Groups | Search | Server Info | Keyboard shortcuts | Login | Register [http] [https] [nntp] [nntps]

Groups > linux.kernel > #1714438 > unrolled thread

[PATCH v2 00/20] Speculative page faults

Back to article view | Back to linux.kernel

Contents

  [PATCH v2 00/20] Speculative page faults Laurent Dufour <ldufour@linux.vnet.ibm.com> - 2017-08-18 00:10 +0200
    [PATCH v2 10/20] mm: Introduce __lru_cache_add_active_or_unevictable Laurent Dufour <ldufour@linux.vnet.ibm.com> - 2017-08-18 00:20 +0200
    [PATCH v2 12/20] mm: Introduce __vm_normal_page() Laurent Dufour <ldufour@linux.vnet.ibm.com> - 2017-08-18 00:20 +0200
    [PATCH v2 09/20] mm/migrate: Pass vm_fault pointer to migrate_misplaced_page() Laurent Dufour <ldufour@linux.vnet.ibm.com> - 2017-08-18 00:20 +0200
    [PATCH v2 14/20] mm: Provide speculative fault infrastructure Laurent Dufour <ldufour@linux.vnet.ibm.com> - 2017-08-18 00:20 +0200
      Re: [PATCH v2 14/20] mm: Provide speculative fault infrastructure Sergey Senozhatsky <sergey.senozhatsky@gmail.com> - 2017-08-20 14:20 +0200
        Re: [PATCH v2 14/20] mm: Provide speculative fault infrastructure Laurent Dufour <ldufour@linux.vnet.ibm.com> - 2017-08-25 11:00 +0200
    [PATCH v2 02/20] mm: Prepare for FAULT_FLAG_SPECULATIVE Laurent Dufour <ldufour@linux.vnet.ibm.com> - 2017-08-18 00:20 +0200
    [PATCH v2 06/20] mm: RCU free VMAs Laurent Dufour <ldufour@linux.vnet.ibm.com> - 2017-08-18 00:20 +0200
    [PATCH v2 08/20] mm: Protect SPF handler against anon_vma changes Laurent Dufour <ldufour@linux.vnet.ibm.com> - 2017-08-18 00:20 +0200
    Re: [PATCH v2 00/20] Speculative page faults Sergey Senozhatsky <sergey.senozhatsky.work@gmail.com> - 2017-08-21 04:30 +0200
    Re: [PATCH v2 00/20] Speculative page faults Anshuman Khandual <khandual@linux.vnet.ibm.com> - 2017-08-21 08:30 +0200
      Re: [PATCH v2 00/20] Speculative page faults "Paul E. McKenney" <paulmck@linux.vnet.ibm.com> - 2017-08-22 02:50 +0200
      Re: [PATCH v2 00/20] Speculative page faults Laurent Dufour <ldufour@linux.vnet.ibm.com> - 2017-08-25 11:50 +0200

#1714438 — [PATCH v2 00/20] Speculative page faults

This is a port on kernel 4.13 of the work done by Peter Zijlstra to
handle page fault without holding the mm semaphore [1].

The idea is to try to handle user space page faults without holding the
mmap_sem. This should allow better concurrency for massively threaded
process since the page fault handler will not wait for other threads memory
layout change to be done, assuming that this change is done in another part
of the process's memory space. This type page fault is named speculative
page fault. If the speculative page fault fails because of a concurrency is
detected or because underlying PMD or PTE tables are not yet allocating, it
is failing its processing and a classic page fault is then tried.

The speculative page fault (SPF) has to look for the VMA matching the fault
address without holding the mmap_sem, so the VMA list is now managed using
SRCU allowing lockless walking. The only impact would be the deferred file
derefencing in the case of a file mapping, since the file pointer is
released once the SRCU cleaning is done.  This patch relies on the change
done recently by Paul McKenney in SRCU which now runs a callback per CPU
instead of per SRCU structure [1].

The VMA's attributes checked during the speculative page fault processing
have to be protected against parallel changes. This is done by using a per
VMA sequence lock. This sequence lock allows the speculative page fault
handler to fast check for parallel changes in progress and to abort the
speculative page fault in that case.

Once the VMA is found, the speculative page fault handler would check for
the VMA's attributes to verify that the page fault has to be handled
correctly or not. Thus the VMA is protected through a sequence lock which
allows fast detection of concurrent VMA changes. If such a change is
detected, the speculative page fault is aborted and a *classic* page fault
is tried.  VMA sequence locks are added when VMA attributes which are
checked during the page fault are modified.

When the PTE is fetched, the VMA is checked to see if it has been changed,
so once the page table is locked, the VMA is valid, so any other changes
leading to touching this PTE will need to lock the page table, so no
parallel change is possible at this time.

Compared to the Peter's initial work, this series introduces a spin_trylock
when dealing with speculative page fault. This is required to avoid dead
lock when handling a page fault while a TLB invalidate is requested by an
other CPU holding the PTE. Another change due to a lock dependency issue
with mapping->i_mmap_rwsem.

In addition some VMA field values which are used once the PTE is unlocked
at the end the page fault path are saved into the vm_fault structure to
used the values matching the VMA at the time the PTE was locked.

This series builds on top of v4.13-rc5 and is functional on x86 and
PowerPC.

Tests have been made using a large commercial in-memory database on a
PowerPC system with 752 CPU using RFC v5. The results are very encouraging
since the loading of the 2TB database was faster by 14% with the
speculative page fault.

Using ebizzy test [3], which spreads a lot of threads, the result are good
when running on both a large or a small system. When using kernbench, the
result are quite similar which expected as not so much multithreaded
processes are involved. But there is no performance degradation neither
which is good.

------------------
Benchmarks results

Note these test have been made on top of 4.13-rc3 with the following patch
from Paul McKenney applied: 
 "srcu: Provide ordering for CPU not involved in grace period" [5]

Ebizzy:
-------
The test is counting the number of records per second it can manage, the
higher is the best. I run it like this 'ebizzy -mTRp'. To get consistent
result I repeated the test 100 times and measure the average result, mean
deviation, max and min.

- 16 CPUs x86 VM
Records/s	4.13-rc5	4.13-rc5-spf
Average		11350.29	21760.36
Mean deviation	396.56		881.40
Max		13773		26194
Min		10567		19223

- 80 CPUs Power 8 node:
Records/s	4.13-rc5	4.13-rc5-spf
Average		33904.67	58847.91
Mean deviation	789.40		1753.19
Max		36703		68958
Min		31759		55125

The number of record per second is far better with the speculative page
fault.
The mean deviation is higher with the speculative page fault, may be
because sometime the fault are not handled in a speculative way leading to
more variation.


Kernbench:
----------
This test is building a 4.12 kernel using platform default config. The
build has been run 5 times each time.

- 16 CPUs x86 VM
Average Half load -j 8 Run (std deviation)
 		 4.13.0-rc5		4.13.0-rc5-spf
Elapsed Time     166.574 (0.340779)	145.754 (0.776325)		
User Time        1080.77 (2.05871)	999.272 (4.12142)		
System Time      204.594 (1.02449)	116.362 (1.22974)		
Percent CPU 	 771.2 (1.30384)	765 (0.707107)
Context Switches 46590.6 (935.591)	66316.4 (744.64)
Sleeps           84421.2 (596.612)	85186 (523.041)		

Average Optimal load -j 16 Run (std deviation)
 		 4.13.0-rc5		4.13.0-rc5-spf
Elapsed Time     85.422 (0.42293)	74.81 (0.419345)
User Time        1031.79 (51.6557)	954.912 (46.8439)
System Time      186.528 (19.0575)	107.514 (9.36902)
Percent CPU 	 1059.2 (303.607)	1056.8 (307.624)
Context Switches 67240.3 (21788.9)	89360.6 (24299.9)
Sleeps           89607.8 (5511.22)	90372.5 (5490.16)

The elapsed time is a bit shorter in the case of the SPF release, but the
impact less important since there are less multithreaded processes involved
here. 

