On 9/27/19 4:39 PM, Leonardo Bras wrote:
It's necessary to monitor lockless pagetable walks, in order to avoid doing
THP splitting/collapsing during them.
Some methods rely on local_irq_{save,restore}, but that can be slow on
cases with a lot of cpus are used for the process.
In order to speedup some cases, I propose a refcount-based approach, that
counts the number of lockless pagetable walks happening on the process.
This method does not exclude the current irq-oriented method. It works as a
complement to skip unnecessary waiting.
start_lockless_pgtbl_walk(mm)
Insert before starting any lockless pgtable walk
end_lockless_pgtbl_walk(mm)
Insert after the end of any lockless pgtable walk
(Mostly after the ptep is last used)
running_lockless_pgtbl_walk(mm)
Returns the number of lockless pgtable walks running
Signed-off-by: Leonardo Bras <leonardo@xxxxxxxxxxxxx>
---
arch/powerpc/include/asm/book3s/64/mmu.h | 3 ++
arch/powerpc/mm/book3s64/mmu_context.c | 1 +
arch/powerpc/mm/book3s64/pgtable.c | 45 ++++++++++++++++++++++++
3 files changed, 49 insertions(+)
diff --git a/arch/powerpc/include/asm/book3s/64/mmu.h b/arch/powerpc/include/asm/book3s/64/mmu.h
index 23b83d3593e2..13b006e7dde4 100644
--- a/arch/powerpc/include/asm/book3s/64/mmu.h
+++ b/arch/powerpc/include/asm/book3s/64/mmu.h
@@ -116,6 +116,9 @@ typedef struct {
/* Number of users of the external (Nest) MMU */
atomic_t copros;
+ /* Number of running instances of lockless pagetable walk*/
+ atomic_t lockless_pgtbl_walk_count;
+
struct hash_mm_context *hash_context;
unsigned long vdso_base;
diff --git a/arch/powerpc/mm/book3s64/mmu_context.c b/arch/powerpc/mm/book3s64/mmu_context.c
index 2d0cb5ba9a47..3dd01c0ca5be 100644
--- a/arch/powerpc/mm/book3s64/mmu_context.c
+++ b/arch/powerpc/mm/book3s64/mmu_context.c
@@ -200,6 +200,7 @@ int init_new_context(struct task_struct *tsk, struct mm_struct *mm)
#endif
atomic_set(&mm->context.active_cpus, 0);
atomic_set(&mm->context.copros, 0);
+ atomic_set(&mm->context.lockless_pgtbl_walk_count, 0);
return 0;
}
diff --git a/arch/powerpc/mm/book3s64/pgtable.c b/arch/powerpc/mm/book3s64/pgtable.c
index 7d0e0d0d22c4..6ba6195bff1b 100644
--- a/arch/powerpc/mm/book3s64/pgtable.c
+++ b/arch/powerpc/mm/book3s64/pgtable.c
@@ -98,6 +98,51 @@ void serialize_against_pte_lookup(struct mm_struct *mm)
smp_call_function_many(mm_cpumask(mm), do_nothing, NULL, 1);
}
+/*
+ * Counting method to monitor lockless pagetable walks:
+ * Uses start_lockless_pgtbl_walk and end_lockless_pgtbl_walk to track the
+ * number of lockless pgtable walks happening, and
+ * running_lockless_pgtbl_walk to return this value.
+ */
+
+/* start_lockless_pgtbl_walk: Must be inserted before a function call that does
+ * lockless pagetable walks, such as __find_linux_pte()
+ */
+void start_lockless_pgtbl_walk(struct mm_struct *mm)
+{
+ atomic_inc(&mm->context.lockless_pgtbl_walk_count);
+ /* Avoid reorder to garantee that the increment will happen before any
+ * part of the walkless pagetable walk after it.
+ */
+ smp_mb();
+}
+EXPORT_SYMBOL(start_lockless_pgtbl_walk);
+
+/*
+ * end_lockless_pgtbl_walk: Must be inserted after the last use of a pointer
+ * returned by a lockless pagetable walk, such as __find_linux_pte()
+*/
+void end_lockless_pgtbl_walk(struct mm_struct *mm)
+{
+ /* Avoid reorder to garantee that it will only decrement after the last
+ * use of the returned ptep from the lockless pagetable walk.
