Hi all,
I've made some progress on mmiotrace and now it can actually work as a
built-in. Previously mmio-mod.c was full of locking bugs, which never
triggered, because a traced (binary) kernel module depended on the
mmiotrace module.
Changes between preview 1 and preview 2 include:
- Bug and style fixes in Kconfig.debug, Makefile, testmmiotrace.c, mutex
usage, and failure path in register probe function.
- RCU read-lock held over single stepping in kmmio.c
- mmio-mod.c is now a built-in with rewritten locking
- Added debugfs file to enable/disable mmiotracing, since this is not a
module anymore. (Temporary solution)
- Marker file moved from /proc into debugfs.
- mmiotrace entrypoints called directly from ioremap.c, which is why
this cannot be a module anymore.
It still uses relay for pushing data into user space, this would be
replaced by ftrace, along with the "enabled" debugfs file.
The kernel argument mmiotrace.enable_now=1 can be used to enable
mmiotracing at boot. I have tried this and it seems to work well,
I get about 80k events from boot until starting the user space
logger program by hand. Buffers are big enough that no event is
missed due to out-of-memory.
I have also tried tracing Nouveau, started X, and disabled
mmiotracing while in X. Works fine, the logger program just starts
to hog cpu. The logger program will become 'cat' when we get to ftrace.
Hooking or modifying kernel module files is not needed anymore, mmiotrace
will trace everything that is ioremapped after it has been enabled.
Well, everything that __ioremap() actually maps, not including e.g.
ISA area.
The only thing checkpatch.pl complains about is the use of "volatile",
but that is part of the original declaration of iounmap().
The next step would be going for ftrace framework.
This patch is against torvalds/linux-2.6.git master. I will reply to
this email with a patch for Ingo's x86/testing branch.
Signed-off-by: Pekka Paalanen <[email protected]>
---
arch/x86/Kconfig.debug | 29 ++
arch/x86/mm/Makefile | 5 +
arch/x86/mm/fault.c | 13 +
arch/x86/mm/ioremap.c | 9 +-
arch/x86/mm/kmmio.c | 533 ++++++++++++++++++++++++++++++++++
arch/x86/mm/mmio-mod.c | 666 +++++++++++++++++++++++++++++++++++++++++++
arch/x86/mm/pageattr.c | 1 +
arch/x86/mm/pf_in.c | 489 +++++++++++++++++++++++++++++++
arch/x86/mm/pf_in.h | 39 +++
arch/x86/mm/testmmiotrace.c | 71 +++++
include/linux/mmiotrace.h | 120 ++++++++
11 files changed, 1974 insertions(+), 1 deletions(-)
diff --git a/arch/x86/Kconfig.debug b/arch/x86/Kconfig.debug
index 702eb39..6ab0ca2 100644
--- a/arch/x86/Kconfig.debug
+++ b/arch/x86/Kconfig.debug
@@ -134,6 +134,35 @@ config IOMMU_LEAK
Add a simple leak tracer to the IOMMU code. This is useful when you
are debugging a buggy device driver that leaks IOMMU mappings.
+config MMIOTRACE_HOOKS
+ bool
+
+config MMIOTRACE
+ bool "Memory mapped IO tracing"
+ depends on DEBUG_KERNEL && RELAY && DEBUG_FS
+ select MMIOTRACE_HOOKS
+ default y
+ help
+ Mmiotrace traces Memory Mapped I/O access and is meant for
+ debugging and reverse engineering. It is called from the ioremap
+ implementation and works via page faults. A user space program is
+ required to collect the MMIO data from debugfs files.
+ Tracing is disabled by default and can be enabled from a debugfs
+ file.
+
+ See http://nouveau.freedesktop.org/wiki/MmioTrace
+ If you are not helping to develop drivers, say N.
+
+config MMIOTRACE_TEST
+ tristate "Test module for mmiotrace"
+ depends on MMIOTRACE && m
+ help
+ This is a dumb module for testing mmiotrace. It is very dangerous
+ as it will write garbage to IO memory starting at a given address.
+ However, it should be safe to use on e.g. unused portion of VRAM.
+
+ Say N, unless you absolutely know what you are doing.
+
#
# IO delay types:
#
diff --git a/arch/x86/mm/Makefile b/arch/x86/mm/Makefile
index 9832910..3423865 100644
--- a/arch/x86/mm/Makefile
+++ b/arch/x86/mm/Makefile
@@ -3,3 +3,8 @@ include ${srctree}/arch/x86/mm/Makefile_32
else
include ${srctree}/arch/x86/mm/Makefile_64
endif
+
+obj-$(CONFIG_MMIOTRACE_HOOKS) += kmmio.o
+obj-$(CONFIG_MMIOTRACE) += mmiotrace.o
+mmiotrace-y := pf_in.o mmio-mod.o
+obj-$(CONFIG_MMIOTRACE_TEST) += testmmiotrace.o
diff --git a/arch/x86/mm/fault.c b/arch/x86/mm/fault.c
index fdc6674..ba6d4c5 100644
--- a/arch/x86/mm/fault.c
+++ b/arch/x86/mm/fault.c
@@ -25,6 +25,7 @@
#include <linux/kprobes.h>
#include <linux/uaccess.h>
#include <linux/kdebug.h>
+#include <linux/mmiotrace.h>
#include <asm/system.h>
#include <asm/desc.h>
@@ -49,6 +50,16 @@
#define PF_RSVD (1<<3)
#define PF_INSTR (1<<4)
+static inline int kmmio_fault(struct pt_regs *regs, unsigned long addr)
+{
+#ifdef CONFIG_MMIOTRACE_HOOKS
+ if (unlikely(is_kmmio_active()))
+ if (kmmio_handler(regs, addr) == 1)
+ return -1;
+#endif
+ return 0;
+}
+
static inline int notify_page_fault(struct pt_regs *regs)
{
#ifdef CONFIG_KPROBES
@@ -603,6 +614,8 @@ void __kprobes do_page_fault(struct pt_regs *regs, unsigned long error_code)
if (notify_page_fault(regs))
return;
+ if (unlikely(kmmio_fault(regs, address)))
+ return;
/*
* We fault-in kernel-space virtual memory on-demand. The
diff --git a/arch/x86/mm/ioremap.c b/arch/x86/mm/ioremap.c
index ac3c959..bfaa51f 100644
--- a/arch/x86/mm/ioremap.c
+++ b/arch/x86/mm/ioremap.c
@@ -12,6 +12,7 @@
#include <linux/module.h>
#include <linux/slab.h>
#include <linux/vmalloc.h>
+#include <linux/mmiotrace.h>
#include <asm/cacheflush.h>
#include <asm/e820.h>
@@ -112,6 +113,7 @@ static void __iomem *__ioremap(unsigned long phys_addr, unsigned long size,
unsigned long pfn, offset, last_addr, vaddr;
struct vm_struct *area;
pgprot_t prot;
+ void __iomem *ret_addr;
/* Don't allow wraparound or zero size */
last_addr = phys_addr + size - 1;
@@ -171,7 +173,10 @@ static void __iomem *__ioremap(unsigned long phys_addr, unsigned long size,
return NULL;
}
- return (void __iomem *) (vaddr + offset);
+ ret_addr = (void __iomem *) (vaddr + offset);
+ mmiotrace_ioremap(phys_addr, size, ret_addr);
+
+ return ret_addr;
}
/**
@@ -229,6 +234,8 @@ void iounmap(volatile void __iomem *addr)
addr < phys_to_virt(ISA_END_ADDRESS))
return;
+ mmiotrace_iounmap(addr);
+
addr = (volatile void __iomem *)
(PAGE_MASK & (unsigned long __force)addr);
diff --git a/arch/x86/mm/kmmio.c b/arch/x86/mm/kmmio.c
new file mode 100644
index 0000000..efb4679
--- /dev/null
+++ b/arch/x86/mm/kmmio.c
@@ -0,0 +1,533 @@
+/* Support for MMIO probes.
+ * Benfit many code from kprobes
+ * (C) 2002 Louis Zhuang <[email protected]>.
+ * 2007 Alexander Eichner
+ * 2008 Pekka Paalanen <[email protected]>
+ */
+
+#include <linux/version.h>
+#include <linux/list.h>
+#include <linux/spinlock.h>
+#include <linux/hash.h>
+#include <linux/init.h>
+#include <linux/module.h>
+#include <linux/slab.h>
+#include <linux/kernel.h>
+#include <linux/mm.h>
+#include <linux/uaccess.h>
+#include <linux/ptrace.h>
+#include <linux/preempt.h>
+#include <linux/percpu.h>
+#include <linux/kdebug.h>
+#include <linux/mutex.h>
+#include <asm/io.h>
+#include <asm/cacheflush.h>
+#include <asm/errno.h>
+#include <asm/tlbflush.h>
+#include <asm/pgtable.h>
+
+#include <linux/mmiotrace.h>
+
+#define KMMIO_PAGE_HASH_BITS 4
+#define KMMIO_PAGE_TABLE_SIZE (1 << KMMIO_PAGE_HASH_BITS)
+
+struct kmmio_fault_page {
+ struct list_head list;
+ struct kmmio_fault_page *release_next;
+ unsigned long page; /* location of the fault page */
+
+ /*
+ * Number of times this page has been registered as a part
+ * of a probe. If zero, page is disarmed and this may be freed.
+ * Used only by writers (RCU).
+ */
+ int count;
+};
+
+struct kmmio_delayed_release {
+ struct rcu_head rcu;
+ struct kmmio_fault_page *release_list;
+};
+
+struct kmmio_context {
+ struct kmmio_fault_page *fpage;
+ struct kmmio_probe *probe;
+ unsigned long saved_flags;
+ unsigned long addr;
+ int active;
+};
+
+static int kmmio_die_notifier(struct notifier_block *nb, unsigned long val,
+ void *args);
+
+static DEFINE_MUTEX(kmmio_init_mutex);
+static DEFINE_SPINLOCK(kmmio_lock);
+
+/* These are protected by kmmio_lock */
+static int kmmio_initialized;
+unsigned int kmmio_count;
+
+/* Read-protected by RCU, write-protected by kmmio_lock. */
+static struct list_head kmmio_page_table[KMMIO_PAGE_TABLE_SIZE];
+static LIST_HEAD(kmmio_probes);
+
+static struct list_head *kmmio_page_list(unsigned long page)
+{
+ return &kmmio_page_table[hash_long(page, KMMIO_PAGE_HASH_BITS)];
+}
+
+/* Accessed per-cpu */
+static DEFINE_PER_CPU(struct kmmio_context, kmmio_ctx);
+
+/* protected by kmmio_init_mutex */
+static struct notifier_block nb_die = {
+ .notifier_call = kmmio_die_notifier
+};
+
+/**
+ * Makes sure kmmio is initialized and usable.
+ * This must be called before any other kmmio function defined here.
+ * May sleep.
+ */
+void reference_kmmio(void)
+{
+ mutex_lock(&kmmio_init_mutex);
+ spin_lock_irq(&kmmio_lock);
+ if (!kmmio_initialized) {
+ int i;
+ for (i = 0; i < KMMIO_PAGE_TABLE_SIZE; i++)
+ INIT_LIST_HEAD(&kmmio_page_table[i]);
+ if (register_die_notifier(&nb_die))
+ BUG();
+ }
+ kmmio_initialized++;
+ spin_unlock_irq(&kmmio_lock);
+ mutex_unlock(&kmmio_init_mutex);
+}
+EXPORT_SYMBOL_GPL(reference_kmmio);
+
+/**
+ * Clean up kmmio after use. This must be called for every call to
+ * reference_kmmio(). All probes registered after the corresponding
+ * reference_kmmio() must have been unregistered when calling this.
+ * May sleep.
+ */
+void unreference_kmmio(void)
+{
+ bool unreg = false;
+
+ mutex_lock(&kmmio_init_mutex);
+ spin_lock_irq(&kmmio_lock);
+
+ if (kmmio_initialized == 1) {
+ BUG_ON(is_kmmio_active());
+ unreg = true;
+ }
+ kmmio_initialized--;
+ BUG_ON(kmmio_initialized < 0);
+ spin_unlock_irq(&kmmio_lock);
+
+ if (unreg)
+ unregister_die_notifier(&nb_die); /* calls sync_rcu() */
+ mutex_unlock(&kmmio_init_mutex);
+}
+EXPORT_SYMBOL(unreference_kmmio);
+
+/*
+ * this is basically a dynamic stabbing problem:
+ * Could use the existing prio tree code or
+ * Possible better implementations:
+ * The Interval Skip List: A Data Structure for Finding All Intervals That
+ * Overlap a Point (might be simple)
+ * Space Efficient Dynamic Stabbing with Fast Queries - Mikkel Thorup
+ */
+/* Get the kmmio at this addr (if any). You must be holding RCU read lock. */
+static struct kmmio_probe *get_kmmio_probe(unsigned long addr)
+{
+ struct kmmio_probe *p;
+ list_for_each_entry_rcu(p, &kmmio_probes, list) {
+ if (addr >= p->addr && addr <= (p->addr + p->len))
+ return p;
+ }
+ return NULL;
+}
+
+/* You must be holding RCU read lock. */
+static struct kmmio_fault_page *get_kmmio_fault_page(unsigned long page)
+{
+ struct list_head *head;
+ struct kmmio_fault_page *p;
+
+ page &= PAGE_MASK;
+ head = kmmio_page_list(page);
+ list_for_each_entry_rcu(p, head, list) {
+ if (p->page == page)
+ return p;
+ }
+ return NULL;
+}
+
+/** Mark the given page as not present. Access to it will trigger a fault. */
+static void arm_kmmio_fault_page(unsigned long page, int *page_level)
+{
+ unsigned long address = page & PAGE_MASK;
+ int level;
+ pte_t *pte = lookup_address(address, &level);
+
+ if (!pte) {
+ pr_err("kmmio: Error in %s: no pte for page 0x%08lx\n",
+ __func__, page);
+ return;
+ }
+
+ if (level == PG_LEVEL_2M) {
+ pmd_t *pmd = (pmd_t *)pte;
+ set_pmd(pmd, __pmd(pmd_val(*pmd) & ~_PAGE_PRESENT));
+ } else {
+ /* PG_LEVEL_4K */
+ set_pte(pte, __pte(pte_val(*pte) & ~_PAGE_PRESENT));
+ }
+
+ if (page_level)
+ *page_level = level;
+
+ __flush_tlb_one(page);
+}
+
+/** Mark the given page as present. */
+static void disarm_kmmio_fault_page(unsigned long page, int *page_level)
+{
+ unsigned long address = page & PAGE_MASK;
+ int level;
+ pte_t *pte = lookup_address(address, &level);
+
+ if (!pte) {
+ pr_err("kmmio: Error in %s: no pte for page 0x%08lx\n",
+ __func__, page);
+ return;
+ }
+
+ if (level == PG_LEVEL_2M) {
+ pmd_t *pmd = (pmd_t *)pte;
+ set_pmd(pmd, __pmd(pmd_val(*pmd) | _PAGE_PRESENT));
+ } else {
+ /* PG_LEVEL_4K */
+ set_pte(pte, __pte(pte_val(*pte) | _PAGE_PRESENT));
+ }
+
+ if (page_level)
+ *page_level = level;
+
+ __flush_tlb_one(page);
+}
+
+/*
+ * This is being called from do_page_fault().
+ *
+ * We may be in an interrupt or a critical section. Also prefecthing may
+ * trigger a page fault. We may be in the middle of process switch.
+ * We cannot take any locks, because we could be executing especially
+ * within a kmmio critical section.
+ *
+ * Local interrupts are disabled, so preemption cannot happen.
+ * Do not enable interrupts, do not sleep, and watch out for other CPUs.
+ */
+/*
+ * Interrupts are disabled on entry as trap3 is an interrupt gate
+ * and they remain disabled thorough out this function.
+ */
+int kmmio_handler(struct pt_regs *regs, unsigned long addr)
+{
+ struct kmmio_context *ctx;
+ struct kmmio_fault_page *faultpage;
+
+ /*
+ * Preemption is now disabled to prevent process switch during
+ * single stepping. We can only handle one active kmmio trace
+ * per cpu, so ensure that we finish it before something else
+ * gets to run. We also hold the RCU read lock over single
+ * stepping to avoid looking up the probe and kmmio_fault_page
+ * again.
