From: Shaohua Li <[email protected]>
Hi,
The IO dispatched to under layer disk by loop block device isn't cloned from
original bio, so the IO loses cgroup information of original bio. These IO
escapes from cgroup control. The patches try to address this issue. The idea is
quite generic, but we currently only make it work for blkcg.
Thanks,
Shaohua
Shaohua Li (3):
kthread: add a mechanism to store cgroup info
block: make blkcg aware of kthread stored original cgroup info
block/loop: make loop cgroup aware
block/bio.c | 31 -------------------------
drivers/block/loop.c | 13 +++++++++++
drivers/block/loop.h | 1 +
include/linux/bio.h | 2 --
include/linux/blk-cgroup.h | 25 ++++++--------------
include/linux/kthread.h | 13 +++++++++++
include/linux/sched.h | 1 +
kernel/kthread.c | 57 +++++++++++++++++++++++++++++++++++++++++++++-
8 files changed, 91 insertions(+), 52 deletions(-)
--
2.9.5
From: Shaohua Li <[email protected]>
loop block device handles IO in a separate thread. The actual IO
dispatched isn't cloned from the IO loop device received, so the
dispatched IO loses the cgroup context.
I'm ignoring buffer IO case now, which is quite complicated. Making the
loop thread aware cgroup context doesn't really help. The loop device
only writes to a single file. In current writeback cgroup
implementation, the file can only belong to one cgroup.
For direct IO case, we could workaround the issue in theory. For
example, say we assign cgroup1 5M/s BW for loop device and cgroup2
10M/s. We can create a special cgroup for loop thread and assign at
least 15M/s for the underlayer disk. In this way, we correctly throttle
the two cgroups. But this is tricky to setup.
This patch tries to address the issue. We record bio's css in loop
command. When loop thread is handling the command, we then use the API
provided in patch 1 to set the css for current task. The bio layer will
use the css for new IO.
Signed-off-by: Shaohua Li <[email protected]>
---
drivers/block/loop.c | 13 +++++++++++++
drivers/block/loop.h | 1 +
2 files changed, 14 insertions(+)
diff --git a/drivers/block/loop.c b/drivers/block/loop.c
index 9d4545f..9850b27 100644
--- a/drivers/block/loop.c
+++ b/drivers/block/loop.c
@@ -482,6 +482,8 @@ static void lo_rw_aio_complete(struct kiocb *iocb, long ret, long ret2)
{
struct loop_cmd *cmd = container_of(iocb, struct loop_cmd, iocb);
+ if (cmd->css)
+ css_put(cmd->css);
cmd->ret = ret > 0 ? 0 : ret;
lo_rw_aio_do_completion(cmd);
}
@@ -541,6 +543,8 @@ static int lo_rw_aio(struct loop_device *lo, struct loop_cmd *cmd,
cmd->iocb.ki_filp = file;
cmd->iocb.ki_complete = lo_rw_aio_complete;
cmd->iocb.ki_flags = IOCB_DIRECT;
+ if (cmd->css)
+ kthread_set_orig_css(cmd->css);
if (rw == WRITE)
ret = call_write_iter(file, &cmd->iocb, &iter);
@@ -548,6 +552,7 @@ static int lo_rw_aio(struct loop_device *lo, struct loop_cmd *cmd,
ret = call_read_iter(file, &cmd->iocb, &iter);
lo_rw_aio_do_completion(cmd);
+ kthread_reset_orig_css();
if (ret != -EIOCBQUEUED)
cmd->iocb.ki_complete(&cmd->iocb, ret, 0);
@@ -1692,6 +1697,14 @@ static blk_status_t loop_queue_rq(struct blk_mq_hw_ctx *hctx,
break;
}
+ /* always use the first bio's css */
+#ifdef CONFIG_CGROUPS
+ if (cmd->use_aio && cmd->rq->bio && cmd->rq->bio->bi_css) {
+ cmd->css = cmd->rq->bio->bi_css;
+ css_get(cmd->css);
+ } else
+#endif
+ cmd->css = NULL;
kthread_queue_work(&lo->worker, &cmd->work);
return BLK_STS_OK;
diff --git a/drivers/block/loop.h b/drivers/block/loop.h
index f68c1d5..d93b669 100644
--- a/drivers/block/loop.h
+++ b/drivers/block/loop.h
@@ -74,6 +74,7 @@ struct loop_cmd {
long ret;
struct kiocb iocb;
struct bio_vec *bvec;
+ struct cgroup_subsys_state *css;
};
/* Support for loadable transfer modules */
--
2.9.5
From: Shaohua Li <[email protected]>
Several blkcg APIs are deprecated. After removing them, bio_blkcg is the
only API to get cgroup info for a bio. If bio_blkcg finds current task
is a kthread and has original css recorded, it will use the css instead
of associating the bio to current task.
