the first 11 patches,
as they are similar to before.
Patches 1-11 are stable, providing self- and external- c/r of a
single process.
Patches 12 and 13 are newer, adding support for c/r of multiple
processes.
The git tree tracking v10, branch 'ckpt-v10' (and older versions):
git://git.ncl.cs.columbia.edu/pub/git/linux-cr.git
Restarting multiple processes requires 'mktree' userspace tool:
git://git.ncl.cs.columbia.edu/pub/git/user-cr.git
Oren.
--
Why do we want it? It allows containers to be moved between physical
machines' kernels in the same way that VMWare can move VMs between
physical machines' hypervisors. There are currently at least two
out-of-tree implementations of this in the commercial world (IBM's
Metacluster and Parallels' OpenVZ/Virtuozzo) and several in the academic
world like Zap.
Why do we need it in mainline now? Because we already have plenty of
out-of-tree ones, and want to know what an in-tree one will be like. :)
What *I* want right now is the extra review and scrutiny that comes with
a mainline submission to make sure we're not going in a direction
contrary to the community.
This only supports pretty simple apps. But, I trust Ingo when he says:
>> > > Generally, if something works for simple apps already (in a robust,
>> > > compatible and supportable way) and users find it "very cool", then
>> > > support for more complex apps is not far in the future. but if you
>> > > want to support more complex apps straight away, it takes forever and
>> > > gets ugly.
We're *certainly* going to be changing the ABI (which is the format of
the checkpoint). I'd like to follow the model that we used for
ext4-dev, which is to make it very clear that this is a development-only
feature for now. Perhaps we do that by making the interface only
available through debugfs or something similar for now. Or, reserving
the syscall numbers but require some runtime switch to be thrown before
they can be used. I'm open to suggestions here.
--
--
Todo:
- Add support for x86-64 and improve ABI
- Refine or change syscall interface
- Handle multiple namespaces in a container (e.g. save the filesystem
namespaces state with the file descriptors)
- Security (without CAPS_SYS_ADMIN files restore may fail)
Changelog:
[2008-Nov-26] v10:
- Grab vfs root of container init, rather than current process
- Acquire dcache_lock around call to __d_path() in cr_fill_name()
- Force end-of-string in cr_read_string() (fix possible DoS)
- Introduce cr_write_buffer(), cr_read_buffer() and cr_read_buf_type()
[2008-Nov-10] v9:
- Support multiple processes c/r
- Extend checkpoint header with archtiecture dependent header
- Misc bug fixes (see individual changelogs)
- Rebase to v2.6.28-rc3.
[2008-Oct-29] v8:
- Support "external" checkpoint
- Include Dave Hansen's 'deny-checkpoint' patch
- Split docs in Documentation/checkpoint/..., and improve contents
[2008-Oct-17] v7:
- Fix save/restore state of FPU
- Fix argument given to kunmap_atomic() in memory dump/restore
[2008-Oct-07] v6:
- Balance all calls to cr_hbuf_get() with matching cr_hbuf_put()
(even though it's not really needed)
- Add assumptions and what's-missing to documentation
- Misc fixes and cleanups
[2008-Sep-11] v5:
- Config is now 'def_bool n' by default
- Improve memory dump/restore code (following Dave Hansen's comments)
- Change dump format (and code) to allow chunks of <vaddrs, pages>
instead of one long list of each
- Fix use of follow_page() to avoid faulting in non-present pages
- Memory restore now maps user pages explicitly to copy data into them,
instead of reading directly to user space; got rid of mprotect_fixup()
- Remove preempt_disable() when restoring debug registers
- Rename headers files s/ckpt/checkpoint/
- Fix misc bugs in files dump/restore
- Fixes and cleanups on some error paths
- Fix misc coding style
[2008-Sep-09] v4:
- Various fixes and clean-ups
- Fix calculation of hash table size
- Fix header structure alignment
- Use stand list_... for cr_pgarr
[2008-Aug-29] v3:
- Various fixes and clean-ups
- Use standard hlist_... for hash table
- Better use of standard kmalloc/kfree
[2008-Aug-20] v2:
- Added Dump and restore of open files (regular and directories)
- Added basic handling of shared objects, and improve handling of
'parent tag' concept
- Added documentation
- Improved ABI, 64bit padding for image data
- Improved locking when saving/restoring memory
- Added UTS information to header (release, version, machine)
- Cleanup extraction of filename from a file pointer
- Refactor to allow easier reviewing
- Remove requirement for CAPS_SYS_ADMIN until we come up with a
security policy (this means that file restore may fail)
- Other cleanup and response to comments for v1
[2008-Jul-29] v1:
- Initial version: support a single task with address space of only
private anonymous or file-mapped VMAs; syscalls ignore pid/crid
argument and act on current process.
--
At the containers mini-conference before OLS, the consensus among
all the stakeholders was that doing checkpoint/restart in the kernel
as much as possible was the best approach. With this approach, the
kernel will export a relatively opaque 'blob' of data to userspace
which can then be handed to the new kernel at restore time.
This is different than what had been proposed before, which was
that a userspace application would be responsible for collecting
all of this data. We were also planning on adding lots of new,
little kernel interfaces for all of the things that needed
checkpointing. This unites those into a single, grand interface.
The 'blob' will contain copies of select portions of kernel
structures such as vmas and mm_structs. It will also contain
copies of the actual memory that the process uses. Any changes
in this blob's format between kernel revisions can be handled by
an in-userspace conversion program.
This is a similar approach to virtually all of the commercial
checkpoint/restart products out there, as well as the research
project Zap.
These patches basically serialize internel kernel state and write
it out to a file descriptor. The checkpoint and restore are done
with two new system calls: sys_checkpoint and sys_restart.
In this incarnation, they can only work checkpoint and restore a
single task. The task's address space may consist of only private,
simple vma's - anonymous or file-mapped. The open files may consist
of only simple files and directories.
--
Restarting of multiple processes expects all restarting tasks to call
sys_restart(). Once inside the system call, each task will restart
itself at the same order that they were saved. The internals of the
syscall will take care of in-kernel synchronization bewteen tasks.
This patch does _not_ create the task tree in the kernel. Instead it
assumes that all tasks are created in some way and then invoke the
restart syscall. You can use the userspace mktree.c program to do
that.
The init task (*) has a special role: it allocates the restart context
(ctx), and coordinates the operation. In particular, it first waits
until all participating tasks enter the kernel, and provides them the
common restart context. Once everyone in ready, it begins to restart
itself.
In contrast, the other tasks enter the kernel, locate the init task (*)
and grab its restart context, and then wait for their turn to restore.
When a task (init or not) completes its restart, it hands the control
over to the next in line, by waking that task.
An array of pids (the one saved during the checkpoint) is used to
synchronize the operation. The first task in the array is the init
task (*). The restart context (ctx) maintain a "current position" in
the array, which indicates which task is currently active. Once the
currently active task completes its own restart, it increments that
position and wakes up the next task.
Restart assumes that userspace provides meaningful data, otherwise
it's garbage-in-garbage-out. In this case, the syscall may block
indefinitely, but in TASK_INTERRUPTIBLE, so the user can ctrl-c or
otherwise kill the stray restarting tasks.
In terms of security, restart runs as the user the invokes it, so it
will not allow a user to do more than is otherwise permitted by the
usual system semantics and policy.
Currently we ignore threads and zombies, as well as session ids.
Add support for multiple processes
(*) For containers, restart should be called inside a fresh container
by the init task of that container. However, it is also possible to
restart applications not necessarily inside a container, and without
restoring the original pids of the processes (that is, provided that
the application can tolerate such behavior). This is useful to allow
multi-process restart of tasks not isolated inside a container, and
also for debugging.
Signed-off-by: Oren Laadan <[email protected]>
---
checkpoint/restart.c | 214 +++++++++++++++++++++++++++++++++++++++++++-
checkpoint/sys.c | 34 ++++++--
include/linux/checkpoint.h | 23 ++++-
include/linux/sched.h | 1 +
4 files changed, 258 insertions(+), 14 deletions(-)
diff --git a/checkpoint/restart.c b/checkpoint/restart.c
index f4f737d..24392ee 100644
--- a/checkpoint/restart.c
+++ b/checkpoint/restart.c
@@ -10,6 +10,7 @@
#include <linux/version.h>
#include <linux/sched.h>
+#include <linux/wait.h>
#include <linux/file.h>
#include <linux/magic.h>
#include <linux/checkpoint.h>
@@ -276,30 +277,235 @@ static int cr_read_task(struct cr_ctx *ctx)
return ret;
}
+/* cr_read_tree - read the tasks tree into the checkpoint context */
+static int cr_read_tree(struct cr_ctx *ctx)
+{
+ struct cr_hdr_tree *hh = cr_hbuf_get(ctx, sizeof(*hh));
+ int parent, size, ret = -EINVAL;
+
+ parent = cr_read_obj_type(ctx, hh, sizeof(*hh), CR_HDR_TREE);
+ if (parent < 0) {
+ ret = parent;
+ goto out;
+ } else if (parent != 0)
+ goto out;
+
+ if (hh->tasks_nr < 0)
+ goto out;
+
+ ctx->pids_nr = hh->tasks_nr;
+ size = sizeof(*ctx->pids_arr) * ctx->pids_nr;
+ if (size < 0) /* overflow ? */
+ goto out;
+
+ ctx->pids_arr = kmalloc(size, GFP_KERNEL);
+ if (!ctx->pids_arr) {
+ ret = -ENOMEM;
+ goto out;
+ }
+ ret = cr_kread(ctx, ctx->pids_arr, size);
+ out:
+ cr_hbuf_put(ctx, sizeof(*hh));
+ return ret;
+}
+
+static int cr_wait_task(struct cr_ctx *ctx)
+{
+ pid_t pid = task_pid_vnr(current);
+
+ cr_debug("pid %d waiting\n", pid);
+ return wait_event_interruptible(ctx->waitq, ctx->pids_active == pid);
+}
+
+static int cr_next_task(struct cr_ctx *ctx)
+{
+ struct task_struct *tsk;
+
+ ctx->pids_pos++;
+
+ cr_debug("pids_pos %d %d\n", ctx->pids_pos, ctx->pids_nr);
+ if (ctx->pids_pos == ctx->pids_nr) {
+ complete(&ctx->complete);
+ return 0;
+ }
+
+ ctx->pids_active = ctx->pids_arr[ctx->pids_pos].vpid;
+
+ cr_debug("pids_next %d\n", ctx->pids_active);
+
+ rcu_read_lock();
+ tsk = find_task_by_pid_ns(ctx->pids_active, ctx->root_nsproxy->pid_ns);
+ if (tsk)
+ wake_up_process(tsk);
+ rcu_read_unlock();
+
+ if (!tsk) {
+ ctx->pids_err = -ESRCH;
+ complete(&ctx->complete);
+ return -ESRCH;
+ }
+
+ return 0;
+}
+
+/* FIXME: this should be per container */
+DECLARE_WAIT_QUEUE_HEAD(cr_restart_waitq);
+
+static int do_restart_task(struct cr_ctx *ctx, pid_t pid)
+{
+ struct task_struct *root_task;
+ int ret;
+
+ rcu_read_lock();
+ root_task = find_task_by_pid_ns(pid, current->nsproxy->pid_ns);
+ if (root_task)
+ get_task_struct(root_task);
+ rcu_read_unlock();
+
+ if (!root_task)
+ return -EINVAL;
+
+ /*
+ * wait for container init to initialize the restart context, then
+ * grab a reference to that context, and if we're the last task to
+ * do it, notify the container init.
+ */
+ ret = wait_event_interruptible(cr_restart_waitq,
+ root_task->checkpoint_ctx);
+ if (ret < 0)
+ goto out;
+
+ task_lock(root_task);
+ ctx = root_task->checkpoint_ctx;
+ if (ctx)
+ cr_ctx_get(ctx);
+ task_unlock(root_task);
+
+ if (!ctx) {
+ ret = -EAGAIN;
+ goto out;
+ }
+
+ if (atomic_dec_and_test(&ctx->tasks_count))
+ complete(&ctx->complete);
+
+ /* wait for our turn, do the restore, and tell next task in line */
+ ret = cr_wait_task(ctx);
+ if (ret < 0)
+ goto out;
+ ret = cr_read_task(ctx);
+ if (ret < 0)
+ goto out;
+ ret = cr_next_task(ctx);
+
+ out:
+ cr_ctx_put(ctx);
+ put_task_struct(root_task);
+ return ret;
+}
+
+static int cr_wait_all_tasks_start(struct cr_ctx *ctx)
+{
+ int ret;
+
+ if (ctx->pids_nr == 1)
+ return 0;
+
+ init_completion(&ctx->complete);
+ current->checkpoint_ctx = ctx;
+
+ wake_up_all(&cr_restart_waitq);
+
+ ret = wait_for_completion_interruptible(&ctx->complete);
+ if (ret < 0)
+ return ret;
+
+ task_lock(current);
+ current->checkpoint_ctx = NULL;
+ task_unlock(current);
+
+ return 0;
+}
+
+static int cr_wait_all_tasks_finish(struct cr_ctx *ctx)
+{
+ int ret;
+
+ if (ctx->pids_nr == 1)
+ return 0;
+
+ init_completion(&ctx->complete);
+
+ ret = cr_next_task(ctx);
+ if (ret < 0)
+ return ret;
+
+ ret = wait_for_completion_interruptible(&ctx->complete);
+ if (ret < 0)
+ return ret;
+
+ return 0;
+}
+
/* setup restart-specific parts of ctx */
static int cr_ctx_restart(struct cr_ctx *ctx, pid_t pid)
{
+ ctx->root_pid = pid;
+ ctx->root_task = current;
+ ctx->root_nsproxy = current->nsproxy;
+
+ get_task_struct(ctx->root_task);
+ get_nsproxy(ctx->root_nsproxy);
+
+ atomic_set(&ctx->tasks_count, ctx->pids_nr - 1);
+
return 0;
}
-int do_restart(struct cr_ctx *ctx, pid_t pid)
+static int do_restart_root(struct cr_ctx *ctx, pid_t pid)
{
int ret;
+ ret = cr_read_head(ctx);
+ if (ret < 0)
+ goto out;
+ ret = cr_read_tree(ctx);
+ if (ret < 0)
+ goto out;
+
ret = cr_ctx_restart(ctx, pid);
if (ret < 0)
goto out;
- ret = cr_read_head(ctx);
+
+ /* wait for all other tasks to enter do_restart_task() */
+ ret = cr_wait_all_tasks_start(ctx);
if (ret < 0)
goto out;
+
ret = cr_read_task(ctx);
if (ret < 0)
goto out;
- ret = cr_read_tail(ctx);
+
+ /* wait for all other tasks to complete do_restart_task() */
+ ret = cr_wait_all_tasks_finish(ctx);
if (ret < 0)
goto out;
- /* on success, adjust the return value if needed [TODO] */
+ ret = cr_read_tail(ctx);
+
out:
return ret;
}
+
+int do_restart(struct cr_ctx *ctx, pid_t pid)
+{
+ int ret;
+
+ if (ctx)
+ ret = do_restart_root(ctx, pid);
+ else
+ ret = do_restart_task(ctx, pid);
+
+ /* on success, adjust the return value if needed [TODO] */
+ return ret;
+}
diff --git a/checkpoint/sys.c b/checkpoint/sys.c
index fbaf683..7745500 100644
--- a/checkpoint/sys.c
+++ b/checkpoint/sys.c
@@ -145,6 +145,8 @@ static void cr_task_arr_free(struct cr_ctx *ctx)
static void cr_ctx_free(struct cr_ctx *ctx)
{
+ BUG_ON(atomic_read(&ctx->refcount));
+
if (ctx->file)
fput(ctx->file);
@@ -164,6 +166,8 @@ static void cr_ctx_free(struct cr_ctx *ctx)
if (ctx->root_task)
put_task_struct(ctx->root_task);
+ kfree(ctx->pids_arr);
+
kfree(ctx);
}
@@ -178,7 +182,9 @@ static struct cr_ctx *cr_ctx_alloc(int fd, unsigned long flags)
ctx->flags = flags;
+ atomic_set(&ctx->refcount, 0);
INIT_LIST_HEAD(&ctx->pgarr_list);
+ init_waitqueue_head(&ctx->waitq);
err = -EBADF;
ctx->file = fget(fd);
@@ -193,6 +199,7 @@ static struct cr_ctx *cr_ctx_alloc(int fd, unsigned long flags)
if (cr_objhash_alloc(ctx) < 0)
goto err;
+ atomic_inc(&ctx->refcount);
return ctx;
err:
@@ -200,6 +207,17 @@ static struct cr_ctx *cr_ctx_alloc(int fd, unsigned long flags)
return ERR_PTR(err);
}
+void cr_ctx_get(struct cr_ctx *ctx)
+{
+ atomic_inc(&ctx->refcount);
+}
+
+void cr_ctx_put(struct cr_ctx *ctx)
+{
+ if (ctx && atomic_dec_and_test(&ctx->refcount))
+ cr_ctx_free(ctx);
+}
+
/**
* sys_checkpoint - checkpoint a container
* @pid: pid of the container init(1) process
@@ -227,7 +245,7 @@ asmlinkage long sys_checkpoint(pid_t pid, int fd, unsigned long flags)
if (!ret)
ret = ctx->crid;
- cr_ctx_free(ctx);
+ cr_ctx_put(ctx);
return ret;
}
@@ -242,7 +260,7 @@ asmlinkage long sys_checkpoint(pid_t pid, int fd, unsigned long flags)
*/
asmlinkage long sys_restart(int crid, int fd, unsigned long flags)
{
- struct cr_ctx *ctx;
+ struct cr_ctx *ctx = NULL;
pid_t pid;
int ret;
@@ -250,15 +268,17 @@ asmlinkage long sys_restart(int crid, int fd, unsigned long flags)
if (flags)
return -EINVAL;
- ctx = cr_ctx_alloc(fd, flags | CR_CTX_RSTR);
- if (IS_ERR(ctx))
- return PTR_ERR(ctx);
-
/* FIXME: for now, we use 'crid' as a pid */
pid = (pid_t) crid;
+ if (pid == task_pid_vnr(current))
+ ctx = cr_ctx_alloc(fd, flags | CR_CTX_RSTR);
+
+ if (IS_ERR(ctx))
+ return PTR_ERR(ctx);
+
ret = do_restart(ctx, pid);
- cr_ctx_free(ctx);
+ cr_ctx_put(ctx);
return ret;
}
diff --git a/include/linux/checkpoint.h b/include/linux/checkpoint.h
index ac330f5..b9ce707 100644
--- a/include/linux/checkpoint.h
+++ b/include/linux/checkpoint.h
@@ -13,10 +13,11 @@
#include <linux/fs.h>
#include <linux/path.h>
#include <linux/sched.h>
+#include <asm/atomic.h>
#ifdef CONFIG_CHECKPOINT_RESTART
-#define CR_VERSION 2
+#define CR_VERSION 3
struct cr_ctx {
int crid; /* unique checkpoint id */
@@ -34,14 +35,27 @@ struct cr_ctx {
void *hbuf; /* temporary buffer for headers */
int hpos; /* position in headers buffer */
- struct task_struct **tasks_arr; /* array of all tasks in container */
- int tasks_nr; /* size of tasks array */
+ atomic_t refcount;
struct cr_objhash *objhash; /* hash for shared objects */
struct list_head pgarr_list; /* page array to dump VMA contents */
struct path *vfsroot; /* container root (FIXME) */
+
+ /* [multi-process checkpoint] */
+ struct task_struct **tasks_arr; /* array of all tasks [checkpoint] */
+ int tasks_nr; /* size of tasks array */
+
+ /* [multi-process restart] */
+ struct cr_hdr_pids *pids_arr; /* array of all pids [restart] */
+ int pids_nr; /* size of pids array */
+ int pids_pos; /* position pids array */
+ int pids_err; /* error occured ? */
+ pid_t pids_active; /* pid of (next) active task */
+ atomic_t tasks_count; /* sync of restarting tasks */
+ struct completion complete; /* sync of restarting tasks */
+ wait_queue_head_t waitq; /* sync of restarting tasks */
};
/* cr_ctx: flags */
@@ -54,6 +68,9 @@ extern int cr_kread(struct cr_ctx *ctx, void *buf, int count);
extern void *cr_hbuf_get(struct cr_ctx *ctx, int n);
extern void cr_hbuf_put(struct cr_ctx *ctx, int n);
+extern void cr_ctx_get(struct cr_ctx *ctx);
+extern void cr_ctx_put(struct cr_ctx *ctx);
+
/* shared objects handling */
enum {
diff --git a/include/linux/sched.h b/include/linux/sched.h
index 0c8d02d..cb39066 100644
--- a/include/linux/sched.h
+++ b/include/linux/sched.h
@@ -1351,6 +1351,7 @@ struct task_struct {
unsigned long default_timer_slack_ns;
#ifdef CONFIG_CHECKPOINT_RESTART
atomic_t may_checkpoint;
+ struct cr_ctx *checkpoint_ctx;
#endif
};
--
1.5.4.3
From: Dave Hansen <[email protected]>
Suggested by Ingo.
Checkpoint/restart is going to be a long effort to get things working.
We're going to have a lot of things that we know just don't work for
a long time. That doesn't mean that it will be useless, it just means
that there's some complicated features that we are going to have to
work incrementally to fix.
This patch introduces a new mechanism to help the checkpoint/restart
developers. A new function pair: task/process_deny_checkpoint() is
created. When called, these tell the kernel that we *know* that the
process has performed some activity that will keep it from being
properly checkpointed.
The 'flag' is an atomic_t for now so that we can have some level
of atomicity and make sure to only warn once.
For now, this is a one-way trip. Once a process is no longer
'may_checkpoint' capable, neither it nor its children ever will be.
This can, of course, be fixed up in the future. We might want to
reset the flag when a new pid namespace is created, for instance.
Signed-off-by: Dave Hansen <[email protected]>
Signed-off-by: Oren Laadan <[email protected]>
---
checkpoint/checkpoint.c | 6 ++++++
include/linux/checkpoint.h | 33 ++++++++++++++++++++++++++++++++-
include/linux/sched.h | 3 +++
kernel/fork.c | 10 ++++++++++
4 files changed, 51 insertions(+), 1 deletions(-)
diff --git a/checkpoint/checkpoint.c b/checkpoint/checkpoint.c
index 08719ba..3d0a366 100644
--- a/checkpoint/checkpoint.c
+++ b/checkpoint/checkpoint.c
@@ -232,6 +232,12 @@ static int cr_write_task(struct cr_ctx *ctx, struct task_struct *t)
return -EAGAIN;
}
+ if (!atomic_read(&t->may_checkpoint)) {
+ pr_warning("c/r: task %d may not checkpoint\n",
+ task_pid_vnr(t));
+ return -EBUSY;
+ }
+
ret = cr_write_task_struct(ctx, t);
cr_debug("task_struct: ret %d\n", ret);
if (ret < 0)
diff --git a/include/linux/checkpoint.h b/include/linux/checkpoint.h
index 93fea2f..fc1873b 100644
--- a/include/linux/checkpoint.h
+++ b/include/linux/checkpoint.h
@@ -10,8 +10,11 @@
* distribution for more details.
*/
-#include <linux/path.h>
#include <linux/fs.h>
+#include <linux/path.h>
+#include <linux/sched.h>
+
+#ifdef CONFIG_CHECKPOINT_RESTART
#define CR_VERSION 2
@@ -95,4 +98,32 @@ extern int cr_read_files(struct cr_ctx *ctx);
#define cr_debug(fmt, args...) \
pr_debug("[%d:c/r:%s] " fmt, task_pid_vnr(current), __func__, ## args)
+static inline void __task_deny_checkpointing(struct task_struct *task,
+ char *file, int line)
+{
+ if (!atomic_dec_and_test(&task->may_checkpoint))
+ return;
+ printk(KERN_INFO "process performed an action that can not be "
+ "checkpointed at: %s:%d\n", file, line);
+ WARN_ON(1);
+}
+#define process_deny_checkpointing(p) \
+ __task_deny_checkpointing(p, __FILE__, __LINE__)
+
+/*
+ * For now, we're not going to have a distinction between
+ * tasks and processes for the purpose of c/r. But, allow
+ * these two calls anyway to make new users at least think
+ * about it.
+ */
+#define task_deny_checkpointing(p) \
+ __task_deny_checkpointing(p, __FILE__, __LINE__)
+
+#else
+
+static inline void task_deny_checkpointing(struct task_struct *task) {}
+static inline void process_deny_checkpointing(struct task_struct *task) {}
+
+#endif
+
#endif /* _CHECKPOINT_CKPT_H_ */
diff --git a/include/linux/sched.h b/include/linux/sched.h
index b483f39..0c8d02d 100644
--- a/include/linux/sched.h
+++ b/include/linux/sched.h
@@ -1349,6 +1349,9 @@ struct task_struct {
*/
unsigned long timer_slack_ns;
unsigned long default_timer_slack_ns;
+#ifdef CONFIG_CHECKPOINT_RESTART
+ atomic_t may_checkpoint;
+#endif
};
/*
diff --git a/kernel/fork.c b/kernel/fork.c
index f608356..693b0cb 100644
--- a/kernel/fork.c
+++ b/kernel/fork.c
@@ -195,6 +195,13 @@ void __init fork_init(unsigned long mempages)
init_task.signal->rlim[RLIMIT_NPROC].rlim_max = max_threads/2;
init_task.signal->rlim[RLIMIT_SIGPENDING] =
init_task.signal->rlim[RLIMIT_NPROC];
+
+#ifdef CONFIG_CHECKPOINT_RESTART
+ /*
+ * This probably won't stay set for long...
+ */
+ atomic_set(&init_task.may_checkpoint, 1);
+#endif
}
int __attribute__((weak)) arch_dup_task_struct(struct task_struct *dst,
@@ -245,6 +252,9 @@ static struct task_struct *dup_task_struct(struct task_struct *orig)
tsk->btrace_seq = 0;
#endif
tsk->splice_pipe = NULL;
+#ifdef CONFIG_CHECKPOINT_RESTART
+ atomic_set(&tsk->may_checkpoint, atomic_read(&orig->may_checkpoint));
+#endif
return tsk;
out:
--
1.5.4.3
Add logic to save and restore architecture specific state, including
thread-specific state, CPU registers and FPU state.
In addition, architecture capabilities are saved in an architecure
specific extension of the header (cr_hdr_head_arch); Currently this
includes only FPU capabilities.
Currently only x86-32 is supported. Compiling on x86-64 will trigger
an explicit error.
Changelog[v9]:
- Add arch-specific header that details architecture capabilities;
split FPU restore to send capabilities only once.
- Test for zero TLS entries in cr_write_thread()
- Fix asm/checkpoint_hdr.h so it can be included from user-space
Changelog[v7]:
- Fix save/restore state of FPU
Changelog[v5]:
- Remove preempt_disable() when restoring debug registers
Changelog[v4]:
- Fix header structure alignment
Changelog[v2]:
- Pad header structures to 64 bits to ensure compatibility
- Follow Dave Hansen's refactoring of the original post
Signed-off-by: Oren Laadan <[email protected]>
Acked-by: Serge Hallyn <[email protected]>
Signed-off-by: Dave Hansen <[email protected]>
---
arch/x86/include/asm/checkpoint_hdr.h | 85 ++++++++++++
arch/x86/mm/Makefile | 2 +
arch/x86/mm/checkpoint.c | 223 +++++++++++++++++++++++++++++++
arch/x86/mm/restart.c | 232 +++++++++++++++++++++++++++++++++
checkpoint/checkpoint.c | 19 +++-
checkpoint/checkpoint_arch.h | 9 ++
checkpoint/restart.c | 17 ++-
include/linux/checkpoint_hdr.h | 2 +
8 files changed, 583 insertions(+), 6 deletions(-)
create mode 100644 arch/x86/include/asm/checkpoint_hdr.h
create mode 100644 arch/x86/mm/checkpoint.c
create mode 100644 arch/x86/mm/restart.c
create mode 100644 checkpoint/checkpoint_arch.h
diff --git a/arch/x86/include/asm/checkpoint_hdr.h b/arch/x86/include/asm/checkpoint_hdr.h
new file mode 100644
index 0000000..6325062
--- /dev/null
+++ b/arch/x86/include/asm/checkpoint_hdr.h
@@ -0,0 +1,85 @@
+#ifndef __ASM_X86_CKPT_HDR_H
+#define __ASM_X86_CKPT_HDR_H
+/*
+ * Checkpoint/restart - architecture specific headers x86
+ *
+ * Copyright (C) 2008 Oren Laadan
+ *
+ * This file is subject to the terms and conditions of the GNU General Public
+ * License. See the file COPYING in the main directory of the Linux
+ * distribution for more details.
+ */
+
+#include <linux/types.h>
+
+/* i387 structure seen from kernel/userspace */
+#ifdef __KERNEL__
+#include <asm/processor.h>
+#else
+#include <sys/user.h>
+#endif
+
+struct cr_hdr_head_arch {
+ /* FIXME: add HAVE_HWFP */
+
+ __u16 has_fxsr;
+ __u16 has_xsave;
+ __u16 xstate_size;
+ __u16 _pading;
+} __attribute__((aligned(8)));
+
+struct cr_hdr_thread {
+ /* FIXME: restart blocks */
+
+ __s16 gdt_entry_tls_entries;
+ __s16 sizeof_tls_array;
+ __s16 ntls; /* number of TLS entries to follow */
+} __attribute__((aligned(8)));
+
+struct cr_hdr_cpu {
+ /* see struct pt_regs (x86-64) */
+ __u64 r15;
+ __u64 r14;
+ __u64 r13;
+ __u64 r12;
+ __u64 bp;
+ __u64 bx;
+ __u64 r11;
+ __u64 r10;
+ __u64 r9;
+ __u64 r8;
+ __u64 ax;
+ __u64 cx;
+ __u64 dx;
+ __u64 si;
+ __u64 di;
+ __u64 orig_ax;
+ __u64 ip;
+ __u64 cs;
+ __u64 flags;
+ __u64 sp;
+ __u64 ss;
+
+ /* segment registers */
+ __u64 ds;
+ __u64 es;
+ __u64 fs;
+ __u64 gs;
+
+ /* debug registers */
+ __u64 debugreg0;
+ __u64 debugreg1;
+ __u64 debugreg2;
+ __u64 debugreg3;
+ __u64 debugreg4;
+ __u64 debugreg5;
+ __u64 debugreg6;
+ __u64 debugreg7;
+
+ __u32 uses_debug;
+ __u32 used_math;
+
+ /* thread_xstate contents follow (if used_math) */
+} __attribute__((aligned(8)));
+
+#endif /* __ASM_X86_CKPT_HDR__H */
diff --git a/arch/x86/mm/Makefile b/arch/x86/mm/Makefile
index 59f89b4..7d5e2ce 100644
--- a/arch/x86/mm/Makefile
+++ b/arch/x86/mm/Makefile
@@ -18,3 +18,5 @@ obj-$(CONFIG_K8_NUMA) += k8topology_64.o
obj-$(CONFIG_ACPI_NUMA) += srat_$(BITS).o
obj-$(CONFIG_MEMTEST) += memtest.o
+
+obj-$(CONFIG_CHECKPOINT_RESTART) += checkpoint.o restart.o
diff --git a/arch/x86/mm/checkpoint.c b/arch/x86/mm/checkpoint.c
new file mode 100644
index 0000000..8dd6d2d
--- /dev/null
+++ b/arch/x86/mm/checkpoint.c
@@ -0,0 +1,223 @@
+/*
+ * Checkpoint/restart - architecture specific support for x86
+ *
+ * Copyright (C) 2008 Oren Laadan
+ *
+ * This file is subject to the terms and conditions of the GNU General Public
+ * License. See the file COPYING in the main directory of the Linux
+ * distribution for more details.
+ */
+
+#include <asm/desc.h>
+#include <asm/i387.h>
+
+#include <linux/checkpoint.h>
+#include <linux/checkpoint_hdr.h>
+
+/* dump the thread_struct of a given task */
+int cr_write_thread(struct cr_ctx *ctx, struct task_struct *t)
+{
+ struct cr_hdr h;
+ struct cr_hdr_thread *hh = cr_hbuf_get(ctx, sizeof(*hh));
+ struct thread_struct *thread;
+ struct desc_struct *desc;
+ int ntls = 0;
+ int n, ret;
+
+ h.type = CR_HDR_THREAD;
+ h.len = sizeof(*hh);
+ h.parent = task_pid_vnr(t);
+
+ thread = &t->thread;
+
+ /* calculate no. of TLS entries that follow */
+ desc = thread->tls_array;
+ for (n = GDT_ENTRY_TLS_ENTRIES; n > 0; n--, desc++) {
+ if (desc->a || desc->b)
+ ntls++;
+ }
+
+ hh->gdt_entry_tls_entries = GDT_ENTRY_TLS_ENTRIES;
+ hh->sizeof_tls_array = sizeof(thread->tls_array);
+ hh->ntls = ntls;
+
+ ret = cr_write_obj(ctx, &h, hh);
+ cr_hbuf_put(ctx, sizeof(*hh));
+ if (ret < 0)
+ return ret;
+
+ cr_debug("ntls %d\n", ntls);
+ if (ntls == 0)
+ return 0;
+
+ /* for simplicity dump the entire array, cherry-pick upon restart */
+ ret = cr_kwrite(ctx, thread->tls_array, sizeof(thread->tls_array));
+
+ /* IGNORE RESTART BLOCKS FOR NOW ... */
+
+ return ret;
+}
+
+#ifdef CONFIG_X86_64
+
+#error "CONFIG_X86_64 unsupported yet."