+ */
+ smp_mb();
+ atomic_dec(&mm->context.lockless_pgtbl_walk_count);
+}
+EXPORT_SYMBOL(end_lockless_pgtbl_walk);
+
+/*
+ * running_lockless_pgtbl_walk: Returns the number of lockless pagetable walks
+ * currently running. If it returns 0, there is no running pagetable walk, and
+ * THP split/collapse can be safely done. This can be used to avoid more
+ * expensive approaches like serialize_against_pte_lookup()
+ */
+int running_lockless_pgtbl_walk(struct mm_struct *mm)
+{
+ return atomic_read(&mm->context.lockless_pgtbl_walk_count);
+}
+
/*
* We use this to invalidate a pmdp entry before switching from a
* hugepte to regular pmd entry.
Hi, Leonardo,
Can we please do it as shown below, instead (compile-tested only)?
This addresses all of the comments that I was going to make about structure
of this patch, which are:
* The lockless synch is tricky, so it should be encapsulated in function
calls if possible.
* This is really a core mm function, so don't hide it away in arch layers.
(If you're changing mm/ files, that's a big hint.)
* Other things need parts of this: gup.c needs the memory barriers; IMHO you'll
be fixing a pre-existing, theoretical (we've never seen bug reports) problem.
* The documentation needs to accurately explain what's going on here.
(Not shown: one or more of the PPC Kconfig files should select
LOCKLESS_PAGE_TABLE_WALK_TRACKING.)
So:
diff --git a/include/linux/mm.h b/include/linux/mm.h
index 294a67b94147..c9e5defb4d7e 100644
--- a/include/linux/mm.h
+++ b/include/linux/mm.h
@@ -1541,6 +1541,9 @@ int account_locked_vm(struct mm_struct *mm, unsigned long pages, bool inc);
int __account_locked_vm(struct mm_struct *mm, unsigned long pages, bool inc,
struct task_struct *task, bool bypass_rlim);
+void register_lockless_page_table_walker(unsigned long *flags);
+void deregister_lockless_page_table_walker(unsigned long *flags);
+
/* Container for pinned pfns / pages */
struct frame_vector {
unsigned int nr_allocated; /* Number of frames we have space for */
diff --git a/include/linux/mm_types.h b/include/linux/mm_types.h
index 5183e0d77dfa..83b7930a995f 100644
--- a/include/linux/mm_types.h
+++ b/include/linux/mm_types.h
@@ -403,6 +403,16 @@ struct mm_struct {
*/
atomic_t mm_count;
+#ifdef LOCKLESS_PAGE_TABLE_WALK_TRACKING
+ /*
+ * Number of callers who are doing a lockless walk of the
+ * page tables. Typically arches might enable this in order to
+ * help optimize performance, by possibly avoiding expensive
+ * IPIs at the wrong times.
+ */
+ atomic_t lockless_pgtbl_nr_walkers;
+#endif
+
#ifdef CONFIG_MMU
atomic_long_t pgtables_bytes; /* PTE page table pages */
#endif
diff --git a/mm/Kconfig b/mm/Kconfig
index a5dae9a7eb51..1cf58f668fe1 100644
--- a/mm/Kconfig
+++ b/mm/Kconfig
@@ -736,4 +736,15 @@ config ARCH_HAS_PTE_SPECIAL
config ARCH_HAS_HUGEPD
bool
+config LOCKLESS_PAGE_TABLE_WALK_TRACKING
+ bool "Tracking (and optimization) of lockless page table walkers"
+ default n
+
+ help
+ Maintain a reference count of active lockless page table
+ walkers. This adds 4 bytes to struct mm size, and two atomic
+ operations to calls such as get_user_pages_fast(). Some
+ architectures can optimize page table operations if this
+ is enabled.