Signed-off-by: Shaohua Li <[email protected]>
---
block/bio.c | 31 -------------------------------
include/linux/bio.h | 2 --
include/linux/blk-cgroup.h | 25 +++++++------------------
3 files changed, 7 insertions(+), 51 deletions(-)
diff --git a/block/bio.c b/block/bio.c
index 6745759..9271fa3 100644
--- a/block/bio.c
+++ b/block/bio.c
@@ -2033,37 +2033,6 @@ int bio_associate_blkcg(struct bio *bio, struct cgroup_subsys_state *blkcg_css)
EXPORT_SYMBOL_GPL(bio_associate_blkcg);
/**
- * bio_associate_current - associate a bio with %current
- * @bio: target bio
- *
- * Associate @bio with %current if it hasn't been associated yet. Block
- * layer will treat @bio as if it were issued by %current no matter which
- * task actually issues it.
- *
- * This function takes an extra reference of @task's io_context and blkcg
- * which will be put when @bio is released. The caller must own @bio,
- * ensure %current->io_context exists, and is responsible for synchronizing
- * calls to this function.
- */
-int bio_associate_current(struct bio *bio)
-{
- struct io_context *ioc;
-
- if (bio->bi_css)
- return -EBUSY;
-
- ioc = current->io_context;
- if (!ioc)
- return -ENOENT;
-
- get_io_context_active(ioc);
- bio->bi_ioc = ioc;
- bio->bi_css = task_get_css(current, io_cgrp_id);
- return 0;
-}
-EXPORT_SYMBOL_GPL(bio_associate_current);
-
-/**
* bio_disassociate_task - undo bio_associate_current()
* @bio: target bio
*/
diff --git a/include/linux/bio.h b/include/linux/bio.h
index a8fe793..d795cdd 100644
--- a/include/linux/bio.h
+++ b/include/linux/bio.h
@@ -514,13 +514,11 @@ do { \
#ifdef CONFIG_BLK_CGROUP
int bio_associate_blkcg(struct bio *bio, struct cgroup_subsys_state *blkcg_css);
-int bio_associate_current(struct bio *bio);
void bio_disassociate_task(struct bio *bio);
void bio_clone_blkcg_association(struct bio *dst, struct bio *src);
#else /* CONFIG_BLK_CGROUP */
static inline int bio_associate_blkcg(struct bio *bio,
struct cgroup_subsys_state *blkcg_css) { return 0; }
-static inline int bio_associate_current(struct bio *bio) { return -ENOENT; }
static inline void bio_disassociate_task(struct bio *bio) { }
static inline void bio_clone_blkcg_association(struct bio *dst,
struct bio *src) { }
diff --git a/include/linux/blk-cgroup.h b/include/linux/blk-cgroup.h
index 9d92153..0cdcf6b 100644
--- a/include/linux/blk-cgroup.h
+++ b/include/linux/blk-cgroup.h
@@ -19,6 +19,7 @@
#include <linux/radix-tree.h>
#include <linux/blkdev.h>
#include <linux/atomic.h>
+#include <linux/kthread.h>
/* percpu_counter batch for blkg_[rw]stats, per-cpu drift doesn't matter */
#define BLKG_STAT_CPU_BATCH (INT_MAX / 2)
@@ -223,22 +224,16 @@ static inline struct blkcg *css_to_blkcg(struct cgroup_subsys_state *css)
return css ? container_of(css, struct blkcg, css) : NULL;
}
-static inline struct blkcg *task_blkcg(struct task_struct *tsk)
-{
- return css_to_blkcg(task_css(tsk, io_cgrp_id));
-}
-
static inline struct blkcg *bio_blkcg(struct bio *bio)
{
+ struct cgroup_subsys_state *css;
+
if (bio && bio->bi_css)