- 80 CPUs Power 8 node:
Average Half load -j 40 Run (std deviation)
 		 4.13.0-rc5		4.13.0-rc5-spf
Elapsed Time     117.176 (0.824093)	116.792 (0.695392)
User Time        4412.34 (24.29)	4396.02 (24.4819)
System Time      131.106 (1.28343)	133.452 (0.708851)
Percent CPU      3876.8 (18.1439)	3877.6 (21.9955)
Context Switches 72470.2 (466.181)	72971 (673.624)
Sleeps           161294 (2284.85)	161946 (2217.9)

Average Optimal load -j 80 Run (std deviation)
 		 4.13.0-rc5		4.13.0-rc5-spf
Elapsed Time     111.176 (1.11123)	111.242 (0.801542)
User Time        5930.03 (1600.07)	5929.89 (1617)
System Time      166.258 (37.0662)	169.337 (37.8419)
Percent CPU      5378.5 (1584.16)	5385.6 (1590.24)
Context Switches 117389 (47350.1)	130132 (60256.3)
Sleeps           163354 (4153.9)	163219 (2251.27)

Here the elapsed time is a bit shorter using the spf release, but we
remain in the error margin. It has to be noted that this system is not
correctly balanced on the NUMA point of view as all the available memory is
attached to one core.

------------------------
Changes since v1:
 - Remove PERF_COUNT_SW_SPF_FAILED perf event.
 - Add tracing events to details speculative page fault failures.
 - Cache VMA fields values which are used once the PTE is unlocked at the
 end of the page fault events.
 - Ensure that fields read during the speculative path are written and read
 using WRITE_ONCE and READ_ONCE.
 - Add checks at the beginning of the speculative path to abort it if the
 VMA is known to not be supported.
Changes since RFC V5 [6]
 - Port to 4.13 kernel
 - Merging patch fixing lock dependency into the original patch
 - Replace the 2 parameters of vma_has_changed() with the vmf pointer
 - In patch 7, don't call __do_fault() in the speculative path as it may
 want to unlock the mmap_sem.
 - In patch 11-12, don't check for vma boundaries when
 page_add_new_anon_rmap() is called during the spf path and protect against
 anon_vma pointer's update.
 - In patch 13-16, add performance events to report number of successful
 and failed speculative events. 

[1] http://linux-kernel.2935.n7.nabble.com/RFC-PATCH-0-6-Another-go-at-speculative-page-faults-tt965642.html#none
[2] https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=da915ad5cf25b5f5d358dd3670c3378d8ae8c03e
[3] http://ebizzy.sourceforge.net/
[4] http://ck.kolivas.org/apps/kernbench/kernbench-0.50/
[5] https://lkml.org/lkml/2017/7/24/829
[6] https://lwn.net/Articles/725607/

Laurent Dufour (14):
  mm: Introduce pte_spinlock for FAULT_FLAG_SPECULATIVE
  mm: Protect VMA modifications using VMA sequence count
  mm: Cache some VMA fields in the vm_fault structure
  mm: Protect SPF handler against anon_vma changes
  mm/migrate: Pass vm_fault pointer to migrate_misplaced_page()
  mm: Introduce __lru_cache_add_active_or_unevictable
  mm: Introduce __maybe_mkwrite()
  mm: Introduce __vm_normal_page()
  mm: Introduce __page_add_new_anon_rmap()
  mm: Try spin lock in speculative path
  mm: Adding speculative page fault failure trace events
  perf: Add a speculative page fault sw event
  perf tools: Add support for the SPF perf event
  powerpc/mm: Add speculative page fault

Peter Zijlstra (6):
  mm: Dont assume page-table invariance during faults
  mm: Prepare for FAULT_FLAG_SPECULATIVE
  mm: VMA sequence count
  mm: RCU free VMAs
  mm: Provide speculative fault infrastructure
  x86/mm: Add speculative pagefault handling

 arch/powerpc/include/asm/book3s/64/pgtable.h |   5 +
 arch/powerpc/mm/fault.c                      |  30 +-
 arch/x86/include/asm/pgtable_types.h         |   7 +
 arch/x86/mm/fault.c                          |  19 ++
 fs/proc/task_mmu.c                           |   5 +-
 fs/userfaultfd.c                             |  17 +-
 include/linux/hugetlb_inline.h               |   2 +-
 include/linux/migrate.h                      |   4 +-
 include/linux/mm.h                           |  21 +-
 include/linux/mm_types.h                     |   3 +
 include/linux/pagemap.h                      |   4 +-
 include/linux/rmap.h                         |  12 +-
 include/linux/swap.h                         |  11 +-
 include/trace/events/pagefault.h             |  87 +++++
 include/uapi/linux/perf_event.h              |   1 +
 kernel/fork.c                                |   1 +
 mm/hugetlb.c                                 |   2 +
 mm/init-mm.c                                 |   1 +
 mm/internal.h                                |  19 ++
 mm/khugepaged.c                              |   5 +
 mm/madvise.c                                 |   6 +-
 mm/memory.c                                  | 474 ++++++++++++++++++++++-----
 mm/mempolicy.c                               |  51 ++-
 mm/migrate.c                                 |   4 +-
 mm/mlock.c                                   |  13 +-
 mm/mmap.c                                    | 138 ++++++--
 mm/mprotect.c                                |   4 +-
 mm/mremap.c                                  |   7 +
 mm/rmap.c                                    |   5 +-
 mm/swap.c                                    |  12 +-
 tools/include/uapi/linux/perf_event.h        |   1 +
 tools/perf/util/evsel.c                      |   1 +
 tools/perf/util/parse-events.c               |   4 +
 tools/perf/util/parse-events.l               |   1 +
 tools/perf/util/python.c                     |   1 +
 35 files changed, 803 insertions(+), 175 deletions(-)
 create mode 100644 include/trace/events/pagefault.h

-- 
2.7.4

[toc] | [next] | [standalone]

#1714446 — [PATCH v2 10/20] mm: Introduce __lru_cache_add_active_or_unevictable

The speculative page fault handler which is run without holding the
mmap_sem is calling lru_cache_add_active_or_unevictable() but the vm_flags
is not guaranteed to remain constant.
Introducing __lru_cache_add_active_or_unevictable() which has the vma flags
value parameter instead of the vma pointer.

Signed-off-by: Laurent Dufour <ldufour@linux.vnet.ibm.com>
---
 include/linux/swap.h | 11 +++++++++--
 mm/memory.c          |  8 ++++----
 mm/swap.c            | 12 ++++++------
 3 files changed, 19 insertions(+), 12 deletions(-)

diff --git a/include/linux/swap.h b/include/linux/swap.h
index d83d28e53e62..fdea932fe10f 100644
--- a/include/linux/swap.h
+++ b/include/linux/swap.h
@@ -285,8 +285,15 @@ extern void swap_setup(void);
 
 extern void add_page_to_unevictable_list(struct page *page);
 
-extern void lru_cache_add_active_or_unevictable(struct page *page,
-						struct vm_area_struct *vma);
+extern void __lru_cache_add_active_or_unevictable(struct page *page,
+						unsigned long vma_flags);
+
+static inline void lru_cache_add_active_or_unevictable(struct page *page,
+						struct vm_area_struct *vma)
+{
+	return __lru_cache_add_active_or_unevictable(page, vma->vm_flags);
+}
+
 