+
+#else /* !CONFIG_X86_64 */
+
+static void cr_save_cpu_regs(struct cr_hdr_cpu *hh, struct task_struct *t)
+{
+ struct thread_struct *thread = &t->thread;
+ struct pt_regs *regs = task_pt_regs(t);
+
+ hh->bp = regs->bp;
+ hh->bx = regs->bx;
+ hh->ax = regs->ax;
+ hh->cx = regs->cx;
+ hh->dx = regs->dx;
+ hh->si = regs->si;
+ hh->di = regs->di;
+ hh->orig_ax = regs->orig_ax;
+ hh->ip = regs->ip;
+ hh->cs = regs->cs;
+ hh->flags = regs->flags;
+ hh->sp = regs->sp;
+ hh->ss = regs->ss;
+
+ hh->ds = regs->ds;
+ hh->es = regs->es;
+
+ /*
+ * for checkpoint in process context (from within a container)
+ * the GS and FS registers should be saved from the hardware;
+ * otherwise they are already sabed on the thread structure
+ */
+ if (t == current) {
+ savesegment(gs, hh->gs);
+ savesegment(fs, hh->fs);
+ } else {
+ hh->gs = thread->gs;
+ hh->fs = thread->fs;
+ }
+
+ /*
+ * for checkpoint in process context (from within a container),
+ * the actual syscall is taking place at this very moment; so
+ * we (optimistically) subtitute the future return value (0) of
+ * this syscall into the orig_eax, so that upon restart it will
+ * succeed (or it will endlessly retry checkpoint...)
+ */
+ if (t == current) {
+ BUG_ON(hh->orig_ax < 0);
+ hh->ax = 0;
+ }
+}
+
+static void cr_save_cpu_debug(struct cr_hdr_cpu *hh, struct task_struct *t)
+{
+ struct thread_struct *thread = &t->thread;
+
+ /* debug regs */
+
+ /*
+ * for checkpoint in process context (from within a container),
+ * get the actual registers; otherwise get the saved values.
+ */
+
+ if (t == current) {
+ get_debugreg(hh->debugreg0, 0);
+ get_debugreg(hh->debugreg1, 1);
+ get_debugreg(hh->debugreg2, 2);
+ get_debugreg(hh->debugreg3, 3);
+ get_debugreg(hh->debugreg6, 6);
+ get_debugreg(hh->debugreg7, 7);
+ } else {
+ hh->debugreg0 = thread->debugreg0;
+ hh->debugreg1 = thread->debugreg1;
+ hh->debugreg2 = thread->debugreg2;
+ hh->debugreg3 = thread->debugreg3;
+ hh->debugreg6 = thread->debugreg6;
+ hh->debugreg7 = thread->debugreg7;
+ }
+
+ hh->debugreg4 = 0;
+ hh->debugreg5 = 0;
+
+ hh->uses_debug = !!(task_thread_info(t)->flags & TIF_DEBUG);
+}
+
+static void cr_save_cpu_fpu(struct cr_hdr_cpu *hh, struct task_struct *t)
+{
+ hh->used_math = tsk_used_math(t) ? 1 : 0;
+}
+
+static int cr_write_cpu_fpu(struct cr_ctx *ctx, struct task_struct *t)
+{
+ void *xstate_buf = cr_hbuf_get(ctx, xstate_size);
+
+ /* i387 + MMU + SSE logic */
+ preempt_disable(); /* needed it (t == current) */
+
+ /*
+ * normally, no need to unlazy_fpu(), since TS_USEDFPU flag
+ * have been cleared when task was context-switched out...
+ * except if we are in process context, in which case we do
+ */
+ if (t == current && (task_thread_info(t)->status & TS_USEDFPU))
+ unlazy_fpu(current);
+
+ memcpy(xstate_buf, t->thread.xstate, xstate_size);
+ preempt_enable(); /* needed it (t == current) */
+
+ return cr_kwrite(ctx, xstate_buf, xstate_size);
+}
+
+#endif /* CONFIG_X86_64 */
+
+/* dump the cpu state and registers of a given task */
+int cr_write_cpu(struct cr_ctx *ctx, struct task_struct *t)
+{
+ struct cr_hdr h;
+ struct cr_hdr_cpu *hh = cr_hbuf_get(ctx, sizeof(*hh));
+ int ret;
+
+ h.type = CR_HDR_CPU;
+ h.len = sizeof(*hh);
+ h.parent = task_pid_vnr(t);
+
+ cr_save_cpu_regs(hh, t);
+ cr_save_cpu_debug(hh, t);
+ cr_save_cpu_fpu(hh, t);
+
+ cr_debug("math %d debug %d\n", hh->used_math, hh->uses_debug);
+
+ ret = cr_write_obj(ctx, &h, hh);
+ if (ret < 0)
+ goto out;
+
+ if (hh->used_math)
+ ret = cr_write_cpu_fpu(ctx, t);
+ out:
+ cr_hbuf_put(ctx, sizeof(*hh));
+ return ret;
+}
+
+int cr_write_head_arch(struct cr_ctx *ctx)
+{
+ struct cr_hdr h;
+ struct cr_hdr_head_arch *hh = cr_hbuf_get(ctx, sizeof(*hh));
+ int ret;
+
+ h.type = CR_HDR_HEAD_ARCH;
+ h.len = sizeof(*hh);
+ h.parent = 0;
+
+ /* FPU capabilities */
+ hh->has_fxsr = cpu_has_fxsr;
+ hh->has_xsave = cpu_has_xsave;
+ hh->xstate_size = xstate_size;
+
+ ret = cr_write_obj(ctx, &h, hh);
+ cr_hbuf_put(ctx, sizeof(*hh));
+
+ return ret;
+}
diff --git a/arch/x86/mm/restart.c b/arch/x86/mm/restart.c
new file mode 100644
index 0000000..45ad790
--- /dev/null
+++ b/arch/x86/mm/restart.c
@@ -0,0 +1,232 @@
+/*
+ * Checkpoint/restart - architecture specific support for x86
+ *
+ * Copyright (C) 2008 Oren Laadan
+ *
+ * This file is subject to the terms and conditions of the GNU General Public
+ * License. See the file COPYING in the main directory of the Linux
+ * distribution for more details.
+ */
+
+#include <asm/desc.h>
+#include <asm/i387.h>
+
+#include <linux/checkpoint.h>
+#include <linux/checkpoint_hdr.h>
+
+/* read the thread_struct into the current task */
+int cr_read_thread(struct cr_ctx *ctx)
+{
+ struct cr_hdr_thread *hh = cr_hbuf_get(ctx, sizeof(*hh));
+ struct task_struct *t = current;
+ struct thread_struct *thread = &t->thread;
+ int parent, ret;
+
+ parent = cr_read_obj_type(ctx, hh, sizeof(*hh), CR_HDR_THREAD);
+ if (parent < 0) {
+ ret = parent;
+ goto out;
+ }
+
+ ret = -EINVAL;
+
+#if 0 /* activate when containers are used */
+ if (parent != task_pid_vnr(t))
+ goto out;
+#endif
+ cr_debug("ntls %d\n", hh->ntls);
+
+ if (hh->gdt_entry_tls_entries != GDT_ENTRY_TLS_ENTRIES ||
+ hh->sizeof_tls_array != sizeof(thread->tls_array) ||
+ hh->ntls < 0 || hh->ntls > GDT_ENTRY_TLS_ENTRIES)
+ goto out;
+
+ if (hh->ntls > 0) {
+ struct desc_struct *desc;
+ int size, cpu;
+
+ /*
+ * restore TLS by hand: why convert to struct user_desc if
+ * sys_set_thread_entry() will convert it back ?
+ */
+
+ size = sizeof(*desc) * GDT_ENTRY_TLS_ENTRIES;
+ desc = kmalloc(size, GFP_KERNEL);
+ if (!desc)
+ return -ENOMEM;
+
+ ret = cr_kread(ctx, desc, size);
+ if (ret >= 0) {
+ /*
+ * FIX: add sanity checks (eg. that values makes
+ * sense, that we don't overwrite old values, etc
+ */
+ cpu = get_cpu();
+ memcpy(thread->tls_array, desc, size);
+ load_TLS(thread, cpu);
+ put_cpu();
+ }
+ kfree(desc);
+ }
+
+ ret = 0;
+ out:
+ cr_hbuf_put(ctx, sizeof(*hh));
+ return ret;
+}
+
+#ifdef CONFIG_X86_64
+
+#error "CONFIG_X86_64 unsupported yet."
+
+#else /* !CONFIG_X86_64 */
+
+static int cr_load_cpu_regs(struct cr_hdr_cpu *hh, struct task_struct *t)
+{
+ struct thread_struct *thread = &t->thread;
+ struct pt_regs *regs = task_pt_regs(t);
+
+ regs->bx = hh->bx;
+ regs->cx = hh->cx;
+ regs->dx = hh->dx;
+ regs->si = hh->si;
+ regs->di = hh->di;
+ regs->bp = hh->bp;
+ regs->ax = hh->ax;
+ regs->ds = hh->ds;
+ regs->es = hh->es;
+ regs->orig_ax = hh->orig_ax;
+ regs->ip = hh->ip;
+ regs->cs = hh->cs;
+ regs->flags = hh->flags;
+ regs->sp = hh->sp;
+ regs->ss = hh->ss;
+
+ thread->gs = hh->gs;
+ thread->fs = hh->fs;
+ loadsegment(gs, hh->gs);
+ loadsegment(fs, hh->fs);
+
+ return 0;
+}
+
+static int cr_load_cpu_debug(struct cr_hdr_cpu *hh, struct task_struct *t)
+{
+ /* debug regs */
+
+ if (hh->uses_debug) {
+ set_debugreg(hh->debugreg0, 0);
+ set_debugreg(hh->debugreg1, 1);
+ /* ignore 4, 5 */
+ set_debugreg(hh->debugreg2, 2);
+ set_debugreg(hh->debugreg3, 3);
+ set_debugreg(hh->debugreg6, 6);
+ set_debugreg(hh->debugreg7, 7);
+ }
+
+ return 0;
+}
+
+static int cr_load_cpu_fpu(struct cr_hdr_cpu *hh, struct task_struct *t)
+{
+ preempt_disable();
+
+ __clear_fpu(t); /* in case we used FPU in user mode */
+
+ if (!hh->used_math)
+ clear_used_math();
+
+ preempt_enable();
+ return 0;
+}
+
+static int cr_read_cpu_fpu(struct cr_ctx *ctx, struct task_struct *t)
+{
+ void *xstate_buf = cr_hbuf_get(ctx, xstate_size);
+ int ret;
+
+ ret = cr_kread(ctx, xstate_buf, xstate_size);
+ if (ret < 0)
+ goto out;
+
+ /* i387 + MMU + SSE */
+ preempt_disable();
+
+ /* init_fpu() also calls set_used_math() */
+ ret = init_fpu(current);
+ if (ret < 0)
+ return ret;
+
+ memcpy(t->thread.xstate, xstate_buf, xstate_size);
+ preempt_enable();
+ out:
+ cr_hbuf_put(ctx, xstate_size);
+ return 0;
+}
+
+#endif /* CONFIG_X86_64 */
+
+/* read the cpu state and registers for the current task */
+int cr_read_cpu(struct cr_ctx *ctx)
+{
+ struct cr_hdr_cpu *hh = cr_hbuf_get(ctx, sizeof(*hh));
+ struct task_struct *t = current;
+ int parent, ret;
+
+ parent = cr_read_obj_type(ctx, hh, sizeof(*hh), CR_HDR_CPU);
+ if (parent < 0) {
+ ret = parent;
+ goto out;
+ }
+
+ ret = -EINVAL;
+
+#if 0 /* activate when containers are used */
+ if (parent != task_pid_vnr(t))
+ goto out;
+#endif
+ /* FIX: sanity check for sensitive registers (eg. eflags) */
+
+ cr_debug("math %d debug %d\n", hh->used_math, hh->uses_debug);
+
+ ret = cr_load_cpu_regs(hh, t);
+ if (ret < 0)
+ goto out;
+ ret = cr_load_cpu_debug(hh, t);
+ if (ret < 0)
+ goto out;
+ ret = cr_load_cpu_fpu(hh, t);
+ if (ret < 0)
+ goto out;
+
+ if (hh->used_math)
+ ret = cr_read_cpu_fpu(ctx, t);
+ out:
+ cr_hbuf_put(ctx, sizeof(*hh));
+ return ret;
+}
+
+int cr_read_head_arch(struct cr_ctx *ctx)
+{
+ struct cr_hdr_head_arch *hh = cr_hbuf_get(ctx, sizeof(*hh));
+ int parent, ret = 0;
+
+ parent = cr_read_obj_type(ctx, hh, sizeof(*hh), CR_HDR_HEAD_ARCH);
+ if (parent < 0) {
+ ret = parent;
+ goto out;
+ } else if (parent != 0)
+ goto out;
+
+ /* FIX: verify compatibility of architecture features */
+
+ /* verify FPU capabilities */
+ if (hh->has_fxsr != cpu_has_fxsr ||
+ hh->has_xsave != cpu_has_xsave ||
+ hh->xstate_size != xstate_size)
+ ret = -EINVAL;
+ out:
+ cr_hbuf_put(ctx, sizeof(*hh));
+
+ return ret;
+}
diff --git a/checkpoint/checkpoint.c b/checkpoint/checkpoint.c
index fccf723..17cc8d2 100644
--- a/checkpoint/checkpoint.c
+++ b/checkpoint/checkpoint.c
@@ -20,6 +20,8 @@
#include <linux/checkpoint.h>
#include <linux/checkpoint_hdr.h>
+#include "checkpoint_arch.h"
+
/* unique checkpoint identifier (FIXME: should be per-container ?) */
static atomic_t cr_ctx_count = ATOMIC_INIT(0);
@@ -105,7 +107,10 @@ static int cr_write_head(struct cr_ctx *ctx)
ret = cr_write_obj(ctx, &h, hh);
cr_hbuf_put(ctx, sizeof(*hh));
- return ret;
+ if (ret < 0)
+ return ret;
+
+ return cr_write_head_arch(ctx);
}
/* write the checkpoint trailer */
@@ -160,8 +165,16 @@ static int cr_write_task(struct cr_ctx *ctx, struct task_struct *t)
int ret;
ret = cr_write_task_struct(ctx, t);
- cr_debug("ret %d\n", ret);
-
+ cr_debug("task_struct: ret %d\n", ret);
+ if (ret < 0)
+ goto out;
+ ret = cr_write_thread(ctx, t);
+ cr_debug("thread: ret %d\n", ret);
+ if (ret < 0)
+ goto out;
+ ret = cr_write_cpu(ctx, t);
+ cr_debug("cpu: ret %d\n", ret);
+ out:
return ret;
}
diff --git a/checkpoint/checkpoint_arch.h b/checkpoint/checkpoint_arch.h
new file mode 100644
index 0000000..ada1369
--- /dev/null
+++ b/checkpoint/checkpoint_arch.h
@@ -0,0 +1,9 @@
+#include <linux/checkpoint.h>
+
+extern int cr_write_head_arch(struct cr_ctx *ctx);
+extern int cr_write_thread(struct cr_ctx *ctx, struct task_struct *t);
+extern int cr_write_cpu(struct cr_ctx *ctx, struct task_struct *t);
+
+extern int cr_read_head_arch(struct cr_ctx *ctx);
+extern int cr_read_thread(struct cr_ctx *ctx);
+extern int cr_read_cpu(struct cr_ctx *ctx);
diff --git a/checkpoint/restart.c b/checkpoint/restart.c
index a95d2e8..d74d755 100644
--- a/checkpoint/restart.c
+++ b/checkpoint/restart.c
@@ -15,6 +15,8 @@
#include <linux/checkpoint.h>
#include <linux/checkpoint_hdr.h>
+#include "checkpoint_arch.h"
+
/**
* cr_read_obj - read a whole record (cr_hdr followed by payload)
* @ctx: checkpoint context
@@ -142,9 +144,9 @@ static int cr_read_head(struct cr_ctx *ctx)
ctx->oflags = hh->flags;
- /* FIX: verify compatibility of release, version and machine */
+ /* FIX: verify compatibility of release, version */
- ret = 0;
+ ret = cr_read_head_arch(ctx);
out:
cr_hbuf_put(ctx, sizeof(*hh));
return ret;
@@ -214,8 +216,17 @@ static int cr_read_task(struct cr_ctx *ctx)
int ret;
ret = cr_read_task_struct(ctx);
- cr_debug("ret %d\n", ret);
+ cr_debug("task_struct: ret %d\n", ret);
+ if (ret < 0)
+ goto out;
+ ret = cr_read_thread(ctx);
+ cr_debug("thread: ret %d\n", ret);
+ if (ret < 0)
+ goto out;
+ ret = cr_read_cpu(ctx);
+ cr_debug("cpu: ret %d\n", ret);
+ out:
return ret;
}
diff --git a/include/linux/checkpoint_hdr.h b/include/linux/checkpoint_hdr.h
index 257f87f..b74b5f9 100644
--- a/include/linux/checkpoint_hdr.h
+++ b/include/linux/checkpoint_hdr.h
@@ -12,6 +12,7 @@
#include <linux/types.h>
#include <linux/utsname.h>
+#include <asm/checkpoint_hdr.h>
/*
* To maintain compatibility between 32-bit and 64-bit architecture flavors,
@@ -30,6 +31,7 @@ struct cr_hdr {
/* header types */
enum {
CR_HDR_HEAD = 1,
+ CR_HDR_HEAD_ARCH,
CR_HDR_BUFFER,
CR_HDR_STRING,
--
1.5.4.3
Now we can do "external" checkpoint, i.e. act on another task.
sys_checkpoint() now looks up the target pid (in our namespace) and
checkpoints that corresponding task. That task should be the root of
a container.
sys_restart() remains the same, as the restart is always done in the
context of the restarting task.
Changelog[v10]:
- Grab vfs root of container init, rather than current process
Signed-off-by: Oren Laadan <[email protected]>
Acked-by: Serge Hallyn <[email protected]>
---
checkpoint/checkpoint.c | 75 ++++++++++++++++++++++++++++++++++++++++++--
checkpoint/restart.c | 4 +-
checkpoint/sys.c | 6 +++
include/linux/checkpoint.h | 2 +
4 files changed, 82 insertions(+), 5 deletions(-)
diff --git a/checkpoint/checkpoint.c b/checkpoint/checkpoint.c
index ae1952b..08719ba 100644
--- a/checkpoint/checkpoint.c
+++ b/checkpoint/checkpoint.c
@@ -10,6 +10,7 @@
#include <linux/version.h>
#include <linux/sched.h>
+#include <linux/ptrace.h>
#include <linux/time.h>
#include <linux/fs.h>
#include <linux/file.h>
@@ -224,6 +225,13 @@ static int cr_write_task(struct cr_ctx *ctx, struct task_struct *t)
{
int ret;
+ /* TODO: verity that the task is frozen (unless self) */
+
+ if (t->state == TASK_DEAD) {
+ pr_warning("c/r: task may not be in state TASK_DEAD\n");
+ return -EAGAIN;
+ }
+
ret = cr_write_task_struct(ctx, t);
cr_debug("task_struct: ret %d\n", ret);
if (ret < 0)
@@ -246,15 +254,76 @@ static int cr_write_task(struct cr_ctx *ctx, struct task_struct *t)
return ret;
}
-static int cr_ctx_checkpoint(struct cr_ctx *ctx, pid_t pid)
+static int cr_get_container(struct cr_ctx *ctx, pid_t pid)
{
+ struct task_struct *task = NULL;
+ struct nsproxy *nsproxy = NULL;
+ int err = -ESRCH;
+
ctx->root_pid = pid;
+ read_lock(&tasklist_lock);
+ task = find_task_by_vpid(pid);
+ if (task)
+ get_task_struct(task);
+ read_unlock(&tasklist_lock);
+
+ if (!task)
+ goto out;
+
+#if 0 /* enable to use containers */
+ if (!is_container_init(task)) {
+ err = -EINVAL;
+ goto out;
+ }
+#endif
+
+ if (!ptrace_may_access(task, PTRACE_MODE_READ)) {
+ err = -EPERM;
+ goto out;
+ }
+
+ rcu_read_lock();
+ if (task_nsproxy(task)) {
+ nsproxy = task_nsproxy(task);
+ get_nsproxy(nsproxy);
+ }
+ rcu_read_unlock();
+
+ if (!nsproxy)
+ goto out;
+
+ /* TODO: verify that the container is frozen */
+
+ ctx->root_task = task;
+ ctx->root_nsproxy = nsproxy;
+
+ return 0;
+
+ out:
+ if (task)
+ put_task_struct(task);
+ return err;
+}
+
+/* setup checkpoint-specific parts of ctx */
+static int cr_ctx_checkpoint(struct cr_ctx *ctx, pid_t pid)
+{
+ int ret;
+
+ ret = cr_get_container(ctx, pid);
+ if (ret < 0)
+ return ret;
+
/*
* assume checkpointer is in container's root vfs
* FIXME: this works for now, but will change with real containers
*/
- ctx->vfsroot = ¤t->fs->root;
+ task_lock(ctx->root_task);
+ ctx->vfsroot = &ctx->root_task->fs->root;
+ task_unlock(ctx->root_task);
+ if (!ctx->vfsroot)
+ return -EAGAIN;
path_get(ctx->vfsroot);
return 0;
@@ -270,7 +339,7 @@ int do_checkpoint(struct cr_ctx *ctx, pid_t pid)
ret = cr_write_head(ctx);
if (ret < 0)
goto out;
- ret = cr_write_task(ctx, current);
+ ret = cr_write_task(ctx, ctx->root_task);
if (ret < 0)
goto out;
ret = cr_write_tail(ctx);
diff --git a/checkpoint/restart.c b/checkpoint/restart.c
index 22e7995..f4f737d 100644
--- a/checkpoint/restart.c
+++ b/checkpoint/restart.c
@@ -277,7 +277,7 @@ static int cr_read_task(struct cr_ctx *ctx)
}
/* setup restart-specific parts of ctx */
-static int cr_ctx_restart(struct cr_ctx *ctx)
+static int cr_ctx_restart(struct cr_ctx *ctx, pid_t pid)
{
return 0;
}
@@ -286,7 +286,7 @@ int do_restart(struct cr_ctx *ctx, pid_t pid)
{
int ret;
- ret = cr_ctx_restart(ctx);
+ ret = cr_ctx_restart(ctx, pid);
if (ret < 0)
goto out;
ret = cr_read_head(ctx);
diff --git a/checkpoint/sys.c b/checkpoint/sys.c
index b640bee..dfe63ca 100644
--- a/checkpoint/sys.c
+++ b/checkpoint/sys.c
@@ -9,6 +9,7 @@
*/
#include <linux/sched.h>
+#include <linux/nsproxy.h>
#include <linux/kernel.h>
#include <linux/fs.h>
#include <linux/file.h>
@@ -142,6 +143,11 @@ static void cr_ctx_free(struct cr_ctx *ctx)
cr_pgarr_free(ctx);
cr_objhash_free(ctx);
+ if (ctx->root_nsproxy)
+ put_nsproxy(ctx->root_nsproxy);
+ if (ctx->root_task)
+ put_task_struct(ctx->root_task);
+
kfree(ctx);
}
diff --git a/include/linux/checkpoint.h b/include/linux/checkpoint.h
index 2675410..93fea2f 100644
--- a/include/linux/checkpoint.h
+++ b/include/linux/checkpoint.h
@@ -19,6 +19,8 @@ struct cr_ctx {
int crid; /* unique checkpoint id */
pid_t root_pid; /* container identifier */
+ struct task_struct *root_task; /* container root task */
+ struct nsproxy *root_nsproxy; /* container root nsproxy */
unsigned long flags;
unsigned long oflags; /* restart: old flags */
--
1.5.4.3
For each VMA, there is a 'struct cr_vma'; if the VMA is file-mapped,
it will be followed by the file name. Then comes the actual contents,
in one or more chunk: each chunk begins with a header that specifies
how many pages it holds, then the virtual addresses of all the dumped
pages in that chunk, followed by the actual contents of all dumped
pages. A header with zero number of pages marks the end of the contents.
Then comes the next VMA and so on.
Changelog[v10]:
- Acquire dcache_lock around call to __d_path() in cr_fill_name()
Changelog[v9]:
- Introduce cr_ctx_checkpoint() for checkpoint-specific ctx setup
- Test if __d_path() changes mnt/dentry (when crossing filesystem
namespace boundary). for now cr_fill_fname() fails the checkpoint.
Changelog[v7]:
- Fix argument given to kunmap_atomic() in memory dump/restore
Changelog[v6]:
- Balance all calls to cr_hbuf_get() with matching cr_hbuf_put()
(even though it's not really needed)
Changelog[v5]:
- Improve memory dump code (following Dave Hansen's comments)
- Change dump format (and code) to allow chunks of <vaddrs, pages>
instead of one long list of each
- Fix use of follow_page() to avoid faulting in non-present pages
Changelog[v4]:
- Use standard list_... for cr_pgarr
Signed-off-by: Oren Laadan <[email protected]>
Acked-by: Serge Hallyn <[email protected]>
Signed-off-by: Dave Hansen <[email protected]>
---
arch/x86/include/asm/checkpoint_hdr.h | 5 +
arch/x86/mm/checkpoint.c | 31 ++
checkpoint/Makefile | 3 +-
checkpoint/checkpoint.c | 81 ++++++
checkpoint/checkpoint_arch.h | 2 +
checkpoint/checkpoint_mem.h | 41 +++
checkpoint/ckpt_mem.c | 500 +++++++++++++++++++++++++++++++++
checkpoint/sys.c | 11 +
include/linux/checkpoint.h | 12 +
include/linux/checkpoint_hdr.h | 32 ++
10 files changed, 717 insertions(+), 1 deletions(-)
create mode 100644 checkpoint/checkpoint_mem.h
create mode 100644 checkpoint/ckpt_mem.c
diff --git a/arch/x86/include/asm/checkpoint_hdr.h b/arch/x86/include/asm/checkpoint_hdr.h
index 6325062..33f4c70 100644
--- a/arch/x86/include/asm/checkpoint_hdr.h
+++ b/arch/x86/include/asm/checkpoint_hdr.h
@@ -82,4 +82,9 @@ struct cr_hdr_cpu {
/* thread_xstate contents follow (if used_math) */
} __attribute__((aligned(8)));
+struct cr_hdr_mm_context {
+ __s16 ldt_entry_size;
+ __s16 nldt;
+} __attribute__((aligned(8)));
+
#endif /* __ASM_X86_CKPT_HDR__H */
diff --git a/arch/x86/mm/checkpoint.c b/arch/x86/mm/checkpoint.c
index 8dd6d2d..757936e 100644
--- a/arch/x86/mm/checkpoint.c
+++ b/arch/x86/mm/checkpoint.c
@@ -221,3 +221,34 @@ int cr_write_head_arch(struct cr_ctx *ctx)
return ret;
}
+
+/* dump the mm->context state */
+int cr_write_mm_context(struct cr_ctx *ctx, struct mm_struct *mm, int parent)
+{
+ struct cr_hdr h;
+ struct cr_hdr_mm_context *hh = cr_hbuf_get(ctx, sizeof(*hh));
+ int ret;
+
+ h.type = CR_HDR_MM_CONTEXT;
+ h.len = sizeof(*hh);
+ h.parent = parent;
+
+ mutex_lock(&mm->context.lock);
+
+ hh->ldt_entry_size = LDT_ENTRY_SIZE;
+ hh->nldt = mm->context.size;
+
+ cr_debug("nldt %d\n", hh->nldt);
+
+ ret = cr_write_obj(ctx, &h, hh);
+ cr_hbuf_put(ctx, sizeof(*hh));
+ if (ret < 0)
+ goto out;
+
+ ret = cr_kwrite(ctx, mm->context.ldt,
+ mm->context.size * LDT_ENTRY_SIZE);
+
+ out:
+ mutex_unlock(&mm->context.lock);
+ return ret;
+}
diff --git a/checkpoint/Makefile b/checkpoint/Makefile
index d2df68c..3a0df6d 100644
--- a/checkpoint/Makefile
+++ b/checkpoint/Makefile
@@ -2,4 +2,5 @@
# Makefile for linux checkpoint/restart.
#
-obj-$(CONFIG_CHECKPOINT_RESTART) += sys.o checkpoint.o restart.o
+obj-$(CONFIG_CHECKPOINT_RESTART) += sys.o checkpoint.o restart.o \
+ ckpt_mem.o
diff --git a/checkpoint/checkpoint.c b/checkpoint/checkpoint.c
index 17cc8d2..6a8f810 100644
--- a/checkpoint/checkpoint.c
+++ b/checkpoint/checkpoint.c
@@ -75,6 +75,66 @@ int cr_write_string(struct cr_ctx *ctx, char *str, int len)
return cr_write_obj(ctx, &h, str);
}
+/**
+ * cr_fill_fname - return pathname of a given file
+ * @path: path name
+ * @root: relative root
+ * @buf: buffer for pathname
+ * @n: buffer length (in) and pathname length (out)
+ */
+static char *
+cr_fill_fname(struct path *path, struct path *root, char *buf, int *n)
+{
+ struct path tmp = *root;
+ char *fname;
+
+ BUG_ON(!buf);
+ spin_lock(&dcache_lock);
+ fname = __d_path(path, &tmp, buf, *n);
+ spin_unlock(&dcache_lock);
+ if (!IS_ERR(fname))
+ *n = (buf + (*n) - fname);
+ /*
+ * FIXME: if __d_path() changed these, it must have stepped out of
+ * init's namespace. Since currently we require a unified namespace
+ * within the container: simply fail.
+ */
+ if (tmp.mnt != root->mnt || tmp.dentry != root->dentry)
+ fname = ERR_PTR(-EBADF);
+
+ return fname;
+}
+
+/**
+ * cr_write_fname - write a file name
+ * @ctx: checkpoint context
+ * @path: path name
+ * @root: relative root
+ */
+int cr_write_fname(struct cr_ctx *ctx, struct path *path, struct path *root)
+{
+ struct cr_hdr h;
+ char *buf, *fname;
+ int ret, flen;
+
+ flen = PATH_MAX;
+ buf = kmalloc(flen, GFP_KERNEL);
+ if (!buf)
+ return -ENOMEM;
+
+ fname = cr_fill_fname(path, root, buf, &flen);
+ if (!IS_ERR(fname)) {
+ h.type = CR_HDR_FNAME;
+ h.len = flen;
+ h.parent = 0;
+ ret = cr_write_obj(ctx, &h, fname);
+ } else
+ ret = PTR_ERR(fname);
+
+ kfree(buf);
+ return ret;
+}
+
/* write the checkpoint header */
static int cr_write_head(struct cr_ctx *ctx)
{
@@ -168,6 +228,10 @@ static int cr_write_task(struct cr_ctx *ctx, struct task_struct *t)
cr_debug("task_struct: ret %d\n", ret);
if (ret < 0)
goto out;
+ ret = cr_write_mm(ctx, t);
+ cr_debug("memory: ret %d\n", ret);
+ if (ret < 0)
+ goto out;
ret = cr_write_thread(ctx, t);
cr_debug("thread: ret %d\n", ret);
if (ret < 0)
@@ -178,10 +242,27 @@ static int cr_write_task(struct cr_ctx *ctx, struct task_struct *t)
return ret;
}
+static int cr_ctx_checkpoint(struct cr_ctx *ctx, pid_t pid)
+{
+ ctx->root_pid = pid;
+
+ /*
+ * assume checkpointer is in container's root vfs
+ * FIXME: this works for now, but will change with real containers
+ */
+ ctx->vfsroot = ¤t->fs->root;
+ path_get(ctx->vfsroot);
+
+ return 0;
+}
+
int do_checkpoint(struct cr_ctx *ctx, pid_t pid)
{
int ret;
+ ret = cr_ctx_checkpoint(ctx, pid);
+ if (ret < 0)
+ goto out;
ret = cr_write_head(ctx);
if (ret < 0)
goto out;
diff --git a/checkpoint/checkpoint_arch.h b/checkpoint/checkpoint_arch.h
index ada1369..f06c7eb 100644
--- a/checkpoint/checkpoint_arch.h
+++ b/checkpoint/checkpoint_arch.h
@@ -3,6 +3,8 @@
extern int cr_write_head_arch(struct cr_ctx *ctx);
extern int cr_write_thread(struct cr_ctx *ctx, struct task_struct *t);
extern int cr_write_cpu(struct cr_ctx *ctx, struct task_struct *t);
+extern int cr_write_mm_context(struct cr_ctx *ctx,
+ struct mm_struct *mm, int parent);
extern int cr_read_head_arch(struct cr_ctx *ctx);
extern int cr_read_thread(struct cr_ctx *ctx);
diff --git a/checkpoint/checkpoint_mem.h b/checkpoint/checkpoint_mem.h
new file mode 100644
index 0000000..85546f4
--- /dev/null
+++ b/checkpoint/checkpoint_mem.h
@@ -0,0 +1,41 @@
+#ifndef _CHECKPOINT_CKPT_MEM_H_
+#define _CHECKPOINT_CKPT_MEM_H_
+/*
+ * Generic container checkpoint-restart
+ *
+ * Copyright (C) 2008 Oren Laadan
+ *
+ * This file is subject to the terms and conditions of the GNU General Public
+ * License. See the file COPYING in the main directory of the Linux
+ * distribution for more details.