return css_to_blkcg(bio->bi_css);
- return task_blkcg(current);
-}
-
-static inline struct cgroup_subsys_state *
-task_get_blkcg_css(struct task_struct *task)
-{
- return task_get_css(task, io_cgrp_id);
+ css = kthread_get_orig_css();
+ if (css)
+ return css_to_blkcg(css);
+ return css_to_blkcg(task_css(current, io_cgrp_id));
}
/**
@@ -735,12 +730,6 @@ struct blkcg_policy {
#define blkcg_root_css ((struct cgroup_subsys_state *)ERR_PTR(-EINVAL))
-static inline struct cgroup_subsys_state *
-task_get_blkcg_css(struct task_struct *task)
-{
- return NULL;
-}
-
#ifdef CONFIG_BLOCK
static inline struct blkcg_gq *blkg_lookup(struct blkcg *blkcg, void *key) { return NULL; }
--
2.9.5
From: Shaohua Li <[email protected]>
kthread usually runs jobs on behalf of other threads. The jobs should be
charged to cgroup of original threads. But the jobs run in a kthread,
where we lose the cgroup context of original threads. The patch adds a
machanism to record cgroup info of original threads in kthread context.
Later we can retrieve the cgroup info and attach the cgroup info to jobs.
Since this mechanism is only required by kthread, we store the cgroup
info in kthread data instead of generic task_struct.
Signed-off-by: Shaohua Li <[email protected]>
---
include/linux/kthread.h | 13 +++++++++++
include/linux/sched.h | 1 +
kernel/kthread.c | 57 ++++++++++++++++++++++++++++++++++++++++++++++++-
3 files changed, 70 insertions(+), 1 deletion(-)
diff --git a/include/linux/kthread.h b/include/linux/kthread.h
index 82e197e..3eb24d1 100644
--- a/include/linux/kthread.h
+++ b/include/linux/kthread.h
@@ -3,6 +3,7 @@
/* Simple interface for creating and stopping kernel threads without mess. */
#include <linux/err.h>
#include <linux/sched.h>
+#include <linux/cgroup.h>
__printf(4, 5)
struct task_struct *kthread_create_on_node(int (*threadfn)(void *data),
@@ -198,4 +199,16 @@ bool kthread_cancel_delayed_work_sync(struct kthread_delayed_work *work);
void kthread_destroy_worker(struct kthread_worker *worker);
+#ifdef CONFIG_CGROUPS
+void kthread_set_orig_css(struct cgroup_subsys_state *css);
+struct cgroup_subsys_state *kthread_get_orig_css(void);
+void kthread_reset_orig_css(void);
+#else
+static inline void kthread_set_orig_css(struct cgroup_subsys_state *css) { }
+static inline struct cgroup_subsys_state *kthread_get_orig_css(void)
+{
+ return NULL;
+}
+static inline void kthread_reset_orig_css(void) { }
+#endif
#endif /* _LINUX_KTHREAD_H */
diff --git a/include/linux/sched.h b/include/linux/sched.h
index c05ac5f..ab2295d 100644
--- a/include/linux/sched.h
+++ b/include/linux/sched.h
@@ -1276,6 +1276,7 @@ extern struct pid *cad_pid;
#define PF_SWAPWRITE 0x00800000 /* Allowed to write to swap */
#define PF_NO_SETAFFINITY 0x04000000 /* Userland is not allowed to meddle with cpus_allowed */
#define PF_MCE_EARLY 0x08000000 /* Early kill for mce process policy */
+#define PF_KTHREAD_HAS_CSS 0x10000000 /* kthread has css info attached */
#define PF_MUTEX_TESTER 0x20000000 /* Thread belongs to the rt mutex tester */