 /* linux/mm/vmscan.c */
 extern unsigned long zone_reclaimable_pages(struct zone *zone);
diff --git a/mm/memory.c b/mm/memory.c
index 53528eeee2b3..c6b18cc87e90 100644
--- a/mm/memory.c
+++ b/mm/memory.c
@@ -2370,7 +2370,7 @@ static int wp_page_copy(struct vm_fault *vmf)
 		ptep_clear_flush_notify(vma, vmf->address, vmf->pte);
 		page_add_new_anon_rmap(new_page, vma, vmf->address, false);
 		mem_cgroup_commit_charge(new_page, memcg, false, false);
-		lru_cache_add_active_or_unevictable(new_page, vma);
+		__lru_cache_add_active_or_unevictable(new_page, vmf->vma_flags);
 		/*
 		 * We call the notify macro here because, when using secondary
 		 * mmu page tables (such as kvm shadow page tables), we want the
@@ -2840,7 +2840,7 @@ int do_swap_page(struct vm_fault *vmf)
 	} else { /* ksm created a completely new copy */
 		page_add_new_anon_rmap(page, vma, vmf->address, false);
 		mem_cgroup_commit_charge(page, memcg, false, false);
-		lru_cache_add_active_or_unevictable(page, vma);
+		__lru_cache_add_active_or_unevictable(page, vmf->vma_flags);
 	}
 
 	swap_free(entry);
@@ -2978,7 +2978,7 @@ static int do_anonymous_page(struct vm_fault *vmf)
 	inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
 	page_add_new_anon_rmap(page, vma, vmf->address, false);
 	mem_cgroup_commit_charge(page, memcg, false, false);
-	lru_cache_add_active_or_unevictable(page, vma);
+	__lru_cache_add_active_or_unevictable(page, vmf->vma_flags);
 setpte:
 	set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
 
@@ -3230,7 +3230,7 @@ int alloc_set_pte(struct vm_fault *vmf, struct mem_cgroup *memcg,
 		inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
 		page_add_new_anon_rmap(page, vma, vmf->address, false);
 		mem_cgroup_commit_charge(page, memcg, false, false);
-		lru_cache_add_active_or_unevictable(page, vma);
+		__lru_cache_add_active_or_unevictable(page, vmf->vma_flags);
 	} else {
 		inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
 		page_add_file_rmap(page, false);
diff --git a/mm/swap.c b/mm/swap.c
index 60b1d2a75852..ece0826a205b 100644
--- a/mm/swap.c
+++ b/mm/swap.c
@@ -470,21 +470,21 @@ void add_page_to_unevictable_list(struct page *page)
 }
 
 /**
- * lru_cache_add_active_or_unevictable
- * @page:  the page to be added to LRU
- * @vma:   vma in which page is mapped for determining reclaimability
+ * __lru_cache_add_active_or_unevictable
+ * @page:	the page to be added to LRU
+ * @vma_flags:  vma in which page is mapped for determining reclaimability
  *
  * Place @page on the active or unevictable LRU list, depending on its
  * evictability.  Note that if the page is not evictable, it goes
  * directly back onto it's zone's unevictable list, it does NOT use a
  * per cpu pagevec.
  */
-void lru_cache_add_active_or_unevictable(struct page *page,
-					 struct vm_area_struct *vma)
+void __lru_cache_add_active_or_unevictable(struct page *page,
+					   unsigned long vma_flags)
 {
 	VM_BUG_ON_PAGE(PageLRU(page), page);
 
-	if (likely((vma->vm_flags & (VM_LOCKED | VM_SPECIAL)) != VM_LOCKED)) {
+	if (likely((vma_flags & (VM_LOCKED | VM_SPECIAL)) != VM_LOCKED)) {
 		SetPageActive(page);
 		lru_cache_add(page);
 		return;
-- 
2.7.4

[toc] | [prev] | [next] | [standalone]

#1714447 — [PATCH v2 12/20] mm: Introduce __vm_normal_page()

When dealing with the speculative fault path we should use the VMA's field
cached value stored in the vm_fault structure.

Currently vm_normal_page() is using the pointer to the VMA to fetch the
vm_flags value. This patch provides a new __vm_normal_page() which is
receiving the vm_flags flags value as parameter.

Note: The speculative path is turned on for architecture providing support
for special PTE flag. So only the first block of vm_normal_page is used
during the speculative path.

Signed-off-by: Laurent Dufour <ldufour@linux.vnet.ibm.com>
---
 mm/memory.c | 25 +++++++++++++++++--------
 1 file changed, 17 insertions(+), 8 deletions(-)

diff --git a/mm/memory.c b/mm/memory.c
index ad7b6372d302..9f9e5bb7a556 100644
--- a/mm/memory.c
+++ b/mm/memory.c
@@ -820,8 +820,9 @@ static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
 #else
 # define HAVE_PTE_SPECIAL 0
 #endif
-struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
-				pte_t pte)
+static struct page *__vm_normal_page(struct vm_area_struct *vma,
+				     unsigned long addr,
+				     pte_t pte, unsigned long vma_flags)
 {
 	unsigned long pfn = pte_pfn(pte);
 
@@ -830,7 +831,7 @@ struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
 			goto check_pfn;
 		if (vma->vm_ops && vma->vm_ops->find_special_page)
 			return vma->vm_ops->find_special_page(vma, addr);
-		if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
+		if (vma_flags & (VM_PFNMAP | VM_MIXEDMAP))
 			return NULL;
 		if (!is_zero_pfn(pfn))
 			print_bad_pte(vma, addr, pte, NULL);
@@ -839,8 +840,8 @@ struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
 
 	/* !HAVE_PTE_SPECIAL case follows: */
 
-	if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
-		if (vma->vm_flags & VM_MIXEDMAP) {
+	if (unlikely(vma_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
+		if (vma_flags & VM_MIXEDMAP) {
 			if (!pfn_valid(pfn))
 				return NULL;
 			goto out;
@@ -849,7 +850,7 @@ struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
 			off = (addr - vma->vm_start) >> PAGE_SHIFT;
 			if (pfn == vma->vm_pgoff + off)
 				return NULL;
-			if (!is_cow_mapping(vma->vm_flags))
+			if (!is_cow_mapping(vma_flags))
 				return NULL;
 		}
 	}
@@ -870,6 +871,13 @@ struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
 	return pfn_to_page(pfn);
 }
 
+struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
+			    pte_t pte)
+{
+	return __vm_normal_page(vma, addr, pte, vma->vm_flags);
+}
+
+
 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
 				pmd_t pmd)
@@ -2548,7 +2556,8 @@ static int do_wp_page(struct vm_fault *vmf)
 {
 	struct vm_area_struct *vma = vmf->vma;
 
-	vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte);
+	vmf->page = __vm_normal_page(vma, vmf->address, vmf->orig_pte,
+				     vmf->vma_flags);
 	if (!vmf->page) {
 		/*
 		 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
@@ -3575,7 +3584,7 @@ static int do_numa_page(struct vm_fault *vmf)
 	ptep_modify_prot_commit(vma->vm_mm, vmf->address, vmf->pte, pte);
 	update_mmu_cache(vma, vmf->address, vmf->pte);
 
-	page = vm_normal_page(vma, vmf->address, pte);
+	page = __vm_normal_page(vma, vmf->address, pte, vmf->vma_flags);
 	if (!page) {
 		pte_unmap_unlock(vmf->pte, vmf->ptl);
 		return 0;
-- 
2.7.4

[toc] | [prev] | [next] | [standalone]

#1714451 — [PATCH v2 09/20] mm/migrate: Pass vm_fault pointer to migrate_misplaced_page()

migrate_misplaced_page() is only called during the page fault handling so
it's better to pass the pointer to the struct vm_fault instead of the vma.

This way during the speculative page fault path the saved vma->vm_flags
could be used.