+ */
+ preempt_disable();
+ rcu_read_lock();
+
+ faultpage = get_kmmio_fault_page(addr);
+ if (!faultpage) {
+ /*
+ * Either this page fault is not caused by kmmio, or
+ * another CPU just pulled the kmmio probe from under
+ * our feet. In the latter case all hell breaks loose.
+ */
+ goto no_kmmio;
+ }
+
+ ctx = &get_cpu_var(kmmio_ctx);
+ if (ctx->active) {
+ /*
+ * Prevent overwriting already in-flight context.
+ * If this page fault really was due to kmmio trap,
+ * all hell breaks loose.
+ */
+ pr_emerg("kmmio: recursive probe hit on CPU %d, "
+ "for address 0x%08lx. Ignoring.\n",
+ smp_processor_id(), addr);
+ goto no_kmmio_ctx;
+ }
+ ctx->active++;
+
+ ctx->fpage = faultpage;
+ ctx->probe = get_kmmio_probe(addr);
+ ctx->saved_flags = (regs->flags & (TF_MASK|IF_MASK));
+ ctx->addr = addr;
+
+ if (ctx->probe && ctx->probe->pre_handler)
+ ctx->probe->pre_handler(ctx->probe, regs, addr);
+
+ /*
+ * Enable single-stepping and disable interrupts for the faulting
+ * context. Local interrupts must not get enabled during stepping.
+ */
+ regs->flags |= TF_MASK;
+ regs->flags &= ~IF_MASK;
+
+ /* Now we set present bit in PTE and single step. */
+ disarm_kmmio_fault_page(ctx->fpage->page, NULL);
+
+ /*
+ * If another cpu accesses the same page while we are stepping,
+ * the access will not be caught. It will simply succeed and the
+ * only downside is we lose the event. If this becomes a problem,
+ * the user should drop to single cpu before tracing.
+ */
+
+ put_cpu_var(kmmio_ctx);
+ return 1;
+
+no_kmmio_ctx:
+ put_cpu_var(kmmio_ctx);
+no_kmmio:
+ rcu_read_unlock();
+ preempt_enable_no_resched();
+ return 0; /* page fault not handled by kmmio */
+}
+
+/*
+ * Interrupts are disabled on entry as trap1 is an interrupt gate
+ * and they remain disabled thorough out this function.
+ * This must always get called as the pair to kmmio_handler().
+ */
+static int post_kmmio_handler(unsigned long condition, struct pt_regs *regs)
+{
+ int ret = 0;
+ struct kmmio_context *ctx = &get_cpu_var(kmmio_ctx);
+
+ if (!ctx->active)
+ goto out;
+
+ if (ctx->probe && ctx->probe->post_handler)
+ ctx->probe->post_handler(ctx->probe, condition, regs);
+
+ arm_kmmio_fault_page(ctx->fpage->page, NULL);
+
+ regs->flags &= ~TF_MASK;
+ regs->flags |= ctx->saved_flags;
+
+ /* These were acquired in kmmio_handler(). */
+ ctx->active--;
+ BUG_ON(ctx->active);
+ rcu_read_unlock();
+ preempt_enable_no_resched();
+
+ /*
+ * if somebody else is singlestepping across a probe point, flags
+ * will have TF set, in which case, continue the remaining processing
+ * of do_debug, as if this is not a probe hit.
+ */
+ if (!(regs->flags & TF_MASK))
+ ret = 1;
+out:
+ put_cpu_var(kmmio_ctx);
+ return ret;
+}
+
+/* You must be holding kmmio_lock. */
+static int add_kmmio_fault_page(unsigned long page)
+{
+ struct kmmio_fault_page *f;
+
+ page &= PAGE_MASK;
+ f = get_kmmio_fault_page(page);
+ if (f) {
+ if (!f->count)
+ arm_kmmio_fault_page(f->page, NULL);
+ f->count++;
+ return 0;
+ }
+
+ f = kmalloc(sizeof(*f), GFP_ATOMIC);
+ if (!f)
+ return -1;
+
+ f->count = 1;
+ f->page = page;
+ list_add_rcu(&f->list, kmmio_page_list(f->page));
+
+ arm_kmmio_fault_page(f->page, NULL);
+
+ return 0;
+}
+
+/* You must be holding kmmio_lock. */
+static void release_kmmio_fault_page(unsigned long page,
+ struct kmmio_fault_page **release_list)
+{
+ struct kmmio_fault_page *f;
+
+ page &= PAGE_MASK;
+ f = get_kmmio_fault_page(page);
+ if (!f)
+ return;
+
+ f->count--;
+ BUG_ON(f->count < 0);
+ if (!f->count) {
+ disarm_kmmio_fault_page(f->page, NULL);
+ f->release_next = *release_list;
+ *release_list = f;
+ }
+}
+
+int register_kmmio_probe(struct kmmio_probe *p)
+{
+ unsigned long flags;
+ int ret = 0;
+ unsigned long size = 0;
+
+ spin_lock_irqsave(&kmmio_lock, flags);
+ if (get_kmmio_probe(p->addr)) {
+ ret = -EEXIST;
+ goto out;
+ }
+ kmmio_count++;
+ list_add_rcu(&p->list, &kmmio_probes);
+ while (size < p->len) {
+ if (add_kmmio_fault_page(p->addr + size))
+ pr_err("kmmio: Unable to set page fault.\n");
+ size += PAGE_SIZE;
+ }
+out:
+ spin_unlock_irqrestore(&kmmio_lock, flags);
+ /*
+ * XXX: What should I do here?
+ * Here was a call to global_flush_tlb(), but it does not exist
+ * anymore. It seems it's not needed after all.
+ */
+ return ret;
+}
+EXPORT_SYMBOL(register_kmmio_probe);
+
+static void rcu_free_kmmio_fault_pages(struct rcu_head *head)
+{
+ struct kmmio_delayed_release *dr = container_of(
+ head,
+ struct kmmio_delayed_release,
+ rcu);
+ struct kmmio_fault_page *p = dr->release_list;
+ while (p) {
+ struct kmmio_fault_page *next = p->release_next;
+ BUG_ON(p->count);
+ kfree(p);
+ p = next;
+ }
+ kfree(dr);
+}
+
+static void remove_kmmio_fault_pages(struct rcu_head *head)
+{
+ struct kmmio_delayed_release *dr = container_of(
+ head,
+ struct kmmio_delayed_release,
+ rcu);
+ struct kmmio_fault_page *p = dr->release_list;
+ struct kmmio_fault_page **prevp = &dr->release_list;
+ unsigned long flags;
+ spin_lock_irqsave(&kmmio_lock, flags);
+ while (p) {
+ if (!p->count)
+ list_del_rcu(&p->list);
+ else
+ *prevp = p->release_next;
+ prevp = &p->release_next;
+ p = p->release_next;
+ }
+ spin_unlock_irqrestore(&kmmio_lock, flags);
+ /* This is the real RCU destroy call. */
+ call_rcu(&dr->rcu, rcu_free_kmmio_fault_pages);
+}
+
+/*
+ * Remove a kmmio probe. You have to synchronize_rcu() before you can be
+ * sure that the callbacks will not be called anymore. Only after that
+ * you may actually release your struct kmmio_probe.
+ *
+ * Unregistering a kmmio fault page has three steps:
+ * 1. release_kmmio_fault_page()
+ * Disarm the page, wait a grace period to let all faults finish.
+ * 2. remove_kmmio_fault_pages()
+ * Remove the pages from kmmio_page_table.
+ * 3. rcu_free_kmmio_fault_pages()
+ * Actally free the kmmio_fault_page structs as with RCU.
+ */
+void unregister_kmmio_probe(struct kmmio_probe *p)
+{
+ unsigned long flags;
+ unsigned long size = 0;
+ struct kmmio_fault_page *release_list = NULL;
+ struct kmmio_delayed_release *drelease;
+
+ spin_lock_irqsave(&kmmio_lock, flags);
+ while (size < p->len) {
+ release_kmmio_fault_page(p->addr + size, &release_list);
+ size += PAGE_SIZE;
+ }
+ list_del_rcu(&p->list);
+ kmmio_count--;
+ spin_unlock_irqrestore(&kmmio_lock, flags);
+
+ drelease = kmalloc(sizeof(*drelease), GFP_ATOMIC);
+ if (!drelease) {
+ pr_crit("kmmio: leaking kmmio_fault_page objects.\n");
+ return;
+ }
+ drelease->release_list = release_list;
+
+ /*
+ * This is not really RCU here. We have just disarmed a set of
+ * pages so that they cannot trigger page faults anymore. However,
+ * we cannot remove the pages from kmmio_page_table,
+ * because a probe hit might be in flight on another CPU. The
+ * pages are collected into a list, and they will be removed from
+ * kmmio_page_table when it is certain that no probe hit related to
+ * these pages can be in flight. RCU grace period sounds like a
+ * good choice.
+ *
+ * If we removed the pages too early, kmmio page fault handler might
+ * not find the respective kmmio_fault_page and determine it's not
+ * a kmmio fault, when it actually is. This would lead to madness.
+ */
+ call_rcu(&drelease->rcu, remove_kmmio_fault_pages);
+}
+EXPORT_SYMBOL(unregister_kmmio_probe);
+
+static int kmmio_die_notifier(struct notifier_block *nb, unsigned long val,
+ void *args)
+{
+ struct die_args *arg = args;
+
+ if (val == DIE_DEBUG)
+ if (post_kmmio_handler(arg->err, arg->regs) == 1)
+ return NOTIFY_STOP;
+
+ return NOTIFY_DONE;
+}
diff --git a/arch/x86/mm/mmio-mod.c b/arch/x86/mm/mmio-mod.c
new file mode 100644
index 0000000..7386440
--- /dev/null
+++ b/arch/x86/mm/mmio-mod.c
@@ -0,0 +1,666 @@
+/*
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; either version 2 of the License, or
+ * (at your option) any later version.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
+ *
+ * Copyright (C) IBM Corporation, 2005
+ * Jeff Muizelaar, 2006, 2007
+ * Pekka Paalanen, 2008 <[email protected]>
+ *
+ * Derived from the read-mod example from relay-examples by Tom Zanussi.
+ */
+#define DEBUG 1
+
+#include <linux/module.h>
+#include <linux/relay.h>
+#include <linux/debugfs.h>
+#include <linux/proc_fs.h>
+#include <asm/io.h>
+#include <linux/version.h>
+#include <linux/kallsyms.h>
+#include <asm/pgtable.h>
+#include <linux/mmiotrace.h>
+#include <asm/e820.h> /* for ISA_START_ADDRESS */
+#include <asm/atomic.h>
+#include <linux/percpu.h>
+
+#include "pf_in.h"
+
+#define NAME "mmiotrace: "
+
+/* This app's relay channel files will appear in /debug/mmio-trace */
+static const char APP_DIR[] = "mmio-trace";
+/* the marker injection file in /debug/APP_DIR */
+static const char MARKER_FILE[] = "mmio-marker";
+
+struct trap_reason {
+ unsigned long addr;
+ unsigned long ip;
+ enum reason_type type;
+ int active_traces;
+};
+
+struct remap_trace {
+ struct list_head list;
+ struct kmmio_probe probe;
+ unsigned long phys;
+ unsigned long id;
+};
+
+static const size_t subbuf_size = 256*1024;
+
+/* Accessed per-cpu. */
+static DEFINE_PER_CPU(struct trap_reason, pf_reason);
+static DEFINE_PER_CPU(struct mm_io_header_rw, cpu_trace);
+
+/* Access to this is not per-cpu. */
+static DEFINE_PER_CPU(atomic_t, dropped);
+
+static struct dentry *dir;
+static struct dentry *enabled_file;
+static struct dentry *marker_file;
+
+static DEFINE_MUTEX(mmiotrace_mutex);
+static DEFINE_SPINLOCK(trace_lock);
+static atomic_t mmiotrace_enabled;
+static LIST_HEAD(trace_list); /* struct remap_trace */
+static struct rchan *chan;
+
+/*
+ * Locking in this file:
+ * - mmiotrace_mutex enforces enable/disable_mmiotrace() critical sections.
+ * - mmiotrace_enabled may be modified only when holding mmiotrace_mutex
+ * and trace_lock.
+ * - Routines depending on is_enabled() must take trace_lock.
+ * - trace_list users must hold trace_lock.
+ * - is_enabled() guarantees that chan is valid.
+ * - pre/post callbacks assume the effect of is_enabled() being true.
+ */
+
+/* module parameters */
+static unsigned int n_subbufs = 32*4;
+static unsigned long filter_offset;
+static int nommiotrace;
+static int ISA_trace;
+static int trace_pc;
+static int enable_now;
+
+module_param(n_subbufs, uint, 0);
+module_param(filter_offset, ulong, 0);
+module_param(nommiotrace, bool, 0);
+module_param(ISA_trace, bool, 0);
+module_param(trace_pc, bool, 0);
+module_param(enable_now, bool, 0);
+
+MODULE_PARM_DESC(n_subbufs, "Number of 256kB buffers, default 128.");
+MODULE_PARM_DESC(filter_offset, "Start address of traced mappings.");
+MODULE_PARM_DESC(nommiotrace, "Disable actual MMIO tracing.");
+MODULE_PARM_DESC(ISA_trace, "Do not exclude the low ISA range.");
+MODULE_PARM_DESC(trace_pc, "Record address of faulting instructions.");
+MODULE_PARM_DESC(enable_now, "Start mmiotracing immediately on module load.");
+
+static bool is_enabled(void)
+{
+ return atomic_read(&mmiotrace_enabled);
+}
+
+static void record_timestamp(struct mm_io_header *header)
+{
+ struct timespec now;
+
+ getnstimeofday(&now);
+ header->sec = now.tv_sec;
+ header->nsec = now.tv_nsec;
+}
+
+/*
+ * Write callback for the debugfs entry:
+ * Read a marker and write it to the mmio trace log
+ */
+static ssize_t write_marker(struct file *file, const char __user *buffer,
+ size_t count, loff_t *ppos)
+{
+ char *event = NULL;
+ struct mm_io_header *headp;
+ ssize_t len = (count > 65535) ? 65535 : count;
+
+ event = kzalloc(sizeof(*headp) + len, GFP_KERNEL);
+ if (!event)
+ return -ENOMEM;
+
+ headp = (struct mm_io_header *)event;
+ headp->type = MMIO_MAGIC | (MMIO_MARKER << MMIO_OPCODE_SHIFT);
+ headp->data_len = len;
+ record_timestamp(headp);
+
+ if (copy_from_user(event + sizeof(*headp), buffer, len)) {
+ kfree(event);
+ return -EFAULT;
+ }
+
+ spin_lock_irq(&trace_lock);
+ if (is_enabled())
+ relay_write(chan, event, sizeof(*headp) + len);
+ else
+ len = -EINVAL;
+ spin_unlock_irq(&trace_lock);
+ kfree(event);
+ return len;
+}
+
+static void print_pte(unsigned long address)
+{
+ int level;
+ pte_t *pte = lookup_address(address, &level);
+
+ if (!pte) {
+ pr_err(NAME "Error in %s: no pte for page 0x%08lx\n",
+ __func__, address);
+ return;
+ }
+
+ if (level == PG_LEVEL_2M) {
+ pr_emerg(NAME "4MB pages are not currently supported: "
+ "0x%08lx\n", address);
+ BUG();
+ }
+ pr_info(NAME "pte for 0x%lx: 0x%lx 0x%lx\n", address, pte_val(*pte),
+ pte_val(*pte) & _PAGE_PRESENT);
+}
+
+/*
+ * For some reason the pre/post pairs have been called in an
+ * unmatched order. Report and die.