+ */
+
+#include <linux/mm_types.h>
+
+/*
+ * page-array chains: each cr_pgarr describes a set of <strcut page *,vaddr>
+ * tuples (where vaddr is the virtual address of a page in a particular mm).
+ * Specifically, we use separate arrays so that all vaddrs can be written
+ * and read at once.
+ */
+
+struct cr_pgarr {
+ unsigned long *vaddrs;
+ struct page **pages;
+ unsigned int nr_used;
+ struct list_head list;
+};
+
+#define CR_PGARR_TOTAL (PAGE_SIZE / sizeof(void *))
+#define CR_PGARR_CHUNK (4 * CR_PGARR_TOTAL)
+
+extern void cr_pgarr_free(struct cr_ctx *ctx);
+extern struct cr_pgarr *cr_pgarr_current(struct cr_ctx *ctx);
+extern void cr_pgarr_reset_all(struct cr_ctx *ctx);
+
+static inline int cr_pgarr_is_full(struct cr_pgarr *pgarr)
+{
+ return (pgarr->nr_used == CR_PGARR_TOTAL);
+}
+
+#endif /* _CHECKPOINT_CKPT_MEM_H_ */
diff --git a/checkpoint/ckpt_mem.c b/checkpoint/ckpt_mem.c
new file mode 100644
index 0000000..66a2c39
--- /dev/null
+++ b/checkpoint/ckpt_mem.c
@@ -0,0 +1,500 @@
+/*
+ * Checkpoint memory contents
+ *
+ * Copyright (C) 2008 Oren Laadan
+ *
+ * This file is subject to the terms and conditions of the GNU General Public
+ * License. See the file COPYING in the main directory of the Linux
+ * distribution for more details.
+ */
+
+#include <linux/kernel.h>
+#include <linux/sched.h>
+#include <linux/slab.h>
+#include <linux/file.h>
+#include <linux/pagemap.h>
+#include <linux/mm_types.h>
+#include <linux/checkpoint.h>
+#include <linux/checkpoint_hdr.h>
+
+#include "checkpoint_arch.h"
+#include "checkpoint_mem.h"
+
+/*
+ * utilities to alloc, free, and handle 'struct cr_pgarr' (page-arrays)
+ * (common to ckpt_mem.c and rstr_mem.c).
+ *
+ * The checkpoint context structure has two members for page-arrays:
+ * ctx->pgarr_list: list head of the page-array chain
+ *
+ * During checkpoint (and restart) the chain tracks the dirty pages (page
+ * pointer and virtual address) of each MM. For a particular MM, these are
+ * always added to the head of the page-array chain (ctx->pgarr_list).
+ * This "current" page-array advances as necessary, and new page-array
+ * descriptors are allocated on-demand. Before the next chunk of pages,
+ * the chain is reset but not freed (that is, dereference page pointers).
+ */
+
+/* return first page-array in the chain */
+static inline struct cr_pgarr *cr_pgarr_first(struct cr_ctx *ctx)
+{
+ if (list_empty(&ctx->pgarr_list))
+ return NULL;
+ return list_first_entry(&ctx->pgarr_list, struct cr_pgarr, list);
+}
+
+/* release pages referenced by a page-array */
+static void cr_pgarr_release_pages(struct cr_pgarr *pgarr)
+{
+ int i;
+
+ cr_debug("nr_used %d\n", pgarr->nr_used);
+ /*
+ * although both checkpoint and restart use 'nr_used', we only
+ * collect pages during checkpoint; in restart we simply return
+ */
+ if (!pgarr->pages)
+ return;
+ for (i = pgarr->nr_used; i--; /**/)
+ page_cache_release(pgarr->pages[i]);
+}
+
+/* free a single page-array object */
+static void cr_pgarr_free_one(struct cr_pgarr *pgarr)
+{
+ cr_pgarr_release_pages(pgarr);
+ kfree(pgarr->pages);
+ kfree(pgarr->vaddrs);
+ kfree(pgarr);
+}
+
+/* free a chain of page-arrays */
+void cr_pgarr_free(struct cr_ctx *ctx)
+{
+ struct cr_pgarr *pgarr, *tmp;
+
+ list_for_each_entry_safe(pgarr, tmp, &ctx->pgarr_list, list) {
+ list_del(&pgarr->list);
+ cr_pgarr_free_one(pgarr);
+ }
+}
+
+/* allocate a single page-array object */
+static struct cr_pgarr *cr_pgarr_alloc_one(unsigned long flags)
+{
+ struct cr_pgarr *pgarr;
+
+ pgarr = kzalloc(sizeof(*pgarr), GFP_KERNEL);
+ if (!pgarr)
+ return NULL;
+
+ pgarr->vaddrs = kmalloc(CR_PGARR_TOTAL * sizeof(unsigned long),
+ GFP_KERNEL);
+ if (!pgarr->vaddrs)
+ goto nomem;
+
+ /* pgarr->pages is needed only for checkpoint */
+ if (flags & CR_CTX_CKPT) {
+ pgarr->pages = kmalloc(CR_PGARR_TOTAL * sizeof(struct page *),
+ GFP_KERNEL);
+ if (!pgarr->pages)
+ goto nomem;
+ }
+
+ return pgarr;
+
+ nomem:
+ cr_pgarr_free_one(pgarr);
+ return NULL;
+}
+
+/* cr_pgarr_current - return the next available page-array in the chain
+ * @ctx: checkpoint context
+ *
+ * Returns the first page-array in the list that has space. Extends the
+ * list if none has space.
+ */
+struct cr_pgarr *cr_pgarr_current(struct cr_ctx *ctx)
+{
+ struct cr_pgarr *pgarr;
+
+ pgarr = cr_pgarr_first(ctx);
+ if (pgarr && !cr_pgarr_is_full(pgarr))
+ goto out;
+ pgarr = cr_pgarr_alloc_one(ctx->flags);
+ if (!pgarr)
+ goto out;
+ list_add(&pgarr->list, &ctx->pgarr_list);
+ out:
+ return pgarr;
+}
+
+/* reset the page-array chain (dropping page references if necessary) */
+void cr_pgarr_reset_all(struct cr_ctx *ctx)
+{
+ struct cr_pgarr *pgarr;
+
+ list_for_each_entry(pgarr, &ctx->pgarr_list, list) {
+ cr_pgarr_release_pages(pgarr);
+ pgarr->nr_used = 0;
+ }
+}
+
+/*
+ * Checkpoint is outside the context of the checkpointee, so one cannot
+ * simply read pages from user-space. Instead, we scan the address space
+ * of the target to cherry-pick pages of interest. Selected pages are
+ * enlisted in a page-array chain (attached to the checkpoint context).
+ * To save their contents, each page is mapped to kernel memory and then
+ * dumped to the file descriptor.
+ */
+
+
+/**
+ * cr_private_follow_page - return page pointer for dirty pages
+ * @vma - target vma
+ * @addr - page address
+ *
+ * Looks up the page that correspond to the address in the vma, and
+ * returns the page if it was modified (and grabs a reference to it),
+ * or otherwise returns NULL (or error).
+ *
+ * This function should _only_ called for private vma's.
+ */
+static struct page *
+cr_private_follow_page(struct vm_area_struct *vma, unsigned long addr)
+{
+ struct page *page;
+
+ BUG_ON(vma->vm_flags & (VM_SHARED | VM_MAYSHARE));
+
+ /*
+ * simplified version of get_user_pages(): already have vma,
+ * only need FOLL_ANON, and (for now) ignore fault stats.
+ *
+ * follow_page() will return NULL if the page is not present
+ * (swapped), ZERO_PAGE(0) if the pte wasn't allocated, and
+ * the actual page pointer otherwise.
+ *
+ * FIXME: consolidate with get_user_pages()
+ */
+
+ cond_resched();
+ while (!(page = follow_page(vma, addr, FOLL_ANON | FOLL_GET))) {
+ int ret;
+
+ /* the page is swapped out - bring it in (optimize ?) */
+ ret = handle_mm_fault(vma->vm_mm, vma, addr, 0);
+ if (ret & VM_FAULT_ERROR) {
+ if (ret & VM_FAULT_OOM)
+ return ERR_PTR(-ENOMEM);
+ else if (ret & VM_FAULT_SIGBUS)
+ return ERR_PTR(-EFAULT);
+ else
+ BUG();
+ break;
+ }
+ cond_resched();
+ }
+
+ if (IS_ERR(page))
+ return page;
+
+ /*
+ * We only care about dirty pages: either non-zero page, or
+ * file-backed (copy-on-write) that were touched. For the latter,
+ * the page_mapping() will be unset because it will no longer be
+ * mapped to the original file after having been modified.
+ */
+ if (page == ZERO_PAGE(0)) {
+ /* this is the zero page: ignore */
+ page_cache_release(page);
+ page = NULL;
+ } else if (vma->vm_file && (page_mapping(page) != NULL)) {
+ /* file backed clean cow: ignore */
+ page_cache_release(page);
+ page = NULL;
+ }
+
+ return page;
+}
+
+/**
+ * cr_private_vma_fill_pgarr - fill a page-array with addr/page tuples
+ * @ctx - checkpoint context
+ * @pgarr - page-array to fill
+ * @vma - vma to scan
+ * @start - start address (updated)
+ *
+ * Returns the number of pages collected
+ */
+static int
+cr_private_vma_fill_pgarr(struct cr_ctx *ctx, struct cr_pgarr *pgarr,
+ struct vm_area_struct *vma, unsigned long *start)
+{
+ unsigned long end = vma->vm_end;
+ unsigned long addr = *start;
+ int orig_used = pgarr->nr_used;
+
+ /* this function is only for private memory (anon or file-mapped) */
+ BUG_ON(vma->vm_flags & (VM_SHARED | VM_MAYSHARE));
+
+ while (addr < end) {
+ struct page *page;
+
+ page = cr_private_follow_page(vma, addr);
+ if (IS_ERR(page))
+ return PTR_ERR(page);
+
+ if (page) {
+ pgarr->pages[pgarr->nr_used] = page;
+ pgarr->vaddrs[pgarr->nr_used] = addr;
+ pgarr->nr_used++;
+ }
+
+ addr += PAGE_SIZE;
+
+ if (cr_pgarr_is_full(pgarr))
+ break;
+ }
+
+ *start = addr;
+ return pgarr->nr_used - orig_used;
+}
+
+/* dump contents of a pages: use kmap_atomic() to avoid TLB flush */
+static int cr_page_write(struct cr_ctx *ctx, struct page *page, char *buf)
+{
+ void *ptr;
+
+ ptr = kmap_atomic(page, KM_USER1);
+ memcpy(buf, ptr, PAGE_SIZE);
+ kunmap_atomic(ptr, KM_USER1);
+
+ return cr_kwrite(ctx, buf, PAGE_SIZE);
+}
+
+/**
+ * cr_vma_dump_pages - dump pages listed in the ctx page-array chain
+ * @ctx - checkpoint context
+ * @total - total number of pages
+ *
+ * First dump all virtual addresses, followed by the contents of all pages
+ */
+static int cr_vma_dump_pages(struct cr_ctx *ctx, int total)
+{
+ struct cr_pgarr *pgarr;
+ char *buf;
+ int i, ret = 0;
+
+ if (!total)
+ return 0;
+
+ list_for_each_entry_reverse(pgarr, &ctx->pgarr_list, list) {
+ ret = cr_kwrite(ctx, pgarr->vaddrs,
+ pgarr->nr_used * sizeof(*pgarr->vaddrs));
+ if (ret < 0)
+ return ret;
+ }
+
+ buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
+ if (!buf)
+ return -ENOMEM;
+
+ list_for_each_entry_reverse(pgarr, &ctx->pgarr_list, list) {
+ for (i = 0; i < pgarr->nr_used; i++) {
+ ret = cr_page_write(ctx, pgarr->pages[i], buf);
+ if (ret < 0)
+ goto out;
+ }
+ }
+
+ out:
+ kfree(buf);
+ return ret;
+}
+
+/**
+ * cr_write_private_vma_contents - dump contents of a VMA with private memory
+ * @ctx - checkpoint context
+ * @vma - vma to scan
+ *
+ * Collect lists of pages that needs to be dumped, and corresponding
+ * virtual addresses into ctx->pgarr_list page-array chain. Then dump
+ * the addresses, followed by the page contents.
+ */
+static int
+cr_write_private_vma_contents(struct cr_ctx *ctx, struct vm_area_struct *vma)
+{
+ struct cr_hdr h;
+ struct cr_hdr_pgarr *hh;
+ unsigned long addr = vma->vm_start;
+ struct cr_pgarr *pgarr;
+ unsigned long cnt = 0;
+ int ret;
+
+ /*
+ * Work iteratively, collecting and dumping at most CR_PGARR_CHUNK
+ * in each round. Each iterations is divided into two steps:
+ *
+ * (1) scan: scan through the PTEs of the vma to collect the pages
+ * to dump (later we'll also make them COW), while keeping a list
+ * of pages and their corresponding addresses on ctx->pgarr_list.
+ *
+ * (2) dump: write out a header specifying how many pages, followed
+ * by the addresses of all pages in ctx->pgarr_list, followed by
+ * the actual contents of all pages. (Then, release the references
+ * to the pages and reset the page-array chain).
+ *
+ * (This split makes the logic simpler by first counting the pages
+ * that need saving. More importantly, it allows for a future
+ * optimization that will reduce application downtime by deferring
+ * the actual write-out of the data to after the application is
+ * allowed to resume execution).
+ *
+ * After dumpting the entire contents, conclude with a header that
+ * specifies 0 pages to mark the end of the contents.
+ */
+
+ h.type = CR_HDR_PGARR;
+ h.len = sizeof(*hh);
+ h.parent = 0;
+
+ while (addr < vma->vm_end) {
+ pgarr = cr_pgarr_current(ctx);
+ if (!pgarr)
+ return -ENOMEM;
+ ret = cr_private_vma_fill_pgarr(ctx, pgarr, vma, &addr);
+ if (ret < 0)
+ return ret;
+ cnt += ret;
+
+ /* did we complete a chunk, or is this the last chunk ? */
+ if (cnt >= CR_PGARR_CHUNK || (cnt && addr == vma->vm_end)) {
+ hh = cr_hbuf_get(ctx, sizeof(*hh));
+ hh->nr_pages = cnt;
+ ret = cr_write_obj(ctx, &h, hh);
+ cr_hbuf_put(ctx, sizeof(*hh));
+ if (ret < 0)
+ return ret;
+
+ ret = cr_vma_dump_pages(ctx, cnt);
+ if (ret < 0)
+ return ret;
+
+ cr_pgarr_reset_all(ctx);
+ }
+ }
+
+ /* mark end of contents with header saying "0" pages */
+ hh = cr_hbuf_get(ctx, sizeof(*hh));
+ hh->nr_pages = 0;
+ ret = cr_write_obj(ctx, &h, hh);
+ cr_hbuf_put(ctx, sizeof(*hh));
+
+ return ret;
+}
+
+static int cr_write_vma(struct cr_ctx *ctx, struct vm_area_struct *vma)
+{
+ struct cr_hdr h;
+ struct cr_hdr_vma *hh = cr_hbuf_get(ctx, sizeof(*hh));
+ int vma_type, ret;
+
+ h.type = CR_HDR_VMA;
+ h.len = sizeof(*hh);
+ h.parent = 0;
+
+ hh->vm_start = vma->vm_start;
+ hh->vm_end = vma->vm_end;
+ hh->vm_page_prot = vma->vm_page_prot.pgprot;
+ hh->vm_flags = vma->vm_flags;
+ hh->vm_pgoff = vma->vm_pgoff;
+
+ if (vma->vm_flags & (VM_SHARED | VM_IO | VM_HUGETLB | VM_NONLINEAR)) {
+ pr_warning("c/r: unsupported VMA %#lx\n", vma->vm_flags);
+ cr_hbuf_put(ctx, sizeof(*hh));
+ return -ENOSYS;
+ }
+
+ /* by default assume anon memory */
+ vma_type = CR_VMA_ANON;
+
+ /*
+ * if there is a backing file, assume private-mapped
+ * (FIXME: check if the file is unlinked)
+ */
+ if (vma->vm_file)
+ vma_type = CR_VMA_FILE;
+
+ hh->vma_type = vma_type;
+
+ ret = cr_write_obj(ctx, &h, hh);
+ cr_hbuf_put(ctx, sizeof(*hh));
+ if (ret < 0)
+ return ret;
+
+ /* save the file name, if relevant */
+ if (vma->vm_file) {
+ ret = cr_write_fname(ctx, &vma->vm_file->f_path, ctx->vfsroot);
+ if (ret < 0)
+ return ret;
+ }
+
+ return cr_write_private_vma_contents(ctx, vma);
+}
+
+int cr_write_mm(struct cr_ctx *ctx, struct task_struct *t)
+{
+ struct cr_hdr h;
+ struct cr_hdr_mm *hh = cr_hbuf_get(ctx, sizeof(*hh));
+ struct mm_struct *mm;
+ struct vm_area_struct *vma;
+ int objref, ret;
+
+ h.type = CR_HDR_MM;
+ h.len = sizeof(*hh);
+ h.parent = task_pid_vnr(t);
+
+ mm = get_task_mm(t);
+
+ objref = 0; /* will be meaningful with multiple processes */
+ hh->objref = objref;
+
+ down_read(&mm->mmap_sem);
+
+ hh->start_code = mm->start_code;
+ hh->end_code = mm->end_code;
+ hh->start_data = mm->start_data;
+ hh->end_data = mm->end_data;
+ hh->start_brk = mm->start_brk;
+ hh->brk = mm->brk;
+ hh->start_stack = mm->start_stack;
+ hh->arg_start = mm->arg_start;
+ hh->arg_end = mm->arg_end;
+ hh->env_start = mm->env_start;
+ hh->env_end = mm->env_end;
+
+ hh->map_count = mm->map_count;
+
+ /* FIX: need also mm->flags */
+
+ ret = cr_write_obj(ctx, &h, hh);
+ cr_hbuf_put(ctx, sizeof(*hh));
+ if (ret < 0)
+ goto out;
+
+ /* write the vma's */
+ for (vma = mm->mmap; vma; vma = vma->vm_next) {
+ ret = cr_write_vma(ctx, vma);
+ if (ret < 0)
+ goto out;
+ }
+
+ ret = cr_write_mm_context(ctx, mm, objref);
+
+ out:
+ up_read(&mm->mmap_sem);
+ mmput(mm);
+ return ret;
+}
diff --git a/checkpoint/sys.c b/checkpoint/sys.c
index bd14ef9..455d124 100644
--- a/checkpoint/sys.c
+++ b/checkpoint/sys.c
@@ -16,6 +16,8 @@
#include <linux/capability.h>
#include <linux/checkpoint.h>
+#include "checkpoint_mem.h"
+
/*
* Helpers to write(read) from(to) kernel space to(from) the checkpoint
* image file descriptor (similar to how a core-dump is performed).
@@ -131,7 +133,14 @@ static void cr_ctx_free(struct cr_ctx *ctx)
{
if (ctx->file)
fput(ctx->file);
+
kfree(ctx->hbuf);
+
+ if (ctx->vfsroot)
+ path_put(ctx->vfsroot);
+
+ cr_pgarr_free(ctx);
+
kfree(ctx);
}
@@ -146,6 +155,8 @@ static struct cr_ctx *cr_ctx_alloc(int fd, unsigned long flags)
ctx->flags = flags;
+ INIT_LIST_HEAD(&ctx->pgarr_list);
+
err = -EBADF;
ctx->file = fget(fd);
if (!ctx->file)
diff --git a/include/linux/checkpoint.h b/include/linux/checkpoint.h
index 63f298f..f7898b1 100644
--- a/include/linux/checkpoint.h
+++ b/include/linux/checkpoint.h
@@ -10,6 +10,9 @@
* distribution for more details.
*/
+#include <linux/path.h>
+#include <linux/fs.h>
+
#define CR_VERSION 1
struct cr_ctx {
@@ -25,6 +28,10 @@ struct cr_ctx {
void *hbuf; /* temporary buffer for headers */
int hpos; /* position in headers buffer */
+
+ struct list_head pgarr_list; /* page array to dump VMA contents */
+
+ struct path *vfsroot; /* container root (FIXME) */
};
/* cr_ctx: flags */
@@ -42,6 +49,8 @@ struct cr_hdr;
extern int cr_write_obj(struct cr_ctx *ctx, struct cr_hdr *h, void *buf);
extern int cr_write_buffer(struct cr_ctx *ctx, void *buf, int len);
extern int cr_write_string(struct cr_ctx *ctx, char *str, int len);
+extern int cr_write_fname(struct cr_ctx *ctx,
+ struct path *path, struct path *root);
extern int cr_read_obj(struct cr_ctx *ctx, struct cr_hdr *h, void *buf, int n);
extern int cr_read_obj_type(struct cr_ctx *ctx, void *buf, int len, int type);
@@ -50,7 +59,10 @@ extern int cr_read_buffer(struct cr_ctx *ctx, void *buf, int *len);
extern int cr_read_string(struct cr_ctx *ctx, char *str, int len);
extern int do_checkpoint(struct cr_ctx *ctx, pid_t pid);
+extern int cr_write_mm(struct cr_ctx *ctx, struct task_struct *t);
+
extern int do_restart(struct cr_ctx *ctx, pid_t pid);
+extern int cr_read_mm(struct cr_ctx *ctx);
#define cr_debug(fmt, args...) \
pr_debug("[%d:c/r:%s] " fmt, task_pid_vnr(current), __func__, ## args)
diff --git a/include/linux/checkpoint_hdr.h b/include/linux/checkpoint_hdr.h
index b74b5f9..d78f0f1 100644
--- a/include/linux/checkpoint_hdr.h
+++ b/include/linux/checkpoint_hdr.h
@@ -34,6 +34,7 @@ enum {
CR_HDR_HEAD_ARCH,
CR_HDR_BUFFER,
CR_HDR_STRING,
+ CR_HDR_FNAME,
CR_HDR_TASK = 101,
CR_HDR_THREAD,
@@ -41,6 +42,7 @@ enum {
CR_HDR_MM = 201,
CR_HDR_VMA,
+ CR_HDR_PGARR,
CR_HDR_MM_CONTEXT,
CR_HDR_TAIL = 5001
@@ -75,4 +77,34 @@ struct cr_hdr_task {
__s32 task_comm_len;
} __attribute__((aligned(8)));
+struct cr_hdr_mm {
+ __u32 objref; /* identifier for shared objects */
+ __u32 map_count;
+
+ __u64 start_code, end_code, start_data, end_data;
+ __u64 start_brk, brk, start_stack;
+ __u64 arg_start, arg_end, env_start, env_end;
+} __attribute__((aligned(8)));
+
+/* vma subtypes */
+enum vm_type {
+ CR_VMA_ANON = 1,
+ CR_VMA_FILE
+};
+
+struct cr_hdr_vma {
+ __u32 vma_type;
+ __u32 _padding;
+
+ __u64 vm_start;
+ __u64 vm_end;
+ __u64 vm_page_prot;
+ __u64 vm_flags;
+ __u64 vm_pgoff;
+} __attribute__((aligned(8)));
+
+struct cr_hdr_pgarr {
+ __u64 nr_pages; /* number of pages to saved */
+} __attribute__((aligned(8)));
+
#endif /* _CHECKPOINT_CKPT_HDR_H_ */
--
1.5.4.3
Restore open file descriptors: for each FD read 'struct cr_hdr_fd_ent'
and lookup objref in the hash table; if not found (first occurence), read
in 'struct cr_hdr_fd_data', create a new FD and register in the hash.
Otherwise attach the file pointer from the hash as an FD.
This patch only handles basic FDs - regular files, directories and also
symbolic links.
Changelog[v6]:
- Balance all calls to cr_hbuf_get() with matching cr_hbuf_put()
(even though it's not really needed)
Signed-off-by: Oren Laadan <[email protected]>
Acked-by: Serge Hallyn <[email protected]>
Signed-off-by: Dave Hansen <[email protected]>
---
checkpoint/Makefile | 2 +-
checkpoint/restart.c | 4 +
checkpoint/rstr_file.c | 246 ++++++++++++++++++++++++++++++++++++++++++++
include/linux/checkpoint.h | 1 +
4 files changed, 252 insertions(+), 1 deletions(-)
create mode 100644 checkpoint/rstr_file.c
diff --git a/checkpoint/Makefile b/checkpoint/Makefile
index 7496695..88bbc10 100644
--- a/checkpoint/Makefile
+++ b/checkpoint/Makefile
@@ -3,4 +3,4 @@
#
obj-$(CONFIG_CHECKPOINT_RESTART) += sys.o checkpoint.o restart.o objhash.o \
- ckpt_mem.o rstr_mem.o ckpt_file.o
+ ckpt_mem.o rstr_mem.o ckpt_file.o rstr_file.o
diff --git a/checkpoint/restart.c b/checkpoint/restart.c
index d90c28a..22e7995 100644
--- a/checkpoint/restart.c
+++ b/checkpoint/restart.c
@@ -261,6 +261,10 @@ static int cr_read_task(struct cr_ctx *ctx)
cr_debug("memory: ret %d\n", ret);
if (ret < 0)
goto out;
+ ret = cr_read_files(ctx);
+ cr_debug("files: ret %d\n", ret);
+ if (ret < 0)
+ goto out;
ret = cr_read_thread(ctx);
cr_debug("thread: ret %d\n", ret);
if (ret < 0)
diff --git a/checkpoint/rstr_file.c b/checkpoint/rstr_file.c
new file mode 100644
index 0000000..d8ce20d
--- /dev/null
+++ b/checkpoint/rstr_file.c
@@ -0,0 +1,246 @@
+/*
+ * Checkpoint file descriptors
+ *
+ * Copyright (C) 2008 Oren Laadan
+ *
+ * This file is subject to the terms and conditions of the GNU General Public
+ * License. See the file COPYING in the main directory of the Linux
+ * distribution for more details.
+ */
+
+#include <linux/kernel.h>
+#include <linux/sched.h>
+#include <linux/fs.h>
+#include <linux/file.h>
+#include <linux/fdtable.h>
+#include <linux/fsnotify.h>
+#include <linux/syscalls.h>
+#include <linux/checkpoint.h>
+#include <linux/checkpoint_hdr.h>
+
+#include "checkpoint_file.h"
+
+static int cr_close_all_fds(struct files_struct *files)
+{
+ int *fdtable;
+ int nfds;
+
+ nfds = cr_scan_fds(files, &fdtable);
+ if (nfds < 0)
+ return nfds;
+ while (nfds--)
+ sys_close(fdtable[nfds]);
+ kfree(fdtable);
+ return 0;
+}
+
+/**
+ * cr_attach_file - attach a lonely file ptr to a file descriptor
+ * @file: lonely file pointer
+ */
+static int cr_attach_file(struct file *file)
+{
+ int fd = get_unused_fd_flags(0);
+
+ if (fd >= 0) {
+ fsnotify_open(file->f_path.dentry);
+ fd_install(fd, file);
+ }
+ return fd;
+}
+
+/**
+ * cr_attach_get_file - attach (and get) lonely file ptr to a file descriptor
+ * @file: lonely file pointer
+ */
+static int cr_attach_get_file(struct file *file)
+{
+ int fd = get_unused_fd_flags(0);
+
+ if (fd >= 0) {
+ fsnotify_open(file->f_path.dentry);
+ fd_install(fd, file);
+ get_file(file);
+ }
+ return fd;
+}
+
+#define CR_SETFL_MASK (O_APPEND|O_NONBLOCK|O_NDELAY|FASYNC|O_DIRECT|O_NOATIME)
+
+/* cr_read_fd_data - restore the state of a given file pointer */
+static int
+cr_read_fd_data(struct cr_ctx *ctx, struct files_struct *files, int rparent)
+{
+ struct cr_hdr_fd_data *hh = cr_hbuf_get(ctx, sizeof(*hh));
+ struct file *file;
+ int parent, ret;
+ int fd = 0; /* pacify gcc warning */
+
+ parent = cr_read_obj_type(ctx, hh, sizeof(*hh), CR_HDR_FD_DATA);
+ cr_debug("rparent %d parent %d flags %#x mode %#x how %d\n",
+ rparent, parent, hh->f_flags, hh->f_mode, hh->fd_type);
+ if (parent < 0) {
+ ret = parent;
+ goto out;
+ }
+
+ ret = -EINVAL;
+
+ if (parent != rparent)
+ goto out;
+
+ /* FIX: more sanity checks on f_flags, f_mode etc */
+
+ switch (hh->fd_type) {
+ case CR_FD_FILE:
+ case CR_FD_DIR:
+ case CR_FD_LINK:
+ file = cr_read_open_fname(ctx, hh->f_flags, hh->f_mode);
+ break;
+ default:
+ goto out;
+ }
+
+ if (IS_ERR(file)) {
+ ret = PTR_ERR(file);
+ goto out;
+ }
+
+ /* FIX: need to restore uid, gid, owner etc */
+
+ fd = cr_attach_file(file); /* no need to cleanup 'file' below */
+ if (fd < 0) {
+ filp_close(file, NULL);
+ ret = fd;
+ goto out;
+ }
+
+ /* register new <objref, file> tuple in hash table */
+ ret = cr_obj_add_ref(ctx, file, parent, CR_OBJ_FILE, 0);
+ if (ret < 0)
+ goto out;
+ ret = sys_fcntl(fd, F_SETFL, hh->f_flags & CR_SETFL_MASK);
+ if (ret < 0)
+ goto out;
+ ret = vfs_llseek(file, hh->f_pos, SEEK_SET);
+ if (ret == -ESPIPE) /* ignore error on non-seekable files */
+ ret = 0;
+
+ ret = 0;
+ out:
+ cr_hbuf_put(ctx, sizeof(*hh));
+ return ret < 0 ? ret : fd;
+}
+
+/**
+ * cr_read_fd_ent - restore the state of a given file descriptor
+ * @ctx: checkpoint context
+ * @files: files_struct pointer
+ * @parent: parent objref
+ *
+ * Restores the state of a file descriptor; looks up the objref (in the
+ * header) in the hash table, and if found picks the matching file and
+ * use it; otherwise calls cr_read_fd_data to restore the file too.