#define PF_FREEZER_SKIP 0x40000000 /* Freezer should not count it as freezable */
#define PF_SUSPEND_TASK 0x80000000 /* This thread called freeze_processes() and should not be frozen */
diff --git a/kernel/kthread.c b/kernel/kthread.c
index 26db528..d084ef3 100644
--- a/kernel/kthread.c
+++ b/kernel/kthread.c
@@ -20,7 +20,6 @@
#include <linux/freezer.h>
#include <linux/ptrace.h>
#include <linux/uaccess.h>
-#include <linux/cgroup.h>
#include <trace/events/sched.h>
static DEFINE_SPINLOCK(kthread_create_lock);
@@ -47,6 +46,7 @@ struct kthread {
void *data;
struct completion parked;
struct completion exited;
+ struct cgroup_subsys_state *orig_css;
};
enum KTHREAD_BITS {
@@ -1153,3 +1153,58 @@ void kthread_destroy_worker(struct kthread_worker *worker)
kfree(worker);
}
EXPORT_SYMBOL(kthread_destroy_worker);
+
+#ifdef CONFIG_CGROUPS
+/**
+ * kthread_set_orig_css - set the original css for current thread
+ * @css: the cgroup info
+ *
+ * Current thread must be a kthread. The thread is running jobs on behalf of
+ * other threads. In some cases, we expect the jobs attach cgroup info of
+ * original threads instead of that of current thread. This function stores
+ * original thread's cgroup info in current kthread context for later
+ * retrieval.
+ */
+void kthread_set_orig_css(struct cgroup_subsys_state *css)
+{
+ struct kthread *kthread = to_kthread(current);
+
+ if (css) {
+ css_get(css);
+ kthread->orig_css = css;
+ current->flags |= PF_KTHREAD_HAS_CSS;
+ }
+}
+EXPORT_SYMBOL(kthread_set_orig_css);
+
+/**
+ * kthread_get_orig_css - get the stored original cgroup info
+ *
+ * Current thread must be a kthread.
+ */
+struct cgroup_subsys_state *kthread_get_orig_css(void)
+{
+ if (current->flags & PF_KTHREAD_HAS_CSS)
+ return to_kthread(current)->orig_css;
+ return NULL;
+}
+EXPORT_SYMBOL(kthread_get_orig_css);
+
+/**
+ * kthread_reset_orig_css - clear stored cgroup info
+ *
+ * Current thread must be a kthread.
+ */
+void kthread_reset_orig_css(void)
+{
+ struct kthread *kthread = to_kthread(current);
+ struct cgroup_subsys_state *css;
+
+ css = kthread->orig_css;
+ if (css)
+ css_put(css);
+ kthread->orig_css = NULL;
+ current->flags &= ~PF_KTHREAD_HAS_CSS;
+}
+EXPORT_SYMBOL(kthread_reset_orig_css);
+#endif
--
2.9.5
Hello,
On Wed, Sep 06, 2017 at 07:00:51PM -0700, Shaohua Li wrote:
> +#ifdef CONFIG_CGROUPS
> +void kthread_set_orig_css(struct cgroup_subsys_state *css);
> +struct cgroup_subsys_state *kthread_get_orig_css(void);
> +void kthread_reset_orig_css(void);
* It's a bit weird to associate a kthread with a css without being
specific. If what's needed is a generic association (this kthread
is temporarily servicing this cgroup), it should be associated with
the cgroup. But, I think it'd be better to make it specific instead
- ie. kthread_set_io_css().
* Do we need the reset function to be separate? Can't we just clear
it when the set function is called with NULL css?
* For the accessor, can we do sth like kthread_orig_css() (or
kthread_io_css())? "get" is overloaded between set/get and get/put,
so it can get confusing.