+ */
+static void die_kmmio_nesting_error(struct pt_regs *regs, unsigned long addr)
+{
+ const struct trap_reason *my_reason = &get_cpu_var(pf_reason);
+ pr_emerg(NAME "unexpected fault for address: 0x%08lx, "
+ "last fault for address: 0x%08lx\n",
+ addr, my_reason->addr);
+ print_pte(addr);
+ print_symbol(KERN_EMERG "faulting IP is at %s\n", regs->ip);
+ print_symbol(KERN_EMERG "last faulting IP was at %s\n", my_reason->ip);
+#ifdef __i386__
+ pr_emerg("eax: %08lx ebx: %08lx ecx: %08lx edx: %08lx\n",
+ regs->ax, regs->bx, regs->cx, regs->dx);
+ pr_emerg("esi: %08lx edi: %08lx ebp: %08lx esp: %08lx\n",
+ regs->si, regs->di, regs->bp, regs->sp);
+#else
+ pr_emerg("rax: %016lx rcx: %016lx rdx: %016lx\n",
+ regs->ax, regs->cx, regs->dx);
+ pr_emerg("rsi: %016lx rdi: %016lx rbp: %016lx rsp: %016lx\n",
+ regs->si, regs->di, regs->bp, regs->sp);
+#endif
+ put_cpu_var(pf_reason);
+ BUG();
+}
+
+static void pre(struct kmmio_probe *p, struct pt_regs *regs,
+ unsigned long addr)
+{
+ struct trap_reason *my_reason = &get_cpu_var(pf_reason);
+ struct mm_io_header_rw *my_trace = &get_cpu_var(cpu_trace);
+ const unsigned long instptr = instruction_pointer(regs);
+ const enum reason_type type = get_ins_type(instptr);
+
+ /* it doesn't make sense to have more than one active trace per cpu */
+ if (my_reason->active_traces)
+ die_kmmio_nesting_error(regs, addr);
+ else
+ my_reason->active_traces++;
+
+ my_reason->type = type;
+ my_reason->addr = addr;
+ my_reason->ip = instptr;
+
+ my_trace->header.type = MMIO_MAGIC;
+ my_trace->header.pid = 0;
+ my_trace->header.data_len = sizeof(struct mm_io_rw);
+ my_trace->rw.address = addr;
+ /*
+ * struct remap_trace *trace = p->user_data;
+ * phys = addr - trace->probe.addr + trace->phys;
+ */
+
+ /*
+ * Only record the program counter when requested.
+ * It may taint clean-room reverse engineering.
+ */
+ if (trace_pc)
+ my_trace->rw.pc = instptr;
+ else
+ my_trace->rw.pc = 0;
+
+ record_timestamp(&my_trace->header);
+
+ switch (type) {
+ case REG_READ:
+ my_trace->header.type |=
+ (MMIO_READ << MMIO_OPCODE_SHIFT) |
+ (get_ins_mem_width(instptr) << MMIO_WIDTH_SHIFT);
+ break;
+ case REG_WRITE:
+ my_trace->header.type |=
+ (MMIO_WRITE << MMIO_OPCODE_SHIFT) |
+ (get_ins_mem_width(instptr) << MMIO_WIDTH_SHIFT);
+ my_trace->rw.value = get_ins_reg_val(instptr, regs);
+ break;
+ case IMM_WRITE:
+ my_trace->header.type |=
+ (MMIO_WRITE << MMIO_OPCODE_SHIFT) |
+ (get_ins_mem_width(instptr) << MMIO_WIDTH_SHIFT);
+ my_trace->rw.value = get_ins_imm_val(instptr);
+ break;
+ default:
+ {
+ unsigned char *ip = (unsigned char *)instptr;
+ my_trace->header.type |=
+ (MMIO_UNKNOWN_OP << MMIO_OPCODE_SHIFT);
+ my_trace->rw.value = (*ip) << 16 | *(ip + 1) << 8 |
+ *(ip + 2);
+ }
+ }
+ put_cpu_var(cpu_trace);
+ put_cpu_var(pf_reason);
+}
+
+static void post(struct kmmio_probe *p, unsigned long condition,
+ struct pt_regs *regs)
+{
+ struct trap_reason *my_reason = &get_cpu_var(pf_reason);
+ struct mm_io_header_rw *my_trace = &get_cpu_var(cpu_trace);
+
+ /* this should always return the active_trace count to 0 */
+ my_reason->active_traces--;
+ if (my_reason->active_traces) {
+ pr_emerg(NAME "unexpected post handler");
+ BUG();
+ }
+
+ switch (my_reason->type) {
+ case REG_READ:
+ my_trace->rw.value = get_ins_reg_val(my_reason->ip, regs);
+ break;
+ default:
+ break;
+ }
+ relay_write(chan, my_trace, sizeof(*my_trace));
+ put_cpu_var(cpu_trace);
+ put_cpu_var(pf_reason);
+}
+
+/*
+ * subbuf_start() relay callback.
+ *
+ * Defined so that we know when events are dropped due to the buffer-full
+ * condition.
+ */
+static int subbuf_start_handler(struct rchan_buf *buf, void *subbuf,
+ void *prev_subbuf, size_t prev_padding)
+{
+ unsigned int cpu = buf->cpu;
+ atomic_t *drop = &per_cpu(dropped, cpu);
+ int count;
+ if (relay_buf_full(buf)) {
+ if (atomic_inc_return(drop) == 1)
+ pr_err(NAME "cpu %d buffer full!\n", cpu);
+ return 0;
+ }
+ count = atomic_read(drop);
+ if (count) {
+ pr_err(NAME "cpu %d buffer no longer full, missed %d events.\n",
+ cpu, count);
+ atomic_sub(count, drop);
+ }
+
+ return 1;
+}
+
+static struct file_operations mmio_fops = {
+ .owner = THIS_MODULE,
+};
+
+/* file_create() callback. Creates relay file in debugfs. */
+static struct dentry *create_buf_file_handler(const char *filename,
+ struct dentry *parent,
+ int mode,
+ struct rchan_buf *buf,
+ int *is_global)
+{
+ struct dentry *buf_file;
+
+ mmio_fops.read = relay_file_operations.read;
+ mmio_fops.open = relay_file_operations.open;
+ mmio_fops.poll = relay_file_operations.poll;
+ mmio_fops.mmap = relay_file_operations.mmap;
+ mmio_fops.release = relay_file_operations.release;
+ mmio_fops.splice_read = relay_file_operations.splice_read;
+
+ buf_file = debugfs_create_file(filename, mode, parent, buf,
+ &mmio_fops);
+
+ return buf_file;
+}
+
+/* file_remove() default callback. Removes relay file in debugfs. */
+static int remove_buf_file_handler(struct dentry *dentry)
+{
+ debugfs_remove(dentry);
+ return 0;
+}
+
+static struct rchan_callbacks relay_callbacks = {
+ .subbuf_start = subbuf_start_handler,
+ .create_buf_file = create_buf_file_handler,
+ .remove_buf_file = remove_buf_file_handler,
+};
+
+static void ioremap_trace_core(unsigned long offset, unsigned long size,
+ void __iomem *addr)
+{
+ static atomic_t next_id;
+ struct remap_trace *trace = kmalloc(sizeof(*trace), GFP_KERNEL);
+ struct mm_io_header_map event = {
+ .header = {
+ .type = MMIO_MAGIC |
+ (MMIO_PROBE << MMIO_OPCODE_SHIFT),
+ .sec = 0,
+ .nsec = 0,
+ .pid = 0,
+ .data_len = sizeof(struct mm_io_map)
+ },
+ .map = {
+ .phys = offset,
+ .addr = (unsigned long)addr,
+ .len = size,
+ .pc = 0
+ }
+ };
+ record_timestamp(&event.header);
+
+ if (!trace) {
+ pr_err(NAME "kmalloc failed in ioremap\n");
+ return;
+ }
+
+ *trace = (struct remap_trace) {
+ .probe = {
+ .addr = (unsigned long)addr,
+ .len = size,
+ .pre_handler = pre,
+ .post_handler = post,
+ .user_data = trace
+ },
+ .phys = offset,
+ .id = atomic_inc_return(&next_id)
+ };
+
+ spin_lock_irq(&trace_lock);
+ if (!is_enabled())
+ goto not_enabled;
+
+ relay_write(chan, &event, sizeof(event));
+ list_add_tail(&trace->list, &trace_list);
+ if (!nommiotrace)
+ register_kmmio_probe(&trace->probe);
+
+not_enabled:
+ spin_unlock_irq(&trace_lock);
+}
+
+void
+mmiotrace_ioremap(unsigned long offset, unsigned long size, void __iomem *addr)
+{
+ if (!is_enabled()) /* recheck and proper locking in *_core() */
+ return;
+
+ pr_debug(NAME "ioremap_*(0x%lx, 0x%lx) = %p\n", offset, size, addr);
+ if ((filter_offset) && (offset != filter_offset))
+ return;
+ ioremap_trace_core(offset, size, addr);
+}
+
+static void iounmap_trace_core(volatile void __iomem *addr)
+{
+ struct mm_io_header_map event = {
+ .header = {
+ .type = MMIO_MAGIC |
+ (MMIO_UNPROBE << MMIO_OPCODE_SHIFT),
+ .sec = 0,
+ .nsec = 0,
+ .pid = 0,
+ .data_len = sizeof(struct mm_io_map)
+ },
+ .map = {
+ .phys = 0,
+ .addr = (unsigned long)addr,
+ .len = 0,
+ .pc = 0
+ }
+ };
+ struct remap_trace *trace;
+ struct remap_trace *tmp;
+ struct remap_trace *found_trace = NULL;
+
+ pr_debug(NAME "Unmapping %p.\n", addr);
+ record_timestamp(&event.header);
+
+ spin_lock_irq(&trace_lock);
+ if (!is_enabled())
+ goto not_enabled;
+
+ list_for_each_entry_safe(trace, tmp, &trace_list, list) {
+ if ((unsigned long)addr == trace->probe.addr) {
+ if (!nommiotrace)
+ unregister_kmmio_probe(&trace->probe);
+ list_del(&trace->list);
+ found_trace = trace;
+ break;
+ }
+ }
+ relay_write(chan, &event, sizeof(event));
+
+not_enabled:
+ spin_unlock_irq(&trace_lock);
+ if (found_trace) {
+ synchronize_rcu(); /* unregister_kmmio_probe() requirement */
+ kfree(found_trace);
+ }
+}
+
+void mmiotrace_iounmap(volatile void __iomem *addr)
+{
+ might_sleep();
+ if (is_enabled()) /* recheck and proper locking in *_core() */
+ iounmap_trace_core(addr);
+}
+
+static void clear_trace_list(void)
+{
+ struct remap_trace *trace;
+ struct remap_trace *tmp;
+
+ /*
+ * No locking required, because the caller ensures we are in a
+ * critical section via mutex, and is_enabled() is false,
+ * i.e. nothing can traverse or modify this list.
+ * Caller also ensures is_enabled() cannot change.
+ */
+ list_for_each_entry(trace, &trace_list, list) {
+ pr_notice(NAME "purging non-iounmapped "
+ "trace @0x%08lx, size 0x%lx.\n",
+ trace->probe.addr, trace->probe.len);
+ if (!nommiotrace)
+ unregister_kmmio_probe(&trace->probe);
+ }
+ synchronize_rcu(); /* unregister_kmmio_probe() requirement */
+
+ list_for_each_entry_safe(trace, tmp, &trace_list, list) {
+ list_del(&trace->list);
+ kfree(trace);
+ }
+}
+
+static ssize_t read_enabled_file_bool(struct file *file,
+ char __user *user_buf, size_t count, loff_t *ppos)
+{
+ char buf[3];
+
+ if (is_enabled())
+ buf[0] = '1';
+ else
+ buf[0] = '0';
+ buf[1] = '\n';
+ buf[2] = '\0';
+ return simple_read_from_buffer(user_buf, count, ppos, buf, 2);
+}
+
+static void enable_mmiotrace(void);
+static void disable_mmiotrace(void);
+
+static ssize_t write_enabled_file_bool(struct file *file,
+ const char __user *user_buf, size_t count, loff_t *ppos)
+{
+ char buf[32];
+ int buf_size = min(count, (sizeof(buf)-1));
+
+ if (copy_from_user(buf, user_buf, buf_size))
+ return -EFAULT;
+
+ switch (buf[0]) {
+ case 'y':
+ case 'Y':
+ case '1':
+ enable_mmiotrace();
+ break;
+ case 'n':
+ case 'N':
+ case '0':
+ disable_mmiotrace();
+ break;
+ }
+
+ return count;
+}
+
+/* this ripped from kernel/kprobes.c */
+static struct file_operations fops_enabled = {
+ .owner = THIS_MODULE,
+ .read = read_enabled_file_bool,
+ .write = write_enabled_file_bool
+};
+
+static struct file_operations fops_marker = {
+ .owner = THIS_MODULE,
+ .write = write_marker
+};
+
+static void enable_mmiotrace(void)
+{
+ mutex_lock(&mmiotrace_mutex);
+ if (is_enabled())
+ goto out;
+
+ chan = relay_open("cpu", dir, subbuf_size, n_subbufs,
+ &relay_callbacks, NULL);
+ if (!chan) {
+ pr_err(NAME "relay app channel creation failed.\n");
+ goto out;
+ }
+
+ reference_kmmio();
+
+ marker_file = debugfs_create_file("marker", 0660, dir, NULL,
+ &fops_marker);
+ if (!marker_file)
+ pr_err(NAME "marker file creation failed.\n");
+
+ if (nommiotrace)
+ pr_info(NAME "MMIO tracing disabled.\n");
+ if (ISA_trace)
+ pr_warning(NAME "Warning! low ISA range will be traced.\n");
+ spin_lock_irq(&trace_lock);
+ atomic_inc(&mmiotrace_enabled);
+ spin_unlock_irq(&trace_lock);
+ pr_info(NAME "enabled.\n");
+out:
+ mutex_unlock(&mmiotrace_mutex);
+}
+
+static void disable_mmiotrace(void)
+{
+ mutex_lock(&mmiotrace_mutex);
+ if (!is_enabled())
+ goto out;
+
+ spin_lock_irq(&trace_lock);
+ atomic_dec(&mmiotrace_enabled);
+ BUG_ON(is_enabled());
+ spin_unlock_irq(&trace_lock);
+
+ clear_trace_list(); /* guarantees: no more kmmio callbacks */
+ unreference_kmmio();
+ if (marker_file) {
+ debugfs_remove(marker_file);
+ marker_file = NULL;
+ }
+ if (chan) {
+ relay_close(chan);
+ chan = NULL;
+ }
+
+ pr_info(NAME "disabled.\n");
+out:
+ mutex_unlock(&mmiotrace_mutex);
+}
+
+static int __init init(void)
+{
+ pr_debug(NAME "load...\n");
+ if (n_subbufs < 2)
+ return -EINVAL;
+
+ dir = debugfs_create_dir(APP_DIR, NULL);
+ if (!dir) {
+ pr_err(NAME "Couldn't create relay app directory.\n");
+ return -ENOMEM;
+ }
+
+ enabled_file = debugfs_create_file("enabled", 0600, dir, NULL,
+ &fops_enabled);
+ if (!enabled_file) {
+ pr_err(NAME "Couldn't create enabled file.\n");
+ debugfs_remove(dir);
+ return -ENOMEM;
+ }
+
+ if (enable_now)
+ enable_mmiotrace();
+
+ return 0;
+}
+
+static void __exit cleanup(void)
+{
+ pr_debug(NAME "unload...\n");
+ if (enabled_file)
+ debugfs_remove(enabled_file);
+ disable_mmiotrace();
+ if (dir)
+ debugfs_remove(dir);
+}
+
+module_init(init);
+module_exit(cleanup);
+MODULE_LICENSE("GPL");
diff --git a/arch/x86/mm/pageattr.c b/arch/x86/mm/pageattr.c
index 14e48b5..7574bee 100644
--- a/arch/x86/mm/pageattr.c
+++ b/arch/x86/mm/pageattr.c
@@ -223,6 +223,7 @@ pte_t *lookup_address(unsigned long address, unsigned int *level)
return pte_offset_kernel(pmd, address);
}
+EXPORT_SYMBOL_GPL(lookup_address);
/*
* Set the new pmd in all the pgds we know about:
diff --git a/arch/x86/mm/pf_in.c b/arch/x86/mm/pf_in.c
new file mode 100644
index 0000000..efa1911
--- /dev/null
+++ b/arch/x86/mm/pf_in.c
@@ -0,0 +1,489 @@
+/*
+ * Fault Injection Test harness (FI)
+ * Copyright (C) Intel Crop.
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License
+ * as published by the Free Software Foundation; either version 2
+ * of the License, or (at your option) any later version.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307,
+ * USA.