Signed-off-by: Laurent Dufour <ldufour@linux.vnet.ibm.com>
---
 include/linux/migrate.h | 4 ++--
 mm/memory.c             | 2 +-
 mm/migrate.c            | 4 ++--
 3 files changed, 5 insertions(+), 5 deletions(-)

diff --git a/include/linux/migrate.h b/include/linux/migrate.h
index 3e0d405dc842..65357105cbab 100644
--- a/include/linux/migrate.h
+++ b/include/linux/migrate.h
@@ -108,14 +108,14 @@ static inline void __ClearPageMovable(struct page *page)
 #ifdef CONFIG_NUMA_BALANCING
 extern bool pmd_trans_migrating(pmd_t pmd);
 extern int migrate_misplaced_page(struct page *page,
-				  struct vm_area_struct *vma, int node);
+				  struct vm_fault *vmf, int node);
 #else
 static inline bool pmd_trans_migrating(pmd_t pmd)
 {
 	return false;
 }
 static inline int migrate_misplaced_page(struct page *page,
-					 struct vm_area_struct *vma, int node)
+					 struct vm_fault *vmf, int node)
 {
 	return -EAGAIN; /* can't migrate now */
 }
diff --git a/mm/memory.c b/mm/memory.c
index 68e4fdcce692..53528eeee2b3 100644
--- a/mm/memory.c
+++ b/mm/memory.c
@@ -3616,7 +3616,7 @@ static int do_numa_page(struct vm_fault *vmf)
 	}
 
 	/* Migrate to the requested node */
-	migrated = migrate_misplaced_page(page, vma, target_nid);
+	migrated = migrate_misplaced_page(page, vmf, target_nid);
 	if (migrated) {
 		page_nid = target_nid;
 		flags |= TNF_MIGRATED;
diff --git a/mm/migrate.c b/mm/migrate.c
index d68a41da6abb..354f74f7dad3 100644
--- a/mm/migrate.c
+++ b/mm/migrate.c
@@ -1847,7 +1847,7 @@ bool pmd_trans_migrating(pmd_t pmd)
  * node. Caller is expected to have an elevated reference count on
  * the page that will be dropped by this function before returning.
  */
-int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
+int migrate_misplaced_page(struct page *page, struct vm_fault *vmf,
 			   int node)
 {
 	pg_data_t *pgdat = NODE_DATA(node);
@@ -1860,7 +1860,7 @@ int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
 	 * with execute permissions as they are probably shared libraries.
 	 */
 	if (page_mapcount(page) != 1 && page_is_file_cache(page) &&
-	    (vma->vm_flags & VM_EXEC))
+	    (vmf->vma_flags & VM_EXEC))
 		goto out;
 
 	/*
-- 
2.7.4

[toc] | [prev] | [next] | [standalone]

#1714452 — [PATCH v2 14/20] mm: Provide speculative fault infrastructure

From: Peter Zijlstra <peterz@infradead.org>

Provide infrastructure to do a speculative fault (not holding
mmap_sem).

The not holding of mmap_sem means we can race against VMA
change/removal and page-table destruction. We use the SRCU VMA freeing
to keep the VMA around. We use the VMA seqcount to detect change
(including umapping / page-table deletion) and we use gup_fast() style
page-table walking to deal with page-table races.

Once we've obtained the page and are ready to update the PTE, we
validate if the state we started the fault with is still valid, if
not, we'll fail the fault with VM_FAULT_RETRY, otherwise we update the
PTE and we're done.

Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>

[Manage the newly introduced pte_spinlock() for speculative page
 fault to fail if the VMA is touched in our back]
[Rename vma_is_dead() to vma_has_changed() and declare it here]
[Call p4d_alloc() as it is safe since pgd is valid]
[Call pud_alloc() as it is safe since p4d is valid]
[Set fe.sequence in __handle_mm_fault()]
[Abort speculative path when handle_userfault() has to be called]
[Add additional VMA's flags checks in handle_speculative_fault()]
[Clear FAULT_FLAG_ALLOW_RETRY in handle_speculative_fault()]
[Don't set vmf->pte and vmf->ptl if pte_map_lock() failed]
[Remove warning comment about waiting for !seq&1 since we don't want
 to wait]
[Remove warning about no huge page support, mention it explictly]
[Don't call do_fault() in the speculative path as __do_fault() calls
 vma->vm_ops->fault() which may want to release mmap_sem]
[Only vm_fault pointer argument for vma_has_changed()]
[Fix check against huge page, calling pmd_trans_huge()]
[Introduce __HAVE_ARCH_CALL_SPF to declare the SPF handler only when
 architecture is supporting it]
[Use READ_ONCE() when reading VMA's fields in the speculative path]
[Explicitly check for __HAVE_ARCH_PTE_SPECIAL as we can't support for
 processing done in vm_normal_page()]
[Check that vma->anon_vma is already set when starting the speculative
 path]
[Check for memory policy as we can't support MPOL_INTERLEAVE case due to
 the processing done in mpol_misplaced()]
[Don't support VMA growing up or down]
[Move check on vm_sequence just before calling handle_pte_fault()]
Signed-off-by: Laurent Dufour <ldufour@linux.vnet.ibm.com>
---
 include/linux/hugetlb_inline.h |   2 +-
 include/linux/mm.h             |   5 +
 include/linux/pagemap.h        |   4 +-
 mm/internal.h                  |  14 +++
 mm/memory.c                    | 237 ++++++++++++++++++++++++++++++++++++++++-
 5 files changed, 254 insertions(+), 8 deletions(-)

diff --git a/include/linux/hugetlb_inline.h b/include/linux/hugetlb_inline.h
index a4e7ca0f3585..6cfdfca4cc2a 100644
--- a/include/linux/hugetlb_inline.h
+++ b/include/linux/hugetlb_inline.h
@@ -7,7 +7,7 @@
 
 static inline bool is_vm_hugetlb_page(struct vm_area_struct *vma)
 {
-	return !!(vma->vm_flags & VM_HUGETLB);
+	return !!(READ_ONCE(vma->vm_flags) & VM_HUGETLB);
 }
 
 #else
diff --git a/include/linux/mm.h b/include/linux/mm.h
index 0f4ddd72b172..0fe0811d304f 100644
--- a/include/linux/mm.h
+++ b/include/linux/mm.h
@@ -315,6 +315,7 @@ struct vm_fault {
 	gfp_t gfp_mask;			/* gfp mask to be used for allocations */
 	pgoff_t pgoff;			/* Logical page offset based on vma */
 	unsigned long address;		/* Faulting virtual address */
+	unsigned int sequence;
 	pmd_t *pmd;			/* Pointer to pmd entry matching
 					 * the 'address' */
 	pud_t *pud;			/* Pointer to pud entry matching
@@ -1297,6 +1298,10 @@ int invalidate_inode_page(struct page *page);
 #ifdef CONFIG_MMU
 extern int handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
 		unsigned int flags);
+#ifdef __HAVE_ARCH_CALL_SPF
+extern int handle_speculative_fault(struct mm_struct *mm,
+				    unsigned long address, unsigned int flags);
+#endif /* __HAVE_ARCH_CALL_SPF */
 extern int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
 			    unsigned long address, unsigned int fault_flags,
 			    bool *unlocked);
diff --git a/include/linux/pagemap.h b/include/linux/pagemap.h
index 79b36f57c3ba..3a9735dfa6b6 100644
--- a/include/linux/pagemap.h
+++ b/include/linux/pagemap.h
@@ -443,8 +443,8 @@ static inline pgoff_t linear_page_index(struct vm_area_struct *vma,
 	pgoff_t pgoff;
 	if (unlikely(is_vm_hugetlb_page(vma)))
 		return linear_hugepage_index(vma, address);
-	pgoff = (address - vma->vm_start) >> PAGE_SHIFT;
-	pgoff += vma->vm_pgoff;
+	pgoff = (address - READ_ONCE(vma->vm_start)) >> PAGE_SHIFT;
+	pgoff += READ_ONCE(vma->vm_pgoff);
 	return pgoff;
 }
 
diff --git a/mm/internal.h b/mm/internal.h
index 736540f15936..9d6347e35747 100644
--- a/mm/internal.h
+++ b/mm/internal.h
@@ -45,6 +45,20 @@ extern struct srcu_struct vma_srcu;
 extern struct vm_area_struct *find_vma_srcu(struct mm_struct *mm,
 					    unsigned long addr);
 