+ */
+static int
+cr_read_fd_ent(struct cr_ctx *ctx, struct files_struct *files, int rparent)
+{
+ struct cr_hdr_fd_ent *hh = cr_hbuf_get(ctx, sizeof(*hh));
+ struct file *file;
+ int newfd, parent, ret;
+
+ parent = cr_read_obj_type(ctx, hh, sizeof(*hh), CR_HDR_FD_ENT);
+ cr_debug("rparent %d parent %d ref %d fd %d c.o.e %d\n",
+ rparent, parent, hh->objref, hh->fd, hh->close_on_exec);
+ if (parent < 0) {
+ ret = parent;
+ goto out;
+ }
+
+ ret = -EINVAL;
+
+ if (parent != rparent)
+ goto out;
+ if (hh->objref <= 0)
+ goto out;
+
+ file = cr_obj_get_by_ref(ctx, hh->objref, CR_OBJ_FILE);
+ if (IS_ERR(file)) {
+ ret = PTR_ERR(file);
+ goto out;
+ }
+
+ if (file) {
+ /* reuse file descriptor found in the hash table */
+ newfd = cr_attach_get_file(file);
+ } else {
+ /* create new file pointer (and register in hash table) */
+ newfd = cr_read_fd_data(ctx, files, hh->objref);
+ }
+
+ if (newfd < 0) {
+ ret = newfd;
+ goto out;
+ }
+
+ cr_debug("newfd got %d wanted %d\n", newfd, hh->fd);
+
+ /* if newfd isn't desired fd then reposition it */
+ if (newfd != hh->fd) {
+ ret = sys_dup2(newfd, hh->fd);
+ if (ret < 0)
+ goto out;
+ sys_close(newfd);
+ }
+
+ if (hh->close_on_exec)
+ set_close_on_exec(hh->fd, 1);
+
+ ret = 0;
+ out:
+ cr_hbuf_put(ctx, sizeof(*hh));
+ return ret;
+}
+
+int cr_read_files(struct cr_ctx *ctx)
+{
+ struct cr_hdr_files *hh = cr_hbuf_get(ctx, sizeof(*hh));
+ struct files_struct *files = current->files;
+ int i, parent, ret;
+
+ parent = cr_read_obj_type(ctx, hh, sizeof(*hh), CR_HDR_FILES);
+ if (parent < 0) {
+ ret = parent;
+ goto out;
+ }
+
+ ret = -EINVAL;
+#if 0 /* activate when containers are used */
+ if (parent != task_pid_vnr(current))
+ goto out;
+#endif
+ cr_debug("objref %d nfds %d\n", hh->objref, hh->nfds);
+ if (hh->objref < 0 || hh->nfds < 0)
+ goto out;
+
+ if (hh->nfds > sysctl_nr_open) {
+ ret = -EMFILE;
+ goto out;
+ }
+
+ /* point of no return -- close all file descriptors */
+ ret = cr_close_all_fds(files);
+ if (ret < 0)
+ goto out;
+
+ for (i = 0; i < hh->nfds; i++) {
+ ret = cr_read_fd_ent(ctx, files, hh->objref);
+ if (ret < 0)
+ break;
+ }
+
+ ret = 0;
+ out:
+ cr_hbuf_put(ctx, sizeof(*hh));
+ return ret;
+}
diff --git a/include/linux/checkpoint.h b/include/linux/checkpoint.h
index 74cac0b..2675410 100644
--- a/include/linux/checkpoint.h
+++ b/include/linux/checkpoint.h
@@ -88,6 +88,7 @@ extern int cr_write_files(struct cr_ctx *ctx, struct task_struct *t);
extern int do_restart(struct cr_ctx *ctx, pid_t pid);
extern int cr_read_mm(struct cr_ctx *ctx);
+extern int cr_read_files(struct cr_ctx *ctx);
#define cr_debug(fmt, args...) \
pr_debug("[%d:c/r:%s] " fmt, task_pid_vnr(current), __func__, ## args)
--
1.5.4.3
Checkpointing of multiple processes works by recording the tasks tree
structure below a given task (usually this task is the container init).
For a given task, do a DFS scan of the tasks tree and collect them
into an array (keeping a reference to each task). Using DFS simplifies
the recreation of tasks either in user space or kernel space. For each
task collected, test if it can be checkpointed, and save its pid, tgid,
and ppid.
The actual work is divided into two passes: a first scan counts the
tasks, then memory is allocated and a second scan fills the array.
The logic is suitable for creation of processes during restart either
in userspace or by the kernel.
Currently we ignore threads and zombies, as well as session ids.
Signed-off-by: Oren Laadan <[email protected]>
Acked-by: Serge Hallyn <[email protected]>
---
checkpoint/checkpoint.c | 228 +++++++++++++++++++++++++++++++++++++---
checkpoint/sys.c | 16 +++
include/linux/checkpoint.h | 3 +
include/linux/checkpoint_hdr.h | 13 ++-
4 files changed, 243 insertions(+), 17 deletions(-)
diff --git a/checkpoint/checkpoint.c b/checkpoint/checkpoint.c
index 3d0a366..b0a757b 100644
--- a/checkpoint/checkpoint.c
+++ b/checkpoint/checkpoint.c
@@ -225,19 +225,6 @@ static int cr_write_task(struct cr_ctx *ctx, struct task_struct *t)
{
int ret;
- /* TODO: verity that the task is frozen (unless self) */
-
- if (t->state == TASK_DEAD) {
- pr_warning("c/r: task may not be in state TASK_DEAD\n");
- return -EAGAIN;
- }
-
- if (!atomic_read(&t->may_checkpoint)) {
- pr_warning("c/r: task %d may not checkpoint\n",
- task_pid_vnr(t));
- return -EBUSY;
- }
-
ret = cr_write_task_struct(ctx, t);
cr_debug("task_struct: ret %d\n", ret);
if (ret < 0)
@@ -260,6 +247,205 @@ static int cr_write_task(struct cr_ctx *ctx, struct task_struct *t)
return ret;
}
+/* dump all tasks in ctx->tasks_arr[] */
+static int cr_write_all_tasks(struct cr_ctx *ctx)
+{
+ int n, ret = 0;
+
+ for (n = 0; n < ctx->tasks_nr; n++) {
+ cr_debug("dumping task #%d\n", n);
+ ret = cr_write_task(ctx, ctx->tasks_arr[n]);
+ if (ret < 0)
+ break;
+ }
+
+ return ret;
+}
+
+static int cr_may_checkpoint_task(struct task_struct *t, struct cr_ctx *ctx)
+{
+ cr_debug("check %d\n", task_pid_nr_ns(t, ctx->root_nsproxy->pid_ns));
+
+ if (t->state == TASK_DEAD) {
+ pr_warning("c/r: task %d is TASK_DEAD\n", task_pid_vnr(t));
+ return -EAGAIN;
+ }
+
+ if (!atomic_read(&t->may_checkpoint)) {
+ pr_warning("c/r: task %d uncheckpointable\n", task_pid_vnr(t));
+ return -EBUSY;
+ }
+
+ if (!ptrace_may_access(t, PTRACE_MODE_READ))
+ return -EPERM;
+
+ /* FIXME: verify that the task is frozen (unless self) */
+
+ /* FIXME: change this for nested containers */
+ if (task_nsproxy(t) != ctx->root_nsproxy)
+ return -EPERM;
+
+ return 0;
+}
+
+#define CR_HDR_PIDS_CHUNK 256
+
+static int cr_write_pids(struct cr_ctx *ctx)
+{
+ struct cr_hdr_pids *hh;
+ struct pid_namespace *ns;
+ struct task_struct *task;
+ struct task_struct **tasks_arr;
+ int tasks_nr, n, ret = 0, pos = 0;
+
+ ns = ctx->root_nsproxy->pid_ns;
+ tasks_arr = ctx->tasks_arr;
+ tasks_nr = ctx->tasks_nr;
+
+ hh = cr_hbuf_get(ctx, sizeof(*hh) * CR_HDR_PIDS_CHUNK);
+
+ while (tasks_nr > 0) {
+ rcu_read_lock();
+ for (n = min(tasks_nr, CR_HDR_PIDS_CHUNK); n; n--) {
+ task = tasks_arr[pos];
+
+ /* is this task cool ? */
+ ret = cr_may_checkpoint_task(task, ctx);
+ if (ret < 0) {
+ rcu_read_unlock();
+ goto out;
+ }
+ hh[pos].vpid = task_pid_nr_ns(task, ns);
+ hh[pos].vtgid = task_tgid_nr_ns(task, ns);
+ hh[pos].vppid = task_tgid_nr_ns(task->real_parent, ns);
+ cr_debug("task[%d]: vpid %d vtgid %d parent %d\n", pos,
+ hh[pos].vpid, hh[pos].vtgid, hh[pos].vppid);
+ pos++;
+ }
+ rcu_read_unlock();
+
+ n = min(tasks_nr, CR_HDR_PIDS_CHUNK);
+ ret = cr_kwrite(ctx, hh, n * sizeof(*hh));
+ if (ret < 0)
+ break;
+
+ tasks_nr -= n;
+ }
+ out:
+ cr_hbuf_put(ctx, sizeof(*hh));
+ return ret;
+}
+
+/* count number of tasks in tree (and optionally fill pid's in array) */
+static int cr_tree_count_tasks(struct cr_ctx *ctx)
+{
+ struct task_struct *root = ctx->root_task;
+ struct task_struct *task = root;
+ struct task_struct *parent = NULL;
+ struct task_struct **tasks_arr = ctx->tasks_arr;
+ int tasks_nr = ctx->tasks_nr;
+ int nr = 0;
+
+ read_lock(&tasklist_lock);
+
+ /* count tasks via DFS scan of the tree */
+ while (1) {
+ if (tasks_arr) {
+ /* unlikely, but ... */
+ if (nr == tasks_nr)
+ return -EBUSY; /* cleanup in cr_ctx_free() */
+ tasks_arr[nr] = task;
+ get_task_struct(task);
+ }
+
+ nr++;
+
+ /* if has children - proceed with child */
+ if (!list_empty(&task->children)) {
+ parent = task;
+ task = list_entry(task->children.next,
+ struct task_struct, sibling);
+ continue;
+ }
+
+ while (task != root) {
+ /* if has sibling - proceed with sibling */
+ if (!list_is_last(&task->sibling, &parent->children)) {
+ task = list_entry(task->sibling.next,
+ struct task_struct, sibling);
+ break;
+ }
+
+ /* else, trace back to parent and proceed */
+ task = parent;
+ parent = parent->real_parent;
+ }
+
+ if (task == root)
+ break;
+ }
+
+ read_unlock(&tasklist_lock);
+ return nr;
+}
+
+/*
+ * cr_build_tree - scan the tasks tree in DFS order and fill in array
+ * @ctx: checkpoint context
+ *
+ * Using DFS order simplifies the restart logic to re-create the tasks.
+ *
+ * On success, ctx->tasks_arr will be allocated and populated with all
+ * tasks (reference taken), and ctx->tasks_nr will hold the total count.
+ * The array is cleaned up by cr_ctx_free().
+ */
+static int cr_build_tree(struct cr_ctx *ctx)
+{
+ int n, m;
+
+ /* count tasks (no side effects) */
+ n = cr_tree_count_tasks(ctx);
+
+ ctx->tasks_nr = n;
+ ctx->tasks_arr = kzalloc(n * sizeof(*ctx->tasks_arr), GFP_KERNEL);
+ if (!ctx->tasks_arr)
+ return -ENOMEM;
+
+ /* count again (now will fill array) */
+ m = cr_tree_count_tasks(ctx);
+
+ /* unlikely, but ... (cleanup in cr_ctx_free) */
+ if (m < 0)
+ return m;
+ else if (m != n)
+ return -EBUSY;
+
+ return 0;
+}
+
+/* dump the array that describes the tasks tree */
+static int cr_write_tree(struct cr_ctx *ctx)
+{
+ struct cr_hdr h;
+ struct cr_hdr_tree *hh;
+ int ret;
+
+ h.type = CR_HDR_TREE;
+ h.len = sizeof(*hh);
+ h.parent = 0;
+
+ hh = cr_hbuf_get(ctx, sizeof(*hh));
+ hh->tasks_nr = ctx->tasks_nr;
+
+ ret = cr_write_obj(ctx, &h, hh);
+ cr_hbuf_put(ctx, sizeof(*hh));
+ if (ret < 0)
+ return ret;
+
+ ret = cr_write_pids(ctx);
+ return ret;
+}
+
static int cr_get_container(struct cr_ctx *ctx, pid_t pid)
{
struct task_struct *task = NULL;
@@ -277,7 +463,7 @@ static int cr_get_container(struct cr_ctx *ctx, pid_t pid)
if (!task)
goto out;
-#if 0 /* enable to use containers */
+#if 0 /* enable with containers */
if (!is_container_init(task)) {
err = -EINVAL;
goto out;
@@ -299,7 +485,7 @@ static int cr_get_container(struct cr_ctx *ctx, pid_t pid)
if (!nsproxy)
goto out;
- /* TODO: verify that the container is frozen */
+ /* FIXME: verify that the container is frozen */
ctx->root_task = task;
ctx->root_nsproxy = nsproxy;
@@ -342,12 +528,22 @@ int do_checkpoint(struct cr_ctx *ctx, pid_t pid)
ret = cr_ctx_checkpoint(ctx, pid);
if (ret < 0)
goto out;
+
+ ret = cr_build_tree(ctx);
+ if (ret < 0)
+ goto out;
+
ret = cr_write_head(ctx);
if (ret < 0)
goto out;
- ret = cr_write_task(ctx, ctx->root_task);
+ ret = cr_write_tree(ctx);
if (ret < 0)
goto out;
+
+ ret = cr_write_all_tasks(ctx);
+ if (ret < 0)
+ goto out;
+
ret = cr_write_tail(ctx);
if (ret < 0)
goto out;
diff --git a/checkpoint/sys.c b/checkpoint/sys.c
index dfe63ca..fbaf683 100644
--- a/checkpoint/sys.c
+++ b/checkpoint/sys.c
@@ -130,6 +130,19 @@ void cr_hbuf_put(struct cr_ctx *ctx, int n)
* restart operation, and persists until the operation is completed.
*/
+static void cr_task_arr_free(struct cr_ctx *ctx)
+{
+ int n;
+
+ for (n = 0; n < ctx->tasks_nr; n++) {
+ if (ctx->tasks_arr[n]) {
+ put_task_struct(ctx->tasks_arr[n]);
+ ctx->tasks_arr[n] = NULL;
+ }
+ }
+ kfree(ctx->tasks_arr);
+}
+
static void cr_ctx_free(struct cr_ctx *ctx)
{
if (ctx->file)
@@ -143,6 +156,9 @@ static void cr_ctx_free(struct cr_ctx *ctx)
cr_pgarr_free(ctx);
cr_objhash_free(ctx);
+ if (ctx->tasks_arr)
+ cr_task_arr_free(ctx);
+
if (ctx->root_nsproxy)
put_nsproxy(ctx->root_nsproxy);
if (ctx->root_task)
diff --git a/include/linux/checkpoint.h b/include/linux/checkpoint.h
index fc1873b..ac330f5 100644
--- a/include/linux/checkpoint.h
+++ b/include/linux/checkpoint.h
@@ -34,6 +34,9 @@ struct cr_ctx {
void *hbuf; /* temporary buffer for headers */
int hpos; /* position in headers buffer */
+ struct task_struct **tasks_arr; /* array of all tasks in container */
+ int tasks_nr; /* size of tasks array */
+
struct cr_objhash *objhash; /* hash for shared objects */
struct list_head pgarr_list; /* page array to dump VMA contents */
diff --git a/include/linux/checkpoint_hdr.h b/include/linux/checkpoint_hdr.h
index e92c37d..7ae2a77 100644
--- a/include/linux/checkpoint_hdr.h
+++ b/include/linux/checkpoint_hdr.h
@@ -36,7 +36,8 @@ enum {
CR_HDR_STRING,
CR_HDR_FNAME,
- CR_HDR_TASK = 101,
+ CR_HDR_TREE = 101,
+ CR_HDR_TASK,
CR_HDR_THREAD,
CR_HDR_CPU,
@@ -72,6 +73,16 @@ struct cr_hdr_tail {
__u64 magic;
} __attribute__((aligned(8)));
+struct cr_hdr_tree {
+ __u32 tasks_nr;
+} __attribute__((aligned(8)));
+
+struct cr_hdr_pids {
+ __s32 vpid;
+ __s32 vtgid;
+ __s32 vppid;
+} __attribute__((aligned(8)));
+
struct cr_hdr_task {
__u32 state;
__u32 exit_state;
--
1.5.4.3
Infrastructure to handle objects that may be shared and referenced by
multiple tasks or other objects, e..g open files, memory address space
etc.
The state of shared objects is saved once. On the first encounter, the
state is dumped and the object is assigned a unique identifier (objref)
and also stored in a hash table (indexed by its physical kenrel address).
>From then on the object will be found in the hash and only its identifier
is saved.
On restart the identifier is looked up in the hash table; if not found
then the state is read, the object is created, and added to the hash
table (this time indexed by its identifier). Otherwise, the object in
the hash table is used.
Changelog[v4]:
- Fix calculation of hash table size
Changelog[v3]:
- Use standard hlist_... for hash table
Signed-off-by: Oren Laadan <[email protected]>
Acked-by: Serge Hallyn <[email protected]>
Signed-off-by: Dave Hansen <[email protected]>
---
checkpoint/Makefile | 2 +-
checkpoint/objhash.c | 268 ++++++++++++++++++++++++++++++++++++++++++++
checkpoint/sys.c | 4 +
include/linux/checkpoint.h | 20 ++++
4 files changed, 293 insertions(+), 1 deletions(-)
create mode 100644 checkpoint/objhash.c
diff --git a/checkpoint/Makefile b/checkpoint/Makefile
index ac35033..9843fb9 100644
--- a/checkpoint/Makefile
+++ b/checkpoint/Makefile
@@ -2,5 +2,5 @@
# Makefile for linux checkpoint/restart.
#
-obj-$(CONFIG_CHECKPOINT_RESTART) += sys.o checkpoint.o restart.o \
+obj-$(CONFIG_CHECKPOINT_RESTART) += sys.o checkpoint.o restart.o objhash.o \
ckpt_mem.o rstr_mem.o
diff --git a/checkpoint/objhash.c b/checkpoint/objhash.c
new file mode 100644
index 0000000..05b1a1b
--- /dev/null
+++ b/checkpoint/objhash.c
@@ -0,0 +1,268 @@
+/*
+ * Checkpoint-restart - object hash infrastructure to manage shared objects
+ *
+ * Copyright (C) 2008 Oren Laadan
+ *
+ * This file is subject to the terms and conditions of the GNU General Public
+ * License. See the file COPYING in the main directory of the Linux
+ * distribution for more details.
+ */
+
+#include <linux/kernel.h>
+#include <linux/file.h>
+#include <linux/hash.h>
+#include <linux/checkpoint.h>
+
+struct cr_objref {
+ int objref;
+ void *ptr;
+ unsigned short type;
+ unsigned short flags;
+ struct hlist_node hash;
+};
+
+struct cr_objhash {
+ struct hlist_head *head;
+ int next_free_objref;
+};
+
+#define CR_OBJHASH_NBITS 10
+#define CR_OBJHASH_TOTAL (1UL << CR_OBJHASH_NBITS)
+
+static void cr_obj_ref_drop(struct cr_objref *obj)
+{
+ switch (obj->type) {
+ case CR_OBJ_FILE:
+ fput((struct file *) obj->ptr);
+ break;
+ default:
+ BUG();
+ }
+}
+
+static void cr_obj_ref_grab(struct cr_objref *obj)
+{
+ switch (obj->type) {
+ case CR_OBJ_FILE:
+ get_file((struct file *) obj->ptr);
+ break;
+ default:
+ BUG();
+ }
+}
+
+static void cr_objhash_clear(struct cr_objhash *objhash)
+{
+ struct hlist_head *h = objhash->head;
+ struct hlist_node *n, *t;
+ struct cr_objref *obj;
+ int i;
+
+ for (i = 0; i < CR_OBJHASH_TOTAL; i++) {
+ hlist_for_each_entry_safe(obj, n, t, &h[i], hash) {
+ cr_obj_ref_drop(obj);
+ kfree(obj);
+ }
+ }
+}
+
+void cr_objhash_free(struct cr_ctx *ctx)
+{
+ struct cr_objhash *objhash = ctx->objhash;
+
+ if (objhash) {
+ cr_objhash_clear(objhash);
+ kfree(objhash->head);
+ kfree(ctx->objhash);
+ ctx->objhash = NULL;
+ }
+}
+
+int cr_objhash_alloc(struct cr_ctx *ctx)
+{
+ struct cr_objhash *objhash;
+ struct hlist_head *head;
+
+ objhash = kzalloc(sizeof(*objhash), GFP_KERNEL);
+ if (!objhash)
+ return -ENOMEM;
+ head = kzalloc(CR_OBJHASH_TOTAL * sizeof(*head), GFP_KERNEL);
+ if (!head) {
+ kfree(objhash);
+ return -ENOMEM;
+ }
+
+ objhash->head = head;
+ objhash->next_free_objref = 1;
+
+ ctx->objhash = objhash;
+ return 0;
+}
+
+static struct cr_objref *cr_obj_find_by_ptr(struct cr_ctx *ctx, void *ptr)
+{
+ struct hlist_head *h;
+ struct hlist_node *n;
+ struct cr_objref *obj;
+
+ h = &ctx->objhash->head[hash_ptr(ptr, CR_OBJHASH_NBITS)];
+ hlist_for_each_entry(obj, n, h, hash)
+ if (obj->ptr == ptr)
+ return obj;
+ return NULL;
+}
+
+static struct cr_objref *cr_obj_find_by_objref(struct cr_ctx *ctx, int objref)
+{
+ struct hlist_head *h;
+ struct hlist_node *n;
+ struct cr_objref *obj;
+
+ h = &ctx->objhash->head[hash_ptr((void *) objref, CR_OBJHASH_NBITS)];
+ hlist_for_each_entry(obj, n, h, hash)
+ if (obj->objref == objref)
+ return obj;
+ return NULL;
+}
+
+/**
+ * cr_obj_new - allocate an object and add to the hash table
+ * @ctx: checkpoint context
+ * @ptr: pointer to object
+ * @objref: unique object reference
+ * @type: object type
+ * @flags: object flags
+ *
+ * Allocate an object referring to @ptr and add to the hash table.
+ * If @objref is zero, assign a unique object reference and use @ptr
+ * as a hash key [checkpoint]. Else use @objref as a key [restart].
+ */
+static struct cr_objref *cr_obj_new(struct cr_ctx *ctx, void *ptr, int objref,
+ unsigned short type, unsigned short flags)
+{
+ struct cr_objref *obj;
+ int i;
+
+ obj = kmalloc(sizeof(*obj), GFP_KERNEL);
+ if (!obj)
+ return NULL;
+
+ obj->ptr = ptr;
+ obj->type = type;
+ obj->flags = flags;
+
+ if (objref) {
+ /* use @objref to index (restart) */
+ obj->objref = objref;
+ i = hash_ptr((void *) objref, CR_OBJHASH_NBITS);
+ } else {
+ /* use @ptr to index, assign objref (checkpoint) */
+ obj->objref = ctx->objhash->next_free_objref++;;
+ i = hash_ptr(ptr, CR_OBJHASH_NBITS);
+ }
+
+ hlist_add_head(&obj->hash, &ctx->objhash->head[i]);
+ cr_obj_ref_grab(obj);
+ return obj;
+}
+
+/**
+ * cr_obj_add_ptr - add an object to the hash table if not already there
+ * @ctx: checkpoint context
+ * @ptr: pointer to object
+ * @objref: unique object reference [output]
+ * @type: object type
+ * @flags: object flags
+ *
+ * Look up the object pointed to by @ptr in the hash table. If it isn't
+ * already found there, then add the object to the table, and allocate a
+ * fresh unique object reference (objref). Fills the unique objref of
+ * the object into @objref.
+ * [This is used during checkpoint].
+ *
+ * Returns 0 if found, 1 if added, < 0 on error
+ */
+int cr_obj_add_ptr(struct cr_ctx *ctx, void *ptr, int *objref,
+ unsigned short type, unsigned short flags)
+{
+ struct cr_objref *obj;
+ int ret = 0;
+
+ obj = cr_obj_find_by_ptr(ctx, ptr);
+ if (!obj) {
+ obj = cr_obj_new(ctx, ptr, 0, type, flags);
+ if (!obj)
+ return -ENOMEM;
+ else
+ ret = 1;
+ } else if (obj->type != type) /* sanity check */
+ return -EINVAL;
+ *objref = obj->objref;
+ return ret;
+}
+
+/**
+ * cr_obj_add_ref - add an object with unique objref to the hash table
+ * @ctx: checkpoint context
+ * @ptr: pointer to object
+ * @objref: unique identifier - object reference
+ * @type: object type
+ * @flags: object flags
+ *
+ * Add the object pointer to by @ptr and identified by unique object
+ * reference given by @objref to the hash table (indexed by @objref).
+ * [This is used during restart].
+ */
+int cr_obj_add_ref(struct cr_ctx *ctx, void *ptr, int objref,
+ unsigned short type, unsigned short flags)
+{
+ struct cr_objref *obj;
+
+ obj = cr_obj_new(ctx, ptr, objref, type, flags);
+ return obj ? 0 : -ENOMEM;
+}
+
+/**
+ * cr_obj_get_by_ptr - find the unique object reference of an object
+ * @ctx: checkpoint context
+ * @ptr: pointer to object
+ * @type: object type
+ *
+ * Look up the unique object reference (objref) of the object pointed
+ * to by @ptr, and return that number, or 0 if not found.
+ * [This is used during checkpoint].
+ */
+int cr_obj_get_by_ptr(struct cr_ctx *ctx, void *ptr, unsigned short type)
+{
+ struct cr_objref *obj;
+
+ obj = cr_obj_find_by_ptr(ctx, ptr);
+ if (!obj)
+ return -ESRCH;
+ if (obj->type != type)
+ return -EINVAL;
+ return obj->objref;
+}
+
+/**
+ * cr_obj_get_by_ref - find an object given its unique object reference
+ * @ctx: checkpoint context
+ * @objref: unique identifier - object reference
+ * @type: object type
+ *
+ * Look up the object who is identified by unique object reference that
+ * is specified by @objref, and return a pointer to that matching object,
+ * or NULL if not found.
+ * [This is used during restart].
+ */
+void *cr_obj_get_by_ref(struct cr_ctx *ctx, int objref, unsigned short type)
+{
+ struct cr_objref *obj;
+
+ obj = cr_obj_find_by_objref(ctx, objref);
+ if (!obj)
+ return NULL;
+ if (obj->type != type)
+ return ERR_PTR(-EINVAL);
+ return obj->ptr;
+}
diff --git a/checkpoint/sys.c b/checkpoint/sys.c
index 455d124..b640bee 100644
--- a/checkpoint/sys.c
+++ b/checkpoint/sys.c
@@ -140,6 +140,7 @@ static void cr_ctx_free(struct cr_ctx *ctx)
path_put(ctx->vfsroot);
cr_pgarr_free(ctx);
+ cr_objhash_free(ctx);
kfree(ctx);
}
@@ -167,6 +168,9 @@ static struct cr_ctx *cr_ctx_alloc(int fd, unsigned long flags)
if (!ctx->hbuf)
goto err;
+ if (cr_objhash_alloc(ctx) < 0)
+ goto err;
+
return ctx;
err:
diff --git a/include/linux/checkpoint.h b/include/linux/checkpoint.h
index 331968d..deccee5 100644
--- a/include/linux/checkpoint.h
+++ b/include/linux/checkpoint.h
@@ -29,6 +29,8 @@ struct cr_ctx {
void *hbuf; /* temporary buffer for headers */
int hpos; /* position in headers buffer */
+ struct cr_objhash *objhash; /* hash for shared objects */
+
struct list_head pgarr_list; /* page array to dump VMA contents */
struct path *vfsroot; /* container root (FIXME) */
@@ -44,6 +46,24 @@ extern int cr_kread(struct cr_ctx *ctx, void *buf, int count);
extern void *cr_hbuf_get(struct cr_ctx *ctx, int n);
extern void cr_hbuf_put(struct cr_ctx *ctx, int n);
+/* shared objects handling */
+
+enum {
+ CR_OBJ_FILE = 1,
+ CR_OBJ_MAX
+};
+
+extern void cr_objhash_free(struct cr_ctx *ctx);
+extern int cr_objhash_alloc(struct cr_ctx *ctx);
+extern void *cr_obj_get_by_ref(struct cr_ctx *ctx,
+ int objref, unsigned short type);
+extern int cr_obj_get_by_ptr(struct cr_ctx *ctx,
+ void *ptr, unsigned short type);
+extern int cr_obj_add_ptr(struct cr_ctx *ctx, void *ptr, int *objref,
+ unsigned short type, unsigned short flags);
+extern int cr_obj_add_ref(struct cr_ctx *ctx, void *ptr, int objref,
+ unsigned short type, unsigned short flags);
+
struct cr_hdr;
extern int cr_write_obj(struct cr_ctx *ctx, struct cr_hdr *h, void *buf);
--
1.5.4.3
Restoring the memory address space begins with nuking the existing one
of the current process, and then reading the VMA state and contents.
Call do_mmap_pgoffset() for each VMA and then read in the data.
Changelog[v9]:
- Introduce cr_ctx_checkpoint() for checkpoint-specific ctx setup
Changelog[v7]:
- Fix argument given to kunmap_atomic() in memory dump/restore
Changelog[v6]:
- Balance all calls to cr_hbuf_get() with matching cr_hbuf_put()
(even though it's not really needed)
Changelog[v5]:
- Improve memory restore code (following Dave Hansen's comments)
- Change dump format (and code) to allow chunks of <vaddrs, pages>
instead of one long list of each
- Memory restore now maps user pages explicitly to copy data into them,
instead of reading directly to user space; got rid of mprotect_fixup()
Changelog[v4]:
- Use standard list_... for cr_pgarr
Signed-off-by: Oren Laadan <[email protected]>
Acked-by: Serge Hallyn <[email protected]>
Signed-off-by: Dave Hansen <[email protected]>
---
arch/x86/include/asm/checkpoint_hdr.h | 5 +
arch/x86/mm/restart.c | 64 ++++++-
checkpoint/Makefile | 2 +-
checkpoint/checkpoint_arch.h | 2 +
checkpoint/checkpoint_mem.h | 5 +
checkpoint/restart.c | 51 +++++
checkpoint/rstr_mem.c | 384 +++++++++++++++++++++++++++++++++
include/linux/checkpoint.h | 4 +
8 files changed, 514 insertions(+), 3 deletions(-)
create mode 100644 checkpoint/rstr_mem.c
diff --git a/arch/x86/include/asm/checkpoint_hdr.h b/arch/x86/include/asm/checkpoint_hdr.h
index 33f4c70..d13db9b 100644
--- a/arch/x86/include/asm/checkpoint_hdr.h
+++ b/arch/x86/include/asm/checkpoint_hdr.h
@@ -87,4 +87,9 @@ struct cr_hdr_mm_context {
__s16 nldt;
} __attribute__((aligned(8)));
+#ifdef __KERNEL__
+/* misc prototypes from kernel (not defined elsewhere) */
+asmlinkage int sys_modify_ldt(int func, void __user *ptr, unsigned long bytecount);
+#endif
+
#endif /* __ASM_X86_CKPT_HDR__H */
diff --git a/arch/x86/mm/restart.c b/arch/x86/mm/restart.c
index 45ad790..ab68b2f 100644
--- a/arch/x86/mm/restart.c
+++ b/arch/x86/mm/restart.c
@@ -52,8 +52,10 @@ int cr_read_thread(struct cr_ctx *ctx)
size = sizeof(*desc) * GDT_ENTRY_TLS_ENTRIES;
desc = kmalloc(size, GFP_KERNEL);
- if (!desc)
- return -ENOMEM;
+ if (!desc) {
+ ret = -ENOMEM;
+ goto out;
+ }
ret = cr_kread(ctx, desc, size);
if (ret >= 0) {
@@ -230,3 +232,61 @@ int cr_read_head_arch(struct cr_ctx *ctx)
return ret;
}
+
+int cr_read_mm_context(struct cr_ctx *ctx, struct mm_struct *mm, int rparent)
+{
+ struct cr_hdr_mm_context *hh = cr_hbuf_get(ctx, sizeof(*hh));
+ int n, parent, ret = -EINVAL;
+
+ parent = cr_read_obj_type(ctx, hh, sizeof(*hh), CR_HDR_MM_CONTEXT);
+ cr_debug("parent %d rparent %d nldt %d\n", parent, rparent, hh->nldt);
+ if (parent < 0) {
+ ret = parent;
+ goto out;
+ }
+ if (parent != rparent)
+ goto out;
+
+ if (hh->nldt < 0 || hh->ldt_entry_size != LDT_ENTRY_SIZE)
+ goto out;
+
+ /*
+ * to utilize the syscall modify_ldt() we first convert the data
+ * in the checkpoint image from 'struct desc_struct' to 'struct
+ * user_desc' with reverse logic of include/asm/desc.h:fill_ldt()
+ */
+
+ for (n = 0; n < hh->nldt; n++) {
+ struct user_desc info;
+ struct desc_struct desc;
+ mm_segment_t old_fs;
+
+ ret = cr_kread(ctx, &desc, LDT_ENTRY_SIZE);
+ if (ret < 0)
+ goto out;
+
+ info.entry_number = n;
+ info.base_addr = desc.base0 | (desc.base1 << 16);
+ info.limit = desc.limit0;
+ info.seg_32bit = desc.d;
+ info.contents = desc.type >> 2;
+ info.read_exec_only = (desc.type >> 1) ^ 1;
+ info.limit_in_pages = desc.g;
+ info.seg_not_present = desc.p ^ 1;
+ info.useable = desc.avl;
+
+ old_fs = get_fs();
+ set_fs(get_ds());
+ ret = sys_modify_ldt(1, (struct user_desc __user *) &info,
+ sizeof(info));
+ set_fs(old_fs);
+
+ if (ret < 0)
+ goto out;
+ }
+
+ ret = 0;
+ out:
+ cr_hbuf_put(ctx, sizeof(*hh));
+ return ret;
+}
diff --git a/checkpoint/Makefile b/checkpoint/Makefile
index 3a0df6d..ac35033 100644
--- a/checkpoint/Makefile
+++ b/checkpoint/Makefile
@@ -3,4 +3,4 @@
#
obj-$(CONFIG_CHECKPOINT_RESTART) += sys.o checkpoint.o restart.o \
- ckpt_mem.o
+ ckpt_mem.o rstr_mem.o
diff --git a/checkpoint/checkpoint_arch.h b/checkpoint/checkpoint_arch.h
index f06c7eb..39c8224 100644
--- a/checkpoint/checkpoint_arch.h
+++ b/checkpoint/checkpoint_arch.h
@@ -9,3 +9,5 @@ extern int cr_write_mm_context(struct cr_ctx *ctx,
extern int cr_read_head_arch(struct cr_ctx *ctx);
extern int cr_read_thread(struct cr_ctx *ctx);
extern int cr_read_cpu(struct cr_ctx *ctx);
+extern int cr_read_mm_context(struct cr_ctx *ctx,
+ struct mm_struct *mm, int parent);
diff --git a/checkpoint/checkpoint_mem.h b/checkpoint/checkpoint_mem.h
index 85546f4..85a5cf3 100644
--- a/checkpoint/checkpoint_mem.h
+++ b/checkpoint/checkpoint_mem.h
@@ -38,4 +38,9 @@ static inline int cr_pgarr_is_full(struct cr_pgarr *pgarr)
return (pgarr->nr_used == CR_PGARR_TOTAL);
}
+static inline int cr_pgarr_nr_free(struct cr_pgarr *pgarr)
+{
+ return CR_PGARR_TOTAL - pgarr->nr_used;
+}
+
#endif /* _CHECKPOINT_CKPT_MEM_H_ */
diff --git a/checkpoint/restart.c b/checkpoint/restart.c
index d74d755..d90c28a 100644
--- a/checkpoint/restart.c
+++ b/checkpoint/restart.c
@@ -120,6 +120,44 @@ int cr_read_string(struct cr_ctx *ctx, char *str, int len)
return ret;
}
+/**
+ * cr_read_fname - read a file name
+ * @ctx: checkpoint context
+ * @fname: buffer
+ * @n: buffer length
+ */
+int cr_read_fname(struct cr_ctx *ctx, char *fname, int flen)
+{
+ return cr_read_buf_type(ctx, fname, &flen, CR_HDR_FNAME);
+}
+
+/**
+ * cr_read_open_fname - read a file name and open a file
+ * @ctx: checkpoint context
+ * @flags: file flags
+ * @mode: file mode
+ */
+struct file *cr_read_open_fname(struct cr_ctx *ctx, int flags, int mode)
+{
+ struct file *file;
+ char *fname;
+ int ret;
+
+ fname = kmalloc(PATH_MAX, GFP_KERNEL);
+ if (!fname)
+ return ERR_PTR(-ENOMEM);
+
+ ret = cr_read_fname(ctx, fname, PATH_MAX);
+ cr_debug("fname '%s' flags %#x mode %#x\n", fname, flags, mode);
+ if (ret >= 0)
+ file = filp_open(fname, flags, mode);
+ else
+ file = ERR_PTR(ret);
+
+ kfree(fname);
+ return file;
+}
+
/* read the checkpoint header */
static int cr_read_head(struct cr_ctx *ctx)
{
@@ -219,6 +257,10 @@ static int cr_read_task(struct cr_ctx *ctx)
cr_debug("task_struct: ret %d\n", ret);
if (ret < 0)
goto out;
+ ret = cr_read_mm(ctx);
+ cr_debug("memory: ret %d\n", ret);
+ if (ret < 0)
+ goto out;
ret = cr_read_thread(ctx);
cr_debug("thread: ret %d\n", ret);
if (ret < 0)
@@ -230,10 +272,19 @@ static int cr_read_task(struct cr_ctx *ctx)
return ret;
}
+/* setup restart-specific parts of ctx */
+static int cr_ctx_restart(struct cr_ctx *ctx)
+{
+ return 0;
+}
+
int do_restart(struct cr_ctx *ctx, pid_t pid)
{
int ret;
+ ret = cr_ctx_restart(ctx);
+ if (ret < 0)
+ goto out;
ret = cr_read_head(ctx);
if (ret < 0)
goto out;
diff --git a/checkpoint/rstr_mem.c b/checkpoint/rstr_mem.c
new file mode 100644
index 0000000..6713f4f
--- /dev/null
+++ b/checkpoint/rstr_mem.c
@@ -0,0 +1,384 @@
+/*
+ * Restart memory contents
+ *
+ * Copyright (C) 2008 Oren Laadan
+ *
+ * This file is subject to the terms and conditions of the GNU General Public
+ * License. See the file COPYING in the main directory of the Linux
+ * distribution for more details.