+ *
+ */
+
+/* Id: pf_in.c,v 1.1.1.1 2002/11/12 05:56:32 brlock Exp
+ * Copyright by Intel Crop., 2002
+ * Louis Zhuang ([email protected])
+ *
+ * Bjorn Steinbrink ([email protected]), 2007
+ */
+
+#include <linux/module.h>
+#include <linux/ptrace.h> /* struct pt_regs */
+#include "pf_in.h"
+
+#ifdef __i386__
+/* IA32 Manual 3, 2-1 */
+static unsigned char prefix_codes[] = {
+ 0xF0, 0xF2, 0xF3, 0x2E, 0x36, 0x3E, 0x26, 0x64,
+ 0x65, 0x2E, 0x3E, 0x66, 0x67
+};
+/* IA32 Manual 3, 3-432*/
+static unsigned int reg_rop[] = {
+ 0x8A, 0x8B, 0xB60F, 0xB70F, 0xBE0F, 0xBF0F
+};
+static unsigned int reg_wop[] = { 0x88, 0x89 };
+static unsigned int imm_wop[] = { 0xC6, 0xC7 };
+/* IA32 Manual 3, 3-432*/
+static unsigned int rw8[] = { 0x88, 0x8A, 0xC6 };
+static unsigned int rw32[] = {
+ 0x89, 0x8B, 0xC7, 0xB60F, 0xB70F, 0xBE0F, 0xBF0F
+};
+static unsigned int mw8[] = { 0x88, 0x8A, 0xC6, 0xB60F, 0xBE0F };
+static unsigned int mw16[] = { 0xB70F, 0xBF0F };
+static unsigned int mw32[] = { 0x89, 0x8B, 0xC7 };
+static unsigned int mw64[] = {};
+#else /* not __i386__ */
+static unsigned char prefix_codes[] = {
+ 0x66, 0x67, 0x2E, 0x3E, 0x26, 0x64, 0x65, 0x36,
+ 0xF0, 0xF3, 0xF2,
+ /* REX Prefixes */
+ 0x40, 0x41, 0x42, 0x43, 0x44, 0x45, 0x46, 0x47,
+ 0x48, 0x49, 0x4a, 0x4b, 0x4c, 0x4d, 0x4e, 0x4f
+};
+/* AMD64 Manual 3, Appendix A*/
+static unsigned int reg_rop[] = {
+ 0x8A, 0x8B, 0xB60F, 0xB70F, 0xBE0F, 0xBF0F
+};
+static unsigned int reg_wop[] = { 0x88, 0x89 };
+static unsigned int imm_wop[] = { 0xC6, 0xC7 };
+static unsigned int rw8[] = { 0xC6, 0x88, 0x8A };
+static unsigned int rw32[] = {
+ 0xC7, 0x89, 0x8B, 0xB60F, 0xB70F, 0xBE0F, 0xBF0F
+};
+/* 8 bit only */
+static unsigned int mw8[] = { 0xC6, 0x88, 0x8A, 0xB60F, 0xBE0F };
+/* 16 bit only */
+static unsigned int mw16[] = { 0xB70F, 0xBF0F };
+/* 16 or 32 bit */
+static unsigned int mw32[] = { 0xC7 };
+/* 16, 32 or 64 bit */
+static unsigned int mw64[] = { 0x89, 0x8B };
+#endif /* not __i386__ */
+
+static int skip_prefix(unsigned char *addr, int *shorted, int *enlarged,
+ int *rexr)
+{
+ int i;
+ unsigned char *p = addr;
+ *shorted = 0;
+ *enlarged = 0;
+ *rexr = 0;
+
+restart:
+ for (i = 0; i < ARRAY_SIZE(prefix_codes); i++) {
+ if (*p == prefix_codes[i]) {
+ if (*p == 0x66)
+ *shorted = 1;
+#ifdef __amd64__
+ if ((*p & 0xf8) == 0x48)
+ *enlarged = 1;
+ if ((*p & 0xf4) == 0x44)
+ *rexr = 1;
+#endif
+ p++;
+ goto restart;
+ }
+ }
+
+ return (p - addr);
+}
+
+static int get_opcode(unsigned char *addr, unsigned int *opcode)
+{
+ int len;
+
+ if (*addr == 0x0F) {
+ /* 0x0F is extension instruction */
+ *opcode = *(unsigned short *)addr;
+ len = 2;
+ } else {
+ *opcode = *addr;
+ len = 1;
+ }
+
+ return len;
+}
+
+#define CHECK_OP_TYPE(opcode, array, type) \
+ for (i = 0; i < ARRAY_SIZE(array); i++) { \
+ if (array[i] == opcode) { \
+ rv = type; \
+ goto exit; \
+ } \
+ }
+
+enum reason_type get_ins_type(unsigned long ins_addr)
+{
+ unsigned int opcode;
+ unsigned char *p;
+ int shorted, enlarged, rexr;
+ int i;
+ enum reason_type rv = OTHERS;
+
+ p = (unsigned char *)ins_addr;
+ p += skip_prefix(p, &shorted, &enlarged, &rexr);
+ p += get_opcode(p, &opcode);
+
+ CHECK_OP_TYPE(opcode, reg_rop, REG_READ);
+ CHECK_OP_TYPE(opcode, reg_wop, REG_WRITE);
+ CHECK_OP_TYPE(opcode, imm_wop, IMM_WRITE);
+
+exit:
+ return rv;
+}
+#undef CHECK_OP_TYPE
+
+static unsigned int get_ins_reg_width(unsigned long ins_addr)
+{
+ unsigned int opcode;
+ unsigned char *p;
+ int i, shorted, enlarged, rexr;
+
+ p = (unsigned char *)ins_addr;
+ p += skip_prefix(p, &shorted, &enlarged, &rexr);
+ p += get_opcode(p, &opcode);
+
+ for (i = 0; i < ARRAY_SIZE(rw8); i++)
+ if (rw8[i] == opcode)
+ return 1;
+
+ for (i = 0; i < ARRAY_SIZE(rw32); i++)
+ if (rw32[i] == opcode)
+ return (shorted ? 2 : (enlarged ? 8 : 4));
+
+ printk(KERN_ERR "mmiotrace: Unknown opcode 0x%02x\n", opcode);
+ return 0;
+}
+
+unsigned int get_ins_mem_width(unsigned long ins_addr)
+{
+ unsigned int opcode;
+ unsigned char *p;
+ int i, shorted, enlarged, rexr;
+
+ p = (unsigned char *)ins_addr;
+ p += skip_prefix(p, &shorted, &enlarged, &rexr);
+ p += get_opcode(p, &opcode);
+
+ for (i = 0; i < ARRAY_SIZE(mw8); i++)
+ if (mw8[i] == opcode)
+ return 1;
+
+ for (i = 0; i < ARRAY_SIZE(mw16); i++)
+ if (mw16[i] == opcode)
+ return 2;
+
+ for (i = 0; i < ARRAY_SIZE(mw32); i++)
+ if (mw32[i] == opcode)
+ return shorted ? 2 : 4;
+
+ for (i = 0; i < ARRAY_SIZE(mw64); i++)
+ if (mw64[i] == opcode)
+ return shorted ? 2 : (enlarged ? 8 : 4);
+
+ printk(KERN_ERR "mmiotrace: Unknown opcode 0x%02x\n", opcode);
+ return 0;
+}
+
+/*
+ * Define register ident in mod/rm byte.
+ * Note: these are NOT the same as in ptrace-abi.h.
+ */
+enum {
+ arg_AL = 0,
+ arg_CL = 1,
+ arg_DL = 2,
+ arg_BL = 3,
+ arg_AH = 4,
+ arg_CH = 5,
+ arg_DH = 6,
+ arg_BH = 7,
+
+ arg_AX = 0,
+ arg_CX = 1,
+ arg_DX = 2,
+ arg_BX = 3,
+ arg_SP = 4,
+ arg_BP = 5,
+ arg_SI = 6,
+ arg_DI = 7,
+#ifdef __amd64__
+ arg_R8 = 8,
+ arg_R9 = 9,
+ arg_R10 = 10,
+ arg_R11 = 11,
+ arg_R12 = 12,
+ arg_R13 = 13,
+ arg_R14 = 14,
+ arg_R15 = 15
+#endif
+};
+
+static unsigned char *get_reg_w8(int no, struct pt_regs *regs)
+{
+ unsigned char *rv = NULL;
+
+ switch (no) {
+ case arg_AL:
+ rv = (unsigned char *)®s->ax;
+ break;
+ case arg_BL:
+ rv = (unsigned char *)®s->bx;
+ break;
+ case arg_CL:
+ rv = (unsigned char *)®s->cx;
+ break;
+ case arg_DL:
+ rv = (unsigned char *)®s->dx;
+ break;
+ case arg_AH:
+ rv = 1 + (unsigned char *)®s->ax;
+ break;
+ case arg_BH:
+ rv = 1 + (unsigned char *)®s->bx;
+ break;
+ case arg_CH:
+ rv = 1 + (unsigned char *)®s->cx;
+ break;
+ case arg_DH:
+ rv = 1 + (unsigned char *)®s->dx;
+ break;
+#ifdef __amd64__
+ case arg_R8:
+ rv = (unsigned char *)®s->r8;
+ break;
+ case arg_R9:
+ rv = (unsigned char *)®s->r9;
+ break;
+ case arg_R10:
+ rv = (unsigned char *)®s->r10;
+ break;
+ case arg_R11:
+ rv = (unsigned char *)®s->r11;
+ break;
+ case arg_R12:
+ rv = (unsigned char *)®s->r12;
+ break;
+ case arg_R13:
+ rv = (unsigned char *)®s->r13;
+ break;
+ case arg_R14:
+ rv = (unsigned char *)®s->r14;
+ break;
+ case arg_R15:
+ rv = (unsigned char *)®s->r15;
+ break;
+#endif
+ default:
+ printk(KERN_ERR "mmiotrace: Error reg no# %d\n", no);
+ break;
+ }
+ return rv;
+}
+
+static unsigned long *get_reg_w32(int no, struct pt_regs *regs)
+{
+ unsigned long *rv = NULL;
+
+ switch (no) {
+ case arg_AX:
+ rv = ®s->ax;
+ break;
+ case arg_BX:
+ rv = ®s->bx;
+ break;
+ case arg_CX:
+ rv = ®s->cx;
+ break;
+ case arg_DX:
+ rv = ®s->dx;
+ break;
+ case arg_SP:
+ rv = ®s->sp;
+ break;
+ case arg_BP:
+ rv = ®s->bp;
+ break;
+ case arg_SI:
+ rv = ®s->si;
+ break;
+ case arg_DI:
+ rv = ®s->di;
+ break;
+#ifdef __amd64__
+ case arg_R8:
+ rv = ®s->r8;
+ break;
+ case arg_R9:
+ rv = ®s->r9;
+ break;
+ case arg_R10:
+ rv = ®s->r10;
+ break;
+ case arg_R11:
+ rv = ®s->r11;
+ break;
+ case arg_R12:
+ rv = ®s->r12;
+ break;
+ case arg_R13:
+ rv = ®s->r13;
+ break;
+ case arg_R14:
+ rv = ®s->r14;
+ break;
+ case arg_R15:
+ rv = ®s->r15;
+ break;
+#endif
+ default:
+ printk(KERN_ERR "mmiotrace: Error reg no# %d\n", no);
+ }
+
+ return rv;
+}
+
+unsigned long get_ins_reg_val(unsigned long ins_addr, struct pt_regs *regs)
+{
+ unsigned int opcode;
+ unsigned char mod_rm;
+ int reg;
+ unsigned char *p;
+ int i, shorted, enlarged, rexr;
+ unsigned long rv;
+
+ p = (unsigned char *)ins_addr;
+ p += skip_prefix(p, &shorted, &enlarged, &rexr);
+ p += get_opcode(p, &opcode);
+ for (i = 0; i < ARRAY_SIZE(reg_rop); i++)
+ if (reg_rop[i] == opcode) {
+ rv = REG_READ;
+ goto do_work;
+ }
+
+ for (i = 0; i < ARRAY_SIZE(reg_wop); i++)
+ if (reg_wop[i] == opcode) {
+ rv = REG_WRITE;
+ goto do_work;
+ }
+
+ printk(KERN_ERR "mmiotrace: Not a register instruction, opcode "
+ "0x%02x\n", opcode);
+ goto err;
+
+do_work:
+ mod_rm = *p;
+ reg = ((mod_rm >> 3) & 0x7) | (rexr << 3);
+ switch (get_ins_reg_width(ins_addr)) {
+ case 1:
+ return *get_reg_w8(reg, regs);
+
+ case 2:
+ return *(unsigned short *)get_reg_w32(reg, regs);
+
+ case 4:
+ return *(unsigned int *)get_reg_w32(reg, regs);
+
+#ifdef __amd64__
+ case 8:
+ return *(unsigned long *)get_reg_w32(reg, regs);
+#endif
+
+ default:
+ printk(KERN_ERR "mmiotrace: Error width# %d\n", reg);
+ }
+
+err:
+ return 0;
+}
+
+unsigned long get_ins_imm_val(unsigned long ins_addr)
+{
+ unsigned int opcode;
+ unsigned char mod_rm;
+ unsigned char mod;
+ unsigned char *p;
+ int i, shorted, enlarged, rexr;
+ unsigned long rv;
+
+ p = (unsigned char *)ins_addr;
+ p += skip_prefix(p, &shorted, &enlarged, &rexr);
+ p += get_opcode(p, &opcode);
+ for (i = 0; i < ARRAY_SIZE(imm_wop); i++)
+ if (imm_wop[i] == opcode) {
+ rv = IMM_WRITE;
+ goto do_work;
+ }
+
+ printk(KERN_ERR "mmiotrace: Not an immediate instruction, opcode "
+ "0x%02x\n", opcode);
+ goto err;
+
+do_work:
+ mod_rm = *p;
+ mod = mod_rm >> 6;
+ p++;
+ switch (mod) {
+ case 0:
+ /* if r/m is 5 we have a 32 disp (IA32 Manual 3, Table 2-2) */
+ /* AMD64: XXX Check for address size prefix? */
+ if ((mod_rm & 0x7) == 0x5)
+ p += 4;
+ break;
+
+ case 1:
+ p += 1;
+ break;
+
+ case 2:
+ p += 4;
+ break;
+
+ case 3:
+ default:
+ printk(KERN_ERR "mmiotrace: not a memory access instruction "
+ "at 0x%lx, rm_mod=0x%02x\n",
+ ins_addr, mod_rm);
+ }
+
+ switch (get_ins_reg_width(ins_addr)) {
+ case 1:
+ return *(unsigned char *)p;
+
+ case 2:
+ return *(unsigned short *)p;
+
+ case 4:
+ return *(unsigned int *)p;
+
+#ifdef __amd64__
+ case 8:
+ return *(unsigned long *)p;
+#endif
+
+ default:
+ printk(KERN_ERR "mmiotrace: Error: width.\n");
+ }
+
+err:
+ return 0;
+}
diff --git a/arch/x86/mm/pf_in.h b/arch/x86/mm/pf_in.h
new file mode 100644
index 0000000..e05341a
--- /dev/null
+++ b/arch/x86/mm/pf_in.h
@@ -0,0 +1,39 @@
+/*
+ * Fault Injection Test harness (FI)
+ * Copyright (C) Intel Crop.
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License
+ * as published by the Free Software Foundation; either version 2
+ * of the License, or (at your option) any later version.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307,
+ * USA.