+static inline bool vma_has_changed(struct vm_fault *vmf)
+{
+	int ret = RB_EMPTY_NODE(&vmf->vma->vm_rb);
+	unsigned seq = ACCESS_ONCE(vmf->vma->vm_sequence.sequence);
+
+	/*
+	 * Matches both the wmb in write_seqlock_{begin,end}() and
+	 * the wmb in vma_rb_erase().
+	 */
+	smp_rmb();
+
+	return ret || seq != vmf->sequence;
+}
+
 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *start_vma,
 		unsigned long floor, unsigned long ceiling);
 
diff --git a/mm/memory.c b/mm/memory.c
index 51bc8315281e..0ba14a5797b2 100644
--- a/mm/memory.c
+++ b/mm/memory.c
@@ -760,7 +760,8 @@ static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
 	if (page)
 		dump_page(page, "bad pte");
 	pr_alert("addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
-		 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
+		 (void *)addr, READ_ONCE(vma->vm_flags), vma->anon_vma,
+		 mapping, index);
 	/*
 	 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
 	 */
@@ -2285,15 +2286,69 @@ static inline void wp_page_reuse(struct vm_fault *vmf)
 
 static bool pte_spinlock(struct vm_fault *vmf)
 {
+	bool ret = false;
+
+	/* Check if vma is still valid */
+	if (!(vmf->flags & FAULT_FLAG_SPECULATIVE)) {
+		vmf->ptl = pte_lockptr(vmf->vma->vm_mm, vmf->pmd);
+		spin_lock(vmf->ptl);
+		return true;
+	}
+
+	local_irq_disable();
+	if (vma_has_changed(vmf))
+		goto out;
+
 	vmf->ptl = pte_lockptr(vmf->vma->vm_mm, vmf->pmd);
 	spin_lock(vmf->ptl);
-	return true;
+
+	if (vma_has_changed(vmf)) {
+		spin_unlock(vmf->ptl);
+		goto out;
+	}
+
+	ret = true;
+out:
+	local_irq_enable();
+	return ret;
 }
 
 static bool pte_map_lock(struct vm_fault *vmf)
 {
-	vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address, &vmf->ptl);
-	return true;
+	bool ret = false;
+	pte_t *pte;
+	spinlock_t *ptl;
+
+	if (!(vmf->flags & FAULT_FLAG_SPECULATIVE)) {
+		vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd,
+					       vmf->address, &vmf->ptl);
+		return true;
+	}
+
+	/*
+	 * The first vma_has_changed() guarantees the page-tables are still
+	 * valid, having IRQs disabled ensures they stay around, hence the
+	 * second vma_has_changed() to make sure they are still valid once
+	 * we've got the lock. After that a concurrent zap_pte_range() will
+	 * block on the PTL and thus we're safe.
+	 */
+	local_irq_disable();
+	if (vma_has_changed(vmf))
+		goto out;
+
+	pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd,
+				  vmf->address, &ptl);
+	if (vma_has_changed(vmf)) {
+		pte_unmap_unlock(pte, ptl);
+		goto out;
+	}
+
+	vmf->pte = pte;
+	vmf->ptl = ptl;
+	ret = true;
+out:
+	local_irq_enable();
+	return ret;
 }
 
 /*
@@ -2939,6 +2994,14 @@ static int do_anonymous_page(struct vm_fault *vmf)
 			return VM_FAULT_RETRY;
 		if (!pte_none(*vmf->pte))
 			goto unlock;
+		/*
+		 * Don't call the userfaultfd during the speculative path.
+		 * We already checked for the VMA to not be managed through
+		 * userfaultfd, but it may be set in our back once we have lock
+		 * the pte. In such a case we can ignore it this time.
+		 */
+		if (vmf->flags & FAULT_FLAG_SPECULATIVE)
+			goto setpte;
 		/* Deliver the page fault to userland, check inside PT lock */
 		if (userfaultfd_missing(vma)) {
 			pte_unmap_unlock(vmf->pte, vmf->ptl);
@@ -2977,7 +3040,7 @@ static int do_anonymous_page(struct vm_fault *vmf)
 		goto release;
 
 	/* Deliver the page fault to userland, check inside PT lock */
-	if (userfaultfd_missing(vma)) {
+	if (!(vmf->flags & FAULT_FLAG_SPECULATIVE) && userfaultfd_missing(vma)) {
 		pte_unmap_unlock(vmf->pte, vmf->ptl);
 		mem_cgroup_cancel_charge(page, memcg, false);
 		put_page(page);
@@ -3748,6 +3811,8 @@ static int handle_pte_fault(struct vm_fault *vmf)
 	if (!vmf->pte) {
 		if (vma_is_anonymous(vmf->vma))
 			return do_anonymous_page(vmf);
+		else if (vmf->flags & FAULT_FLAG_SPECULATIVE)
+			return VM_FAULT_RETRY;
 		else
 			return do_fault(vmf);
 	}
@@ -3845,6 +3910,7 @@ static int __handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
 	vmf.pmd = pmd_alloc(mm, vmf.pud, address);
 	if (!vmf.pmd)
 		return VM_FAULT_OOM;
+	vmf.sequence = raw_read_seqcount(&vma->vm_sequence);
 	if (pmd_none(*vmf.pmd) && transparent_hugepage_enabled(vma)) {
 		ret = create_huge_pmd(&vmf);
 		if (!(ret & VM_FAULT_FALLBACK))
@@ -3872,6 +3938,167 @@ static int __handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
 	return handle_pte_fault(&vmf);
 }
 