+
endmenu
diff --git a/mm/gup.c b/mm/gup.c
index 60c3915c8ee6..7b1be8ed1e8f 100644
--- a/mm/gup.c
+++ b/mm/gup.c
@@ -2302,6 +2302,62 @@ static bool gup_fast_permitted(unsigned long start, unsigned long end)
}
#endif
+/*
+ * register_lockless_page_table_walker() - start a lockless page table walk
+ *
+ * @flags: for saving and restoring irq state
+ *
+ * Lockless page table walking still requires synchronization against freeing
+ * of the page tables, and against splitting of huge pages. This is done by
+ * interacting with interrupts, as first described in the struct mmu_table_batch
+ * comments in include/asm-generic/tlb.h.
+ *
+ * In order to do the right thing, code that walks page tables in the style of
+ * get_user_pages_fast() should call register_lockless_page_table_walker()
+ * before starting the walk, and deregister_lockless_page_table_walker() upon
+ * finishing.
+ */
+void register_lockless_page_table_walker(unsigned long *flags)
+{
+#ifdef LOCKLESS_PAGE_TABLE_WALK_TRACKING
+ atomic_inc(¤t->mm->lockless_pgtbl_nr_walkers);
+#endif
+ /*
+ * This memory barrier pairs with any code that is either trying to
+ * delete page tables, or split huge pages. In order for that to work,
+ * interrupts must also be disabled during the lockless page table
+ * walk. That's because the deleting or splitting involves flushing
+ * TLBs, which in turn issues interrupts, which will block here.
+ * However, without memory barriers, the page tables could be
+ * read speculatively outside of interrupt disabling.
+ */
+ smp_mb();
+ local_irq_save(*flags);
+}
+EXPORT_SYMBOL_GPL(gup_fast_lock_acquire);
+
+/*
+ * register_lockless_page_table_walker() - finish a lockless page table walk
+ *
+ * This is the complement to register_lockless_page_table_walker().
+ *
+ * @flags: for saving and restoring irq state
+ */
+void deregister_lockless_page_table_walker(unsigned long *flags)
+{
+ local_irq_restore(flags);
+ /*
+ * This memory barrier pairs with any code that is either trying to
+ * delete page tables, or split huge pages. See the comments in
+ * gup_fast_lock_acquire() for details.
+ */
+ smp_mb();
+#ifdef LOCKLESS_PAGE_TABLE_WALK_TRACKING
+ atomic_dec(¤t->mm->lockless_pgtbl_nr_walkers);
+#endif
+}
+EXPORT_SYMBOL_GPL(gup_fast_lock_release);
+
/*
* Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
* the regular GUP.
@@ -2341,9 +2397,9 @@ int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
if (IS_ENABLED(CONFIG_HAVE_FAST_GUP) &&
gup_fast_permitted(start, end)) {
- local_irq_save(flags);
+ register_lockless_page_table_walker(&flags);
gup_pgd_range(start, end, write ? FOLL_WRITE : 0, pages, &nr);
- local_irq_restore(flags);
+ deregister_lockless_page_table_walker(&flags);
}
return nr;
@@ -2392,7 +2448,7 @@ static int __gup_longterm_unlocked(unsigned long start, int nr_pages,
int get_user_pages_fast(unsigned long start, int nr_pages,
unsigned int gup_flags, struct page **pages)
{
- unsigned long addr, len, end;
+ unsigned long addr, len, end, flags;
int nr = 0, ret = 0;
if (WARN_ON_ONCE(gup_flags & ~(FOLL_WRITE | FOLL_LONGTERM)))
@@ -2410,9 +2466,9 @@ int get_user_pages_fast(unsigned long start, int nr_pages,
if (IS_ENABLED(CONFIG_HAVE_FAST_GUP) &&
gup_fast_permitted(start, end)) {
- local_irq_disable();
+ register_lockless_page_table_walker(&flags);
gup_pgd_range(addr, end, gup_flags, pages, &nr);
- local_irq_enable();
+ deregister_lockless_page_table_walker(&flags);
ret = nr;
}
thanks,
--
John Hubbard
NVIDIA