+ *
+ */
+
+#ifndef __PF_H_
+#define __PF_H_
+
+enum reason_type {
+ NOT_ME, /* page fault is not in regions */
+ NOTHING, /* access others point in regions */
+ REG_READ, /* read from addr to reg */
+ REG_WRITE, /* write from reg to addr */
+ IMM_WRITE, /* write from imm to addr */
+ OTHERS /* Other instructions can not intercept */
+};
+
+enum reason_type get_ins_type(unsigned long ins_addr);
+unsigned int get_ins_mem_width(unsigned long ins_addr);
+unsigned long get_ins_reg_val(unsigned long ins_addr, struct pt_regs *regs);
+unsigned long get_ins_imm_val(unsigned long ins_addr);
+
+#endif /* __PF_H_ */
diff --git a/arch/x86/mm/testmmiotrace.c b/arch/x86/mm/testmmiotrace.c
new file mode 100644
index 0000000..cfa60b2
--- /dev/null
+++ b/arch/x86/mm/testmmiotrace.c
@@ -0,0 +1,71 @@
+/*
+ * Written by Pekka Paalanen, 2008 <[email protected]>
+ */
+#include <linux/module.h>
+#include <asm/io.h>
+
+#define MODULE_NAME "testmmiotrace"
+
+static unsigned long mmio_address;
+module_param(mmio_address, ulong, 0);
+MODULE_PARM_DESC(mmio_address, "Start address of the mapping of 16 kB.");
+
+static void do_write_test(void __iomem *p)
+{
+ unsigned int i;
+ for (i = 0; i < 256; i++)
+ iowrite8(i, p + i);
+ for (i = 1024; i < (5 * 1024); i += 2)
+ iowrite16(i * 12 + 7, p + i);
+ for (i = (5 * 1024); i < (16 * 1024); i += 4)
+ iowrite32(i * 212371 + 13, p + i);
+}
+
+static void do_read_test(void __iomem *p)
+{
+ unsigned int i;
+ for (i = 0; i < 256; i++)
+ ioread8(p + i);
+ for (i = 1024; i < (5 * 1024); i += 2)
+ ioread16(p + i);
+ for (i = (5 * 1024); i < (16 * 1024); i += 4)
+ ioread32(p + i);
+}
+
+static void do_test(void)
+{
+ void __iomem *p = ioremap_nocache(mmio_address, 0x4000);
+ if (!p) {
+ pr_err(MODULE_NAME ": could not ioremap, aborting.\n");
+ return;
+ }
+ do_write_test(p);
+ do_read_test(p);
+ iounmap(p);
+}
+
+static int __init init(void)
+{
+ if (mmio_address == 0) {
+ pr_err(MODULE_NAME ": you have to use the module argument "
+ "mmio_address.\n");
+ pr_err(MODULE_NAME ": DO NOT LOAD THIS MODULE UNLESS"
+ " YOU REALLY KNOW WHAT YOU ARE DOING!\n");
+ return -ENXIO;
+ }
+
+ pr_warning(MODULE_NAME ": WARNING: mapping 16 kB @ 0x%08lx "
+ "in PCI address space, and writing "
+ "rubbish in there.\n", mmio_address);
+ do_test();
+ return 0;
+}
+
+static void __exit cleanup(void)
+{
+ pr_debug(MODULE_NAME ": unloaded.\n");
+}
+
+module_init(init);
+module_exit(cleanup);
+MODULE_LICENSE("GPL");
diff --git a/include/linux/mmiotrace.h b/include/linux/mmiotrace.h
new file mode 100644
index 0000000..cb5efd0
--- /dev/null
+++ b/include/linux/mmiotrace.h
@@ -0,0 +1,120 @@
+#ifndef MMIOTRACE_H
+#define MMIOTRACE_H
+
+#include <asm/types.h>
+
+#ifdef __KERNEL__
+
+#include <linux/list.h>
+
+struct kmmio_probe;
+struct pt_regs;
+
+typedef void (*kmmio_pre_handler_t)(struct kmmio_probe *,
+ struct pt_regs *, unsigned long addr);
+typedef void (*kmmio_post_handler_t)(struct kmmio_probe *,
+ unsigned long condition, struct pt_regs *);
+
+struct kmmio_probe {
+ struct list_head list; /* kmmio internal list */
+ unsigned long addr; /* start location of the probe point */
+ unsigned long len; /* length of the probe region */
+ kmmio_pre_handler_t pre_handler; /* Called before addr is executed. */
+ kmmio_post_handler_t post_handler; /* Called after addr is executed */
+ void *user_data;
+};
+
+/* kmmio is active by some kmmio_probes? */
+static inline int is_kmmio_active(void)
+{
+ extern unsigned int kmmio_count;
+ return kmmio_count;
+}
+
+extern void reference_kmmio(void);
+extern void unreference_kmmio(void);
+extern int register_kmmio_probe(struct kmmio_probe *p);
+extern void unregister_kmmio_probe(struct kmmio_probe *p);
+
+/* Called from page fault handler. */
+extern int kmmio_handler(struct pt_regs *regs, unsigned long addr);
+
+/* Called from ioremap.c */
+#ifdef CONFIG_MMIOTRACE
+extern void
+mmiotrace_ioremap(unsigned long offset, unsigned long size, void __iomem *addr);
+extern void mmiotrace_iounmap(volatile void __iomem *addr);
+#else
+static inline void
+mmiotrace_ioremap(unsigned long offset, unsigned long size, void __iomem *addr)
+{
+}
+static inline void mmiotrace_iounmap(volatile void __iomem *addr)
+{
+}
+#endif /* CONFIG_MMIOTRACE_HOOKS */
+
+#endif /* __KERNEL__ */
+
+
+/*
+ * If you change anything here, you must bump MMIO_VERSION.
+ * This is the relay data format for user space.
+ */
+#define MMIO_VERSION 0x04
+
+/* mm_io_header.type */
+#define MMIO_OPCODE_MASK 0xff
+#define MMIO_OPCODE_SHIFT 0
+#define MMIO_WIDTH_MASK 0xff00
+#define MMIO_WIDTH_SHIFT 8
+#define MMIO_MAGIC (0x6f000000 | (MMIO_VERSION<<16))
+#define MMIO_MAGIC_MASK 0xffff0000
+
+enum mm_io_opcode { /* payload type: */
+ MMIO_READ = 0x1, /* struct mm_io_rw */
+ MMIO_WRITE = 0x2, /* struct mm_io_rw */
+ MMIO_PROBE = 0x3, /* struct mm_io_map */
+ MMIO_UNPROBE = 0x4, /* struct mm_io_map */
+ MMIO_MARKER = 0x5, /* raw char data */
+ MMIO_UNKNOWN_OP = 0x6, /* struct mm_io_rw */
+};
+
+struct mm_io_header {
+ __u32 type; /* see MMIO_* macros above */
+ __u32 sec; /* timestamp */
+ __u32 nsec;
+ __u32 pid; /* PID of the process, or 0 for kernel core */
+ __u16 data_len; /* length of the following payload */
+};
+
+struct mm_io_rw {
+ __u64 address; /* virtual address of register */
+ __u64 value;
+ __u64 pc; /* optional program counter */
+};
+
+struct mm_io_map {
+ __u64 phys; /* base address in PCI space */
+ __u64 addr; /* base virtual address */
+ __u64 len; /* mapping size */
+ __u64 pc; /* optional program counter */
+};
+
+
+/*
+ * These structures are used to allow a single relay_write()
+ * call to write a full packet.
+ */
+
+struct mm_io_header_rw {
+ struct mm_io_header header;
+ struct mm_io_rw rw;
+} __attribute__((packed));
+
+struct mm_io_header_map {
+ struct mm_io_header header;
+ struct mm_io_map map;
+} __attribute__((packed));
+
+#endif /* MMIOTRACE_H */
>From bc32dcfb4048bb74c774fd3de724bd38e8b98c2f Mon Sep 17 00:00:00 2001
From: Pekka Paalanen <[email protected]>
Date: Sun, 9 Mar 2008 13:52:25 +0200
Subject: [PATCH] x86 mmiotrace: update preview 1 to preview 2
Kconfig.debug, Makefile and testmmiotrace.c style fixes.
Use real mutex instead of mutex.
Fix failure path in register probe func.
kmmio: RCU read-locked over single stepping.
Generate mapping id's.
Make mmio-mod.c built-in and rewrite its locking.
Add debugfs file to enable/disable mmiotracing.
kmmio: use irqsave spinlocks.
Lots of cleanups in mmio-mod.c
Marker file moved from /proc into debugfs.
Call mmiotrace entrypoints directly from ioremap.c.
Signed-off-by: Pekka Paalanen <[email protected]>
---
arch/x86/Kconfig.debug | 20 +--
arch/x86/mm/Makefile | 2 +-
arch/x86/mm/ioremap.c | 9 +-
arch/x86/mm/kmmio.c | 72 ++++-----
arch/x86/mm/mmio-mod.c | 397 ++++++++++++++++++++++++++++---------------
arch/x86/mm/testmmiotrace.c | 15 +-
include/linux/mmiotrace.h | 18 ++-
7 files changed, 332 insertions(+), 201 deletions(-)
diff --git a/arch/x86/Kconfig.debug b/arch/x86/Kconfig.debug
index 177fd06..4c4fea0 100644
--- a/arch/x86/Kconfig.debug
+++ b/arch/x86/Kconfig.debug
@@ -183,22 +183,19 @@ config SYSPROF
config MMIOTRACE_HOOKS
bool
- default n
config MMIOTRACE
- tristate "Memory mapped IO tracing"
+ bool "Memory mapped IO tracing"
depends on DEBUG_KERNEL && RELAY && DEBUG_FS
select MMIOTRACE_HOOKS
- default n
+ default y
help
- This will build a kernel module called mmiotrace.
- Making this a built-in is heavily discouraged.
-
- Mmiotrace traces Memory Mapped I/O access and is meant for debugging
- and reverse engineering. The kernel module offers wrapped
- versions of the ioremap family of functions. The driver to be traced
- must be modified to call these wrappers. A user space program is
- required to collect the MMIO data.
+ Mmiotrace traces Memory Mapped I/O access and is meant for
+ debugging and reverse engineering. It is called from the ioremap
+ implementation and works via page faults. A user space program is
+ required to collect the MMIO data from debugfs files.
+ Tracing is disabled by default and can be enabled from a debugfs
+ file.
See http://nouveau.freedesktop.org/wiki/MmioTrace
If you are not helping to develop drivers, say N.
@@ -206,7 +203,6 @@ config MMIOTRACE
config MMIOTRACE_TEST
tristate "Test module for mmiotrace"
depends on MMIOTRACE && m
- default n
help
This is a dumb module for testing mmiotrace. It is very dangerous
as it will write garbage to IO memory starting at a given address.
diff --git a/arch/x86/mm/Makefile b/arch/x86/mm/Makefile
index e8f6e34..e6821ed 100644
--- a/arch/x86/mm/Makefile
+++ b/arch/x86/mm/Makefile
@@ -5,7 +5,7 @@ obj-$(CONFIG_X86_32) += pgtable_32.o
obj-$(CONFIG_MMIOTRACE_HOOKS) += kmmio.o
obj-$(CONFIG_MMIOTRACE) += mmiotrace.o
-mmiotrace-objs := pf_in.o mmio-mod.o
+mmiotrace-y := pf_in.o mmio-mod.o
obj-$(CONFIG_MMIOTRACE_TEST) += testmmiotrace.o
obj-$(CONFIG_HUGETLB_PAGE) += hugetlbpage.o
diff --git a/arch/x86/mm/ioremap.c b/arch/x86/mm/ioremap.c
index 26b75e4..1253c35 100644
--- a/arch/x86/mm/ioremap.c
+++ b/arch/x86/mm/ioremap.c
@@ -12,6 +12,7 @@
#include <linux/module.h>
#include <linux/slab.h>
#include <linux/vmalloc.h>
+#include <linux/mmiotrace.h>
#include <asm/cacheflush.h>
#include <asm/e820.h>
@@ -124,6 +125,7 @@ static void __iomem *__ioremap(unsigned long phys_addr, unsigned long size,
unsigned long pfn, offset, last_addr, vaddr;
struct vm_struct *area;
pgprot_t prot;
+ void __iomem *ret_addr;
/* Don't allow wraparound or zero size */
last_addr = phys_addr + size - 1;
@@ -191,7 +193,10 @@ static void __iomem *__ioremap(unsigned long phys_addr, unsigned long size,
return NULL;
}
- return (void __iomem *) (vaddr + offset);
+ ret_addr = (void __iomem *) (vaddr + offset);
+ mmiotrace_ioremap(phys_addr, size, ret_addr);
+
+ return ret_addr;
}
/**
@@ -249,6 +254,8 @@ void iounmap(volatile void __iomem *addr)
addr < phys_to_virt(ISA_END_ADDRESS))
return;
+ mmiotrace_iounmap(addr);
+
addr = (volatile void __iomem *)
(PAGE_MASK & (unsigned long __force)addr);
diff --git a/arch/x86/mm/kmmio.c b/arch/x86/mm/kmmio.c
index 539a9b1..efb4679 100644
--- a/arch/x86/mm/kmmio.c
+++ b/arch/x86/mm/kmmio.c
@@ -19,6 +19,7 @@
#include <linux/preempt.h>
#include <linux/percpu.h>
#include <linux/kdebug.h>
+#include <linux/mutex.h>
#include <asm/io.h>
#include <asm/cacheflush.h>
#include <asm/errno.h>
@@ -59,7 +60,7 @@ struct kmmio_context {
static int kmmio_die_notifier(struct notifier_block *nb, unsigned long val,
void *args);
-static DECLARE_MUTEX(kmmio_init_mutex);
+static DEFINE_MUTEX(kmmio_init_mutex);
static DEFINE_SPINLOCK(kmmio_lock);
/* These are protected by kmmio_lock */
@@ -90,7 +91,7 @@ static struct notifier_block nb_die = {
*/
void reference_kmmio(void)
{
- down(&kmmio_init_mutex);
+ mutex_lock(&kmmio_init_mutex);
spin_lock_irq(&kmmio_lock);
if (!kmmio_initialized) {
int i;
@@ -101,7 +102,7 @@ void reference_kmmio(void)
}
kmmio_initialized++;
spin_unlock_irq(&kmmio_lock);
- up(&kmmio_init_mutex);
+ mutex_unlock(&kmmio_init_mutex);
}
EXPORT_SYMBOL_GPL(reference_kmmio);
@@ -115,7 +116,7 @@ void unreference_kmmio(void)
{
bool unreg = false;
- down(&kmmio_init_mutex);
+ mutex_lock(&kmmio_init_mutex);
spin_lock_irq(&kmmio_lock);
if (kmmio_initialized == 1) {
@@ -128,7 +129,7 @@ void unreference_kmmio(void)
if (unreg)
unregister_die_notifier(&nb_die); /* calls sync_rcu() */
- up(&kmmio_init_mutex);
+ mutex_unlock(&kmmio_init_mutex);
}
EXPORT_SYMBOL(unreference_kmmio);
@@ -244,17 +245,13 @@ int kmmio_handler(struct pt_regs *regs, unsigned long addr)
* Preemption is now disabled to prevent process switch during
* single stepping. We can only handle one active kmmio trace
* per cpu, so ensure that we finish it before something else
- * gets to run.
- *
- * XXX what if an interrupt occurs between returning from
- * do_page_fault() and entering the single-step exception handler?
- * And that interrupt triggers a kmmio trap?
- * XXX If we tracing an interrupt service routine or whatever, is
- * this enough to keep it on the current cpu?
+ * gets to run. We also hold the RCU read lock over single
+ * stepping to avoid looking up the probe and kmmio_fault_page
+ * again.
*/
preempt_disable();
-
rcu_read_lock();
+
faultpage = get_kmmio_fault_page(addr);
if (!faultpage) {
/*
@@ -287,14 +284,24 @@ int kmmio_handler(struct pt_regs *regs, unsigned long addr)
if (ctx->probe && ctx->probe->pre_handler)
ctx->probe->pre_handler(ctx->probe, regs, addr);
+ /*
+ * Enable single-stepping and disable interrupts for the faulting
+ * context. Local interrupts must not get enabled during stepping.
+ */
regs->flags |= TF_MASK;
regs->flags &= ~IF_MASK;
/* Now we set present bit in PTE and single step. */
disarm_kmmio_fault_page(ctx->fpage->page, NULL);
+ /*
+ * If another cpu accesses the same page while we are stepping,
+ * the access will not be caught. It will simply succeed and the
+ * only downside is we lose the event. If this becomes a problem,
+ * the user should drop to single cpu before tracing.
+ */
+
put_cpu_var(kmmio_ctx);
- rcu_read_unlock();
return 1;
no_kmmio_ctx:
@@ -313,32 +320,15 @@ no_kmmio:
static int post_kmmio_handler(unsigned long condition, struct pt_regs *regs)
{
int ret = 0;
- struct kmmio_probe *probe;
- struct kmmio_fault_page *faultpage;
struct kmmio_context *ctx = &get_cpu_var(kmmio_ctx);
if (!ctx->active)
goto out;
- rcu_read_lock();
-
- faultpage = get_kmmio_fault_page(ctx->addr);
- probe = get_kmmio_probe(ctx->addr);
- if (faultpage != ctx->fpage || probe != ctx->probe) {
- /*
- * The trace setup changed after kmmio_handler() and before
- * running this respective post handler. User does not want
- * the result anymore.