+#ifdef __HAVE_ARCH_CALL_SPF
+
+#ifndef __HAVE_ARCH_PTE_SPECIAL
+/* This is required by vm_normal_page() */
+#error "Speculative page fault handler requires __HAVE_ARCH_PTE_SPECIAL"
+#endif
+
+/*
+ * vm_normal_page() adds some processing which should be done while
+ * hodling the mmap_sem.
+ */
+int handle_speculative_fault(struct mm_struct *mm, unsigned long address,
+			     unsigned int flags)
+{
+	struct vm_fault vmf = {
+		.address = address,
+	};
+	pgd_t *pgd;
+	p4d_t *p4d;
+	pud_t *pud;
+	pmd_t *pmd;
+	int dead, seq, idx, ret = VM_FAULT_RETRY;
+	struct vm_area_struct *vma;
+	struct mempolicy *pol;
+
+	/* Clear flags that may lead to release the mmap_sem to retry */
+	flags &= ~(FAULT_FLAG_ALLOW_RETRY|FAULT_FLAG_KILLABLE);
+	flags |= FAULT_FLAG_SPECULATIVE;
+
+	idx = srcu_read_lock(&vma_srcu);
+	vma = find_vma_srcu(mm, address);
+	if (!vma)
+		goto unlock;
+
+	/*
+	 * Validate the VMA found by the lockless lookup.
+	 */
+	dead = RB_EMPTY_NODE(&vma->vm_rb);
+	seq = raw_read_seqcount(&vma->vm_sequence); /* rmb <-> seqlock,vma_rb_erase() */
+	if ((seq & 1) || dead)
+		goto unlock;
+
+	/*
+	 * Can't call vm_ops service has we don't know what they would do
+	 * with the VMA.
+	 * This include huge page from hugetlbfs.
+	 */
+	if (vma->vm_ops)
+		goto unlock;
+
+	if (unlikely(!vma->anon_vma))
+		goto unlock;
+
+	vmf.vma_flags = READ_ONCE(vma->vm_flags);
+	vmf.vma_page_prot = READ_ONCE(vma->vm_page_prot);
+
+	/* Can't call userland page fault handler in the speculative path */
+	if (unlikely(vmf.vma_flags & VM_UFFD_MISSING))
+		goto unlock;
+
+	/*
+	 * MPOL_INTERLEAVE implies additional check in mpol_misplaced() which
+	 * are not compatible with the speculative page fault processing.
+	 */
+	pol = __get_vma_policy(vma, address);
+	if (!pol)
+		pol = get_task_policy(current);
+	if (pol && pol->mode == MPOL_INTERLEAVE)
+		goto unlock;
+
+	if (vmf.vma_flags & VM_GROWSDOWN || vmf.vma_flags & VM_GROWSUP)
+		/*
+		 * This could be detected by the check address against VMA's
+		 * boundaries but we want to trace it as not supported instead
+		 * of changed.
+		 */
+		goto unlock;
+
+	if (address < READ_ONCE(vma->vm_start)
+	    || READ_ONCE(vma->vm_end) <= address)
+		goto unlock;
+
+	/*
+	 * The three following checks are copied from access_error from
+	 * arch/x86/mm/fault.c
+	 */
+	if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE,
+				       flags & FAULT_FLAG_INSTRUCTION,
+				       flags & FAULT_FLAG_REMOTE))
+		goto unlock;
+
+	/* This is one is required to check that the VMA has write access set */
+	if (flags & FAULT_FLAG_WRITE) {
+		if (unlikely(!(vmf.vma_flags & VM_WRITE)))
+			goto unlock;
+	} else {
+		if (unlikely(!(vmf.vma_flags & (VM_READ | VM_EXEC | VM_WRITE))))
+			goto unlock;
+	}
+
+	/*
+	 * Do a speculative lookup of the PTE entry.
+	 */
+	local_irq_disable();
+	pgd = pgd_offset(mm, address);
+	if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
+		goto out_walk;
+
+	p4d = p4d_alloc(mm, pgd, address);
+	if (p4d_none(*p4d) || unlikely(p4d_bad(*p4d)))
+		goto out_walk;
+
+	pud = pud_alloc(mm, p4d, address);
+	if (pud_none(*pud) || unlikely(pud_bad(*pud)))
+		goto out_walk;
+
+	pmd = pmd_offset(pud, address);
+	if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
+		goto out_walk;
+
+	/*
+	 * The above does not allocate/instantiate page-tables because doing so
+	 * would lead to the possibility of instantiating page-tables after
+	 * free_pgtables() -- and consequently leaking them.
+	 *
+	 * The result is that we take at least one !speculative fault per PMD
+	 * in order to instantiate it.
+	 */
+
+	/* Transparent huge pages are not supported. */
+	if (unlikely(pmd_trans_huge(*pmd)))
+		goto out_walk;
+
+	vmf.vma = vma;
+	vmf.pmd = pmd;
+	vmf.pgoff = linear_page_index(vma, address);
+	vmf.gfp_mask = __get_fault_gfp_mask(vma);
+	vmf.sequence = seq;
+	vmf.flags = flags;
+
+	local_irq_enable();
+
+	/*
+	 * We need to re-validate the VMA after checking the bounds, otherwise
+	 * we might have a false positive on the bounds.
+	 */
+	if (read_seqcount_retry(&vma->vm_sequence, seq))
+		goto unlock;
+
+	ret = handle_pte_fault(&vmf);
+
+unlock:
+	srcu_read_unlock(&vma_srcu, idx);
+	return ret;
+
+out_walk:
+	local_irq_enable();
+	goto unlock;
+}
+#endif /* __HAVE_ARCH_CALL_SPF */
+
 /*
  * By the time we get here, we already hold the mm semaphore
  *
-- 
2.7.4

[toc] | [prev] | [next] | [standalone]

#1715929 — Re: [PATCH v2 14/20] mm: Provide speculative fault infrastructure

On (08/18/17 00:05), Laurent Dufour wrote:
[..]
> +	/*
> +	 * MPOL_INTERLEAVE implies additional check in mpol_misplaced() which
> +	 * are not compatible with the speculative page fault processing.
> +	 */
> +	pol = __get_vma_policy(vma, address);
> +	if (!pol)
> +		pol = get_task_policy(current);
> +	if (pol && pol->mode == MPOL_INTERLEAVE)
> +		goto unlock;

include/linux/mempolicy.h defines

struct mempolicy *get_task_policy(struct task_struct *p);
struct mempolicy *__get_vma_policy(struct vm_area_struct *vma,
		unsigned long addr);

only for CONFIG_NUMA configs.

	-ss

[toc] | [prev] | [next] | [standalone]

#1719839 — Re: [PATCH v2 14/20] mm: Provide speculative fault infrastructure

On 20/08/2017 14:11, Sergey Senozhatsky wrote:
> On (08/18/17 00:05), Laurent Dufour wrote:
> [..]
>> +	/*
>> +	 * MPOL_INTERLEAVE implies additional check in mpol_misplaced() which
>> +	 * are not compatible with the speculative page fault processing.
>> +	 */
>> +	pol = __get_vma_policy(vma, address);
>> +	if (!pol)
>> +		pol = get_task_policy(current);
>> +	if (pol && pol->mode == MPOL_INTERLEAVE)
>> +		goto unlock;
> 
> include/linux/mempolicy.h defines
> 
> struct mempolicy *get_task_policy(struct task_struct *p);
> struct mempolicy *__get_vma_policy(struct vm_area_struct *vma,
> 		unsigned long addr);
> 
> only for CONFIG_NUMA configs.

Thanks Sergey, I'll add #ifdef around this block.

[toc] | [prev] | [next] | [standalone]

#1714460 — [PATCH v2 02/20] mm: Prepare for FAULT_FLAG_SPECULATIVE

From: Peter Zijlstra <peterz@infradead.org>

When speculating faults (without holding mmap_sem) we need to validate
that the vma against which we loaded pages is still valid when we're
ready to install the new PTE.

Therefore, replace the pte_offset_map_lock() calls that (re)take the
PTL with pte_map_lock() which can fail in case we find the VMA changed
since we started the fault.

Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>

[Port to 4.12 kernel]
[Remove the comment about the fault_env structure which has been
 implemented as the vm_fault structure in the kernel]
Signed-off-by: Laurent Dufour <ldufour@linux.vnet.ibm.com>
---
 include/linux/mm.h |  1 +
 mm/memory.c        | 55 ++++++++++++++++++++++++++++++++++++++----------------
 2 files changed, 40 insertions(+), 16 deletions(-)

diff --git a/include/linux/mm.h b/include/linux/mm.h
index 46b9ac5e8569..8763ec96dc78 100644
--- a/include/linux/mm.h
+++ b/include/linux/mm.h
@@ -286,6 +286,7 @@ extern pgprot_t protection_map[16];
 #define FAULT_FLAG_USER		0x40	/* The fault originated in userspace */
 #define FAULT_FLAG_REMOTE	0x80	/* faulting for non current tsk/mm */
 #define FAULT_FLAG_INSTRUCTION  0x100	/* The fault was during an instruction fetch */
+#define FAULT_FLAG_SPECULATIVE	0x200	/* Speculative fault, not holding mmap_sem */
 
 #define FAULT_FLAG_TRACE \
 	{ FAULT_FLAG_WRITE,		"WRITE" }, \
diff --git a/mm/memory.c b/mm/memory.c
index 36609c082256..3ed1b00ca841 100644
--- a/mm/memory.c
+++ b/mm/memory.c
@@ -2269,6 +2269,12 @@ static inline void wp_page_reuse(struct vm_fault *vmf)
 	pte_unmap_unlock(vmf->pte, vmf->ptl);
 }
 
+static bool pte_map_lock(struct vm_fault *vmf)
+{
+	vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address, &vmf->ptl);
+	return true;
+}
+
 /*
  * Handle the case of a page which we actually need to copy to a new page.
  *
@@ -2296,6 +2302,7 @@ static int wp_page_copy(struct vm_fault *vmf)
 	const unsigned long mmun_start = vmf->address & PAGE_MASK;
 	const unsigned long mmun_end = mmun_start + PAGE_SIZE;
 	struct mem_cgroup *memcg;
+	int ret = VM_FAULT_OOM;
 
 	if (unlikely(anon_vma_prepare(vma)))
 		goto oom;
@@ -2323,7 +2330,11 @@ static int wp_page_copy(struct vm_fault *vmf)
 	/*
 	 * Re-check the pte - we dropped the lock
 	 */
-	vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl);
+	if (!pte_map_lock(vmf)) {
+		mem_cgroup_cancel_charge(new_page, memcg, false);
+		ret = VM_FAULT_RETRY;
+		goto oom_free_new;
+	}
 	if (likely(pte_same(*vmf->pte, vmf->orig_pte))) {
 		if (old_page) {
 			if (!PageAnon(old_page)) {
@@ -2411,7 +2422,7 @@ static int wp_page_copy(struct vm_fault *vmf)
 oom:
 	if (old_page)
 		put_page(old_page);
-	return VM_FAULT_OOM;
+	return ret;
 }
 