+ */
+
+#include <linux/kernel.h>
+#include <linux/sched.h>
+#include <linux/fcntl.h>
+#include <linux/file.h>
+#include <linux/fs.h>
+#include <linux/pagemap.h>
+#include <linux/mm_types.h>
+#include <linux/mman.h>
+#include <linux/mm.h>
+#include <linux/err.h>
+#include <linux/checkpoint.h>
+#include <linux/checkpoint_hdr.h>
+
+#include "checkpoint_arch.h"
+#include "checkpoint_mem.h"
+
+/*
+ * Unlike checkpoint, restart is executed in the context of each restarting
+ * process: vma regions are restored via a call to mmap(), and the data is
+ * read into the address space of the current process.
+ */
+
+
+/**
+ * cr_read_pages_vaddrs - read addresses of pages to page-array chain
+ * @ctx - restart context
+ * @nr_pages - number of address to read
+ */
+static int cr_read_pages_vaddrs(struct cr_ctx *ctx, unsigned long nr_pages)
+{
+ struct cr_pgarr *pgarr;
+ unsigned long *vaddrp;
+ int nr, ret;
+
+ while (nr_pages) {
+ pgarr = cr_pgarr_current(ctx);
+ if (!pgarr)
+ return -ENOMEM;
+ nr = cr_pgarr_nr_free(pgarr);
+ if (nr > nr_pages)
+ nr = nr_pages;
+ vaddrp = &pgarr->vaddrs[pgarr->nr_used];
+ ret = cr_kread(ctx, vaddrp, nr * sizeof(unsigned long));
+ if (ret < 0)
+ return ret;
+ pgarr->nr_used += nr;
+ nr_pages -= nr;
+ }
+ return 0;
+}
+
+static int cr_page_read(struct cr_ctx *ctx, struct page *page, char *buf)
+{
+ void *ptr;
+ int ret;
+
+ ret = cr_kread(ctx, buf, PAGE_SIZE);
+ if (ret < 0)
+ return ret;
+
+ ptr = kmap_atomic(page, KM_USER1);
+ memcpy(ptr, buf, PAGE_SIZE);
+ kunmap_atomic(ptr, KM_USER1);
+
+ return 0;
+}
+
+/**
+ * cr_read_pages_contents - read in data of pages in page-array chain
+ * @ctx - restart context
+ */
+static int cr_read_pages_contents(struct cr_ctx *ctx)
+{
+ struct mm_struct *mm = current->mm;
+ struct cr_pgarr *pgarr;
+ unsigned long *vaddrs;
+ char *buf;
+ int i, ret = 0;
+
+ buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
+ if (!buf)
+ return -ENOMEM;
+
+ down_read(&mm->mmap_sem);
+ list_for_each_entry_reverse(pgarr, &ctx->pgarr_list, list) {
+ vaddrs = pgarr->vaddrs;
+ for (i = 0; i < pgarr->nr_used; i++) {
+ struct page *page;
+
+ ret = get_user_pages(current, mm, vaddrs[i],
+ 1, 1, 1, &page, NULL);
+ if (ret < 0)
+ goto out;
+
+ ret = cr_page_read(ctx, page, buf);
+ page_cache_release(page);
+
+ if (ret < 0)
+ goto out;
+ }
+ }
+
+ out:
+ up_read(&mm->mmap_sem);
+ kfree(buf);
+ return 0;
+}
+
+/**
+ * cr_read_private_vma_contents - restore contents of a VMA with private memory
+ * @ctx - restart context
+ *
+ * Reads a header that specifies how many pages will follow, then reads
+ * a list of virtual addresses into ctx->pgarr_list page-array chain,
+ * followed by the actual contents of the corresponding pages. Iterates
+ * these steps until reaching a header specifying "0" pages, which marks
+ * the end of the contents.
+ */
+static int cr_read_private_vma_contents(struct cr_ctx *ctx)
+{
+ struct cr_hdr_pgarr *hh;
+ unsigned long nr_pages;
+ int parent, ret = 0;
+
+ while (1) {
+ hh = cr_hbuf_get(ctx, sizeof(*hh));
+ parent = cr_read_obj_type(ctx, hh, sizeof(*hh), CR_HDR_PGARR);
+ if (parent != 0) {
+ if (parent < 0)
+ ret = parent;
+ else
+ ret = -EINVAL;
+ cr_hbuf_put(ctx, sizeof(*hh));
+ break;
+ }
+
+ cr_debug("nr_pages %ld\n", (unsigned long) hh->nr_pages);
+
+ nr_pages = hh->nr_pages;
+ cr_hbuf_put(ctx, sizeof(*hh));
+
+ if (!nr_pages)
+ break;
+
+ ret = cr_read_pages_vaddrs(ctx, nr_pages);
+ if (ret < 0)
+ break;
+ ret = cr_read_pages_contents(ctx);
+ if (ret < 0)
+ break;
+ cr_pgarr_reset_all(ctx);
+ }
+
+ return ret;
+}
+
+/**
+ * cr_calc_map_prot_bits - convert vm_flags to mmap protection
+ * orig_vm_flags: source vm_flags
+ */
+static unsigned long cr_calc_map_prot_bits(unsigned long orig_vm_flags)
+{
+ unsigned long vm_prot = 0;
+
+ if (orig_vm_flags & VM_READ)
+ vm_prot |= PROT_READ;
+ if (orig_vm_flags & VM_WRITE)
+ vm_prot |= PROT_WRITE;
+ if (orig_vm_flags & VM_EXEC)
+ vm_prot |= PROT_EXEC;
+ if (orig_vm_flags & PROT_SEM) /* only (?) with IPC-SHM */
+ vm_prot |= PROT_SEM;
+
+ return vm_prot;
+}
+
+/**
+ * cr_calc_map_flags_bits - convert vm_flags to mmap flags
+ * orig_vm_flags: source vm_flags
+ */
+static unsigned long cr_calc_map_flags_bits(unsigned long orig_vm_flags)
+{
+ unsigned long vm_flags = 0;
+
+ vm_flags = MAP_FIXED;
+ if (orig_vm_flags & VM_GROWSDOWN)
+ vm_flags |= MAP_GROWSDOWN;
+ if (orig_vm_flags & VM_DENYWRITE)
+ vm_flags |= MAP_DENYWRITE;
+ if (orig_vm_flags & VM_EXECUTABLE)
+ vm_flags |= MAP_EXECUTABLE;
+ if (orig_vm_flags & VM_MAYSHARE)
+ vm_flags |= MAP_SHARED;
+ else
+ vm_flags |= MAP_PRIVATE;
+
+ return vm_flags;
+}
+
+static int cr_read_vma(struct cr_ctx *ctx, struct mm_struct *mm)
+{
+ struct cr_hdr_vma *hh = cr_hbuf_get(ctx, sizeof(*hh));
+ unsigned long vm_size, vm_start, vm_flags, vm_prot, vm_pgoff;
+ unsigned long addr;
+ struct file *file = NULL;
+ int parent, ret = -EINVAL;
+
+ parent = cr_read_obj_type(ctx, hh, sizeof(*hh), CR_HDR_VMA);
+ if (parent < 0) {
+ ret = parent;
+ goto err;
+ } else if (parent != 0)
+ goto err;
+
+ cr_debug("vma %#lx-%#lx type %d\n", (unsigned long) hh->vm_start,
+ (unsigned long) hh->vm_end, (int) hh->vma_type);
+
+ if (hh->vm_end < hh->vm_start)
+ goto err;
+
+ vm_start = hh->vm_start;
+ vm_pgoff = hh->vm_pgoff;
+ vm_size = hh->vm_end - hh->vm_start;
+ vm_prot = cr_calc_map_prot_bits(hh->vm_flags);
+ vm_flags = cr_calc_map_flags_bits(hh->vm_flags);
+
+ switch (hh->vma_type) {
+
+ case CR_VMA_ANON: /* anonymous private mapping */
+ if (vm_flags & VM_SHARED)
+ goto err;
+ /*
+ * vm_pgoff for anonymous mapping is the "global" page
+ * offset (namely from addr 0x0), so we force a zero
+ */
+ vm_pgoff = 0;
+ break;
+
+ case CR_VMA_FILE: /* private mapping from a file */
+ if (vm_flags & VM_SHARED)
+ goto err;
+ /*
+ * for private mapping using 'read-only' is sufficient
+ */
+ file = cr_read_open_fname(ctx, O_RDONLY, 0);
+ if (IS_ERR(file)) {
+ ret = PTR_ERR(file);
+ goto err;
+ }
+ break;
+
+ default:
+ goto err;
+
+ }
+
+ cr_hbuf_put(ctx, sizeof(*hh));
+
+ down_write(&mm->mmap_sem);
+ addr = do_mmap_pgoff(file, vm_start, vm_size,
+ vm_prot, vm_flags, vm_pgoff);
+ up_write(&mm->mmap_sem);
+ cr_debug("size %#lx prot %#lx flag %#lx pgoff %#lx => %#lx\n",
+ vm_size, vm_prot, vm_flags, vm_pgoff, addr);
+
+ /* the file (if opened) is now referenced by the vma */
+ if (file)
+ filp_close(file, NULL);
+
+ if (IS_ERR((void *) addr))
+ return PTR_ERR((void *) addr);
+
+ /*
+ * CR_VMA_ANON: read in memory as is
+ * CR_VMA_FILE: read in memory as is
+ * (more to follow ...)
+ */
+
+ switch (hh->vma_type) {
+ case CR_VMA_ANON:
+ case CR_VMA_FILE:
+ /* standard case: read the data into the memory */
+ ret = cr_read_private_vma_contents(ctx);
+ break;
+ }
+
+ if (ret < 0)
+ return ret;
+
+ cr_debug("vma retval %d\n", ret);
+ return 0;
+
+ err:
+ cr_hbuf_put(ctx, sizeof(*hh));
+ return ret;
+}
+
+static int cr_destroy_mm(struct mm_struct *mm)
+{
+ struct vm_area_struct *vmnext = mm->mmap;
+ struct vm_area_struct *vma;
+ int ret;
+
+ while (vmnext) {
+ vma = vmnext;
+ vmnext = vmnext->vm_next;
+ ret = do_munmap(mm, vma->vm_start, vma->vm_end-vma->vm_start);
+ if (ret < 0) {
+ pr_debug("c/r: restart failed do_munmap (%d)\n", ret);
+ return ret;
+ }
+ }
+ return 0;
+}
+
+int cr_read_mm(struct cr_ctx *ctx)
+{
+ struct cr_hdr_mm *hh = cr_hbuf_get(ctx, sizeof(*hh));
+ struct mm_struct *mm;
+ int nr, parent, ret;
+
+ parent = cr_read_obj_type(ctx, hh, sizeof(*hh), CR_HDR_MM);
+ if (parent < 0) {
+ ret = parent;
+ goto out;
+ }
+
+ ret = -EINVAL;
+#if 0 /* activate when containers are used */
+ if (parent != task_pid_vnr(current))
+ goto out;
+#endif
+ cr_debug("map_count %d\n", hh->map_count);
+
+ /* XXX need more sanity checks */
+ if (hh->start_code > hh->end_code ||
+ hh->start_data > hh->end_data || hh->map_count < 0)
+ goto out;
+
+ mm = current->mm;
+
+ /* point of no return -- destruct current mm */
+ down_write(&mm->mmap_sem);
+ ret = cr_destroy_mm(mm);
+ if (ret < 0) {
+ up_write(&mm->mmap_sem);
+ goto out;
+ }
+ mm->start_code = hh->start_code;
+ mm->end_code = hh->end_code;
+ mm->start_data = hh->start_data;
+ mm->end_data = hh->end_data;
+ mm->start_brk = hh->start_brk;
+ mm->brk = hh->brk;
+ mm->start_stack = hh->start_stack;
+ mm->arg_start = hh->arg_start;
+ mm->arg_end = hh->arg_end;
+ mm->env_start = hh->env_start;
+ mm->env_end = hh->env_end;
+ up_write(&mm->mmap_sem);
+
+ /* FIX: need also mm->flags */
+
+ for (nr = hh->map_count; nr; nr--) {
+ ret = cr_read_vma(ctx, mm);
+ if (ret < 0)
+ goto out;
+ }
+
+ ret = cr_read_mm_context(ctx, mm, hh->objref);
+ out:
+ cr_hbuf_put(ctx, sizeof(*hh));
+ return ret;
+}
diff --git a/include/linux/checkpoint.h b/include/linux/checkpoint.h
index f7898b1..331968d 100644
--- a/include/linux/checkpoint.h
+++ b/include/linux/checkpoint.h
@@ -58,6 +58,10 @@ extern int cr_read_buf_type(struct cr_ctx *ctx, void *buf, int *len, int type);
extern int cr_read_buffer(struct cr_ctx *ctx, void *buf, int *len);
extern int cr_read_string(struct cr_ctx *ctx, char *str, int len);
+extern int cr_read_fname(struct cr_ctx *ctx, char *fname, int n);
+extern struct file *cr_read_open_fname(struct cr_ctx *ctx,
+ int flags, int mode);
+
extern int do_checkpoint(struct cr_ctx *ctx, pid_t pid);
extern int cr_write_mm(struct cr_ctx *ctx, struct task_struct *t);
--
1.5.4.3
Create trivial sys_checkpoint and sys_restore system calls. They will
enable to checkpoint and restart an entire container, to and from a
checkpoint image file descriptor.
The syscalls take a file descriptor (for the image file) and flags as
arguments. For sys_checkpoint the first argument identifies the target
container; for sys_restart it will identify the checkpoint image.
A checkpoint, much like a process coredump, dumps the state of multiple
processes at once, including the state of the container. The checkpoint
image is written to (and read from) the file descriptor directly from
the kernel. This way the data is generated and then pushed out naturally
as resources and tasks are scanned to save their state. This is the
approach taken by, e.g., Zap and OpenVZ.
By using a return value and not a file descriptor, we can distinguish
between a return from checkpoint, a return from restart (in case of a
checkpoint that includes self, i.e. a task checkpointing its own
container, or itself), and an error condition, in a manner analogous
to a fork() call.
We don't use copyin()/copyout() because it requires holding the entire
image in user space, and does not make sense for restart. Also, we
don't use a pipe, pseudo-fs file and the like, because they work by
generating data on demand as the user pulls it (unless the entire
image is buffered in the kernel) and would require more complex logic.
They also would significantly complicate checkpoint that includes self.
Changelog[v5]:
- Config is 'def_bool n' by default
Signed-off-by: Oren Laadan <[email protected]>
Acked-by: Serge Hallyn <[email protected]>
Signed-off-by: Dave Hansen <[email protected]>
---
arch/x86/include/asm/unistd_32.h | 2 +
arch/x86/kernel/syscall_table_32.S | 2 +
checkpoint/Kconfig | 11 +++++++++
checkpoint/Makefile | 5 ++++
checkpoint/sys.c | 41 ++++++++++++++++++++++++++++++++++++
include/linux/syscalls.h | 2 +
init/Kconfig | 2 +
kernel/sys_ni.c | 4 +++
8 files changed, 69 insertions(+), 0 deletions(-)
create mode 100644 checkpoint/Kconfig
create mode 100644 checkpoint/Makefile
create mode 100644 checkpoint/sys.c
diff --git a/arch/x86/include/asm/unistd_32.h b/arch/x86/include/asm/unistd_32.h
index f2bba78..a5f9e09 100644
--- a/arch/x86/include/asm/unistd_32.h
+++ b/arch/x86/include/asm/unistd_32.h
@@ -338,6 +338,8 @@
#define __NR_dup3 330
#define __NR_pipe2 331
#define __NR_inotify_init1 332
+#define __NR_checkpoint 333
+#define __NR_restart 334
#ifdef __KERNEL__
diff --git a/arch/x86/kernel/syscall_table_32.S b/arch/x86/kernel/syscall_table_32.S
index d44395f..5543136 100644
--- a/arch/x86/kernel/syscall_table_32.S
+++ b/arch/x86/kernel/syscall_table_32.S
@@ -332,3 +332,5 @@ ENTRY(sys_call_table)
.long sys_dup3 /* 330 */
.long sys_pipe2
.long sys_inotify_init1
+ .long sys_checkpoint
+ .long sys_restart
diff --git a/checkpoint/Kconfig b/checkpoint/Kconfig
new file mode 100644
index 0000000..ffaa635
--- /dev/null
+++ b/checkpoint/Kconfig
@@ -0,0 +1,11 @@
+config CHECKPOINT_RESTART
+ prompt "Enable checkpoint/restart (EXPERIMENTAL)"
+ def_bool n
+ depends on X86_32 && EXPERIMENTAL
+ help
+ Application checkpoint/restart is the ability to save the
+ state of a running application so that it can later resume
+ its execution from the time at which it was checkpointed.
+
+ Turning this option on will enable checkpoint and restart
+ functionality in the kernel.
diff --git a/checkpoint/Makefile b/checkpoint/Makefile
new file mode 100644
index 0000000..07d018b
--- /dev/null
+++ b/checkpoint/Makefile
@@ -0,0 +1,5 @@
+#
+# Makefile for linux checkpoint/restart.
+#
+
+obj-$(CONFIG_CHECKPOINT_RESTART) += sys.o
diff --git a/checkpoint/sys.c b/checkpoint/sys.c
new file mode 100644
index 0000000..375129c
--- /dev/null
+++ b/checkpoint/sys.c
@@ -0,0 +1,41 @@
+/*
+ * Generic container checkpoint-restart
+ *
+ * Copyright (C) 2008 Oren Laadan
+ *
+ * This file is subject to the terms and conditions of the GNU General Public
+ * License. See the file COPYING in the main directory of the Linux
+ * distribution for more details.
+ */
+
+#include <linux/sched.h>
+#include <linux/kernel.h>
+
+/**
+ * sys_checkpoint - checkpoint a container
+ * @pid: pid of the container init(1) process
+ * @fd: file to which dump the checkpoint image
+ * @flags: checkpoint operation flags
+ *
+ * Returns positive identifier on success, 0 when returning from restart
+ * or negative value on error
+ */
+asmlinkage long sys_checkpoint(pid_t pid, int fd, unsigned long flags)
+{
+ pr_debug("sys_checkpoint not implemented yet\n");
+ return -ENOSYS;
+}
+/**
+ * sys_restart - restart a container
+ * @crid: checkpoint image identifier
+ * @fd: file from which read the checkpoint image
+ * @flags: restart operation flags
+ *
+ * Returns negative value on error, or otherwise returns in the realm
+ * of the original checkpoint
+ */
+asmlinkage long sys_restart(int crid, int fd, unsigned long flags)
+{
+ pr_debug("sys_restart not implemented yet\n");
+ return -ENOSYS;
+}
diff --git a/include/linux/syscalls.h b/include/linux/syscalls.h
index d6ff145..edc218b 100644
--- a/include/linux/syscalls.h
+++ b/include/linux/syscalls.h
@@ -622,6 +622,8 @@ asmlinkage long sys_timerfd_gettime(int ufd, struct itimerspec __user *otmr);
asmlinkage long sys_eventfd(unsigned int count);
asmlinkage long sys_eventfd2(unsigned int count, int flags);
asmlinkage long sys_fallocate(int fd, int mode, loff_t offset, loff_t len);
+asmlinkage long sys_checkpoint(pid_t pid, int fd, unsigned long flags);
+asmlinkage long sys_restart(int crid, int fd, unsigned long flags);
int kernel_execve(const char *filename, char *const argv[], char *const envp[]);
diff --git a/init/Kconfig b/init/Kconfig
index 86b00c5..743e2ad 100644
--- a/init/Kconfig
+++ b/init/Kconfig
@@ -814,6 +814,8 @@ config MARKERS
source "arch/Kconfig"
+source "checkpoint/Kconfig"
+
endmenu # General setup
config HAVE_GENERIC_DMA_COHERENT
diff --git a/kernel/sys_ni.c b/kernel/sys_ni.c
index a77b27b..e4e289e 100644
--- a/kernel/sys_ni.c
+++ b/kernel/sys_ni.c
@@ -174,3 +174,7 @@ cond_syscall(compat_sys_timerfd_settime);
cond_syscall(compat_sys_timerfd_gettime);
cond_syscall(sys_eventfd);
cond_syscall(sys_eventfd2);
+
+/* checkpoint/restart */
+cond_syscall(sys_checkpoint);
+cond_syscall(sys_restart);
--
1.5.4.3
Covers application checkpoint/restart, overall design, interfaces,
usage, shared objects, and and checkpoint image format.
Changelog[v8]:
- Split into multiple files in Documentation/checkpoint/...
- Extend documentation, fix typos and comments from feedback
Signed-off-by: Oren Laadan <[email protected]>
Acked-by: Serge Hallyn <[email protected]>
Signed-off-by: Dave Hansen <[email protected]>
---
Documentation/checkpoint/ckpt.c | 32 +++++++
Documentation/checkpoint/internals.txt | 125 ++++++++++++++++++++++++++
Documentation/checkpoint/readme.txt | 104 ++++++++++++++++++++++
Documentation/checkpoint/rstr.c | 20 ++++
Documentation/checkpoint/security.txt | 38 ++++++++
Documentation/checkpoint/self.c | 57 ++++++++++++
Documentation/checkpoint/test.c | 48 ++++++++++
Documentation/checkpoint/usage.txt | 153 ++++++++++++++++++++++++++++++++
8 files changed, 577 insertions(+), 0 deletions(-)
create mode 100644 Documentation/checkpoint/ckpt.c
create mode 100644 Documentation/checkpoint/internals.txt
create mode 100644 Documentation/checkpoint/readme.txt
create mode 100644 Documentation/checkpoint/rstr.c
create mode 100644 Documentation/checkpoint/security.txt
create mode 100644 Documentation/checkpoint/self.c
create mode 100644 Documentation/checkpoint/test.c
create mode 100644 Documentation/checkpoint/usage.txt
diff --git a/Documentation/checkpoint/ckpt.c b/Documentation/checkpoint/ckpt.c
new file mode 100644
index 0000000..094408c
--- /dev/null
+++ b/Documentation/checkpoint/ckpt.c
@@ -0,0 +1,32 @@
+#include <stdio.h>
+#include <stdlib.h>
+#include <errno.h>
+#include <unistd.h>
+#include <sys/syscall.h>
+
+int main(int argc, char *argv[])
+{
+ pid_t pid;
+ int ret;
+
+ if (argc != 2) {
+ printf("usage: ckpt PID\n");
+ exit(1);
+ }
+
+ pid = atoi(argv[1]);
+ if (pid <= 0) {
+ printf("invalid pid\n");
+ exit(1);
+ }
+
+ ret = syscall(__NR_checkpoint, pid, STDOUT_FILENO, 0);
+
+ if (ret < 0)
+ perror("checkpoint");
+ else
+ printf("checkpoint id %d\n", ret);
+
+ return (ret > 0 ? 0 : 1);
+}
+
diff --git a/Documentation/checkpoint/internals.txt b/Documentation/checkpoint/internals.txt
new file mode 100644
index 0000000..03a92a2
--- /dev/null
+++ b/Documentation/checkpoint/internals.txt
@@ -0,0 +1,125 @@
+
+ ===== Internals of Checkpoint-Restart =====
+
+
+(1) Order of state dump
+
+The order of operations, both save and restore, is as follows:
+
+* Header section: header, container information, etc.
+
+* Global section: [TBD] global resources such as IPC, UTS, etc.
+
+* Process forest: [TBD] tasks and their relationships
+
+* Per task data (for each task):
+ -> task state: elements of task_struct
+ -> thread state: elements of thread_struct and thread_info
+ -> CPU state: registers etc, including FPU
+ -> memory state: memory address space layout and contents
+ -> filesystem state: [TBD] filesystem namespace state, chroot, cwd, etc
+ -> files state: open file descriptors and their state
+ -> signals state: [TBD] pending signals and signal handling state
+ -> credentials state: [TBD] user and group state, statistics
+
+
+(2) Checkpoint image format
+
+The checkpoint image format is composed of records consisting of a
+pre-header that identifies its contents, followed by a payload. (The
+idea here is to enable parallel checkpointing in the future in which
+multiple threads interleave data from multiple processes into a single
+stream).
+
+The pre-header is defined by "struct cr_hdr" as follows:
+
+struct cr_hdr {
+ __s16 type;
+ __s16 len;
+ __u32 parent;
+};
+
+'type' identifies the type of the payload, 'len' tells its length in
+bytes, and 'parent' identifies the owner object instance. The meaning
+of 'parent' varies depending on the type. For example, for CR_HDR_MM,
+'parent' identifies the task to which this MM belongs. The payload
+also varies depending on the type, for instance, the data describing a
+task_struct is given by a 'struct cr_hdr_task' (type CR_HDR_TASK) and
+so on.
+
+The format of the memory dump is as follows: for each VMA, there is a
+'struct cr_vma'; if the VMA is file-mapped, it is followed by the file
+name. Following comes the actual contents, in one or more chunks: each
+chunk begins with a header that specifies how many pages it holds,
+then the virtual addresses of all the dumped pages in that chunk,
+followed by the actual contents of all the dumped pages. A header with
+zero number of pages marks the end of the contents for a particular
+VMA. Then comes the next VMA and so on.
+
+To illustrate this, consider a single simple task with two VMAs: one
+is file mapped with two dumped pages, and the other is anonymous with
+three dumped pages. The checkpoint image will look like this:
+
+cr_hdr + cr_hdr_head
+cr_hdr + cr_hdr_task
+ cr_hdr + cr_hdr_mm
+ cr_hdr + cr_hdr_vma + cr_hdr + string
+ cr_hdr_pgarr (nr_pages = 2)
+ addr1, addr2
+ page1, page2
+ cr_hdr_pgarr (nr_pages = 0)
+ cr_hdr + cr_hdr_vma
+ cr_hdr_pgarr (nr_pages = 3)
+ addr3, addr4, addr5
+ page3, page4, page5
+ cr_hdr_pgarr (nr_pages = 0)
+ cr_hdr + cr_mm_context
+ cr_hdr + cr_hdr_thread
+ cr_hdr + cr_hdr_cpu
+cr_hdr + cr_hdr_tail
+
+
+(3) Shared resources (objects)
+
+Many resources used by tasks may be shared by more than one task (e.g.
+file descriptors, memory address space, etc), or even have multiple
+references from other resources (e.g. a single inode that represents
+two ends of a pipe).
+
+Clearly, the state of shared objects need only be saved once, even if
+they occur multiple times. We use a hash table (ctx->objhash) to keep
+track of shared objects and whether they were already saved. Shared
+objects are stored in a hash table as they appear, indexed by their
+kernel address. (The hash table itself is not saved as part of the
+checkpoint image: it is constructed dynamically during both checkpoint
+and restart, and discarded at the end of the operation).
+
+Each shared object that is found is first looked up in the hash table.
+On the first encounter, the object will not be found, so its state is
+dumped, and the object is assigned a unique identifier and also stored
+in the hash table. Subsequent lookups of that object in the hash table
+will yield that entry, and then only the unique identifier is saved,
+as opposed the entire state of the object.
+
+During restart, shared objects are seen by their unique identifiers as
+assigned during the checkpoint. Each shared object that it read in is
+first looked up in the hash table. On the first encounter it will not
+be found, meaning that the object needs to be created and its state
+read in and restored. Then the object is added to the hash table, this
+time indexed by its unique identifier. Subsequent lookups of the same
+unique identifier in the hash table will yield that entry, and then
+the existing object instance is reused instead of creating another one.
+
+The interface for the hash table is the following:
+
+cr_obj_get_by_ptr() - find the unique object reference (objref)
+ of the object that is pointer to by ptr [checkpoint]
+
+cr_obj_add_ptr() - add the object pointed to by ptr to the hash table
+ if not already there, and fill its unique object reference (objref)
+
+cr_obj_get_by_ref() - return the pointer to the object whose unique
+ object reference is equal to objref [restart]
+
+cr_obj_add_ref() - add the object with given unique object reference
+ (objref), pointed to by ptr to the hash table. [restart]
diff --git a/Documentation/checkpoint/readme.txt b/Documentation/checkpoint/readme.txt
new file mode 100644
index 0000000..c7f79e9
--- /dev/null
+++ b/Documentation/checkpoint/readme.txt
@@ -0,0 +1,104 @@
+
+ ===== Checkpoint-Restart support in the Linux kernel =====
+
+Copyright (C) 2008 Oren Laadan
+
+Author: Oren Laadan <[email protected]>
+
+License: The GNU Free Documentation License, Version 1.2
+ (dual licensed under the GPL v2)
+
+Reviewers: Serge Hallyn <[email protected]>
+ Dave Hansen <[email protected]>
+
+Application checkpoint/restart [C/R] is the ability to save the state
+of a running application so that it can later resume its execution
+from the time at which it was checkpointed. An application can be
+migrated by checkpointing it on one machine and restarting it on
+another. C/R can provide many potential benefits:
+
+* Failure recovery: by rolling back to a previous checkpoint
+
+* Improved response time: by restarting applications from checkpoints
+ instead of from scratch.