- */
- ctx->probe = NULL;
- ctx->fpage = NULL;
- }
-
if (ctx->probe && ctx->probe->post_handler)
ctx->probe->post_handler(ctx->probe, condition, regs);
- if (ctx->fpage)
- arm_kmmio_fault_page(ctx->fpage->page, NULL);
+ arm_kmmio_fault_page(ctx->fpage->page, NULL);
regs->flags &= ~TF_MASK;
regs->flags |= ctx->saved_flags;
@@ -346,6 +336,7 @@ static int post_kmmio_handler(unsigned long condition, struct pt_regs *regs)
/* These were acquired in kmmio_handler(). */
ctx->active--;
BUG_ON(ctx->active);
+ rcu_read_unlock();
preempt_enable_no_resched();
/*
@@ -355,8 +346,6 @@ static int post_kmmio_handler(unsigned long condition, struct pt_regs *regs)
*/
if (!(regs->flags & TF_MASK))
ret = 1;
-
- rcu_read_unlock();
out:
put_cpu_var(kmmio_ctx);
return ret;
@@ -411,15 +400,16 @@ static void release_kmmio_fault_page(unsigned long page,
int register_kmmio_probe(struct kmmio_probe *p)
{
+ unsigned long flags;
int ret = 0;
unsigned long size = 0;
- spin_lock_irq(&kmmio_lock);
- kmmio_count++;
+ spin_lock_irqsave(&kmmio_lock, flags);
if (get_kmmio_probe(p->addr)) {
ret = -EEXIST;
goto out;
}
+ kmmio_count++;
list_add_rcu(&p->list, &kmmio_probes);
while (size < p->len) {
if (add_kmmio_fault_page(p->addr + size))
@@ -427,7 +417,7 @@ int register_kmmio_probe(struct kmmio_probe *p)
size += PAGE_SIZE;
}
out:
- spin_unlock_irq(&kmmio_lock);
+ spin_unlock_irqrestore(&kmmio_lock, flags);
/*
* XXX: What should I do here?
* Here was a call to global_flush_tlb(), but it does not exist
@@ -478,7 +468,8 @@ static void remove_kmmio_fault_pages(struct rcu_head *head)
/*
* Remove a kmmio probe. You have to synchronize_rcu() before you can be
- * sure that the callbacks will not be called anymore.
+ * sure that the callbacks will not be called anymore. Only after that
+ * you may actually release your struct kmmio_probe.
*
* Unregistering a kmmio fault page has three steps:
* 1. release_kmmio_fault_page()
@@ -490,18 +481,19 @@ static void remove_kmmio_fault_pages(struct rcu_head *head)
*/
void unregister_kmmio_probe(struct kmmio_probe *p)
{
+ unsigned long flags;
unsigned long size = 0;
struct kmmio_fault_page *release_list = NULL;
struct kmmio_delayed_release *drelease;
- spin_lock_irq(&kmmio_lock);
+ spin_lock_irqsave(&kmmio_lock, flags);
while (size < p->len) {
release_kmmio_fault_page(p->addr + size, &release_list);
size += PAGE_SIZE;
}
list_del_rcu(&p->list);
kmmio_count--;
- spin_unlock_irq(&kmmio_lock);
+ spin_unlock_irqrestore(&kmmio_lock, flags);
drelease = kmalloc(sizeof(*drelease), GFP_ATOMIC);
if (!drelease) {
diff --git a/arch/x86/mm/mmio-mod.c b/arch/x86/mm/mmio-mod.c
index e1a5085..7386440 100644
--- a/arch/x86/mm/mmio-mod.c
+++ b/arch/x86/mm/mmio-mod.c
@@ -19,6 +19,8 @@
*
* Derived from the read-mod example from relay-examples by Tom Zanussi.
*/
+#define DEBUG 1
+
#include <linux/module.h>
#include <linux/relay.h>
#include <linux/debugfs.h>
@@ -34,12 +36,12 @@
#include "pf_in.h"
-/* This app's relay channel files will appear in /debug/mmio-trace */
-#define APP_DIR "mmio-trace"
-/* the marker injection file in /proc */
-#define MARKER_FILE "mmio-marker"
+#define NAME "mmiotrace: "
-#define MODULE_NAME "mmiotrace"
+/* This app's relay channel files will appear in /debug/mmio-trace */
+static const char APP_DIR[] = "mmio-trace";
+/* the marker injection file in /debug/APP_DIR */
+static const char MARKER_FILE[] = "mmio-marker";
struct trap_reason {
unsigned long addr;
@@ -48,6 +50,15 @@ struct trap_reason {
int active_traces;
};
+struct remap_trace {
+ struct list_head list;
+ struct kmmio_probe probe;
+ unsigned long phys;
+ unsigned long id;
+};
+
+static const size_t subbuf_size = 256*1024;
+
/* Accessed per-cpu. */
static DEFINE_PER_CPU(struct trap_reason, pf_reason);
static DEFINE_PER_CPU(struct mm_io_header_rw, cpu_trace);
@@ -55,33 +66,53 @@ static DEFINE_PER_CPU(struct mm_io_header_rw, cpu_trace);
/* Access to this is not per-cpu. */
static DEFINE_PER_CPU(atomic_t, dropped);
-static struct file_operations mmio_fops = {
- .owner = THIS_MODULE,
-};
+static struct dentry *dir;
+static struct dentry *enabled_file;
+static struct dentry *marker_file;
-static const size_t subbuf_size = 256*1024;
+static DEFINE_MUTEX(mmiotrace_mutex);
+static DEFINE_SPINLOCK(trace_lock);
+static atomic_t mmiotrace_enabled;
+static LIST_HEAD(trace_list); /* struct remap_trace */
static struct rchan *chan;
-static struct dentry *dir;
-static struct proc_dir_entry *proc_marker_file;
+
+/*
+ * Locking in this file:
+ * - mmiotrace_mutex enforces enable/disable_mmiotrace() critical sections.
+ * - mmiotrace_enabled may be modified only when holding mmiotrace_mutex
+ * and trace_lock.
+ * - Routines depending on is_enabled() must take trace_lock.
+ * - trace_list users must hold trace_lock.
+ * - is_enabled() guarantees that chan is valid.
+ * - pre/post callbacks assume the effect of is_enabled() being true.
+ */
/* module parameters */
-static unsigned int n_subbufs = 32*4;
-static unsigned long filter_offset;
-static int nommiotrace;
-static int ISA_trace;
-static int trace_pc;
+static unsigned int n_subbufs = 32*4;
+static unsigned long filter_offset;
+static int nommiotrace;
+static int ISA_trace;
+static int trace_pc;
+static int enable_now;
module_param(n_subbufs, uint, 0);
module_param(filter_offset, ulong, 0);
module_param(nommiotrace, bool, 0);
module_param(ISA_trace, bool, 0);
module_param(trace_pc, bool, 0);
+module_param(enable_now, bool, 0);
MODULE_PARM_DESC(n_subbufs, "Number of 256kB buffers, default 128.");
MODULE_PARM_DESC(filter_offset, "Start address of traced mappings.");
MODULE_PARM_DESC(nommiotrace, "Disable actual MMIO tracing.");
MODULE_PARM_DESC(ISA_trace, "Do not exclude the low ISA range.");
MODULE_PARM_DESC(trace_pc, "Record address of faulting instructions.");
+MODULE_PARM_DESC(enable_now, "Start mmiotracing immediately on module load.");
+
+static bool is_enabled(void)
+{
+ return atomic_read(&mmiotrace_enabled);
+}
static void record_timestamp(struct mm_io_header *header)
{
@@ -93,15 +124,15 @@ static void record_timestamp(struct mm_io_header *header)
}
/*
- * Write callback for the /proc entry:
+ * Write callback for the debugfs entry:
* Read a marker and write it to the mmio trace log
*/
-static int write_marker(struct file *file, const char __user *buffer,
- unsigned long count, void *data)
+static ssize_t write_marker(struct file *file, const char __user *buffer,
+ size_t count, loff_t *ppos)
{
char *event = NULL;
struct mm_io_header *headp;
- int len = (count > 65535) ? 65535 : count;
+ ssize_t len = (count > 65535) ? 65535 : count;
event = kzalloc(sizeof(*headp) + len, GFP_KERNEL);
if (!event)
@@ -117,7 +148,12 @@ static int write_marker(struct file *file, const char __user *buffer,
return -EFAULT;
}
- relay_write(chan, event, sizeof(*headp) + len);
+ spin_lock_irq(&trace_lock);
+ if (is_enabled())
+ relay_write(chan, event, sizeof(*headp) + len);
+ else
+ len = -EINVAL;
+ spin_unlock_irq(&trace_lock);
kfree(event);
return len;
}
@@ -128,19 +164,18 @@ static void print_pte(unsigned long address)
pte_t *pte = lookup_address(address, &level);
if (!pte) {
- pr_err(MODULE_NAME ": Error in %s: no pte for page 0x%08lx\n",
+ pr_err(NAME "Error in %s: no pte for page 0x%08lx\n",
__func__, address);
return;
}
if (level == PG_LEVEL_2M) {
- pr_emerg(MODULE_NAME ": 4MB pages are not currently "
- "supported: %lx\n", address);
+ pr_emerg(NAME "4MB pages are not currently supported: "
+ "0x%08lx\n", address);
BUG();
}
- pr_info(MODULE_NAME ": pte for 0x%lx: 0x%lx 0x%lx\n",
- address, pte_val(*pte),
- pte_val(*pte) & _PAGE_PRESENT);
+ pr_info(NAME "pte for 0x%lx: 0x%lx 0x%lx\n", address, pte_val(*pte),
+ pte_val(*pte) & _PAGE_PRESENT);
}
/*
@@ -150,22 +185,18 @@ static void print_pte(unsigned long address)
static void die_kmmio_nesting_error(struct pt_regs *regs, unsigned long addr)
{
const struct trap_reason *my_reason = &get_cpu_var(pf_reason);
- pr_emerg(MODULE_NAME ": unexpected fault for address: %lx, "
- "last fault for address: %lx\n",
+ pr_emerg(NAME "unexpected fault for address: 0x%08lx, "
+ "last fault for address: 0x%08lx\n",
addr, my_reason->addr);
print_pte(addr);
+ print_symbol(KERN_EMERG "faulting IP is at %s\n", regs->ip);
+ print_symbol(KERN_EMERG "last faulting IP was at %s\n", my_reason->ip);
#ifdef __i386__
- print_symbol(KERN_EMERG "faulting EIP is at %s\n", regs->ip);
- print_symbol(KERN_EMERG "last faulting EIP was at %s\n",
- my_reason->ip);
pr_emerg("eax: %08lx ebx: %08lx ecx: %08lx edx: %08lx\n",
regs->ax, regs->bx, regs->cx, regs->dx);
pr_emerg("esi: %08lx edi: %08lx ebp: %08lx esp: %08lx\n",
regs->si, regs->di, regs->bp, regs->sp);
#else
- print_symbol(KERN_EMERG "faulting RIP is at %s\n", regs->ip);
- print_symbol(KERN_EMERG "last faulting RIP was at %s\n",
- my_reason->ip);
pr_emerg("rax: %016lx rcx: %016lx rdx: %016lx\n",
regs->ax, regs->cx, regs->dx);
pr_emerg("rsi: %016lx rdi: %016lx rbp: %016lx rsp: %016lx\n",
@@ -197,6 +228,10 @@ static void pre(struct kmmio_probe *p, struct pt_regs *regs,
my_trace->header.pid = 0;
my_trace->header.data_len = sizeof(struct mm_io_rw);
my_trace->rw.address = addr;
+ /*
+ * struct remap_trace *trace = p->user_data;
+ * phys = addr - trace->probe.addr + trace->phys;
+ */
/*
* Only record the program counter when requested.
@@ -246,15 +281,10 @@ static void post(struct kmmio_probe *p, unsigned long condition,
struct trap_reason *my_reason = &get_cpu_var(pf_reason);
struct mm_io_header_rw *my_trace = &get_cpu_var(cpu_trace);
- /*
- * XXX: This might not get called, if the probe is removed while
- * trace hit is on flight.
- */
-
/* this should always return the active_trace count to 0 */
my_reason->active_traces--;
if (my_reason->active_traces) {
- pr_emerg(MODULE_NAME ": unexpected post handler");
+ pr_emerg(NAME "unexpected post handler");
BUG();
}
@@ -284,20 +314,23 @@ static int subbuf_start_handler(struct rchan_buf *buf, void *subbuf,
int count;
if (relay_buf_full(buf)) {
if (atomic_inc_return(drop) == 1)
- pr_err(MODULE_NAME ": cpu %d buffer full!\n", cpu);
+ pr_err(NAME "cpu %d buffer full!\n", cpu);
return 0;
}
count = atomic_read(drop);
if (count) {
- pr_err(MODULE_NAME ": cpu %d buffer no longer full, "
- "missed %d events.\n",
- cpu, count);
+ pr_err(NAME "cpu %d buffer no longer full, missed %d events.\n",
+ cpu, count);
atomic_sub(count, drop);
}
return 1;
}
+static struct file_operations mmio_fops = {
+ .owner = THIS_MODULE,
+};
+
/* file_create() callback. Creates relay file in debugfs. */
static struct dentry *create_buf_file_handler(const char *filename,
struct dentry *parent,
@@ -333,34 +366,10 @@ static struct rchan_callbacks relay_callbacks = {
.remove_buf_file = remove_buf_file_handler,
};
-/*
- * create_channel - creates channel /debug/APP_DIR/cpuXXX
- * Returns channel on success, NULL otherwise
- */
-static struct rchan *create_channel(unsigned size, unsigned n)
-{
- return relay_open("cpu", dir, size, n, &relay_callbacks, NULL);
-}
-
-/* destroy_channel - destroys channel /debug/APP_DIR/cpuXXX */
-static void destroy_channel(void)
-{
- if (chan) {
- relay_close(chan);
- chan = NULL;
- }
-}
-
-struct remap_trace {
- struct list_head list;
- struct kmmio_probe probe;
-};
-static LIST_HEAD(trace_list);
-static DEFINE_SPINLOCK(trace_list_lock);
-
-static void do_ioremap_trace_core(unsigned long offset, unsigned long size,
+static void ioremap_trace_core(unsigned long offset, unsigned long size,
void __iomem *addr)
{
+ static atomic_t next_id;
struct remap_trace *trace = kmalloc(sizeof(*trace), GFP_KERNEL);
struct mm_io_header_map event = {
.header = {
@@ -380,61 +389,49 @@ static void do_ioremap_trace_core(unsigned long offset, unsigned long size,
};
record_timestamp(&event.header);
+ if (!trace) {
+ pr_err(NAME "kmalloc failed in ioremap\n");
+ return;
+ }
+
*trace = (struct remap_trace) {
.probe = {
.addr = (unsigned long)addr,
.len = size,
.pre_handler = pre,
.post_handler = post,
- }
+ .user_data = trace
+ },
+ .phys = offset,
+ .id = atomic_inc_return(&next_id)
};
+ spin_lock_irq(&trace_lock);
+ if (!is_enabled())
+ goto not_enabled;
+
relay_write(chan, &event, sizeof(event));
- spin_lock(&trace_list_lock);
list_add_tail(&trace->list, &trace_list);
- spin_unlock(&trace_list_lock);
if (!nommiotrace)
register_kmmio_probe(&trace->probe);
+
+not_enabled:
+ spin_unlock_irq(&trace_lock);
}
-static void ioremap_trace_core(unsigned long offset, unsigned long size,
- void __iomem *addr)
+void
+mmiotrace_ioremap(unsigned long offset, unsigned long size, void __iomem *addr)
{
- if ((filter_offset) && (offset != filter_offset))
+ if (!is_enabled()) /* recheck and proper locking in *_core() */
return;
- /* Don't trace the low PCI/ISA area, it's always mapped.. */
- if (!ISA_trace && (offset < ISA_END_ADDRESS) &&
- (offset + size > ISA_START_ADDRESS)) {
- pr_notice(MODULE_NAME ": Ignoring map of low PCI/ISA area "
- "(0x%lx-0x%lx)\n",
- offset, offset + size);
+ pr_debug(NAME "ioremap_*(0x%lx, 0x%lx) = %p\n", offset, size, addr);
+ if ((filter_offset) && (offset != filter_offset))
return;
- }
- do_ioremap_trace_core(offset, size, addr);
-}
-
-void __iomem *ioremap_cache_trace(unsigned long offset, unsigned long size)
-{
- void __iomem *p = ioremap_cache(offset, size);
- pr_debug(MODULE_NAME ": ioremap_cache(0x%lx, 0x%lx) = %p\n",
- offset, size, p);
- ioremap_trace_core(offset, size, p);
- return p;
+ ioremap_trace_core(offset, size, addr);
}
-EXPORT_SYMBOL(ioremap_cache_trace);
-void __iomem *ioremap_nocache_trace(unsigned long offset, unsigned long size)
-{
- void __iomem *p = ioremap_nocache(offset, size);
- pr_debug(MODULE_NAME ": ioremap_nocache(0x%lx, 0x%lx) = %p\n",
- offset, size, p);
- ioremap_trace_core(offset, size, p);
- return p;
-}
-EXPORT_SYMBOL(ioremap_nocache_trace);
-
-void iounmap_trace(volatile void __iomem *addr)
+static void iounmap_trace_core(volatile void __iomem *addr)
{
struct mm_io_header_map event = {
.header = {
@@ -454,84 +451,212 @@ void iounmap_trace(volatile void __iomem *addr)
};
struct remap_trace *trace;
struct remap_trace *tmp;
- pr_debug(MODULE_NAME ": Unmapping %p.\n", addr);
+ struct remap_trace *found_trace = NULL;
+
+ pr_debug(NAME "Unmapping %p.\n", addr);
record_timestamp(&event.header);
- spin_lock(&trace_list_lock);
+ spin_lock_irq(&trace_lock);
+ if (!is_enabled())
+ goto not_enabled;
+
list_for_each_entry_safe(trace, tmp, &trace_list, list) {
if ((unsigned long)addr == trace->probe.addr) {
if (!nommiotrace)
unregister_kmmio_probe(&trace->probe);
list_del(&trace->list);
- kfree(trace);
+ found_trace = trace;
break;
}
}
- spin_unlock(&trace_list_lock);
relay_write(chan, &event, sizeof(event));
- iounmap(addr);
+
+not_enabled:
+ spin_unlock_irq(&trace_lock);
+ if (found_trace) {
+ synchronize_rcu(); /* unregister_kmmio_probe() requirement */
+ kfree(found_trace);
+ }
+}
+
+void mmiotrace_iounmap(volatile void __iomem *addr)
+{
+ might_sleep();
+ if (is_enabled()) /* recheck and proper locking in *_core() */
+ iounmap_trace_core(addr);
}
-EXPORT_SYMBOL(iounmap_trace);
static void clear_trace_list(void)
{
struct remap_trace *trace;
struct remap_trace *tmp;
- spin_lock(&trace_list_lock);
- list_for_each_entry_safe(trace, tmp, &trace_list, list) {
- pr_warning(MODULE_NAME ": purging non-iounmapped "
+ /*
+ * No locking required, because the caller ensures we are in a
+ * critical section via mutex, and is_enabled() is false,
+ * i.e. nothing can traverse or modify this list.