 /**
@@ -2432,8 +2443,8 @@ static int wp_page_copy(struct vm_fault *vmf)
 int finish_mkwrite_fault(struct vm_fault *vmf)
 {
 	WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED));
-	vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address,
-				       &vmf->ptl);
+	if (!pte_map_lock(vmf))
+		return VM_FAULT_RETRY;
 	/*
 	 * We might have raced with another page fault while we released the
 	 * pte_offset_map_lock.
@@ -2551,8 +2562,11 @@ static int do_wp_page(struct vm_fault *vmf)
 			get_page(vmf->page);
 			pte_unmap_unlock(vmf->pte, vmf->ptl);
 			lock_page(vmf->page);
-			vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
-					vmf->address, &vmf->ptl);
+			if (!pte_map_lock(vmf)) {
+				unlock_page(vmf->page);
+				put_page(vmf->page);
+				return VM_FAULT_RETRY;
+			}
 			if (!pte_same(*vmf->pte, vmf->orig_pte)) {
 				unlock_page(vmf->page);
 				pte_unmap_unlock(vmf->pte, vmf->ptl);
@@ -2710,8 +2724,10 @@ int do_swap_page(struct vm_fault *vmf)
 			 * Back out if somebody else faulted in this pte
 			 * while we released the pte lock.
 			 */
-			vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
-					vmf->address, &vmf->ptl);
+			if (!pte_map_lock(vmf)) {
+				delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
+				return VM_FAULT_RETRY;
+			}
 			if (likely(pte_same(*vmf->pte, vmf->orig_pte)))
 				ret = VM_FAULT_OOM;
 			delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
@@ -2767,8 +2783,11 @@ int do_swap_page(struct vm_fault *vmf)
 	/*
 	 * Back out if somebody else already faulted in this pte.
 	 */
-	vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
-			&vmf->ptl);
+	if (!pte_map_lock(vmf)) {
+		ret = VM_FAULT_RETRY;
+		mem_cgroup_cancel_charge(page, memcg, false);
+		goto out_page;
+	}
 	if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte)))
 		goto out_nomap;
 
@@ -2894,8 +2913,8 @@ static int do_anonymous_page(struct vm_fault *vmf)
 			!mm_forbids_zeropage(vma->vm_mm)) {
 		entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address),
 						vma->vm_page_prot));
-		vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
-				vmf->address, &vmf->ptl);
+		if (!pte_map_lock(vmf))
+			return VM_FAULT_RETRY;
 		if (!pte_none(*vmf->pte))
 			goto unlock;
 		/* Deliver the page fault to userland, check inside PT lock */
@@ -2927,8 +2946,11 @@ static int do_anonymous_page(struct vm_fault *vmf)
 	if (vma->vm_flags & VM_WRITE)
 		entry = pte_mkwrite(pte_mkdirty(entry));
 
-	vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
-			&vmf->ptl);
+	if (!pte_map_lock(vmf)) {
+		mem_cgroup_cancel_charge(page, memcg, false);
+		put_page(page);
+		return VM_FAULT_RETRY;
+	}
 	if (!pte_none(*vmf->pte))
 		goto release;
 
@@ -3048,8 +3070,9 @@ static int pte_alloc_one_map(struct vm_fault *vmf)
 	 * pte_none() under vmf->ptl protection when we return to
 	 * alloc_set_pte().
 	 */
-	vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
-			&vmf->ptl);
+	if (!pte_map_lock(vmf))
+		return VM_FAULT_RETRY;
+
 	return 0;
 }
 
-- 
2.7.4

[toc] | [prev] | [next] | [standalone]

#1714461 — [PATCH v2 06/20] mm: RCU free VMAs

From: Peter Zijlstra <peterz@infradead.org>

Manage the VMAs with SRCU such that we can do a lockless VMA lookup.

We put the fput(vma->vm_file) in the SRCU callback, this keeps files
valid during speculative faults, this is possible due to the delayed
fput work by Al Viro -- do we need srcu_barrier() in unmount
someplace?

We guard the mm_rb tree with a seqlock (this could be a seqcount but
we'd have to disable preemption around the write side in order to make
the retry loop in __read_seqcount_begin() work) such that we can know
if the rb tree walk was correct. We cannot trust the restult of a
lockless tree walk in the face of concurrent tree rotations; although
we can trust on the termination of such walks -- tree rotations
guarantee the end result is a tree again after all.

Furthermore, we rely on the WMB implied by the
write_seqlock/count_begin() to separate the VMA initialization and the
publishing stores, analogous to the RELEASE in rcu_assign_pointer().
We also rely on the RMB from read_seqretry() to separate the vma load
from further loads like the smp_read_barrier_depends() in regular
RCU.

We must not touch the vmacache while doing SRCU lookups as that is not
properly serialized against changes. We update gap information after
publishing the VMA, but A) we don't use that and B) the seqlock
read side would fix that anyhow.

We clear vma->vm_rb for nodes removed from the vma tree such that we
can easily detect such 'dead' nodes, we rely on the WMB from
write_sequnlock() to separate the tree removal and clearing the node.

Provide find_vma_srcu() which wraps the required magic.

Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>

[Remove the warnings in description about the SRCU global lock which
 has been removed now]
[Rename vma_is_dead() to vma_has_changed() and move its adding to the next
 patch]
Signed-off-by: Laurent Dufour <ldufour@linux.vnet.ibm.com>
---
 include/linux/mm_types.h |   2 +
 kernel/fork.c            |   1 +
 mm/init-mm.c             |   1 +
 mm/internal.h            |   5 +++
 mm/mmap.c                | 100 +++++++++++++++++++++++++++++++++++------------
 5 files changed, 83 insertions(+), 26 deletions(-)

diff --git a/include/linux/mm_types.h b/include/linux/mm_types.h
index 642aad26b32f..f3851b250fde 100644
--- a/include/linux/mm_types.h
+++ b/include/linux/mm_types.h
@@ -343,6 +343,7 @@ struct vm_area_struct {
 #endif
 	struct vm_userfaultfd_ctx vm_userfaultfd_ctx;
 	seqcount_t vm_sequence;
+	struct rcu_head vm_rcu_head;
 } __randomize_layout;
 
 struct core_thread {
@@ -360,6 +361,7 @@ struct kioctx_table;
 struct mm_struct {
 	struct vm_area_struct *mmap;		/* list of VMAs */
 	struct rb_root mm_rb;
+	seqlock_t mm_seq;
 	u32 vmacache_seqnum;                   /* per-thread vmacache */
 #ifdef CONFIG_MMU
 	unsigned long (*get_unmapped_area) (struct file *filp,
diff --git a/kernel/fork.c b/kernel/fork.c
index e075b7780421..f28aa54c668c 100644
--- a/kernel/fork.c
+++ b/kernel/fork.c
@@ -791,6 +791,7 @@ static struct mm_struct *mm_init(struct mm_struct *mm, struct task_struct *p,
 	mm->mmap = NULL;
 	mm->mm_rb = RB_ROOT;
 	mm->vmacache_seqnum = 0;
+	seqlock_init(&mm->mm_seq);
 	atomic_set(&mm->mm_users, 1);
 	atomic_set(&mm->mm_count, 1);
 	init_rwsem(&mm->mmap_sem);
diff --git a/mm/init-mm.c b/mm/init-mm.c
index 975e49f00f34..2b1fa061684f 100644
--- a/mm/init-mm.c
+++ b/mm/init-mm.c
@@ -16,6 +16,7 @@
 
 struct mm_struct init_mm = {
 	.mm_rb		= RB_ROOT,
+	.mm_seq		= __SEQLOCK_UNLOCKED(init_mm.mm_seq),
 	.pgd		= swapper_pg_dir,
 	.mm_users	= ATOMIC_INIT(2),
 	.mm_count	= ATOMIC_INIT(1),
diff --git a/mm/internal.h b/mm/internal.h
index 4ef49fc55e58..736540f15936 100644
--- a/mm/internal.h
+++ b/mm/internal.h
@@ -40,6 +40,11 @@ void page_writeback_init(void);
 
 int do_swap_page(struct vm_fault *vmf);
 