> diff --git a/include/linux/sched.h b/include/linux/sched.h
> index c05ac5f..ab2295d 100644
> --- a/include/linux/sched.h
> +++ b/include/linux/sched.h
> @@ -1276,6 +1276,7 @@ extern struct pid *cad_pid;
> #define PF_SWAPWRITE 0x00800000 /* Allowed to write to swap */
> #define PF_NO_SETAFFINITY 0x04000000 /* Userland is not allowed to meddle with cpus_allowed */
> #define PF_MCE_EARLY 0x08000000 /* Early kill for mce process policy */
> +#define PF_KTHREAD_HAS_CSS 0x10000000 /* kthread has css info attached */
Do we need a separate flag for this? kthreads already have PF_KTHREAD
set. I'm not sure why we'd need another flag. Once we know it's a
kthread, we can just access its css pointer field, right?
Thanks.
--
tejun
On Fri, Sep 08, 2017 at 07:35:37AM -0700, Tejun Heo wrote:
> * It's a bit weird to associate a kthread with a css without being
> specific. If what's needed is a generic association (this kthread
> is temporarily servicing this cgroup), it should be associated with
> the cgroup. But, I think it'd be better to make it specific instead
> - ie. kthread_set_io_css().
kthread[_set]_blkcg_css() probably is a more consistent name.
Thanks.
--
tejun
On Fri, Sep 08, 2017 at 07:38:46AM -0700, Tejun Heo wrote:
> On Fri, Sep 08, 2017 at 07:35:37AM -0700, Tejun Heo wrote:
> > * It's a bit weird to associate a kthread with a css without being
> > specific. If what's needed is a generic association (this kthread
> > is temporarily servicing this cgroup), it should be associated with
> > the cgroup. But, I think it'd be better to make it specific instead
> > - ie. kthread_set_io_css().
>
> kthread[_set]_blkcg_css() probably is a more consistent name.
Sorry, I meant kthread[_associate]_blkcg() so that it's consistent
with bio[_associate]_blkcg().
Thanks.
--
tejun
Hello,
On Wed, Sep 06, 2017 at 07:00:52PM -0700, Shaohua Li wrote:
> From: Shaohua Li <[email protected]>
>
> Several blkcg APIs are deprecated. After removing them, bio_blkcg is the
> only API to get cgroup info for a bio. If bio_blkcg finds current task
> is a kthread and has original css recorded, it will use the css instead
> of associating the bio to current task.
Can you please separate out removal of unused APIs from modifications
of existing ones? The patch is a bit confusing to follow.
Thanks.
--
tejun
Hello,
On Wed, Sep 06, 2017 at 07:00:53PM -0700, Shaohua Li wrote:
> diff --git a/drivers/block/loop.c b/drivers/block/loop.c
> index 9d4545f..9850b27 100644
> --- a/drivers/block/loop.c
> +++ b/drivers/block/loop.c
> @@ -482,6 +482,8 @@ static void lo_rw_aio_complete(struct kiocb *iocb, long ret, long ret2)
> {
> struct loop_cmd *cmd = container_of(iocb, struct loop_cmd, iocb);
>
> + if (cmd->css)
> + css_put(cmd->css);
> cmd->ret = ret > 0 ? 0 : ret;
> lo_rw_aio_do_completion(cmd);
The fact that we're forwarding explicitly in loop still bothers me a
bit. Can you please elaborate why we don't want to do this
generically through aio?
Thanks.
--
tejun
On Fri, Sep 08, 2017 at 07:35:37AM -0700, Tejun Heo wrote:
> Hello,
>
> On Wed, Sep 06, 2017 at 07:00:51PM -0700, Shaohua Li wrote:
> > +#ifdef CONFIG_CGROUPS
> > +void kthread_set_orig_css(struct cgroup_subsys_state *css);
> > +struct cgroup_subsys_state *kthread_get_orig_css(void);
> > +void kthread_reset_orig_css(void);
>
> * It's a bit weird to associate a kthread with a css without being
> specific. If what's needed is a generic association (this kthread
> is temporarily servicing this cgroup), it should be associated with
> the cgroup. But, I think it'd be better to make it specific instead
> - ie. kthread_set_io_css().