+ * Caller also ensures is_enabled() cannot change.
+ */
+ list_for_each_entry(trace, &trace_list, list) {
+ pr_notice(NAME "purging non-iounmapped "
"trace @0x%08lx, size 0x%lx.\n",
trace->probe.addr, trace->probe.len);
if (!nommiotrace)
unregister_kmmio_probe(&trace->probe);
+ }
+ synchronize_rcu(); /* unregister_kmmio_probe() requirement */
+
+ list_for_each_entry_safe(trace, tmp, &trace_list, list) {
list_del(&trace->list);
kfree(trace);
+ }
+}
+
+static ssize_t read_enabled_file_bool(struct file *file,
+ char __user *user_buf, size_t count, loff_t *ppos)
+{
+ char buf[3];
+
+ if (is_enabled())
+ buf[0] = '1';
+ else
+ buf[0] = '0';
+ buf[1] = '\n';
+ buf[2] = '\0';
+ return simple_read_from_buffer(user_buf, count, ppos, buf, 2);
+}
+
+static void enable_mmiotrace(void);
+static void disable_mmiotrace(void);
+
+static ssize_t write_enabled_file_bool(struct file *file,
+ const char __user *user_buf, size_t count, loff_t *ppos)
+{
+ char buf[32];
+ int buf_size = min(count, (sizeof(buf)-1));
+
+ if (copy_from_user(buf, user_buf, buf_size))
+ return -EFAULT;
+
+ switch (buf[0]) {
+ case 'y':
+ case 'Y':
+ case '1':
+ enable_mmiotrace();
+ break;
+ case 'n':
+ case 'N':
+ case '0':
+ disable_mmiotrace();
break;
}
- spin_unlock(&trace_list_lock);
+
+ return count;
+}
+
+/* this ripped from kernel/kprobes.c */
+static struct file_operations fops_enabled = {
+ .owner = THIS_MODULE,
+ .read = read_enabled_file_bool,
+ .write = write_enabled_file_bool
+};
+
+static struct file_operations fops_marker = {
+ .owner = THIS_MODULE,
+ .write = write_marker
+};
+
+static void enable_mmiotrace(void)
+{
+ mutex_lock(&mmiotrace_mutex);
+ if (is_enabled())
+ goto out;
+
+ chan = relay_open("cpu", dir, subbuf_size, n_subbufs,
+ &relay_callbacks, NULL);
+ if (!chan) {
+ pr_err(NAME "relay app channel creation failed.\n");
+ goto out;
+ }
+
+ reference_kmmio();
+
+ marker_file = debugfs_create_file("marker", 0660, dir, NULL,
+ &fops_marker);
+ if (!marker_file)
+ pr_err(NAME "marker file creation failed.\n");
+
+ if (nommiotrace)
+ pr_info(NAME "MMIO tracing disabled.\n");
+ if (ISA_trace)
+ pr_warning(NAME "Warning! low ISA range will be traced.\n");
+ spin_lock_irq(&trace_lock);
+ atomic_inc(&mmiotrace_enabled);
+ spin_unlock_irq(&trace_lock);
+ pr_info(NAME "enabled.\n");
+out:
+ mutex_unlock(&mmiotrace_mutex);
+}
+
+static void disable_mmiotrace(void)
+{
+ mutex_lock(&mmiotrace_mutex);
+ if (!is_enabled())
+ goto out;
+
+ spin_lock_irq(&trace_lock);
+ atomic_dec(&mmiotrace_enabled);
+ BUG_ON(is_enabled());
+ spin_unlock_irq(&trace_lock);
+
+ clear_trace_list(); /* guarantees: no more kmmio callbacks */
+ unreference_kmmio();
+ if (marker_file) {
+ debugfs_remove(marker_file);
+ marker_file = NULL;
+ }
+ if (chan) {
+ relay_close(chan);
+ chan = NULL;
+ }
+
+ pr_info(NAME "disabled.\n");
+out:
+ mutex_unlock(&mmiotrace_mutex);
}
static int __init init(void)
{
+ pr_debug(NAME "load...\n");
if (n_subbufs < 2)
return -EINVAL;
dir = debugfs_create_dir(APP_DIR, NULL);
if (!dir) {
- pr_err(MODULE_NAME ": Couldn't create relay app directory.\n");
+ pr_err(NAME "Couldn't create relay app directory.\n");
return -ENOMEM;
}
- chan = create_channel(subbuf_size, n_subbufs);
- if (!chan) {
+ enabled_file = debugfs_create_file("enabled", 0600, dir, NULL,
+ &fops_enabled);
+ if (!enabled_file) {
+ pr_err(NAME "Couldn't create enabled file.\n");
debugfs_remove(dir);
- pr_err(MODULE_NAME ": relay app channel creation failed\n");
return -ENOMEM;
}
- reference_kmmio();
-
- proc_marker_file = create_proc_entry(MARKER_FILE, 0, NULL);
- if (proc_marker_file)
- proc_marker_file->write_proc = write_marker;
+ if (enable_now)
+ enable_mmiotrace();
- pr_debug(MODULE_NAME ": loaded.\n");
- if (nommiotrace)
- pr_info(MODULE_NAME ": MMIO tracing disabled.\n");
- if (ISA_trace)
- pr_warning(MODULE_NAME ": Warning! low ISA range will be "
- "traced.\n");
return 0;
}
static void __exit cleanup(void)
{
- pr_debug(MODULE_NAME ": unload...\n");
- clear_trace_list();
- unreference_kmmio();
- remove_proc_entry(MARKER_FILE, NULL);
- destroy_channel();
+ pr_debug(NAME "unload...\n");
+ if (enabled_file)
+ debugfs_remove(enabled_file);
+ disable_mmiotrace();
if (dir)
debugfs_remove(dir);
}
diff --git a/arch/x86/mm/testmmiotrace.c b/arch/x86/mm/testmmiotrace.c
index 5ecff57..cfa60b2 100644
--- a/arch/x86/mm/testmmiotrace.c
+++ b/arch/x86/mm/testmmiotrace.c
@@ -4,10 +4,6 @@
#include <linux/module.h>
#include <asm/io.h>
-extern void __iomem *ioremap_nocache_trace(unsigned long offset,
- unsigned long size);
-extern void iounmap_trace(volatile void __iomem *addr);
-
#define MODULE_NAME "testmmiotrace"
static unsigned long mmio_address;
@@ -28,25 +24,24 @@ static void do_write_test(void __iomem *p)
static void do_read_test(void __iomem *p)
{
unsigned int i;
- volatile unsigned int v;
for (i = 0; i < 256; i++)
- v = ioread8(p + i);
+ ioread8(p + i);
for (i = 1024; i < (5 * 1024); i += 2)
- v = ioread16(p + i);
+ ioread16(p + i);
for (i = (5 * 1024); i < (16 * 1024); i += 4)
- v = ioread32(p + i);
+ ioread32(p + i);
}
static void do_test(void)
{
- void __iomem *p = ioremap_nocache_trace(mmio_address, 0x4000);
+ void __iomem *p = ioremap_nocache(mmio_address, 0x4000);
if (!p) {
pr_err(MODULE_NAME ": could not ioremap, aborting.\n");
return;
}
do_write_test(p);
do_read_test(p);
- iounmap_trace(p);
+ iounmap(p);
}
static int __init init(void)
diff --git a/include/linux/mmiotrace.h b/include/linux/mmiotrace.h
index d87a6cd..cb5efd0 100644
--- a/include/linux/mmiotrace.h
+++ b/include/linux/mmiotrace.h
@@ -16,11 +16,12 @@ typedef void (*kmmio_post_handler_t)(struct kmmio_probe *,
unsigned long condition, struct pt_regs *);
struct kmmio_probe {
- struct list_head list;
+ struct list_head list; /* kmmio internal list */
unsigned long addr; /* start location of the probe point */
unsigned long len; /* length of the probe region */
kmmio_pre_handler_t pre_handler; /* Called before addr is executed. */
kmmio_post_handler_t post_handler; /* Called after addr is executed */
+ void *user_data;
};
/* kmmio is active by some kmmio_probes? */
@@ -38,6 +39,21 @@ extern void unregister_kmmio_probe(struct kmmio_probe *p);
/* Called from page fault handler. */
extern int kmmio_handler(struct pt_regs *regs, unsigned long addr);
+/* Called from ioremap.c */
+#ifdef CONFIG_MMIOTRACE
+extern void
+mmiotrace_ioremap(unsigned long offset, unsigned long size, void __iomem *addr);
+extern void mmiotrace_iounmap(volatile void __iomem *addr);
+#else
+static inline void
+mmiotrace_ioremap(unsigned long offset, unsigned long size, void __iomem *addr)
+{
+}
+static inline void mmiotrace_iounmap(volatile void __iomem *addr)
+{
+}
+#endif /* CONFIG_MMIOTRACE_HOOKS */
+
#endif /* __KERNEL__ */
--
1.5.3.7
Hi,
I may of course be wrong, but... Shouldn't the post_kmmio_handler(),
called from the die notifier chain, check for the DR_STEP condition?
This makes sure that the function is not called in the cases where the
source of the debug exception was not a single-stepping event. Though
I guess you'll also have other checks in place to notice that the
interrupt was not the one you were expecting. I guess a little extra
safety won't hurt though?
On Sun, Mar 9, 2008 at 3:40 PM, Pekka Paalanen <[email protected]> wrote:
> +/*
> + * Interrupts are disabled on entry as trap1 is an interrupt gate
> + * and they remain disabled thorough out this function.
> + * This must always get called as the pair to kmmio_handler().
> + */
> +static int post_kmmio_handler(unsigned long condition, struct pt_regs *regs)
> +{
> + int ret = 0;
> + struct kmmio_context *ctx = &get_cpu_var(kmmio_ctx);
if (!(condition & DR_STEP))
return;
Vegard
On Fri, 28 Mar 2008 00:13:48 +0100
"Vegard Nossum" <[email protected]> wrote:
> I may of course be wrong, but... Shouldn't the post_kmmio_handler(),
> called from the die notifier chain, check for the DR_STEP condition?
> This makes sure that the function is not called in the cases where the
> source of the debug exception was not a single-stepping event. Though
> I guess you'll also have other checks in place to notice that the
> interrupt was not the one you were expecting. I guess a little extra
> safety won't hurt though?
>
> On Sun, Mar 9, 2008 at 3:40 PM, Pekka Paalanen <[email protected]> wrote:
> > +/*
> > + * Interrupts are disabled on entry as trap1 is an interrupt gate
> > + * and they remain disabled thorough out this function.
> > + * This must always get called as the pair to kmmio_handler().
> > + */
> > +static int post_kmmio_handler(unsigned long condition, struct pt_regs *regs)
> > +{
> > + int ret = 0;
> > + struct kmmio_context *ctx = &get_cpu_var(kmmio_ctx);
>
> if (!(condition & DR_STEP))
> return;
I guess that would be appropriate. I think it should go into this
function:
+static int kmmio_die_notifier(struct notifier_block *nb, unsigned long val,
+ void *args)
+{
+ struct die_args *arg = args;
+
+ if (val == DIE_DEBUG)
+ if (post_kmmio_handler(arg->err, arg->regs) == 1)
+ return NOTIFY_STOP;
+
+ return NOTIFY_DONE;
+}
On the other hand I am thinking of not using the die notifier chain
at all and adding a direct call from do_debug() or something.
This is the last dynamic hook remaining from the out-of-tree module
era of mmiotrace.
I guess with the notifier list there is a possibility that another
module intercepts my single step trap, so that this is never called,
which would leave mmiotrace half blind, and also trigger a recursive
probe hit.
btw. what if someone uses kmemcheck and mmiotrace at the same time?
Mmiotrace will not fiddle with any other pages than returned via
__ioremap(), but can kmemcheck "hide" the same mmio pages?
Also keeping in mind, that some day I'd like to make mmiotrace able
to catch mmio accesses originating in user space.
Thanks.
--
Pekka Paalanen
http://www.iki.fi/pq/
Hi all,
jb17some reported a recursive probe hit when mmiotracing the nvidia
proprietary driver (the blob) on SMP. I'll introduce the basic mechanisms
in mmiotrace, provide some test results and then ask the question:
Why does disarm_kmmio_fault_page() occasionally not do its job on SMP?
Mmiotrace works by hooking into __ioremap() and marks the returned pages
as not present. Any access to these pages, specifically memory mapped IO
access, triggers a page fault. The fault handler, kmmio_handler(), marks
the hit page as present by calling disarm_kmmio_fault_page(), and sets
the TR flag. It also clears IF flag to prevent interrupts being enabled
when returning from the fault. The faulting instruction is re-executed,
this time without a fault, and the debug trap is hit. The debug trap
handler, post_kmmio_handler() re-arms the page and restores cpu flags.
kmmio_handler() and post_kmmio_handler() need to share some state, and
they must be called strictly in pairs, hence clearing IF and holding locks
from the first handler to the second. This is a per-cpu requirement,
different cpus are allowed to run concurrently here.
Here is the relevant code, quoted from my preview 2 email.