+
+* Improved system utilization: by suspending long running CPU
+ intensive jobs and resuming them when load decreases.
+
+* Fault resilience: by migrating applications off faulty hosts.
+
+* Dynamic load balancing: by migrating applications to less loaded
+ hosts.
+
+* Improved service availability and administration: by migrating
+ applications before host maintenance so that they continue to run
+ with minimal downtime
+
+* Time-travel: by taking periodic checkpoints and restarting from
+ any previous checkpoint.
+
+
+=== Overall design
+
+Checkpoint and restart is done in the kernel as much as possible. The
+kernel exports a relatively opaque 'blob' of data to userspace which can
+then be handed to the new kernel at restore time. The 'blob' contains
+data and state of select portions of kernel structures such as VMAs
+and mm_structs, as well as copies of the actual memory that the tasks
+use. Any changes in this blob's format between kernel revisions can be
+handled by an in-userspace conversion program. The approach is similar
+to virtually all of the commercial C/R products out there, as well as
+the research project Zap.
+
+Two new system calls are introduced to provide C/R: sys_checkpoint()
+and sys_restart(). The checkpoint code basically serializes internal
+kernel state and writes it out to a file descriptor, and the resulting
+image is stream-able. More specifically, it consists of 5 steps:
+
+1. Pre-dump
+2. Freeze the container
+3. Dump
+4. Thaw (or kill) the container
+5. Post-dump
+
+Steps 1 and 5 are an optimization to reduce application downtime. In
+particular, "pre-dump" works before freezing the container, e.g. the
+pre-copy for live migration, and "post-dump" works after the container
+resumes execution, e.g. write-back the data to secondary storage.
+
+The restart code basically reads the saved kernel state from a file
+descriptor, and re-creates the tasks and the resources they need to
+resume execution. The restart code is executed by each task that is
+restored in a new container to reconstruct its own state.
+
+
+=== Current Implementation
+
+* How useful is this code as it stands in real-world usage?
+
+Right now, the application must be a single process that does not
+share any resources with other processes. The only file descriptors
+that may be open are simple files and directories, they may not
+include devices, sockets or pipes.
+
+For an "external" checkpoint, the caller must first freeze (or stop)
+the target process. For "self" checkpoint, the application must be
+specifically written to use the new system calls. The restart does not
+yet preserve the pid of the original process, but will use whatever
+pid it was given by the kernel.
+
+What this means in practice is that it is useful for a simple
+application doing computational work and input/output from/to files.
+
+Currently, namespaces are not saved or restored. They will be treated
+as a class of a shared object. In particular, it is assumed that the
+task's file system namespace is the "root" for the entire container.
+It is also assumed that the same file system view is available for the
+restart task(s). Otherwise, a file system snapshot is required.
+
+* What additional work needs to be done to it?
+
+We know this design can work. We have two commercial products and a
+horde of academic projects doing it today using this basic design.
+We're early in this particular implementation because we're trying to
+release early and often.
diff --git a/Documentation/checkpoint/rstr.c b/Documentation/checkpoint/rstr.c
new file mode 100644
index 0000000..288209d
--- /dev/null
+++ b/Documentation/checkpoint/rstr.c
@@ -0,0 +1,20 @@
+#include <stdio.h>
+#include <stdlib.h>
+#include <unistd.h>
+#include <errno.h>
+#include <sys/syscall.h>
+
+int main(int argc, char *argv[])
+{
+ pid_t pid = getpid();
+ int ret;
+
+ ret = syscall(__NR_restart, pid, STDIN_FILENO, 0);
+ if (ret < 0)
+ perror("restart");
+
+ printf("should not reach here !\n");
+
+ return 0;
+}
+
diff --git a/Documentation/checkpoint/security.txt b/Documentation/checkpoint/security.txt
new file mode 100644
index 0000000..e5b4107
--- /dev/null
+++ b/Documentation/checkpoint/security.txt
@@ -0,0 +1,38 @@
+
+ ===== Security consideration for Checkpoint-Restart =====
+
+The main question is whether sys_checkpoint() and sys_restart()
+require privileged or unprivileged operation.
+
+Early versions checked capable(CAP_SYS_ADMIN) assuming that we would
+attempt to remove the need for privilege, so that all users could
+safely use it. Arnd Bergmann pointed out that it'd make more sense to
+let unprivileged users use them now, so that we'll be more careful
+about the security as patches roll in.
+
+Checkpoint: the main concern is whether a task that performs the
+checkpoint of another task has sufficient privileges to access its
+state. We address this by requiring that the checkpointer task will be
+able to ptrace the target task, by means of ptrace_may_access() with
+read mode.
+
+Restart: the main concern is that we may allow an unprivileged user to
+feed the kernel with random data. To this end, the restart works in a
+way that does not skip the usual security checks. Task credentials,
+i.e. euid, reuid, and LSM security contexts currently come from the
+caller, not the checkpoint image. When restoration of credentials
+becomes supported, then definitely the ability of the task that calls
+sys_restore() to setresuid/setresgid to those values must be checked.
+
+Keeping the restart procedure to operate within the limits of the
+caller's credentials means that there various scenarios that cannot
+be supported. For instance, a setuid program that opened a protected
+log file and then dropped privileges will fail the restart, because
+the user won't have enough credentials to reopen the file. In these
+cases, we should probably treat restarting like inserting a kernel
+module: surely the user can cause havoc by providing incorrect data,
+but then again we must trust the root account.
+
+So that's why we don't want CAP_SYS_ADMIN required up-front. That way
+we will be forced to more carefully review each of those features.
+
diff --git a/Documentation/checkpoint/self.c b/Documentation/checkpoint/self.c
new file mode 100644
index 0000000..febb888
--- /dev/null
+++ b/Documentation/checkpoint/self.c
@@ -0,0 +1,57 @@
+#include <stdio.h>
+#include <stdlib.h>
+#include <unistd.h>
+#include <string.h>
+#include <errno.h>
+#include <math.h>
+#include <sys/syscall.h>
+
+#define OUTFILE "/tmp/cr-test.out"
+
+int main(int argc, char *argv[])
+{
+ pid_t pid = getpid();
+ FILE *file;
+ int i, ret;
+ float a;
+
+ close(0);
+ close(2);
+
+ unlink(OUTFILE);
+ file = fopen(OUTFILE, "w+");
+ if (!file) {
+ perror("open");
+ exit(1);
+ }
+ if (dup2(0, 2) < 0) {
+ perror("dup2");
+ exit(1);
+ }
+
+ a = sqrt(2.53 * (getpid() / 1.21));
+
+ fprintf(file, "hello, world (%.2f)!\n", a);
+ fflush(file);
+
+ for (i = 0; i < 1000; i++) {
+ sleep(1);
+ /* make the fpu work -> a = a + i/10 */
+ a = sqrt(a*a + 2*a*(i/10.0) + i*i/100.0);
+ fprintf(file, "count %d (%.2f)!\n", i, a);
+ fflush(file);
+
+ if (i == 2) {
+ ret = syscall(__NR_checkpoint, pid, STDOUT_FILENO, 0);
+ if (ret < 0) {
+ fprintf(file, "ckpt: %s\n", strerror(errno));
+ exit(2);
+ }
+ fprintf(file, "checkpoint ret: %d\n", ret);
+ fflush(file);
+ }
+ }
+
+ return 0;
+}
+
diff --git a/Documentation/checkpoint/test.c b/Documentation/checkpoint/test.c
new file mode 100644
index 0000000..1183655
--- /dev/null
+++ b/Documentation/checkpoint/test.c
@@ -0,0 +1,48 @@
+#include <stdio.h>
+#include <stdlib.h>
+#include <unistd.h>
+#include <errno.h>
+#include <math.h>
+
+#define OUTFILE "/tmp/cr-test.out"
+
+int main(int argc, char *argv[])
+{
+ FILE *file;
+ float a;
+ int i;
+
+ close(0);
+ close(1);
+ close(2);
+
+ unlink(OUTFILE);
+ file = fopen(OUTFILE, "w+");
+ if (!file) {
+ perror("open");
+ exit(1);
+ }
+ if (dup2(0, 2) < 0) {
+ perror("dup2");
+ exit(1);
+ }
+
+ a = sqrt(2.53 * (getpid() / 1.21));
+
+ fprintf(file, "hello, world (%.2f)!\n", a);
+ fflush(file);
+
+ for (i = 0; i < 1000; i++) {
+ sleep(1);
+ /* make the fpu work -> a = a + i/10 */
+ a = sqrt(a*a + 2*a*(i/10.0) + i*i/100.0);
+ fprintf(file, "count %d (%.2f)!\n", i, a);
+ fflush(file);
+ }
+
+ fprintf(file, "world, hello (%.2f) !\n", a);
+ fflush(file);
+
+ return 0;
+}
+
diff --git a/Documentation/checkpoint/usage.txt b/Documentation/checkpoint/usage.txt
new file mode 100644
index 0000000..6bea42d
--- /dev/null
+++ b/Documentation/checkpoint/usage.txt
@@ -0,0 +1,153 @@
+
+ ===== How to use Checkpoint-Restart =====
+
+The API consists of two new system calls:
+
+* int sys_checkpoint(pid_t pid, int fd, unsigned long flag);
+
+ Checkpoint a container whose init task is identified by pid, to
+ the file designated by fd. 'flags' will have future meaning (must
+ be 0 for now).
+
+ Returns: a positive checkpoint identifier (crid) upon success, 0
+ if it returns from a restart, and -1 if an error occurs.
+
+ 'crid' uniquely identifies a checkpoint image. For each checkpoint
+ the kernel allocates a unique 'crid', that remains valid for as
+ long as the checkpoint is kept in the kernel (for instance, when a
+ checkpoint, or a partial checkpoint, may reside in kernel memory).
+
+* int sys_restart(int crid, int fd, unsigned long flags);
+
+ Restart a container from a checkpoint image that is read from the
+ blob stored in the file designated by fd. 'crid' will have future
+ meaning (must be 0 for now). 'flags' will have future meaning
+ (must be 0 for now).
+
+ The role of 'crid' is to identify the checkpoint image in the case
+ that it remains in kernel memory. This will be useful to restart
+ from a checkpoint image that remains in kernel memory.
+
+ Returns: -1 if an error occurs, 0 on success when restarting from
+ a "self" checkpoint, and return value of system call at the time
+ of the checkpoint when restarting from an "external" checkpoint.
+
+ If restarting from an "external" checkpoint, tasks that were
+ executing a system call will observe the return value of that
+ system call (as it was when interrupted for the act of taking the
+ checkpoint), and tasks that were executing in user space will be
+ ready to return there.
+
+ Upon successful "external" restart, the container will end up in a
+ frozen state.
+
+The granularity of a checkpoint usually is a whole container. The
+'pid' argument is interpreted in the caller's pid namespace. So to
+checkpoint a container whose init task (pid 1 in that pidns) appears
+as pid 3497 the caller's pidns, the caller must use pid 3497. Passing
+pid 1 will attempt to checkpoint the caller's container, and if the
+caller isn't privileged and init is owned by root, it will fail.
+
+If the caller passes a pid which does not refer to a container's init
+task, then sys_checkpoint() would return -EINVAL. (This is because
+with nested containers a task may belong to more than one container).
+
+Here is a code snippet that illustrates how a checkpoint is initiated
+by a process in a container - the logic is similar to fork():
+ ...
+ crid = checkpoint(1, ...);
+ switch (crid) {
+ case -1:
+ perror("checkpoint failed");
+ break;
+ default:
+ fprintf(stderr, "checkpoint succeeded, CRID=%d\n", ret);
+ /* proceed with execution after checkpoint */
+ ...
+ break;
+ case 0:
+ fprintf(stderr, "returned after restart\n");
+ /* proceed with action required following a restart */
+ ...
+ break;
+ }
+ ...
+
+And to initiate a restart, the process in an empty container can use
+logic similar to execve():
+ ...
+ if (restart(crid, ...) < 0)
+ perror("restart failed");
+ /* only get here if restart failed */
+ ...
+
+Note, that the code also supports "self" checkpoint, where a process
+can checkpoint itself. This mode does not capture the relationships
+of the task with other tasks, or any shared resources. It is useful
+for application that wish to be able to save and restore their state.
+They will either not use (or care about) shared resources, or they
+will be aware of the operations and adapt suitably after a restart.
+The code above can also be used for "self" checkpoint.
+
+To illustrate how the API works, refer to these sample programs:
+
+* ckpt.c: accepts a 'pid' argument and checkpoint that task to stdout
+* rstr.c: restarts a checkpoint image from stdin
+* self.c: a simple test program doing self-checkpoint
+* test.c: a simple test program to checkpoint
+
+"External" checkpoint:
+---------------------
+To do "external" checkpoint, you need to first freeze that other task
+either using the freezer cgroup, or by sending SIGSTOP.
+
+Restart does not preserve the original PID yet, (because we haven't
+solved yet the fork-with-specific-pid issue). In a real scenario, you
+probably want to first create a new names space, and have the init
+task there call 'sys_restart()'.
+
+I tested it this way:
+ $ ./test &
+ [1] 3493
+
+ $ kill -STOP 3493
+ $ ./ckpt 3493 > ckpt.image
+
+ $ mv /tmp/cr-test.out /tmp/cr-test.out.orig
+ $ cp /tmp/cr-test.out.orig /tmp/cr-test.out
+
+ $ kill -CONT 3493
+
+ $ ./rstr < ckpt.image
+Now compare the output of the two output files.
+
+"Self checkpoint:
+----------------
+To do "self" checkpoint, you can incorporate the code from ckpt.c into
+your application.
+
+Here is how to test the "self" checkpoint:
+ $ ./self > self.image &
+ [1] 3512
+
+ $ sleep 3
+ $ mv /tmp/cr-test.out /tmp/cr-test.out.orig
+ $ cp /tmp/cr-test.out.orig /tmp/cr-test.out
+
+ $ cat /tmp/cr-rest.out
+ hello, world (85.46)!
+ count 0 (85.46)!
+ count 1 (85.56)!
+ count 2 (85.76)!
+ count 3 (86.46)!
+
+ $ sed -i 's/count/xxxx/g' /tmp/cr-rest.out
+
+ $ ./rstr < self.image &
+Now compare the output of the two output files.
+
+Note how in test.c we close stdin, stdout, stderr - that's because
+currently we only support regular files (not ttys/ptys).
+
+If you check the output of ps, you'll see that "rstr" changed its name
+to "test" or "self", as expected.
--
1.5.4.3
Dump the files_struct of a task with 'struct cr_hdr_files', followed by
all open file descriptors. Since FDs can be shared, they are assigned an
objref and registered in the object hash.
For each open FD there is a 'struct cr_hdr_fd_ent' with the FD, its objref
and its close-on-exec property. If the FD is to be saved (first time)
then this is followed by a 'struct cr_hdr_fd_data' with the FD state.
Then will come the next FD and so on.
This patch only handles basic FDs - regular files, directories and also
symbolic links.
Changelog[v9]:
- Fix a couple of leaks in cr_write_files()
- Drop useless kfree from cr_scan_fds()
Changelog[v8]:
- initialize 'coe' to workaround gcc false warning
Changelog[v6]:
- Balance all calls to cr_hbuf_get() with matching cr_hbuf_put()
(even though it's not really needed)
Signed-off-by: Oren Laadan <[email protected]>
Acked-by: Serge Hallyn <[email protected]>
Signed-off-by: Dave Hansen <[email protected]>
---
checkpoint/Makefile | 2 +-
checkpoint/checkpoint.c | 4 +
checkpoint/checkpoint_file.h | 17 +++
checkpoint/ckpt_file.c | 231 ++++++++++++++++++++++++++++++++++++++++
include/linux/checkpoint.h | 3 +-
include/linux/checkpoint_hdr.h | 32 ++++++-
6 files changed, 286 insertions(+), 3 deletions(-)
create mode 100644 checkpoint/checkpoint_file.h
create mode 100644 checkpoint/ckpt_file.c
diff --git a/checkpoint/Makefile b/checkpoint/Makefile
index 9843fb9..7496695 100644
--- a/checkpoint/Makefile
+++ b/checkpoint/Makefile
@@ -3,4 +3,4 @@
#
obj-$(CONFIG_CHECKPOINT_RESTART) += sys.o checkpoint.o restart.o objhash.o \
- ckpt_mem.o rstr_mem.o
+ ckpt_mem.o rstr_mem.o ckpt_file.o
diff --git a/checkpoint/checkpoint.c b/checkpoint/checkpoint.c
index 6a8f810..ae1952b 100644
--- a/checkpoint/checkpoint.c
+++ b/checkpoint/checkpoint.c
@@ -232,6 +232,10 @@ static int cr_write_task(struct cr_ctx *ctx, struct task_struct *t)
cr_debug("memory: ret %d\n", ret);
if (ret < 0)
goto out;
+ ret = cr_write_files(ctx, t);
+ cr_debug("files: ret %d\n", ret);
+ if (ret < 0)
+ goto out;
ret = cr_write_thread(ctx, t);
cr_debug("thread: ret %d\n", ret);
if (ret < 0)
diff --git a/checkpoint/checkpoint_file.h b/checkpoint/checkpoint_file.h
new file mode 100644
index 0000000..9dc3eba
--- /dev/null
+++ b/checkpoint/checkpoint_file.h
@@ -0,0 +1,17 @@
+#ifndef _CHECKPOINT_CKPT_FILE_H_
+#define _CHECKPOINT_CKPT_FILE_H_
+/*
+ * Checkpoint file descriptors
+ *
+ * Copyright (C) 2008 Oren Laadan
+ *
+ * This file is subject to the terms and conditions of the GNU General Public
+ * License. See the file COPYING in the main directory of the Linux
+ * distribution for more details.
+ */
+
+#include <linux/fdtable.h>
+
+int cr_scan_fds(struct files_struct *files, int **fdtable);
+
+#endif /* _CHECKPOINT_CKPT_FILE_H_ */
diff --git a/checkpoint/ckpt_file.c b/checkpoint/ckpt_file.c
new file mode 100644
index 0000000..9198650
--- /dev/null
+++ b/checkpoint/ckpt_file.c
@@ -0,0 +1,231 @@
+/*
+ * Checkpoint file descriptors
+ *
+ * Copyright (C) 2008 Oren Laadan
+ *
+ * This file is subject to the terms and conditions of the GNU General Public
+ * License. See the file COPYING in the main directory of the Linux
+ * distribution for more details.
+ */
+
+#include <linux/kernel.h>
+#include <linux/sched.h>
+#include <linux/file.h>
+#include <linux/fdtable.h>
+#include <linux/checkpoint.h>
+#include <linux/checkpoint_hdr.h>
+
+#include "checkpoint_file.h"
+
+#define CR_DEFAULT_FDTABLE 256 /* an initial guess */
+
+/**
+ * cr_scan_fds - scan file table and construct array of open fds
+ * @files: files_struct pointer
+ * @fdtable: (output) array of open fds
+ *
+ * Returns the number of open fds found, and also the file table
+ * array via *fdtable. The caller should free the array.
+ *
+ * The caller must validate the file descriptors collected in the
+ * array before using them, e.g. by using fcheck_files(), in case
+ * the task's fdtable changes in the meantime.
+ */
+int cr_scan_fds(struct files_struct *files, int **fdtable)
+{
+ struct fdtable *fdt;
+ int *fds;
+ int i, n = 0;
+ int tot = CR_DEFAULT_FDTABLE;
+
+ fds = kmalloc(tot * sizeof(*fds), GFP_KERNEL);
+ if (!fds)
+ return -ENOMEM;
+
+ /*
+ * We assume that the target task is frozen (or that we checkpoint
+ * ourselves), so we can safely proceed after krealloc() from where
+ * we left off; in the worst cases restart will fail.
+ */
+
+ spin_lock(&files->file_lock);
+ rcu_read_lock();
+ fdt = files_fdtable(files);
+ for (i = 0; i < fdt->max_fds; i++) {
+ if (!fcheck_files(files, i))
+ continue;
+ if (n == tot) {
+ /*
+ * fcheck_files() is safe with drop/re-acquire
+ * of the lock, because it tests: fd < max_fds
+ */
+ spin_unlock(&files->file_lock);
+ rcu_read_unlock();
+ tot *= 2; /* won't overflow: kmalloc will fail */
+ fds = krealloc(fds, tot * sizeof(*fds), GFP_KERNEL);
+ if (!fds)
+ return -ENOMEM;
+ rcu_read_lock();
+ spin_lock(&files->file_lock);
+ }
+ fds[n++] = i;
+ }
+ rcu_read_unlock();
+ spin_unlock(&files->file_lock);
+
+ *fdtable = fds;
+ return n;
+}
+
+/* cr_write_fd_data - dump the state of a given file pointer */
+static int cr_write_fd_data(struct cr_ctx *ctx, struct file *file, int parent)
+{
+ struct cr_hdr h;
+ struct cr_hdr_fd_data *hh = cr_hbuf_get(ctx, sizeof(*hh));
+ struct dentry *dent = file->f_dentry;
+ struct inode *inode = dent->d_inode;
+ enum fd_type fd_type;
+ int ret;
+
+ h.type = CR_HDR_FD_DATA;
+ h.len = sizeof(*hh);
+ h.parent = parent;
+
+ hh->f_flags = file->f_flags;
+ hh->f_mode = file->f_mode;
+ hh->f_pos = file->f_pos;
+ hh->f_version = file->f_version;
+ /* FIX: need also file->uid, file->gid, file->f_owner, etc */
+
+ switch (inode->i_mode & S_IFMT) {
+ case S_IFREG:
+ fd_type = CR_FD_FILE;
+ break;
+ case S_IFDIR:
+ fd_type = CR_FD_DIR;
+ break;
+ case S_IFLNK:
+ fd_type = CR_FD_LINK;
+ break;
+ default:
+ cr_hbuf_put(ctx, sizeof(*hh));
+ return -EBADF;
+ }
+
+ /* FIX: check if the file/dir/link is unlinked */
+ hh->fd_type = fd_type;
+
+ ret = cr_write_obj(ctx, &h, hh);
+ cr_hbuf_put(ctx, sizeof(*hh));
+ if (ret < 0)
+ return ret;
+
+ return cr_write_fname(ctx, &file->f_path, ctx->vfsroot);
+}
+
+/**
+ * cr_write_fd_ent - dump the state of a given file descriptor
+ * @ctx: checkpoint context
+ * @files: files_struct pointer
+ * @fd: file descriptor
+ *
+ * Saves the state of the file descriptor; looks up the actual file
+ * pointer in the hash table, and if found saves the matching objref,
+ * otherwise calls cr_write_fd_data to dump the file pointer too.
+ */
+static int
+cr_write_fd_ent(struct cr_ctx *ctx, struct files_struct *files, int fd)
+{
+ struct cr_hdr h;
+ struct cr_hdr_fd_ent *hh = cr_hbuf_get(ctx, sizeof(*hh));
+ struct file *file;
+ struct fdtable *fdt;
+ int objref, new, ret;
+ int coe = 0; /* avoid gcc warning */
+
+ rcu_read_lock();
+ fdt = files_fdtable(files);
+ file = fcheck_files(files, fd);
+ if (file) {
+ coe = FD_ISSET(fd, fdt->close_on_exec);
+ get_file(file);
+ }
+ rcu_read_unlock();
+
+ /* sanity check (although this shouldn't happen) */
+ if (!file) {
+ ret = -EBADF;
+ goto out;
+ }
+
+ new = cr_obj_add_ptr(ctx, file, &objref, CR_OBJ_FILE, 0);
+ cr_debug("fd %d objref %d file %p c-o-e %d)\n", fd, objref, file, coe);
+
+ if (new < 0) {
+ ret = new;
+ goto out;
+ }
+
+ h.type = CR_HDR_FD_ENT;
+ h.len = sizeof(*hh);
+ h.parent = 0;
+
+ hh->objref = objref;
+ hh->fd = fd;
+ hh->close_on_exec = coe;
+
+ ret = cr_write_obj(ctx, &h, hh);
+ if (ret < 0)
+ goto out;
+
+ /* new==1 if-and-only-if file was newly added to hash */
+ if (new)
+ ret = cr_write_fd_data(ctx, file, objref);
+
+out:
+ cr_hbuf_put(ctx, sizeof(*hh));
+ if (file)
+ fput(file);
+ return ret;
+}
+
+int cr_write_files(struct cr_ctx *ctx, struct task_struct *t)
+{
+ struct cr_hdr h;
+ struct cr_hdr_files *hh = cr_hbuf_get(ctx, sizeof(*hh));
+ struct files_struct *files;
+ int *fdtable = NULL;
+ int nfds, n, ret;
+
+ h.type = CR_HDR_FILES;
+ h.len = sizeof(*hh);
+ h.parent = task_pid_vnr(t);
+
+ files = get_files_struct(t);
+
+ nfds = cr_scan_fds(files, &fdtable);
+ if (nfds < 0) {
+ ret = nfds;
+ goto out;
+ }
+
+ hh->objref = 0; /* will be meaningful with multiple processes */
+ hh->nfds = nfds;
+
+ ret = cr_write_obj(ctx, &h, hh);
+ cr_hbuf_put(ctx, sizeof(*hh));
+ if (ret < 0)
+ goto out;
+
+ cr_debug("nfds %d\n", nfds);
+ for (n = 0; n < nfds; n++) {
+ ret = cr_write_fd_ent(ctx, files, fdtable[n]);
+ if (ret < 0)
+ break;
+ }
+
+ out:
+ kfree(fdtable);
+ put_files_struct(files);
+ return ret;
+}
diff --git a/include/linux/checkpoint.h b/include/linux/checkpoint.h
index deccee5..74cac0b 100644
--- a/include/linux/checkpoint.h
+++ b/include/linux/checkpoint.h
@@ -13,7 +13,7 @@
#include <linux/path.h>
#include <linux/fs.h>
-#define CR_VERSION 1
+#define CR_VERSION 2
struct cr_ctx {
int crid; /* unique checkpoint id */
@@ -84,6 +84,7 @@ extern struct file *cr_read_open_fname(struct cr_ctx *ctx,
extern int do_checkpoint(struct cr_ctx *ctx, pid_t pid);
extern int cr_write_mm(struct cr_ctx *ctx, struct task_struct *t);
+extern int cr_write_files(struct cr_ctx *ctx, struct task_struct *t);
extern int do_restart(struct cr_ctx *ctx, pid_t pid);
extern int cr_read_mm(struct cr_ctx *ctx);
diff --git a/include/linux/checkpoint_hdr.h b/include/linux/checkpoint_hdr.h
index d78f0f1..e92c37d 100644
--- a/include/linux/checkpoint_hdr.h
+++ b/include/linux/checkpoint_hdr.h
@@ -17,7 +17,7 @@
/*
* To maintain compatibility between 32-bit and 64-bit architecture flavors,
* keep data 64-bit aligned: use padding for structure members, and use
- * __attribute__ ((aligned (8))) for the entire structure.
+ * __attribute__((aligned(8))) for the entire structure.
*/
/* records: generic header */
@@ -45,6 +45,10 @@ enum {
CR_HDR_PGARR,
CR_HDR_MM_CONTEXT,
+ CR_HDR_FILES = 301,
+ CR_HDR_FD_ENT,
+ CR_HDR_FD_DATA,
+
CR_HDR_TAIL = 5001
};
@@ -107,4 +111,30 @@ struct cr_hdr_pgarr {
__u64 nr_pages; /* number of pages to saved */
} __attribute__((aligned(8)));
+struct cr_hdr_files {
+ __u32 objref; /* identifier for shared objects */
+ __u32 nfds;
+} __attribute__((aligned(8)));
+
+struct cr_hdr_fd_ent {
+ __u32 objref; /* identifier for shared objects */
+ __s32 fd;
+ __u32 close_on_exec;
+} __attribute__((aligned(8)));
+
+/* fd types */
+enum fd_type {
+ CR_FD_FILE = 1,
+ CR_FD_DIR,
+ CR_FD_LINK
+};
+
+struct cr_hdr_fd_data {
+ __u16 fd_type;
+ __u16 f_mode;
+ __u32 f_flags;
+ __u64 f_pos;
+ __u64 f_version;
+} __attribute__((aligned(8)));
+
#endif /* _CHECKPOINT_CKPT_HDR_H_ */
--
1.5.4.3
Add those interfaces, as well as helpers needed to easily manage the
file format. The code is roughly broken out as follows:
checkpoint/sys.c - user/kernel data transfer, as well as setup of the
CR context (a per-checkpoint data structure for housekeeping)
checkpoint/checkpoint.c - output wrappers and basic checkpoint handling
checkpoint/restart.c - input wrappers and basic restart handling
For now, we can only checkpoint the 'current' task ("self" checkpoint),
and the 'pid' argument to to the syscall is ignored.
Patches to add the per-architecture support as well as the actual
work to do the memory checkpoint follow in subsequent patches.
Changelog[v10]:
- add cr_write_buffer(), cr_read_buffer() and cr_read_buf_type()
- force end-of-string in cr_read_string() (fix possible DoS)
Changelog[v9]:
- cr_kwrite/cr_kread() use file->f_op->write() directly
- Drop cr_uwrite/cr_uread() since they aren't used anywhere
Changelog[v6]:
- Balance all calls to cr_hbuf_get() with matching cr_hbuf_put()
(although it's not really needed)
Changelog[v5]:
- Rename headers files s/ckpt/checkpoint/
Changelog[v2]:
- Added utsname->{release,version,machine} to checkpoint header
- Pad header structures to 64 bits to ensure compatibility
Signed-off-by: Oren Laadan <[email protected]>
Acked-by: Serge Hallyn <[email protected]>
Signed-off-by: Dave Hansen <[email protected]>
---
Makefile | 2 +-
checkpoint/Makefile | 2 +-
checkpoint/checkpoint.c | 188 +++++++++++++++++++++++++++++++
checkpoint/restart.c | 239 ++++++++++++++++++++++++++++++++++++++++
checkpoint/sys.c | 194 +++++++++++++++++++++++++++++++-
include/linux/checkpoint.h | 58 ++++++++++
include/linux/checkpoint_hdr.h | 76 +++++++++++++
include/linux/magic.h | 3 +
8 files changed, 756 insertions(+), 6 deletions(-)
create mode 100644 checkpoint/checkpoint.c
create mode 100644 checkpoint/restart.c
create mode 100644 include/linux/checkpoint.h
create mode 100644 include/linux/checkpoint_hdr.h
diff --git a/Makefile b/Makefile
index 29abe62..90dc50f 100644
--- a/Makefile
+++ b/Makefile
@@ -619,7 +619,7 @@ export mod_strip_cmd
ifeq ($(KBUILD_EXTMOD),)
-core-y += kernel/ mm/ fs/ ipc/ security/ crypto/ block/
+core-y += kernel/ mm/ fs/ ipc/ security/ crypto/ block/ checkpoint/
vmlinux-dirs := $(patsubst %/,%,$(filter %/, $(init-y) $(init-m) \
$(core-y) $(core-m) $(drivers-y) $(drivers-m) \
diff --git a/checkpoint/Makefile b/checkpoint/Makefile
index 07d018b..d2df68c 100644
--- a/checkpoint/Makefile
+++ b/checkpoint/Makefile
@@ -2,4 +2,4 @@
# Makefile for linux checkpoint/restart.
#
-obj-$(CONFIG_CHECKPOINT_RESTART) += sys.o
+obj-$(CONFIG_CHECKPOINT_RESTART) += sys.o checkpoint.o restart.o
diff --git a/checkpoint/checkpoint.c b/checkpoint/checkpoint.c
new file mode 100644
index 0000000..fccf723
--- /dev/null
+++ b/checkpoint/checkpoint.c
@@ -0,0 +1,188 @@
+/*
+ * Checkpoint logic and helpers
+ *
+ * Copyright (C) 2008 Oren Laadan
+ *
+ * This file is subject to the terms and conditions of the GNU General Public
+ * License. See the file COPYING in the main directory of the Linux
+ * distribution for more details.