>
> * Do we need the reset function to be separate? Can't we just clear
> it when the set function is called with NULL css?
>
> * For the accessor, can we do sth like kthread_orig_css() (or
> kthread_io_css())? "get" is overloaded between set/get and get/put,
> so it can get confusing.
>
> > diff --git a/include/linux/sched.h b/include/linux/sched.h
> > index c05ac5f..ab2295d 100644
> > --- a/include/linux/sched.h
> > +++ b/include/linux/sched.h
> > @@ -1276,6 +1276,7 @@ extern struct pid *cad_pid;
> > #define PF_SWAPWRITE 0x00800000 /* Allowed to write to swap */
> > #define PF_NO_SETAFFINITY 0x04000000 /* Userland is not allowed to meddle with cpus_allowed */
> > #define PF_MCE_EARLY 0x08000000 /* Early kill for mce process policy */
> > +#define PF_KTHREAD_HAS_CSS 0x10000000 /* kthread has css info attached */
>
> Do we need a separate flag for this? kthreads already have PF_KTHREAD
> set. I'm not sure why we'd need another flag. Once we know it's a
> kthread, we can just access its css pointer field, right?
Ok, all suggestions are good, will update.
Thanks,
Shaohua
On Fri, Sep 08, 2017 at 07:48:09AM -0700, Tejun Heo wrote:
> Hello,
>
> On Wed, Sep 06, 2017 at 07:00:53PM -0700, Shaohua Li wrote:
> > diff --git a/drivers/block/loop.c b/drivers/block/loop.c
> > index 9d4545f..9850b27 100644
> > --- a/drivers/block/loop.c
> > +++ b/drivers/block/loop.c
> > @@ -482,6 +482,8 @@ static void lo_rw_aio_complete(struct kiocb *iocb, long ret, long ret2)
> > {
> > struct loop_cmd *cmd = container_of(iocb, struct loop_cmd, iocb);
> >
> > + if (cmd->css)
> > + css_put(cmd->css);
> > cmd->ret = ret > 0 ? 0 : ret;
> > lo_rw_aio_do_completion(cmd);
>
> The fact that we're forwarding explicitly in loop still bothers me a
> bit. Can you please elaborate why we don't want to do this
> generically through aio?
I think we must forward in loop, because each cmd could come from different
cgroup, so we must explicitly forward for each cmd.
The main reason not to do the forward in aio is complexity. We at least have 3
different implementations for dio:
- __blockdev_direct_IO for ext4 and btrfs
- iomap dio for xfs
- blockdev dio implementation
Forwarding in dio means hooking the cgroup association for each bio dispatched
in the implementations, which is a little messy. I'd like to avoid this if
there is no strong reason to do it.
Thanks,
Shaohua
Hello, Shaohua.
On Fri, Sep 08, 2017 at 10:07:15AM -0700, Shaohua Li wrote:
> > The fact that we're forwarding explicitly in loop still bothers me a
> > bit. Can you please elaborate why we don't want to do this
> > generically through aio?
>
> I think we must forward in loop, because each cmd could come from different
> cgroup, so we must explicitly forward for each cmd.
>
> The main reason not to do the forward in aio is complexity. We at least have 3
> different implementations for dio:
> - __blockdev_direct_IO for ext4 and btrfs
> - iomap dio for xfs
> - blockdev dio implementation
>
> Forwarding in dio means hooking the cgroup association for each bio dispatched
> in the implementations, which is a little messy. I'd like to avoid this if
> there is no strong reason to do it.
I see. I think the important questions is whether we're failing to
forward io cgroup membership propagation on some aios? If we are,
that is an obvious bug which should be addressed one way or the other,
and there's a fair chance that we wouldn't need to do anything special
for loop.
Given how simple the loop changes are, we sure can go with loop
specific changes for now; however, I'm a bit unconvinced that aio
changes would be that much more complex. Can you please look into it?
If it is actually complex, we sure can do it later but I'd much prefer
to plug the hole as soon as possible.
Thanks.
--
tejun