> +/** Mark the given page as not present. Access to it will trigger a fault. */
> +static void arm_kmmio_fault_page(unsigned long page, int *page_level)
> +{
> + unsigned long address = page & PAGE_MASK;
> + int level;
> + pte_t *pte = lookup_address(address, &level);
> +
> + if (!pte) {
> + pr_err("kmmio: Error in %s: no pte for page 0x%08lx\n",
> + __func__, page);
> + return;
> + }
> +
> + if (level == PG_LEVEL_2M) {
> + pmd_t *pmd = (pmd_t *)pte;
> + set_pmd(pmd, __pmd(pmd_val(*pmd) & ~_PAGE_PRESENT));
> + } else {
> + /* PG_LEVEL_4K */
> + set_pte(pte, __pte(pte_val(*pte) & ~_PAGE_PRESENT));
> + }
> +
> + if (page_level)
> + *page_level = level;
> +
> + __flush_tlb_one(page);
> +}
> +
> +/** Mark the given page as present. */
> +static void disarm_kmmio_fault_page(unsigned long page, int *page_level)
> +{
> + unsigned long address = page & PAGE_MASK;
> + int level;
> + pte_t *pte = lookup_address(address, &level);
> +
> + if (!pte) {
> + pr_err("kmmio: Error in %s: no pte for page 0x%08lx\n",
> + __func__, page);
> + return;
> + }
> +
> + if (level == PG_LEVEL_2M) {
> + pmd_t *pmd = (pmd_t *)pte;
> + set_pmd(pmd, __pmd(pmd_val(*pmd) | _PAGE_PRESENT));
> + } else {
> + /* PG_LEVEL_4K */
> + set_pte(pte, __pte(pte_val(*pte) | _PAGE_PRESENT));
> + }
> +
> + if (page_level)
> + *page_level = level;
> +
> + __flush_tlb_one(page);
> +}
> +
> +/*
> + * This is being called from do_page_fault().
> + *
> + * We may be in an interrupt or a critical section. Also prefecthing may
> + * trigger a page fault. We may be in the middle of process switch.
> + * We cannot take any locks, because we could be executing especially
> + * within a kmmio critical section.
> + *
> + * Local interrupts are disabled, so preemption cannot happen.
> + * Do not enable interrupts, do not sleep, and watch out for other CPUs.
> + */
> +/*
> + * Interrupts are disabled on entry as trap3 is an interrupt gate
> + * and they remain disabled thorough out this function.
> + */
> +int kmmio_handler(struct pt_regs *regs, unsigned long addr)
> +{
> + struct kmmio_context *ctx;
> + struct kmmio_fault_page *faultpage;
> +
> + /*
> + * Preemption is now disabled to prevent process switch during
> + * single stepping. We can only handle one active kmmio trace
> + * per cpu, so ensure that we finish it before something else
> + * gets to run. We also hold the RCU read lock over single
> + * stepping to avoid looking up the probe and kmmio_fault_page
> + * again.
> + */
> + preempt_disable();
> + rcu_read_lock();
> +
> + faultpage = get_kmmio_fault_page(addr);
> + if (!faultpage) {
> + /*
> + * Either this page fault is not caused by kmmio, or
> + * another CPU just pulled the kmmio probe from under
> + * our feet. In the latter case all hell breaks loose.
> + */
> + goto no_kmmio;
> + }
> +
> + ctx = &get_cpu_var(kmmio_ctx);
> + if (ctx->active) {
> + /*
> + * Prevent overwriting already in-flight context.
> + * If this page fault really was due to kmmio trap,
> + * all hell breaks loose.
> + */
> + pr_emerg("kmmio: recursive probe hit on CPU %d, "
> + "for address 0x%08lx. Ignoring.\n",
> + smp_processor_id(), addr);
> + goto no_kmmio_ctx;
> + }
> + ctx->active++;
> +
> + ctx->fpage = faultpage;
> + ctx->probe = get_kmmio_probe(addr);
> + ctx->saved_flags = (regs->flags & (TF_MASK|IF_MASK));
> + ctx->addr = addr;
> +
> + if (ctx->probe && ctx->probe->pre_handler)
> + ctx->probe->pre_handler(ctx->probe, regs, addr);
> +
> + /*
> + * Enable single-stepping and disable interrupts for the faulting
> + * context. Local interrupts must not get enabled during stepping.
> + */
> + regs->flags |= TF_MASK;
> + regs->flags &= ~IF_MASK;
> +
> + /* Now we set present bit in PTE and single step. */
> + disarm_kmmio_fault_page(ctx->fpage->page, NULL);
> +
> + /*
> + * If another cpu accesses the same page while we are stepping,
> + * the access will not be caught. It will simply succeed and the
> + * only downside is we lose the event. If this becomes a problem,
> + * the user should drop to single cpu before tracing.
> + */
> +
> + put_cpu_var(kmmio_ctx);
> + return 1;
> +
> +no_kmmio_ctx:
> + put_cpu_var(kmmio_ctx);
> +no_kmmio:
> + rcu_read_unlock();
> + preempt_enable_no_resched();
> + return 0; /* page fault not handled by kmmio */
> +}
> +
> +/*
> + * Interrupts are disabled on entry as trap1 is an interrupt gate
> + * and they remain disabled thorough out this function.
> + * This must always get called as the pair to kmmio_handler().
> + */
> +static int post_kmmio_handler(unsigned long condition, struct pt_regs *regs)
> +{
> + int ret = 0;
> + struct kmmio_context *ctx = &get_cpu_var(kmmio_ctx);
> +
> + if (!ctx->active)
> + goto out;
> +
> + if (ctx->probe && ctx->probe->post_handler)
> + ctx->probe->post_handler(ctx->probe, condition, regs);
> +
> + arm_kmmio_fault_page(ctx->fpage->page, NULL);
> +
> + regs->flags &= ~TF_MASK;
> + regs->flags |= ctx->saved_flags;
> +
> + /* These were acquired in kmmio_handler(). */
> + ctx->active--;
> + BUG_ON(ctx->active);
> + rcu_read_unlock();
> + preempt_enable_no_resched();
> +
> + /*
> + * if somebody else is singlestepping across a probe point, flags
> + * will have TF set, in which case, continue the remaining processing
> + * of do_debug, as if this is not a probe hit.
> + */
> + if (!(regs->flags & TF_MASK))
> + ret = 1;
> +out:
> + put_cpu_var(kmmio_ctx);
> + return ret;
> +}
A recursive probe hit means that kmmio_handler() is called twice without a
a call to post_kmmio_handler() in between. This situation is explicitly
checked for (if (ctx->active)), and the current solution is to ignore the
fault and fall through to do_page_fault() triggering an error there.
According to experience, this does not happen on a uniprocessor machine.
However, on an SMP machine this can occasionally occur. I have reproduced
it on my Core 2 Duo laptop while tracing the blob. Recursive probe hit
is very rare compared to the events logged, I can run two glxgears at the
same time for half an hour generating at least millions of events and never
hit it. Repeatedly start and stop a single glxgears, and I have a fairly
good chance of hitting it. It is random, but reproducible.
Here is a kernel oops log via netconsole:
kmmio: recursive probe hit on CPU 0, for address 0xffffc20004500140. Ignoring.
BUG: unable to handle kernel paging request at ffffc20004500140
IP: [<ffffffffa04e7b38>] :nvidia:_nv003832rm+0x1d/0x22
PGD 3e8d9067 PUD 3e8da067 PMD 3cd58067 PTE 80000000d600017a
Oops: 0000 [1] PREEMPT SMP DEBUG_PAGEALLOC
CPU 0
Modules linked in: nvidia(P) netconsole snd_seq_oss snd_seq_midi_event snd_seq
snd_seq_device snd_pcm_oss snd_mixer_oss sha256_generic cpufreq_stats
acpi_cpufreq video output fan button arc4 ecb snd_hda_intel snd_pcm snd_timer
snd snd_page_alloc iwl4965 mac80211 ehci_hcd uhci_hcd usbcore ohci1394
ieee1394 psmouse e1000 pcspkr sg i2c_i801 i2c_core yenta_socket
rsrc_nonstatic [last unloaded: nvidia]
Pid: 0, comm: swapper Tainted: P 2.6.25-rc6-sched-devel.git-x86-latest.git #2
RIP: 0010:[<ffffffffa04e7b38>] [<ffffffffa04e7b38>] :nvidia:_nv003832rm+0x1d/0x22
RSP: 0018:ffffffff806cfe38 EFLAGS: 00010106
RAX: 0000000000000050 RBX: ffff810019a49000 RCX: 0000000000000140
RDX: ffffc20004500000 RSI: ffff810019a49000 RDI: ffff81001ae5a7f0
RBP: ffff810019b1f148 R08: ffff81001ad6a000 R09: ffffffff806cfe18
R10: ffffffff80428f60 R11: ffff81003cc63140 R12: ffff81001ae5a7f0
R13: 0000000000000140 R14: 0000000000000001 R15: ffff8100194b3000
FS: 0000000000000000(0000) GS:ffffffff80662000(0000) knlGS:0000000000000000
CS: 0010 DS: 0018 ES: 0018 CR0: 000000008005003b
CR2: ffffc20004500140 CR3: 000000001af2a000 CR4: 00000000000006e0
DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
DR3: 0000000000000000 DR6: 00000000ffff4ff0 DR7: 0000000000000400
Process swapper (pid: 0, threadinfo ffffffff80684000, task ffffffff8061f4e0)
Stack: ffffffffa03b978e ffff810019a49000 ffff81001ad6ebf0 ffffc200040b0000
0000000000000001 ffff8100194b3000 ffffffffa04e8655 ffffc200040b0000
ffff810019a49000 ffffc200040b0000 ffff810019b1c180 ffffffff806cff18
Call Trace:
<IRQ> [<ffffffffa03b978e>] ? :nvidia:_nv008123rm+0x2a/0xc7
[<ffffffffa04e8655>] ? :nvidia:_nv003822rm+0x163/0x256
[<ffffffffa04e6b82>] ? :nvidia:_nv003815rm+0x4f/0x76
[<ffffffffa04edc13>] ? :nvidia:rm_isr+0xaa/0x10a
[<ffffffffa059ac61>] ? :nvidia:nv_kern_isr+0x7d/0xb1
[<ffffffff8025f567>] ? handle_IRQ_event+0x20/0x55
[<ffffffff802609c5>] ? handle_fasteoi_irq+0x9c/0xdc
[<ffffffff8020e95a>] ? do_IRQ+0x105/0x17f
[<ffffffff8020b7e6>] ? ret_from_intr+0x0/0xf
<EOI> [<ffffffff80421c12>] ? poll_idle+0x16/0x63
[<ffffffff80421c17>] ? poll_idle+0x1b/0x63
[<ffffffff80421c12>] ? poll_idle+0x16/0x63
[<ffffffff80422039>] ? cpuidle_idle_call+0x7e/0xad
[<ffffffff80421fbb>] ? cpuidle_idle_call+0x0/0xad
[<ffffffff8020a1f5>] ? cpu_idle+0x92/0xd0
[<ffffffff804c6025>] ? rest_init+0x69/0x6b
Code: 89 c8 d1 e8 89 c0 0f b7 04 42 0f b7 c0 c3 89 d1 b8 00 00 00 00 39 96
0c 02 00 00 76 11 48 8b 96 40 02 00 00 89 c8 c1 e8 02 89 c0 <8b> 04 82 f3
c3 39 96 14 02 00 00 76 0c 89 d2 48 8b 86 58 02 00
RIP [<ffffffffa04e7b38>] :nvidia:_nv003832rm+0x1d/0x22
RSP <ffffffff806cfe38>
CR2: ffffc20004500140
---[ end trace 0a8c36f3081eaf7e ]---
Kernel panic - not syncing: Aiee, killing interrupt handler!
Next, after discussion with Enberg and Nossum, I tried the following patch:
@@ -272,6 +272,9 @@ int kmmio_handler(struct pt_regs *regs, unsigned long addr)
pr_emerg("kmmio: recursive probe hit on CPU %d, "
"for address 0x%08lx. Ignoring.\n",
smp_processor_id(), addr);
+ pr_emerg("kmmio: previous hit was at 0x%08lx.\n",
+ ctx->addr);
+ disarm_kmmio_fault_page(faultpage->page, NULL);
goto no_kmmio_ctx;
}
ctx->active++;
@@ -322,8 +325,11 @@ static int post_kmmio_handler(unsigned long condition, struct pt_regs *regs)
int ret = 0;
struct kmmio_context *ctx = &get_cpu_var(kmmio_ctx);
- if (!ctx->active)
+ if (!ctx->active) {
+ pr_notice("kmmio: spurious debug trap on CPU %d.\n",
+ smp_processor_id());
goto out;
+ }
The result is occasional two lines:
kmmio: recursive probe hit on CPU 1, for address 0xffffc20004400140. Ignoring.
kmmio: previous hit was at 0xffffc20004400140.
The addresses vary, but are always the same for a pair of lines.
The "spurious debug trap" is never seen, and no kernel oopses. Note, that I
(unintentionally) still fall through to do_page_fault() after disarming.
Hypothesis 1:
Something gets to run on this cpu between exiting kmmio_handler() and
entering post_kmmio_handler(), and it triggers a new kmmio probe hit.
This is unlikely, as there should be a spurious debug trap afterwards.
Hypothesis 2:
disarm_kmmio_fault_page() does not work.
But it works most of the time, and on UP it seems to work always.
If hypothesis 2 is true, when the faulted instruction is re-executed,
it faults again. This is supported by the fault address being the same
at both times.
Could something else (apart from NMI) get to run in between fault and
debug handlers?
Is there something wrong in arm/disarm_kmmio_fault_page() functions?
Could it be a race between cpus?
Does the fall through to do_page_fault() trigger a proper fix to page
tables in case my disarming is bad?
Is the debugging hack an acceptable workaround?
I hope using a binary blob as a test case does not offend anyone, but
the main purpose of mmiotrace is to reverse engineer binary blobs.
I would appreciate any insight to the problem.
Thanks,
Pekka.
--
Pekka Paalanen
http://www.iki.fi/pq/
On Fri, 28 Mar 2008 22:25:00 +0200
Pekka Paalanen <[email protected]> wrote:
> A recursive probe hit means that kmmio_handler() is called twice without a
> a call to post_kmmio_handler() in between. This situation is explicitly
> checked for (if (ctx->active)), and the current solution is to ignore the
> fault and fall through to do_page_fault() triggering an error there.
> According to experience, this does not happen on a uniprocessor machine.
>
> However, on an SMP machine this can occasionally occur. I have reproduced
> it on my Core 2 Duo laptop while tracing the blob. Recursive probe hit
> is very rare compared to the events logged, I can run two glxgears at the
> same time for half an hour generating at least millions of events and never
> hit it. Repeatedly start and stop a single glxgears, and I have a fairly
> good chance of hitting it. It is random, but reproducible.
It appears this happens:
CPU 0 CPU 1
,---> fault fault
| disarm disarm
| single step
| arm
| single step
'--------'
arm
and the both cpus are faulting on the same page. I guess one cpu is running
an nvidia interrupt service.
I see three possible solutions:
A) Like in this patch, just disarm again and hope for the best.
Seems to work ok. I also compare the fault address to the saved address
ctx->addr. If they are equal, it is a "double probe hit" and harmless.
If they are not equal, it is a real "recursive probe hit" and something
more is wrong. With these definitions, recursive probe hits are gone in
my experiments on Intel Core 2 Duo.
> Next, after discussion with Enberg and Nossum, I tried the following patch:
>
> @@ -272,6 +272,9 @@ int kmmio_handler(struct pt_regs *regs, unsigned long addr)
> pr_emerg("kmmio: recursive probe hit on CPU %d, "
> "for address 0x%08lx. Ignoring.\n",
> smp_processor_id(), addr);
> + pr_emerg("kmmio: previous hit was at 0x%08lx.\n",
> + ctx->addr);
> + disarm_kmmio_fault_page(faultpage->page, NULL);
> goto no_kmmio_ctx;
> }
> ctx->active++;
B) Acquire a spinlock in kmmio_handler() and release it in
post_kmmio_handler(). I don't like this one since I spent some effort
making the fault path spinlockless, but at least this would be a
completely separate spinlock. Or we could use per-page spinlocks.
C) Vegard mentioned something about per-cpu page tables for kmemcheck.
This would be the ultimate solution, because it would solve two problems:
- recursive probe hits
- missed events due to another cpu disarming the page for single stepping
Would it be possible to have a single temporary per-cpu pte?
I understood kmemcheck has similar issues. Of course, one could force the
system down to a single running CPU, but that feels nasty.
Which way to go?
I choose A) as the current workaround, keeping in mind that I will loose
events on SMP. C) would be the only reliable SMP solution on tracing point
of view.
Thanks.
--
Pekka Paalanen
http://www.iki.fi/pq/