+ */
+
+#include <linux/version.h>
+#include <linux/sched.h>
+#include <linux/time.h>
+#include <linux/fs.h>
+#include <linux/file.h>
+#include <linux/dcache.h>
+#include <linux/mount.h>
+#include <linux/utsname.h>
+#include <linux/magic.h>
+#include <linux/checkpoint.h>
+#include <linux/checkpoint_hdr.h>
+
+/* unique checkpoint identifier (FIXME: should be per-container ?) */
+static atomic_t cr_ctx_count = ATOMIC_INIT(0);
+
+/**
+ * cr_write_obj - write a record described by a cr_hdr
+ * @ctx: checkpoint context
+ * @h: record descriptor
+ * @buf: record buffer
+ */
+int cr_write_obj(struct cr_ctx *ctx, struct cr_hdr *h, void *buf)
+{
+ int ret;
+
+ ret = cr_kwrite(ctx, h, sizeof(*h));
+ if (ret < 0)
+ return ret;
+ return cr_kwrite(ctx, buf, h->len);
+}
+
+/**
+ * cr_write_buffer - write a buffer
+ * @ctx: checkpoint context
+ * @str: buffer pointer
+ * @len: buffer size
+ */
+int cr_write_buffer(struct cr_ctx *ctx, void *buf, int len)
+{
+ struct cr_hdr h;
+
+ h.type = CR_HDR_BUFFER;
+ h.len = len;
+ h.parent = 0;
+
+ return cr_write_obj(ctx, &h, buf);
+}
+
+/**
+ * cr_write_string - write a string
+ * @ctx: checkpoint context
+ * @str: string pointer
+ * @len: string length
+ */
+int cr_write_string(struct cr_ctx *ctx, char *str, int len)
+{
+ struct cr_hdr h;
+
+ h.type = CR_HDR_STRING;
+ h.len = len;
+ h.parent = 0;
+
+ return cr_write_obj(ctx, &h, str);
+}
+
+/* write the checkpoint header */
+static int cr_write_head(struct cr_ctx *ctx)
+{
+ struct cr_hdr h;
+ struct cr_hdr_head *hh = cr_hbuf_get(ctx, sizeof(*hh));
+ struct new_utsname *uts;
+ struct timeval ktv;
+ int ret;
+
+ h.type = CR_HDR_HEAD;
+ h.len = sizeof(*hh);
+ h.parent = 0;
+
+ do_gettimeofday(&ktv);
+
+ hh->magic = CHECKPOINT_MAGIC_HEAD;
+ hh->major = (LINUX_VERSION_CODE >> 16) & 0xff;
+ hh->minor = (LINUX_VERSION_CODE >> 8) & 0xff;
+ hh->patch = (LINUX_VERSION_CODE) & 0xff;
+
+ hh->rev = CR_VERSION;
+
+ hh->flags = ctx->flags;
+ hh->time = ktv.tv_sec;
+
+ uts = utsname();
+ memcpy(hh->release, uts->release, __NEW_UTS_LEN);
+ memcpy(hh->version, uts->version, __NEW_UTS_LEN);
+ memcpy(hh->machine, uts->machine, __NEW_UTS_LEN);
+
+ ret = cr_write_obj(ctx, &h, hh);
+ cr_hbuf_put(ctx, sizeof(*hh));
+ return ret;
+}
+
+/* write the checkpoint trailer */
+static int cr_write_tail(struct cr_ctx *ctx)
+{
+ struct cr_hdr h;
+ struct cr_hdr_tail *hh = cr_hbuf_get(ctx, sizeof(*hh));
+ int ret;
+
+ h.type = CR_HDR_TAIL;
+ h.len = sizeof(*hh);
+ h.parent = 0;
+
+ hh->magic = CHECKPOINT_MAGIC_TAIL;
+
+ ret = cr_write_obj(ctx, &h, hh);
+ cr_hbuf_put(ctx, sizeof(*hh));
+ return ret;
+}
+
+/* dump the task_struct of a given task */
+static int cr_write_task_struct(struct cr_ctx *ctx, struct task_struct *t)
+{
+ struct cr_hdr h;
+ struct cr_hdr_task *hh = cr_hbuf_get(ctx, sizeof(*hh));
+ int ret;
+
+ h.type = CR_HDR_TASK;
+ h.len = sizeof(*hh);
+ h.parent = 0;
+
+ hh->state = t->state;
+ hh->exit_state = t->exit_state;
+ hh->exit_code = t->exit_code;
+ hh->exit_signal = t->exit_signal;
+
+ hh->task_comm_len = TASK_COMM_LEN;
+
+ /* FIXME: save remaining relevant task_struct fields */
+
+ ret = cr_write_obj(ctx, &h, hh);
+ cr_hbuf_put(ctx, sizeof(*hh));
+ if (ret < 0)
+ return ret;
+
+ return cr_write_string(ctx, t->comm, TASK_COMM_LEN);
+}
+
+/* dump the entire state of a given task */
+static int cr_write_task(struct cr_ctx *ctx, struct task_struct *t)
+{
+ int ret;
+
+ ret = cr_write_task_struct(ctx, t);
+ cr_debug("ret %d\n", ret);
+
+ return ret;
+}
+
+int do_checkpoint(struct cr_ctx *ctx, pid_t pid)
+{
+ int ret;
+
+ ret = cr_write_head(ctx);
+ if (ret < 0)
+ goto out;
+ ret = cr_write_task(ctx, current);
+ if (ret < 0)
+ goto out;
+ ret = cr_write_tail(ctx);
+ if (ret < 0)
+ goto out;
+
+ ctx->crid = atomic_inc_return(&cr_ctx_count);
+
+ /* on success, return (unique) checkpoint identifier */
+ ret = ctx->crid;
+ out:
+ return ret;
+}
diff --git a/checkpoint/restart.c b/checkpoint/restart.c
new file mode 100644
index 0000000..a95d2e8
--- /dev/null
+++ b/checkpoint/restart.c
@@ -0,0 +1,239 @@
+/*
+ * Restart logic and helpers
+ *
+ * Copyright (C) 2008 Oren Laadan
+ *
+ * This file is subject to the terms and conditions of the GNU General Public
+ * License. See the file COPYING in the main directory of the Linux
+ * distribution for more details.
+ */
+
+#include <linux/version.h>
+#include <linux/sched.h>
+#include <linux/file.h>
+#include <linux/magic.h>
+#include <linux/checkpoint.h>
+#include <linux/checkpoint_hdr.h>
+
+/**
+ * cr_read_obj - read a whole record (cr_hdr followed by payload)
+ * @ctx: checkpoint context
+ * @h: record descriptor
+ * @buf: record buffer
+ * @len: available buffer size
+ */
+int cr_read_obj(struct cr_ctx *ctx, struct cr_hdr *h, void *buf, int len)
+{
+ int ret;
+
+ ret = cr_kread(ctx, h, sizeof(*h));
+ if (ret < 0)
+ return ret;
+
+ cr_debug("type %d len %d parent %d\n", h->type, h->len, h->parent);
+
+ if (h->len < 0 || h->len > len)
+ return -EINVAL;
+
+ return cr_kread(ctx, buf, h->len);
+}
+
+/**
+ * cr_read_obj_type - read a whole record of expected type and size
+ * @ctx: checkpoint context
+ * @buf: record buffer
+ * @n: expected record size
+ * @type: expected record type
+ *
+ * Returns objref of the parent object
+ */
+int cr_read_obj_type(struct cr_ctx *ctx, void *buf, int len, int type)
+{
+ struct cr_hdr h;
+ int ret;
+
+ ret = cr_read_obj(ctx, &h, buf, len);
+ if (ret < 0)
+ return ret;
+
+ if (h.len != len || h.type != type)
+ return -EINVAL;
+
+ return h.parent;
+}
+
+/**
+ * cr_read_buf_type - read a whole record of expected type (unknown size)
+ * @ctx: checkpoint context
+ * @buf: record buffer
+ * @n: availabe buffer size (output: actual record size)
+ * @type: expected record type
+ *
+ * Returns objref of the parent object
+ */
+int cr_read_buf_type(struct cr_ctx *ctx, void *buf, int *len, int type)
+{
+ struct cr_hdr h;
+ int ret;
+
+ ret = cr_read_obj(ctx, &h, buf, *len);
+ if (ret < 0)
+ return ret;
+
+ if (h.type != type)
+ return -EINVAL;
+
+ *len = h.len;
+ return h.parent;
+}
+
+/**
+ * cr_read_buffer - read a buffer
+ * @ctx: checkpoint context
+ * @buf: buffer
+ * @len: buffer size (output actual record size)
+ */
+int cr_read_buffer(struct cr_ctx *ctx, void *buf, int *len)
+{
+ return cr_read_buf_type(ctx, buf, len, CR_HDR_BUFFER);
+}
+
+/**
+ * cr_read_string - read a string
+ * @ctx: checkpoint context
+ * @str: string buffer
+ * @len: string length
+ */
+int cr_read_string(struct cr_ctx *ctx, char *str, int len)
+{
+ int ret;
+
+ ret = cr_read_buf_type(ctx, str, &len, CR_HDR_STRING);
+ if (ret < 0)
+ return ret;
+
+ if (len > 0)
+ str[len - 1] = '\0'; /* always play it safe */
+
+ return ret;
+}
+
+/* read the checkpoint header */
+static int cr_read_head(struct cr_ctx *ctx)
+{
+ struct cr_hdr_head *hh = cr_hbuf_get(ctx, sizeof(*hh));
+ int parent, ret = -EINVAL;
+
+ parent = cr_read_obj_type(ctx, hh, sizeof(*hh), CR_HDR_HEAD);
+ if (parent < 0) {
+ ret = parent;
+ goto out;
+ } else if (parent != 0)
+ goto out;
+
+ if (hh->magic != CHECKPOINT_MAGIC_HEAD || hh->rev != CR_VERSION ||
+ hh->major != ((LINUX_VERSION_CODE >> 16) & 0xff) ||
+ hh->minor != ((LINUX_VERSION_CODE >> 8) & 0xff) ||
+ hh->patch != ((LINUX_VERSION_CODE) & 0xff))
+ goto out;
+
+ if (hh->flags & ~CR_CTX_CKPT)
+ goto out;
+
+ ctx->oflags = hh->flags;
+
+ /* FIX: verify compatibility of release, version and machine */
+
+ ret = 0;
+ out:
+ cr_hbuf_put(ctx, sizeof(*hh));
+ return ret;
+}
+
+/* read the checkpoint trailer */
+static int cr_read_tail(struct cr_ctx *ctx)
+{
+ struct cr_hdr_tail *hh = cr_hbuf_get(ctx, sizeof(*hh));
+ int parent, ret = -EINVAL;
+
+ parent = cr_read_obj_type(ctx, hh, sizeof(*hh), CR_HDR_TAIL);
+ if (parent < 0) {
+ ret = parent;
+ goto out;
+ } else if (parent != 0)
+ goto out;
+
+ if (hh->magic != CHECKPOINT_MAGIC_TAIL)
+ goto out;
+
+ ret = 0;
+ out:
+ cr_hbuf_put(ctx, sizeof(*hh));
+ return ret;
+}
+
+/* read the task_struct into the current task */
+static int cr_read_task_struct(struct cr_ctx *ctx)
+{
+ struct cr_hdr_task *hh = cr_hbuf_get(ctx, sizeof(*hh));
+ struct task_struct *t = current;
+ char *buf;
+ int parent, ret = -EINVAL;
+
+ parent = cr_read_obj_type(ctx, hh, sizeof(*hh), CR_HDR_TASK);
+ if (parent < 0) {
+ ret = parent;
+ goto out;
+ } else if (parent != 0)
+ goto out;
+
+ /* upper limit for task_comm_len to prevent DoS */
+ if (hh->task_comm_len < 0 || hh->task_comm_len > PAGE_SIZE)
+ goto out;
+
+ buf = kmalloc(hh->task_comm_len, GFP_KERNEL);
+ if (!buf)
+ goto out;
+ ret = cr_read_string(ctx, buf, hh->task_comm_len);
+ if (!ret) {
+ /* if t->comm is too long, silently truncate */
+ memset(t->comm, 0, TASK_COMM_LEN);
+ memcpy(t->comm, buf, min(hh->task_comm_len, TASK_COMM_LEN));
+ }
+ kfree(buf);
+
+ /* FIXME: restore remaining relevant task_struct fields */
+ out:
+ cr_hbuf_put(ctx, sizeof(*hh));
+ return ret;
+}
+
+/* read the entire state of the current task */
+static int cr_read_task(struct cr_ctx *ctx)
+{
+ int ret;
+
+ ret = cr_read_task_struct(ctx);
+ cr_debug("ret %d\n", ret);
+
+ return ret;
+}
+
+int do_restart(struct cr_ctx *ctx, pid_t pid)
+{
+ int ret;
+
+ ret = cr_read_head(ctx);
+ if (ret < 0)
+ goto out;
+ ret = cr_read_task(ctx);
+ if (ret < 0)
+ goto out;
+ ret = cr_read_tail(ctx);
+ if (ret < 0)
+ goto out;
+
+ /* on success, adjust the return value if needed [TODO] */
+ out:
+ return ret;
+}
diff --git a/checkpoint/sys.c b/checkpoint/sys.c
index 375129c..bd14ef9 100644
--- a/checkpoint/sys.c
+++ b/checkpoint/sys.c
@@ -10,6 +10,158 @@
#include <linux/sched.h>
#include <linux/kernel.h>
+#include <linux/fs.h>
+#include <linux/file.h>
+#include <linux/uaccess.h>
+#include <linux/capability.h>
+#include <linux/checkpoint.h>
+
+/*
+ * Helpers to write(read) from(to) kernel space to(from) the checkpoint
+ * image file descriptor (similar to how a core-dump is performed).
+ *
+ * cr_kwrite() - write a kernel-space buffer to the checkpoint image
+ * cr_kread() - read from the checkpoint image to a kernel-space buffer
+ */
+
+int cr_kwrite(struct cr_ctx *ctx, void *addr, int count)
+{
+ struct file *file = ctx->file;
+ mm_segment_t fs;
+ ssize_t nwrite;
+ int nleft;
+
+ fs = get_fs();
+ set_fs(KERNEL_DS);
+ for (nleft = count; nleft; nleft -= nwrite) {
+ nwrite = file->f_op->write(file, addr, nleft, &file->f_pos);
+ if (nwrite < 0) {
+ if (nwrite == -EAGAIN)
+ nwrite = 0;
+ else
+ return nwrite;
+ }
+ addr += nwrite;
+ }
+ set_fs(fs);
+ ctx->total += count;
+ return 0;
+}
+
+int cr_kread(struct cr_ctx *ctx, void *addr, int count)
+{
+ struct file *file = ctx->file;
+ mm_segment_t fs;
+ ssize_t nread;
+ int nleft;
+
+ fs = get_fs();
+ set_fs(KERNEL_DS);
+ for (nleft = count; nleft; nleft -= nread) {
+ nread = file->f_op->read(file, addr, nleft, &file->f_pos);
+ if (nread <= 0) {
+ if (nread == -EAGAIN) {
+ nread = 0;
+ continue;
+ } else if (nread == 0)
+ nread = -EPIPE; /* unexecpted EOF */
+ return nread;
+ }
+ addr += nread;
+ }
+ set_fs(fs);
+ ctx->total += count;
+ return 0;
+}
+
+/*
+ * During checkpoint and restart the code writes outs/reads in data
+ * to/from the checkpoint image from/to a temporary buffer (ctx->hbuf).
+ * Because operations can be nested, use cr_hbuf_get() to reserve space
+ * in the buffer, then cr_hbuf_put() when you no longer need that space.
+ */
+
+/*
+ * ctx->hbuf is used to hold headers and data of known (or bound),
+ * static sizes. In some cases, multiple headers may be allocated in
+ * a nested manner. The size should accommodate all headers, nested
+ * or not, on all archs.
+ */
+#define CR_HBUF_TOTAL (8 * 4096)
+
+/**
+ * cr_hbuf_get - reserve space on the hbuf
+ * @ctx: checkpoint context
+ * @n: number of bytes to reserve
+ *
+ * Returns pointer to reserved space
+ */
+void *cr_hbuf_get(struct cr_ctx *ctx, int n)
+{
+ void *ptr;
+
+ /*
+ * Since requests depend on logic and static header sizes (not on
+ * user data), space should always suffice, unless someone either
+ * made a structure bigger or call path deeper than expected.
+ */
+ BUG_ON(ctx->hpos + n > CR_HBUF_TOTAL);
+ ptr = ctx->hbuf + ctx->hpos;
+ ctx->hpos += n;
+ return ptr;
+}
+
+/**
+ * cr_hbuf_put - unreserve space on the hbuf
+ * @ctx: checkpoint context
+ * @n: number of bytes to reserve
+ */
+void cr_hbuf_put(struct cr_ctx *ctx, int n)
+{
+ BUG_ON(ctx->hpos < n);
+ ctx->hpos -= n;
+}
+
+/*
+ * helpers to manage C/R contexts: allocated for each checkpoint and/or
+ * restart operation, and persists until the operation is completed.
+ */
+
+static void cr_ctx_free(struct cr_ctx *ctx)
+{
+ if (ctx->file)
+ fput(ctx->file);
+ kfree(ctx->hbuf);
+ kfree(ctx);
+}
+
+static struct cr_ctx *cr_ctx_alloc(int fd, unsigned long flags)
+{
+ struct cr_ctx *ctx;
+ int err;
+
+ ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
+ if (!ctx)
+ return ERR_PTR(-ENOMEM);
+
+ ctx->flags = flags;
+
+ err = -EBADF;
+ ctx->file = fget(fd);
+ if (!ctx->file)
+ goto err;
+
+ err = -ENOMEM;
+ ctx->hbuf = kmalloc(CR_HBUF_TOTAL, GFP_KERNEL);
+ if (!ctx->hbuf)
+ goto err;
+
+ return ctx;
+
+ err:
+ cr_ctx_free(ctx);
+ return ERR_PTR(err);
+}
/**
* sys_checkpoint - checkpoint a container
@@ -22,9 +174,26 @@
*/
asmlinkage long sys_checkpoint(pid_t pid, int fd, unsigned long flags)
{
- pr_debug("sys_checkpoint not implemented yet\n");
- return -ENOSYS;
+ struct cr_ctx *ctx;
+ int ret;
+
+ /* no flags for now */
+ if (flags)
+ return -EINVAL;
+
+ ctx = cr_ctx_alloc(fd, flags | CR_CTX_CKPT);
+ if (IS_ERR(ctx))
+ return PTR_ERR(ctx);
+
+ ret = do_checkpoint(ctx, pid);
+
+ if (!ret)
+ ret = ctx->crid;
+
+ cr_ctx_free(ctx);
+ return ret;
}
+
/**
* sys_restart - restart a container
* @crid: checkpoint image identifier
@@ -36,6 +205,23 @@ asmlinkage long sys_checkpoint(pid_t pid, int fd, unsigned long flags)
*/
asmlinkage long sys_restart(int crid, int fd, unsigned long flags)
{
- pr_debug("sys_restart not implemented yet\n");
- return -ENOSYS;
+ struct cr_ctx *ctx;
+ pid_t pid;
+ int ret;
+
+ /* no flags for now */
+ if (flags)
+ return -EINVAL;
+
+ ctx = cr_ctx_alloc(fd, flags | CR_CTX_RSTR);
+ if (IS_ERR(ctx))
+ return PTR_ERR(ctx);
+
+ /* FIXME: for now, we use 'crid' as a pid */
+ pid = (pid_t) crid;
+
+ ret = do_restart(ctx, pid);
+
+ cr_ctx_free(ctx);
+ return ret;
}
diff --git a/include/linux/checkpoint.h b/include/linux/checkpoint.h
new file mode 100644
index 0000000..63f298f
--- /dev/null
+++ b/include/linux/checkpoint.h
@@ -0,0 +1,58 @@
+#ifndef _CHECKPOINT_CKPT_H_
+#define _CHECKPOINT_CKPT_H_
+/*
+ * Generic container checkpoint-restart
+ *
+ * Copyright (C) 2008 Oren Laadan
+ *
+ * This file is subject to the terms and conditions of the GNU General Public
+ * License. See the file COPYING in the main directory of the Linux
+ * distribution for more details.
+ */
+
+#define CR_VERSION 1
+
+struct cr_ctx {
+ int crid; /* unique checkpoint id */
+
+ pid_t root_pid; /* container identifier */
+
+ unsigned long flags;
+ unsigned long oflags; /* restart: old flags */
+
+ struct file *file;
+ int total; /* total read/written */
+
+ void *hbuf; /* temporary buffer for headers */
+ int hpos; /* position in headers buffer */
+};
+
+/* cr_ctx: flags */
+#define CR_CTX_CKPT 0x1
+#define CR_CTX_RSTR 0x2
+
+extern int cr_kwrite(struct cr_ctx *ctx, void *buf, int count);
+extern int cr_kread(struct cr_ctx *ctx, void *buf, int count);
+
+extern void *cr_hbuf_get(struct cr_ctx *ctx, int n);
+extern void cr_hbuf_put(struct cr_ctx *ctx, int n);
+
+struct cr_hdr;
+
+extern int cr_write_obj(struct cr_ctx *ctx, struct cr_hdr *h, void *buf);
+extern int cr_write_buffer(struct cr_ctx *ctx, void *buf, int len);
+extern int cr_write_string(struct cr_ctx *ctx, char *str, int len);
+
+extern int cr_read_obj(struct cr_ctx *ctx, struct cr_hdr *h, void *buf, int n);
+extern int cr_read_obj_type(struct cr_ctx *ctx, void *buf, int len, int type);
+extern int cr_read_buf_type(struct cr_ctx *ctx, void *buf, int *len, int type);
+extern int cr_read_buffer(struct cr_ctx *ctx, void *buf, int *len);
+extern int cr_read_string(struct cr_ctx *ctx, char *str, int len);
+
+extern int do_checkpoint(struct cr_ctx *ctx, pid_t pid);
+extern int do_restart(struct cr_ctx *ctx, pid_t pid);
+
+#define cr_debug(fmt, args...) \
+ pr_debug("[%d:c/r:%s] " fmt, task_pid_vnr(current), __func__, ## args)
+
+#endif /* _CHECKPOINT_CKPT_H_ */
diff --git a/include/linux/checkpoint_hdr.h b/include/linux/checkpoint_hdr.h
new file mode 100644
index 0000000..257f87f
--- /dev/null
+++ b/include/linux/checkpoint_hdr.h
@@ -0,0 +1,76 @@
+#ifndef _CHECKPOINT_CKPT_HDR_H_
+#define _CHECKPOINT_CKPT_HDR_H_
+/*
+ * Generic container checkpoint-restart
+ *
+ * Copyright (C) 2008 Oren Laadan
+ *
+ * This file is subject to the terms and conditions of the GNU General Public
+ * License. See the file COPYING in the main directory of the Linux
+ * distribution for more details.
+ */
+
+#include <linux/types.h>
+#include <linux/utsname.h>
+
+/*
+ * To maintain compatibility between 32-bit and 64-bit architecture flavors,
+ * keep data 64-bit aligned: use padding for structure members, and use
+ * __attribute__ ((aligned (8))) for the entire structure.
+ */
+
+/* records: generic header */
+
+struct cr_hdr {
+ __s16 type;
+ __s16 len;
+ __u32 parent;
+};
+
+/* header types */
+enum {
+ CR_HDR_HEAD = 1,
+ CR_HDR_BUFFER,
+ CR_HDR_STRING,
+
+ CR_HDR_TASK = 101,
+ CR_HDR_THREAD,
+ CR_HDR_CPU,
+
+ CR_HDR_MM = 201,
+ CR_HDR_VMA,
+ CR_HDR_MM_CONTEXT,
+
+ CR_HDR_TAIL = 5001
+};
+
+struct cr_hdr_head {
+ __u64 magic;
+
+ __u16 major;
+ __u16 minor;
+ __u16 patch;
+ __u16 rev;
+
+ __u64 time; /* when checkpoint taken */
+ __u64 flags; /* checkpoint options */
+
+ char release[__NEW_UTS_LEN];
+ char version[__NEW_UTS_LEN];
+ char machine[__NEW_UTS_LEN];
+} __attribute__((aligned(8)));
+
+struct cr_hdr_tail {
+ __u64 magic;
+} __attribute__((aligned(8)));
+
+struct cr_hdr_task {
+ __u32 state;
+ __u32 exit_state;
+ __u32 exit_code;
+ __u32 exit_signal;
+
+ __s32 task_comm_len;
+} __attribute__((aligned(8)));
+
+#endif /* _CHECKPOINT_CKPT_HDR_H_ */
diff --git a/include/linux/magic.h b/include/linux/magic.h
index f7f3fdd..5939bbe 100644
--- a/include/linux/magic.h
+++ b/include/linux/magic.h
@@ -46,4 +46,7 @@
#define FUTEXFS_SUPER_MAGIC 0xBAD1DEA
#define INOTIFYFS_SUPER_MAGIC 0x2BAD1DEA
+#define CHECKPOINT_MAGIC_HEAD 0x00feed0cc0a2d200LL
+#define CHECKPOINT_MAGIC_TAIL 0x002d2a0cc0deef00LL
+
#endif /* __LINUX_MAGIC_H__ */
--
1.5.4.3
On Wed, Nov 26, 2008 at 08:04:39PM -0500, Oren Laadan wrote:
> +int cr_scan_fds(struct files_struct *files, int **fdtable)
> +{
> + struct fdtable *fdt;
> + int *fds;
> + int i, n = 0;
> + int tot = CR_DEFAULT_FDTABLE;
> +
> + fds = kmalloc(tot * sizeof(*fds), GFP_KERNEL);
> + if (!fds)
> + return -ENOMEM;
> +
> + /*
> + * We assume that the target task is frozen (or that we checkpoint
> + * ourselves), so we can safely proceed after krealloc() from where
> + * we left off; in the worst cases restart will fail.
> + */
Task may be frozen, but it may share the table with any number of other
tasks...
> + spin_lock(&files->file_lock);
> + rcu_read_lock();
> + fdt = files_fdtable(files);
> + for (i = 0; i < fdt->max_fds; i++) {
> + if (!fcheck_files(files, i))
> + continue;
> + if (n == tot) {
> + /*
> + * fcheck_files() is safe with drop/re-acquire
> + * of the lock, because it tests: fd < max_fds
> + */
> + spin_unlock(&files->file_lock);
> + rcu_read_unlock();
> + tot *= 2; /* won't overflow: kmalloc will fail */
> + fds = krealloc(fds, tot * sizeof(*fds), GFP_KERNEL);
> + if (!fds)
> + return -ENOMEM;
> + rcu_read_lock();
> + spin_lock(&files->file_lock);
> + }
> + fds[n++] = i;
> + }
> + rcu_read_unlock();
> + spin_unlock(&files->file_lock);
> +
> + *fdtable = fds;
> + return n;
> +}
> + switch (inode->i_mode & S_IFMT) {
> + case S_IFREG:
> + fd_type = CR_FD_FILE;
> + break;
> + case S_IFDIR:
> + fd_type = CR_FD_DIR;
> + break;
> + case S_IFLNK:
> + fd_type = CR_FD_LINK;
Opened symlinks? May I have whatever you'd been smoking, please?
On Wed, Nov 26, 2008 at 08:04:33PM -0500, Oren Laadan wrote:
> +Currently, namespaces are not saved or restored. They will be treated
> +as a class of a shared object. In particular, it is assumed that the
> +task's file system namespace is the "root" for the entire container.
> +It is also assumed that the same file system view is available for the
> +restart task(s). Otherwise, a file system snapshot is required.
That is to say, bindings are not handled at all.
> +* What additional work needs to be done to it?
> +We know this design can work. We have two commercial products and a
> +horde of academic projects doing it today using this basic design.
May I use that for a t-shirt, please? With that quote in foreground, and
pus-yellow-greenish "MACH" serving as background. With the names of products
and projects dripping from it...
On Wed, Nov 26, 2008 at 08:04:36PM -0500, Oren Laadan wrote:
> For each VMA, there is a 'struct cr_vma'; if the VMA is file-mapped,
> it will be followed by the file name. Then comes the actual contents,
> in one or more chunk: each chunk begins with a header that specifies
> how many pages it holds, then the virtual addresses of all the dumped
> pages in that chunk, followed by the actual contents of all dumped
> pages. A header with zero number of pages marks the end of the contents.
> Then comes the next VMA and so on.
>
> Changelog[v10]:
> - Acquire dcache_lock around call to __d_path() in cr_fill_name()
>
> Changelog[v9]:
> - Introduce cr_ctx_checkpoint() for checkpoint-specific ctx setup
> - Test if __d_path() changes mnt/dentry (when crossing filesystem
> namespace boundary). for now cr_fill_fname() fails the checkpoint.
>
> Changelog[v7]:
> - Fix argument given to kunmap_atomic() in memory dump/restore
>
> Changelog[v6]:
> - Balance all calls to cr_hbuf_get() with matching cr_hbuf_put()
> (even though it's not really needed)
>
> Changelog[v5]:
> - Improve memory dump code (following Dave Hansen's comments)
> - Change dump format (and code) to allow chunks of <vaddrs, pages>
> instead of one long list of each
> - Fix use of follow_page() to avoid faulting in non-present pages
>
> Changelog[v4]:
> - Use standard list_... for cr_pgarr
>
> Signed-off-by: Oren Laadan <[email protected]>
> Acked-by: Serge Hallyn <[email protected]>
> Signed-off-by: Dave Hansen <[email protected]>
> ---
> arch/x86/include/asm/checkpoint_hdr.h | 5 +
> arch/x86/mm/checkpoint.c | 31 ++
> checkpoint/Makefile | 3 +-
> checkpoint/checkpoint.c | 81 ++++++
> checkpoint/checkpoint_arch.h | 2 +
> checkpoint/checkpoint_mem.h | 41 +++
> checkpoint/ckpt_mem.c | 500 +++++++++++++++++++++++++++++++++
> checkpoint/sys.c | 11 +
> include/linux/checkpoint.h | 12 +
> include/linux/checkpoint_hdr.h | 32 ++
> 10 files changed, 717 insertions(+), 1 deletions(-)
> create mode 100644 checkpoint/checkpoint_mem.h
> create mode 100644 checkpoint/ckpt_mem.c
>
> diff --git a/arch/x86/include/asm/checkpoint_hdr.h b/arch/x86/include/asm/checkpoint_hdr.h
> index 6325062..33f4c70 100644
> --- a/arch/x86/include/asm/checkpoint_hdr.h
> +++ b/arch/x86/include/asm/checkpoint_hdr.h
> @@ -82,4 +82,9 @@ struct cr_hdr_cpu {
> /* thread_xstate contents follow (if used_math) */
> } __attribute__((aligned(8)));
>
> +struct cr_hdr_mm_context {
> + __s16 ldt_entry_size;
> + __s16 nldt;
> +} __attribute__((aligned(8)));
> +
> #endif /* __ASM_X86_CKPT_HDR__H */
> diff --git a/arch/x86/mm/checkpoint.c b/arch/x86/mm/checkpoint.c
> index 8dd6d2d..757936e 100644
> --- a/arch/x86/mm/checkpoint.c
> +++ b/arch/x86/mm/checkpoint.c
> @@ -221,3 +221,34 @@ int cr_write_head_arch(struct cr_ctx *ctx)
>
> return ret;
> }
> +
> +/* dump the mm->context state */
> +int cr_write_mm_context(struct cr_ctx *ctx, struct mm_struct *mm, int parent)
> +{
> + struct cr_hdr h;
> + struct cr_hdr_mm_context *hh = cr_hbuf_get(ctx, sizeof(*hh));
> + int ret;
> +
> + h.type = CR_HDR_MM_CONTEXT;
> + h.len = sizeof(*hh);
> + h.parent = parent;
> +
> + mutex_lock(&mm->context.lock);
> +
> + hh->ldt_entry_size = LDT_ENTRY_SIZE;
> + hh->nldt = mm->context.size;
> +
> + cr_debug("nldt %d\n", hh->nldt);
> +
> + ret = cr_write_obj(ctx, &h, hh);
> + cr_hbuf_put(ctx, sizeof(*hh));
> + if (ret < 0)
> + goto out;
> +
> + ret = cr_kwrite(ctx, mm->context.ldt,
> + mm->context.size * LDT_ENTRY_SIZE);
> +
> + out:
> + mutex_unlock(&mm->context.lock);
> + return ret;
> +}
> diff --git a/checkpoint/Makefile b/checkpoint/Makefile
> index d2df68c..3a0df6d 100644
> --- a/checkpoint/Makefile
> +++ b/checkpoint/Makefile
> @@ -2,4 +2,5 @@
> # Makefile for linux checkpoint/restart.
> #
>
> -obj-$(CONFIG_CHECKPOINT_RESTART) += sys.o checkpoint.o restart.o
> +obj-$(CONFIG_CHECKPOINT_RESTART) += sys.o checkpoint.o restart.o \
> + ckpt_mem.o
> diff --git a/checkpoint/checkpoint.c b/checkpoint/checkpoint.c
> index 17cc8d2..6a8f810 100644
> --- a/checkpoint/checkpoint.c
> +++ b/checkpoint/checkpoint.c
> @@ -75,6 +75,66 @@ int cr_write_string(struct cr_ctx *ctx, char *str, int len)
> return cr_write_obj(ctx, &h, str);
> }
>
> +/**
> + * cr_fill_fname - return pathname of a given file
> + * @path: path name
> + * @root: relative root
> + * @buf: buffer for pathname
> + * @n: buffer length (in) and pathname length (out)
> + */
> +static char *
> +cr_fill_fname(struct path *path, struct path *root, char *buf, int *n)
> +{
> + struct path tmp = *root;
> + char *fname;
> +
> + BUG_ON(!buf);
> + spin_lock(&dcache_lock);
> + fname = __d_path(path, &tmp, buf, *n);
> + spin_unlock(&dcache_lock);
> + if (!IS_ERR(fname))
> + *n = (buf + (*n) - fname);
> + /*
> + * FIXME: if __d_path() changed these, it must have stepped out of
> + * init's namespace. Since currently we require a unified namespace
> + * within the container: simply fail.
> + */
> + if (tmp.mnt != root->mnt || tmp.dentry != root->dentry)
> + fname = ERR_PTR(-EBADF);
> +
> + return fname;
> +}
> +
> +/**
> + * cr_write_fname - write a file name
> + * @ctx: checkpoint context
> + * @path: path name
> + * @root: relative root
> + */
> +int cr_write_fname(struct cr_ctx *ctx, struct path *path, struct path *root)
> +{
> + struct cr_hdr h;
> + char *buf, *fname;
> + int ret, flen;
> +
> + flen = PATH_MAX;
> + buf = kmalloc(flen, GFP_KERNEL);
> + if (!buf)
> + return -ENOMEM;
> +
> + fname = cr_fill_fname(path, root, buf, &flen);
> + if (!IS_ERR(fname)) {
> + h.type = CR_HDR_FNAME;
> + h.len = flen;
> + h.parent = 0;
> + ret = cr_write_obj(ctx, &h, fname);
> + } else
> + ret = PTR_ERR(fname);
> +
> + kfree(buf);
> + return ret;
> +}
> +
> /* write the checkpoint header */
> static int cr_write_head(struct cr_ctx *ctx)
> {
> @@ -168,6 +228,10 @@ static int cr_write_task(struct cr_ctx *ctx, struct task_struct *t)
> cr_debug("task_struct: ret %d\n", ret);
> if (ret < 0)
> goto out;
> + ret = cr_write_mm(ctx, t);
> + cr_debug("memory: ret %d\n", ret);
> + if (ret < 0)
> + goto out;
> ret = cr_write_thread(ctx, t);
> cr_debug("thread: ret %d\n", ret);
> if (ret < 0)
> @@ -178,10 +242,27 @@ static int cr_write_task(struct cr_ctx *ctx, struct task_struct *t)
> return ret;
> }
>
> +static int cr_ctx_checkpoint(struct cr_ctx *ctx, pid_t pid)
> +{
> + ctx->root_pid = pid;
> +
> + /*
> + * assume checkpointer is in container's root vfs
> + * FIXME: this works for now, but will change with real containers
> + */
> + ctx->vfsroot = ¤t->fs->root;
> + path_get(ctx->vfsroot);
This is going to break as soon as you get another thread doing e.g. chroot(2)
while you are in there. And it's a really, _really_ bad idea to take a
pointer to shared object, increment refcount on the current *contents* of
said object and assume that dropping refcount on the later contents of the
same will balance out.
On Wed, Nov 26, 2008 at 08:04:40PM -0500, Oren Laadan wrote:
> +/**
> + * cr_attach_get_file - attach (and get) lonely file ptr to a file descriptor
> + * @file: lonely file pointer
> + */
> +static int cr_attach_get_file(struct file *file)
> +{
> + int fd = get_unused_fd_flags(0);
> +
> + if (fd >= 0) {
> + fsnotify_open(file->f_path.dentry);
> + fd_install(fd, file);
> + get_file(file);
> + }
> + return fd;
> +}
What happens if another thread closes the descriptor in question between
fd_install() and get_file()?
> + fd = cr_attach_file(file); /* no need to cleanup 'file' below */
> + if (fd < 0) {
> + filp_close(file, NULL);
> + ret = fd;
> + goto out;
> + }
> +
> + /* register new <objref, file> tuple in hash table */
> + ret = cr_obj_add_ref(ctx, file, parent, CR_OBJ_FILE, 0);
> + if (ret < 0)
> + goto out;
Who said that file still exists at that point?
BTW, there are shitloads of races here - references to fd and struct file *
are mixed in a way that breaks *badly* if descriptor table is played with
by another thread.
On Fri, 2008-11-28 at 10:19 +0000, Al Viro wrote:
> On Wed, Nov 26, 2008 at 08:04:39PM -0500, Oren Laadan wrote:
> > +int cr_scan_fds(struct files_struct *files, int **fdtable)
> > +{
> > + struct fdtable *fdt;
> > + int *fds;
> > + int i, n = 0;
> > + int tot = CR_DEFAULT_FDTABLE;
> > +
> > + fds = kmalloc(tot * sizeof(*fds), GFP_KERNEL);
> > + if (!fds)
> > + return -ENOMEM;
> > +
> > + /*
> > + * We assume that the target task is frozen (or that we checkpoint
> > + * ourselves), so we can safely proceed after krealloc() from where
> > + * we left off; in the worst cases restart will fail.
> > + */
>
> Task may be frozen, but it may share the table with any number of other
> tasks...
First of all, thanks for looking at this, Al.
I think Oren's assumption here is that all tasks possibly sharing the
table would be frozen. I don't think that's a good assumption,
either. :)
This would be a lot safer and bulletproof if we size the allocation
ahead of time, take all the locks, then retry if the size has changed.
I think that will just plain work of we do this:
> > + spin_lock(&files->file_lock);
> > + rcu_read_lock();
> > + fdt = files_fdtable(files);
> > + for (i = 0; i < fdt->max_fds; i++) {
> > + if (!fcheck_files(files, i))
> > + continue;
> > + if (n == tot) {
> > + /*
> > + * fcheck_files() is safe with drop/re-acquire
> > + * of the lock, because it tests: fd < max_fds
> > + */
> > + spin_unlock(&files->file_lock);
> > + rcu_read_unlock();
> > + tot *= 2; /* won't overflow: kmalloc will fail */
free(fds);
goto first_kmalloc_in_this_function;
> > + }
> > + fds[n++] = i;
> > + }
> > + rcu_read_unlock();
> > + spin_unlock(&files->file_lock);
> > +
> > + *fdtable = fds;
> > + return n;
> > +}
Right?
> > + switch (inode->i_mode & S_IFMT) {
> > + case S_IFREG:
> > + fd_type = CR_FD_FILE;
> > + break;
> > + case S_IFDIR:
> > + fd_type = CR_FD_DIR;
> > + break;
> > + case S_IFLNK:
> > + fd_type = CR_FD_LINK;
>
> Opened symlinks? May I have whatever you'd been smoking, please?
Ugh, that certainly doesn't have any place here. I do wonder if Oren
had some use for that in the fully put together code, but it can
certainly go for now.
I'll send patches for these shortly.
-- Dave
On Fri, 2008-11-28 at 10:53 +0000, Al Viro wrote:
>
> > +static int cr_ctx_checkpoint(struct cr_ctx *ctx, pid_t pid)
> > +{
> > + ctx->root_pid = pid;
> > +
> > + /*
> > + * assume checkpointer is in container's root vfs
> > + * FIXME: this works for now, but will change with real containers
> > + */
> > + ctx->vfsroot = ¤t->fs->root;
> > + path_get(ctx->vfsroot);
>
> This is going to break as soon as you get another thread doing e.g. chroot(2)
> while you are in there.
Yeah, we do need at least a read_lock(¤t->fs->lock) to keep people
from chroot()'ing underneath us.
> And it's a really, _really_ bad idea to take a
> pointer to shared object, increment refcount on the current *contents* of
> said object and assume that dropping refcount on the later contents of the
> same will balance out.
Absolutely. I assume you mean get_fs_struct(current) instead of
path_get().
-- Dave
On Fri, 2008-11-28 at 10:45 +0000, Al Viro wrote:
> On Wed, Nov 26, 2008 at 08:04:33PM -0500, Oren Laadan wrote:
> > +Currently, namespaces are not saved or restored. They will be treated
> > +as a class of a shared object. In particular, it is assumed that the
> > +task's file system namespace is the "root" for the entire container.
> > +It is also assumed that the same file system view is available for the
> > +restart task(s). Otherwise, a file system snapshot is required.
>
> That is to say, bindings are not handled at all.
Sadly, no. I'm trying to convince Oren that this is important, but I've
been unable to do so thus far. I'd appreciate any particularly
pathological cases you can think of.
There are two cases here that worry me. One is the case where we're
checkpointing a container that has been off in its own mount namespace
doing bind mounts to its little heart's content. We want to checkpoint
and restore the sucker on the same machine. In this case, we almost
certainly want the kernel to be doing the restoration of the binds.
The other case is when we're checkpointing, and moving to a completely
different machine. The new machine may have a completely different disk
layout and *need* the admin doing the sys_restore() to set up those
binds differently because of space constraints or whatever.
For now, we're completely assuming the second case, where the admin
(userspace) is responsible for it, and punting. Do you think this is
something we should take care of now, early in the process?
> > +* What additional work needs to be done to it?
>
> > +We know this design can work. We have two commercial products and a
> > +horde of academic projects doing it today using this basic design.
>
> May I use that for a t-shirt, please? With that quote in foreground, and
> pus-yellow-greenish "MACH" serving as background. With the names of products
> and projects dripping from it...
*Functionally* we know this design can work. Practically, in Linux, I
have no freaking idea. The hard part isn't getting it working, of
course. The hard part is getting something that doesn't step on top of
everyone else's toes in the kernel and, at the same time, is actually
*maintainable*.
-- Dave
On Fri, 2008-11-28 at 11:27 +0000, Al Viro wrote:
> On Wed, Nov 26, 2008 at 08:04:40PM -0500, Oren Laadan wrote:
> > +/**
> > + * cr_attach_get_file - attach (and get) lonely file ptr to a file descriptor
> > + * @file: lonely file pointer
> > + */
> > +static int cr_attach_get_file(struct file *file)
> > +{
> > + int fd = get_unused_fd_flags(0);
> > +
> > + if (fd >= 0) {
> > + fsnotify_open(file->f_path.dentry);
> > + fd_install(fd, file);
> > + get_file(file);
> > + }
> > + return fd;
> > +}
>
> What happens if another thread closes the descriptor in question between
> fd_install() and get_file()?
You're just saying to flip the get_file() and fd_install()?
> > + fd = cr_attach_file(file); /* no need to cleanup 'file' below */
> > + if (fd < 0) {
> > + filp_close(file, NULL);
> > + ret = fd;
> > + goto out;
> > + }
> > +
> > + /* register new <objref, file> tuple in hash table */
> > + ret = cr_obj_add_ref(ctx, file, parent, CR_OBJ_FILE, 0);
> > + if (ret < 0)
> > + goto out;
>
> Who said that file still exists at that point?
Ahhh. We're depending on the 'struct file' reference that comes from
the fd table. That's why there is supposedly "no need to cleanup 'file'
below". But, some other thread can come along and close() the fd, which
will __fput() our poor 'struct file' and will make it go away. Next
time we go and pull it out of the hash table, we go boom.
As a quick fix, I think we can just take another get_file() here. But,
as Al notes, there are some much larger issues that we face with the
fd_table and multi-thread access. They haven't "mattered" to us so far
because we assume everything is either single-threaded or frozen.
Sounds like Al isn't comfortable with this being integrated until a much
more detailed look has been taken.
> BTW, there are shitloads of races here - references to fd and struct file *
> are mixed in a way that breaks *badly* if descriptor table is played with
> by another thread.
One of the things about this that bothers me is that it shares too
little with existing VFS code. It calls into a ton of existing stuff
but doesn't refactor anything that is currently there. Surely there are
some common bits somewhere in the VFS that could be consolidated here.
-- Dave
Dave Hansen wrote:
> On Fri, 2008-11-28 at 10:19 +0000, Al Viro wrote:
>> On Wed, Nov 26, 2008 at 08:04:39PM -0500, Oren Laadan wrote:
>>> +int cr_scan_fds(struct files_struct *files, int **fdtable)
>>> +{
>>> + struct fdtable *fdt;
>>> + int *fds;
>>> + int i, n = 0;
>>> + int tot = CR_DEFAULT_FDTABLE;
>>> +
>>> + fds = kmalloc(tot * sizeof(*fds), GFP_KERNEL);
>>> + if (!fds)
>>> + return -ENOMEM;
>>> +
>>> + /*
>>> + * We assume that the target task is frozen (or that we checkpoint
>>> + * ourselves), so we can safely proceed after krealloc() from where
>>> + * we left off; in the worst cases restart will fail.
>>> + */
>> Task may be frozen, but it may share the table with any number of other
>> tasks...
>
> First of all, thanks for looking at this, Al.
That goes for me too.
>
> I think Oren's assumption here is that all tasks possibly sharing the
> table would be frozen. I don't think that's a good assumption,
> either. :)
>
> This would be a lot safer and bulletproof if we size the allocation
> ahead of time, take all the locks, then retry if the size has changed.
Yes, I assume that all tasks possibly sharing the table are frozen for
this to work "right"; by "right" I mean that if checkpoint completes
then a matching restart would complete successfully.
Verifying that the size doesn't change does not ensure that the table's
contents remained the same, so we can still end up with obsolete data.
For that, you'll need to take the lock over a very long period of time,
while you capture and write data about the files.
So back to the assumption: the idea is that if the user does not adhere
to this assumption - ensuring that all tasks are frozen and that there
are no sharing otherwise - then the results are undefined (just as, e.g.
not calling execve() immediately after vfork() gives undefined results).
By "undefined" I mean that it may produce a checkpoint image that can't
be restarted, or fail. If the assumption doesn't hold, then the table may
change, and we will get a list of fd's which is "incorrect". The comment
before this function explicitly says that "The caller must validate the
file descriptors...", so the checkpoint may will fail subsequently, eg.
while trying to access a bad fd. Otherwise, the image file will reflect
a state that is inconsistent, and restart will fail.
Note that in both cases, the code will not crash the kernel (unless you
think I missed something...)
(BTW, the alternative would be to test for such sharing and abort with
an error. The only way I can think of is to loop through the all the
tasks, for each files_struct that we find. But we don't really need to,
given the argument above).
>
> I think that will just plain work of we do this:
>
>>> + spin_lock(&files->file_lock);
>>> + rcu_read_lock();
>>> + fdt = files_fdtable(files);
>>> + for (i = 0; i < fdt->max_fds; i++) {
>>> + if (!fcheck_files(files, i))
>>> + continue;
>>> + if (n == tot) {
>>> + /*
>>> + * fcheck_files() is safe with drop/re-acquire
>>> + * of the lock, because it tests: fd < max_fds
>>> + */
>>> + spin_unlock(&files->file_lock);
>>> + rcu_read_unlock();
>>> + tot *= 2; /* won't overflow: kmalloc will fail */
>
> free(fds);
> goto first_kmalloc_in_this_function;
>
>>> + }
>>> + fds[n++] = i;
>>> + }
>>> + rcu_read_unlock();
>>> + spin_unlock(&files->file_lock);
>>> +
>>> + *fdtable = fds;
>>> + return n;
>>> +}
>
> Right?
Lol .. that was actually in my original post, and changed in response to
comments that explicitly asked to not count in advance :o
Actually, I think the current code is cleaner, and I don't see how counting
in advance with the lock give us any advantage here.
>>> + switch (inode->i_mode & S_IFMT) {
>>> + case S_IFREG:
>>> + fd_type = CR_FD_FILE;
>>> + break;
>>> + case S_IFDIR:
>>> + fd_type = CR_FD_DIR;
>>> + break;
>>> + case S_IFLNK:
>>> + fd_type = CR_FD_LINK;
>> Opened symlinks? May I have whatever you'd been smoking, please?
>
> Ugh, that certainly doesn't have any place here. I do wonder if Oren
> had some use for that in the fully put together code, but it can
> certainly go for now.
Not anymore. Indeed ugh...
Thanks,
Oren.
>
> I'll send patches for these shortly.
>
> -- Dave
>
>
Dave Hansen wrote:
> On Fri, 2008-11-28 at 11:27 +0000, Al Viro wrote:
>> On Wed, Nov 26, 2008 at 08:04:40PM -0500, Oren Laadan wrote:
>>> +/**
>>> + * cr_attach_get_file - attach (and get) lonely file ptr to a file descriptor
>>> + * @file: lonely file pointer
>>> + */
>>> +static int cr_attach_get_file(struct file *file)
>>> +{
>>> + int fd = get_unused_fd_flags(0);
>>> +
>>> + if (fd >= 0) {
>>> + fsnotify_open(file->f_path.dentry);
>>> + fd_install(fd, file);
>>> + get_file(file);
>>> + }
>>> + return fd;
>>> +}
>> What happens if another thread closes the descriptor in question between
>> fd_install() and get_file()?
>
> You're just saying to flip the get_file() and fd_install()?
Indeed.
>
>>> + fd = cr_attach_file(file); /* no need to cleanup 'file' below */
>>> + if (fd < 0) {
>>> + filp_close(file, NULL);
>>> + ret = fd;
>>> + goto out;
>>> + }
>>> +
>>> + /* register new <objref, file> tuple in hash table */
>>> + ret = cr_obj_add_ref(ctx, file, parent, CR_OBJ_FILE, 0);
>>> + if (ret < 0)
>>> + goto out;
>> Who said that file still exists at that point?
Correct. This call should move higher up befor ethe call to cr_attach_file()
>
> Ahhh. We're depending on the 'struct file' reference that comes from
> the fd table. That's why there is supposedly "no need to cleanup 'file'
> below". But, some other thread can come along and close() the fd, which
> will __fput() our poor 'struct file' and will make it go away. Next
> time we go and pull it out of the hash table, we go boom.
>
> As a quick fix, I think we can just take another get_file() here. But,
> as Al notes, there are some much larger issues that we face with the
> fd_table and multi-thread access. They haven't "mattered" to us so far
> because we assume everything is either single-threaded or frozen.
> Sounds like Al isn't comfortable with this being integrated until a much
> more detailed look has been taken.
>
>> BTW, there are shitloads of races here - references to fd and struct file *
>> are mixed in a way that breaks *badly* if descriptor table is played with
>> by another thread.
The assumption about tasks being frozen and no additional sharing is generally
more strict, more likely to hold, and easier to enforce for the restart.
Besides the race pointed above which would crash the kernel, the other races
are "ok" - if the user abuses the interface, then the results are "undefined"
(refer to my reply to "..PATCH 808/13] Dump open file descriptors").
Here, too, by "undefined" I mean that the restart syscall may fail, and if it
completes successfully the resulting set of tasks is not guaranteed to behave
correctly. In contrast, if the user uses the interface correctly (ensuring
that the assumption holds), then restart is guaranteed to succeed. Note that
even when the outcome is undefined, there are no security issues - all actions
are limited to what the initiating user can do.
> One of the things about this that bothers me is that it shares too
> little with existing VFS code. It calls into a ton of existing stuff
> but doesn't refactor anything that is currently there. Surely there are
> some common bits somewhere in the VFS that could be consolidated here.
Actually, the code alternates between "file" and "fd", in attempt to resuse
existing code and not do things ourselves:
ret = sys_fcntl(fd, F_SETFL, hh->f_flags & CR_SETFL_MASK);
if (ret < 0)
goto out;
ret = vfs_llseek(file, hh->f_pos, SEEK_SET);
if (ret == -ESPIPE) /* ignore error on non-seekable files */
ret = 0;
This is still safe: the file struct is protected with a reference count. If
the fd no longer points to the same struct file, then either it will fail
(e.g. if the fd is invalid) or the restart will eventually succeed but the
resulting state of the tasks will be incorrect (that is: undefined behavior).
Oren.
On Mon, 2008-12-01 at 15:23 -0500, Oren Laadan wrote:
> Verifying that the size doesn't change does not ensure that the table's
> contents remained the same, so we can still end up with obsolete data.
With the realloc() scheme, we have virtually no guarantees about how the
fdtable that we read relates to the source. All that we know is that
the n'th fd was at this value at *some* time.
Using the scheme that I just suggested (and you evidently originally
used) at least guarantees that we have an atomic copy of the fdtable.
Why is this done in two steps? It first grabs a list of fd numbers
which needs to be validated, then goes back and turns those into 'struct
file's which it saves off. Is there a problem with doing that
fd->'struct file' conversion under the files->file_lock?
-- Dave
On Mon, 2008-12-01 at 15:41 -0500, Oren Laadan wrote:
> >>> + fd = cr_attach_file(file); /* no need to cleanup 'file' below */
> >>> + if (fd < 0) {
> >>> + filp_close(file, NULL);
> >>> + ret = fd;
> >>> + goto out;
> >>> + }
> >>> +
> >>> + /* register new <objref, file> tuple in hash table */
> >>> + ret = cr_obj_add_ref(ctx, file, parent, CR_OBJ_FILE, 0);
> >>> + if (ret < 0)
> >>> + goto out;
> >> Who said that file still exists at that point?
>
> Correct. This call should move higher up befor ethe call to cr_attach_file()
Is that sufficient? It seems like we're depending on the fd's reference
to the 'struct file' to keep it valid in the hash. If something happens
to the fd (like the other thread messing with it) the 'struct file' can
still go away.
Shouldn't we do another get_file() for the hash's reference?
-- Dave
Dave Hansen wrote:
> On Fri, 2008-11-28 at 10:53 +0000, Al Viro wrote:
>>> +static int cr_ctx_checkpoint(struct cr_ctx *ctx, pid_t pid)
>>> +{
>>> + ctx->root_pid = pid;
>>> +
>>> + /*
>>> + * assume checkpointer is in container's root vfs
>>> + * FIXME: this works for now, but will change with real containers
>>> + */
>>> + ctx->vfsroot = ¤t->fs->root;
>>> + path_get(ctx->vfsroot);
>> This is going to break as soon as you get another thread doing e.g. chroot(2)
>> while you are in there.
>
> Yeah, we do need at least a read_lock(¤t->fs->lock) to keep people
> from chroot()'ing underneath us.
True.
(while adapting older and safer code I omitted these tests with no reason).
>
>> And it's a really, _really_ bad idea to take a
>> pointer to shared object, increment refcount on the current *contents* of
>> said object and assume that dropping refcount on the later contents of the
>> same will balance out.
>
> Absolutely. I assume you mean get_fs_struct(current) instead of
> path_get().
True.
Should change the type of ctx->vfsroot to not be a pointer, and do:
>>> + ctx->vfsroot = *current->fs->root;
>>> + path_get(&ctx->vfsroot);
and adjust accordingly in where the refcount is dropped.
What we need here is a reference point (this will change later when we handle
multiple fs-namespaces), which is the path of the "container root". Assuming
locking is correct so that current->fs does not change under us, it's enough
to get that path and later release that path.
BW, the current->fs is assumed to not change during the checkpoint; if it does,
then it's a mis-use of the checkpoint interface, and the resulting behavior
is undefined - restart is guaranteed to restore the exact old state even if
checkpoint succeeds.
Thanks,
Oren.
Dave Hansen wrote:
> On Mon, 2008-12-01 at 15:41 -0500, Oren Laadan wrote:
>>>>> + fd = cr_attach_file(file); /* no need to cleanup 'file' below */
>>>>> + if (fd < 0) {
>>>>> + filp_close(file, NULL);
>>>>> + ret = fd;
>>>>> + goto out;
>>>>> + }
>>>>> +
>>>>> + /* register new <objref, file> tuple in hash table */
>>>>> + ret = cr_obj_add_ref(ctx, file, parent, CR_OBJ_FILE, 0);
>>>>> + if (ret < 0)
>>>>> + goto out;
>>>> Who said that file still exists at that point?
>> Correct. This call should move higher up befor ethe call to cr_attach_file()
>
> Is that sufficient? It seems like we're depending on the fd's reference
> to the 'struct file' to keep it valid in the hash. If something happens
> to the fd (like the other thread messing with it) the 'struct file' can
> still go away.
>
> Shouldn't we do another get_file() for the hash's reference?
When a shared object is inserted to the hash we automatically take another
reference to it (according to its type) for as long as it remains in the
hash. See: 'cr_obj_ref_grab()' and 'cr_obj_ref_drop()'. So by moving that
call higher up, we protect the struct file.
Oren.
On Mon, 1 Dec 2008, Dave Hansen wrote:
>
> Why is this done in two steps? It first grabs a list of fd numbers
> which needs to be validated, then goes back and turns those into 'struct
> file's which it saves off. Is there a problem with doing that
> fd->'struct file' conversion under the files->file_lock?
Umm, why do we even worry about this?
Wouldn't it be much better to make sure that all other threads are
stopped before we snapshot, and if we cannot account for some thread (ie
there's some elevated count in the fs/files/mm structures that we cannot
see from the threads we've stopped), just refuse to dump.
There is no sane dump from a multi-threaded app that shares resources
without that kind of serialization _anyway_, so why even try?
In other words: any races in dumping are fundamental _bugs_ in the dumping
at a much higher level. There's absolutely no point in trying to make
something like "dump open fd's" be race-free, because if there are other
people that are actively accessing the 'files' structure concurrently, you
had a much more fundamental bug in the first place!
So do things more like the core-dumping does: make sure that all other
threads are quiescent first!
Linus
On Mon, 2008-12-01 at 16:00 -0500, Oren Laadan wrote:
> > Is that sufficient? It seems like we're depending on the fd's reference
> > to the 'struct file' to keep it valid in the hash. If something happens
> > to the fd (like the other thread messing with it) the 'struct file' can
> > still go away.
> >
> > Shouldn't we do another get_file() for the hash's reference?
>
> When a shared object is inserted to the hash we automatically take another
> reference to it (according to its type) for as long as it remains in the
> hash. See: 'cr_obj_ref_grab()' and 'cr_obj_ref_drop()'. So by moving that
> call higher up, we protect the struct file.
That's kinda (and by kinda I mean really) disgusting. Hiding that two
levels deep in what is effectively the hash table code where no one will
ever see it is really bad. It also makes you lazy thinking that the
hash code will just know how to take references on whatever you give to
it.
I think cr_obj_ref_grab() is hideous obfuscation and needs to die.
Let's just do the get_file() directly, please.
-- Dave
Dave Hansen wrote:
> On Mon, 2008-12-01 at 15:23 -0500, Oren Laadan wrote:
>> Verifying that the size doesn't change does not ensure that the table's
>> contents remained the same, so we can still end up with obsolete data.
>
> With the realloc() scheme, we have virtually no guarantees about how the
> fdtable that we read relates to the source. All that we know is that
> the n'th fd was at this value at *some* time.
>
> Using the scheme that I just suggested (and you evidently originally
> used) at least guarantees that we have an atomic copy of the fdtable.
It doesn't matter if the fdtable changes while we scan the fd's or
after we're done scanning the fd's, does it ? the end result is the
same - inconsistency.
> Why is this done in two steps? It first grabs a list of fd numbers
> which needs to be validated, then goes back and turns those into 'struct
> file's which it saves off. Is there a problem with doing that
> fd->'struct file' conversion under the files->file_lock?
Again, say we get all the struct file for all those fd's. Now we have to
drop the lock, and start processing the struct files. So what if at this
point the fdtable changes ? we still end up with potentially inconsistent
state.
We can do it atomically or non-atomically, I couldn't care less.
My argument is the following:
1) Ideally, the state should not change while we are checkpointing
2) If the state changes, then the behavior is undefined (but the kernel
should not crash, of course).
3) It should be the userspace responsibility to ensure that (1) holds
(for instance, that all tasks sharing these resources are frozen
during the checkpoint)
#2 will be an uncommon event - a programmer/user/administrator mistake.
Kernel may help enforce this rule, but catching all such cases will be
expensive as it means looping through all tasks for each new resource.
As a first step, we can ensure that all threads of a process are frozen
at checkpoint time (and prevent unfreeze until checkpoint is over). But
that will not cover a clone() with CLONE_FS. And this will be the same
for other namespaces later.
Oren.
On Mon, 2008-12-01 at 13:02 -0800, Linus Torvalds wrote:
> On Mon, 1 Dec 2008, Dave Hansen wrote:
> >
> > Why is this done in two steps? It first grabs a list of fd numbers
> > which needs to be validated, then goes back and turns those into 'struct
> > file's which it saves off. Is there a problem with doing that
> > fd->'struct file' conversion under the files->file_lock?
>
> Umm, why do we even worry about this?
>
> Wouldn't it be much better to make sure that all other threads are
> stopped before we snapshot, and if we cannot account for some thread (ie
> there's some elevated count in the fs/files/mm structures that we cannot
> see from the threads we've stopped), just refuse to dump.
My guess is that the mm is probably ok here, but we'll need some work on
the vfs structures, at least eventually.
The mm is nice that it has ->count separated from ->users. We can
easily compare the sum of mm->users to the number of tasks we've frozen
since mm->users has nice defined behavior.
But, we've got suckers like proc_fd_info() that do a
get/put_files_struct(). So, somebody just doing lots of looks in /proc
could stop us from ever checkpointing. Unfortunately, I know of a
number of "monitoring" programs that do just that. :)
I guess we can use the plain counts for now, and add something like
mm->users for the vfs structures in a bit.
-- Dave
On Mon, 2008-12-01 at 16:00 -0500, Oren Laadan wrote:
> Dave Hansen wrote:
> > On Mon, 2008-12-01 at 15:41 -0500, Oren Laadan wrote:
> >>>>> + fd = cr_attach_file(file); /* no need to cleanup 'file' below */
> >>>>> + if (fd < 0) {
> >>>>> + filp_close(file, NULL);
> >>>>> + ret = fd;
> >>>>> + goto out;
> >>>>> + }
> >>>>> +
> >>>>> + /* register new <objref, file> tuple in hash table */
> >>>>> + ret = cr_obj_add_ref(ctx, file, parent, CR_OBJ_FILE, 0);
> >>>>> + if (ret < 0)
> >>>>> + goto out;
> >>>> Who said that file still exists at that point?
> >> Correct. This call should move higher up befor ethe call to cr_attach_file()
> >
> > Is that sufficient? It seems like we're depending on the fd's reference
> > to the 'struct file' to keep it valid in the hash. If something happens
> > to the fd (like the other thread messing with it) the 'struct file' can
> > still go away.
> >
> > Shouldn't we do another get_file() for the hash's reference?
>
> When a shared object is inserted to the hash we automatically take another
> reference to it (according to its type) for as long as it remains in the
> hash. See: 'cr_obj_ref_grab()' and 'cr_obj_ref_drop()'. So by moving that
> call higher up, we protect the struct file.
We also need to document that we depend on this reference in the hash to
keep the object around. Take a look at cr_read_fd_data(). Once that
cr_attach_file() has been performed, the only thing keeping the 'file'
around is the hash reference. If someone happened to remove it from the
hash, the vfs_llseek() below would be bogus.
I don't know how we document that the hash is one-way: writes only and
no later deletions.
-- Dave
On Mon, 2008-12-01 at 13:07 -0800, Dave Hansen wrote:
> > When a shared object is inserted to the hash we automatically take another
> > reference to it (according to its type) for as long as it remains in the
> > hash. See: 'cr_obj_ref_grab()' and 'cr_obj_ref_drop()'. So by moving that
> > call higher up, we protect the struct file.
>
> That's kinda (and by kinda I mean really) disgusting. Hiding that two
> levels deep in what is effectively the hash table code where no one will
> ever see it is really bad. It also makes you lazy thinking that the
> hash code will just know how to take references on whatever you give to
> it.
>
> I think cr_obj_ref_grab() is hideous obfuscation and needs to die.
> Let's just do the get_file() directly, please.
Well, I at least see why you need it now. The objhash thing is trying
to be a pretty generic hash implementation and it does need to free the
references up when it is destroyed. Instead of keeping a "hash of
files" and a "hash of pipes" or other shared objects, there's just a
single hash for everything.
One alternative here would be to have an ops-style release function that
gets called instead of what we have now:
static void cr_obj_ref_drop(struct cr_objref *obj)
{
switch (obj->type) {
case CR_OBJ_FILE:
fput((struct file *) obj->ptr);
break;
default:
BUG();
}
}
static void cr_obj_ref_grab(struct cr_objref *obj)
{
switch (obj->type) {
case CR_OBJ_FILE:
get_file((struct file *) obj->ptr);
break;
default:
BUG();
}
}
That would make it something like:
struct cr_obj_ops {
int type;
void (*release)(struct cr_objref *obj);
};
void cr_release_file(struct cr_objref *obj)
{
struct file *file = obj->ptr;
put_file(file);
}
struct cr_obj_ops cr_file_ops = {
.type = CR_OBJ_FILE,
.release = cr_release_file,
};
And the add operation becomes:
get_file(file);
new = cr_obj_add_ptr(ctx, file, &objref, &cr_file_ops, 0);
with 'cr_file_ops' basically replacing the CR_OBJ_FILE that got passed
before.
I like that because it only obfuscates what truly needs to be abstracted
out: the release side. Hiding that get_file() is really tricky.
But, I guess we could also just kill cr_obj_ref_grab(), do the
get_file() explicitly and still keep cr_obj_ref_drop() as it is now.
-- Dave
Quoting Oren Laadan ([email protected]):
> Checkpoint-restart (c/r): fixes a couple of bugs and a DoS issue
> (tested against v2.6.28-rc3).
Tests well for me. Haven't tested mktree yet, will wait until you
re-send addressing feedback to do that.
thanks,
-serge