+extern struct srcu_struct vma_srcu;
+
+extern struct vm_area_struct *find_vma_srcu(struct mm_struct *mm,
+					    unsigned long addr);
+
 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *start_vma,
 		unsigned long floor, unsigned long ceiling);
 
diff --git a/mm/mmap.c b/mm/mmap.c
index b480043e38fb..34a7f1bdffe4 100644
--- a/mm/mmap.c
+++ b/mm/mmap.c
@@ -159,6 +159,23 @@ void unlink_file_vma(struct vm_area_struct *vma)
 	}
 }
 
+DEFINE_SRCU(vma_srcu);
+
+static void __free_vma(struct rcu_head *head)
+{
+	struct vm_area_struct *vma =
+		container_of(head, struct vm_area_struct, vm_rcu_head);
+
+	if (vma->vm_file)
+		fput(vma->vm_file);
+	kmem_cache_free(vm_area_cachep, vma);
+}
+
+static void free_vma(struct vm_area_struct *vma)
+{
+	call_srcu(&vma_srcu, &vma->vm_rcu_head, __free_vma);
+}
+
 /*
  * Close a vm structure and free it, returning the next.
  */
@@ -169,10 +186,8 @@ static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
 	might_sleep();
 	if (vma->vm_ops && vma->vm_ops->close)
 		vma->vm_ops->close(vma);
-	if (vma->vm_file)
-		fput(vma->vm_file);
 	mpol_put(vma_policy(vma));
-	kmem_cache_free(vm_area_cachep, vma);
+	free_vma(vma);
 	return next;
 }
 
@@ -410,26 +425,37 @@ static void vma_gap_update(struct vm_area_struct *vma)
 }
 
 static inline void vma_rb_insert(struct vm_area_struct *vma,
-				 struct rb_root *root)
+				 struct mm_struct *mm)
 {
+	struct rb_root *root = &mm->mm_rb;
+
 	/* All rb_subtree_gap values must be consistent prior to insertion */
 	validate_mm_rb(root, NULL);
 
 	rb_insert_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
 }
 
-static void __vma_rb_erase(struct vm_area_struct *vma, struct rb_root *root)
+static void __vma_rb_erase(struct vm_area_struct *vma, struct mm_struct *mm)
 {
+	struct rb_root *root = &mm->mm_rb;
 	/*
 	 * Note rb_erase_augmented is a fairly large inline function,
 	 * so make sure we instantiate it only once with our desired
 	 * augmented rbtree callbacks.
 	 */
+	write_seqlock(&mm->mm_seq);
 	rb_erase_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
+	write_sequnlock(&mm->mm_seq); /* wmb */
+
+	/*
+	 * Ensure the removal is complete before clearing the node.
+	 * Matched by vma_has_changed()/handle_speculative_fault().
+	 */
+	RB_CLEAR_NODE(&vma->vm_rb);
 }
 
 static __always_inline void vma_rb_erase_ignore(struct vm_area_struct *vma,
-						struct rb_root *root,
+						struct mm_struct *mm,
 						struct vm_area_struct *ignore)
 {
 	/*
@@ -437,21 +463,21 @@ static __always_inline void vma_rb_erase_ignore(struct vm_area_struct *vma,
 	 * with the possible exception of the "next" vma being erased if
 	 * next->vm_start was reduced.
 	 */
-	validate_mm_rb(root, ignore);
+	validate_mm_rb(&mm->mm_rb, ignore);
 
-	__vma_rb_erase(vma, root);
+	__vma_rb_erase(vma, mm);
 }
 
 static __always_inline void vma_rb_erase(struct vm_area_struct *vma,
-					 struct rb_root *root)
+					 struct mm_struct *mm)
 {
 	/*
 	 * All rb_subtree_gap values must be consistent prior to erase,
 	 * with the possible exception of the vma being erased.
 	 */
-	validate_mm_rb(root, vma);
+	validate_mm_rb(&mm->mm_rb, vma);
 
-	__vma_rb_erase(vma, root);
+	__vma_rb_erase(vma, mm);
 }
 
 /*
@@ -568,10 +594,12 @@ void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
 	 * immediately update the gap to the correct value. Finally we
 	 * rebalance the rbtree after all augmented values have been set.
 	 */
+	write_seqlock(&mm->mm_seq);
 	rb_link_node(&vma->vm_rb, rb_parent, rb_link);
 	vma->rb_subtree_gap = 0;
 	vma_gap_update(vma);
-	vma_rb_insert(vma, &mm->mm_rb);
+	vma_rb_insert(vma, mm);
+	write_sequnlock(&mm->mm_seq);
 }
 
 static void __vma_link_file(struct vm_area_struct *vma)
@@ -647,7 +675,7 @@ static __always_inline void __vma_unlink_common(struct mm_struct *mm,
 {
 	struct vm_area_struct *next;
 
-	vma_rb_erase_ignore(vma, &mm->mm_rb, ignore);
+	vma_rb_erase_ignore(vma, mm, ignore);
 	next = vma->vm_next;
 	if (has_prev)
 		prev->vm_next = next;
@@ -901,15 +929,13 @@ int __vma_adjust(struct vm_area_struct *vma, unsigned long start,
 	}
 
 	if (remove_next) {
-		if (file) {
+		if (file)
 			uprobe_munmap(next, next->vm_start, next->vm_end);
-			fput(file);
-		}
 		if (next->anon_vma)
 			anon_vma_merge(vma, next);
 		mm->map_count--;
 		mpol_put(vma_policy(next));
-		kmem_cache_free(vm_area_cachep, next);
+		free_vma(next);
 		write_seqcount_end(&next->vm_sequence);
 		/*
 		 * In mprotect's case 6 (see comments on vma_merge),
@@ -2130,15 +2156,10 @@ get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
 EXPORT_SYMBOL(get_unmapped_area);
 
 /* Look up the first VMA which satisfies  addr < vm_end,  NULL if none. */
-struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
+static struct vm_area_struct *__find_vma(struct mm_struct *mm, unsigned long addr)
 {
 	struct rb_node *rb_node;
-	struct vm_area_struct *vma;
-
-	/* Check the cache first. */
-	vma = vmacache_find(mm, addr);
-	if (likely(vma))
-		return vma;
+	struct vm_area_struct *vma = NULL;
 
 	rb_node = mm->mm_rb.rb_node;
 
@@ -2156,13 +2177,40 @@ struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
 			rb_node = rb_node->rb_right;
 	}
 
+	return vma;
+}
+
+struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
+{
+	struct vm_area_struct *vma;
+
+	/* Check the cache first. */
+	vma = vmacache_find(mm, addr);
+	if (likely(vma))
+		return vma;
+
+	vma = __find_vma(mm, addr);
 	if (vma)
 		vmacache_update(addr, vma);
 	return vma;
 }
-
 EXPORT_SYMBOL(find_vma);
 
+struct vm_area_struct *find_vma_srcu(struct mm_struct *mm, unsigned long addr)
+{
+	struct vm_area_struct *vma;
+	unsigned int seq;
+
+	WARN_ON_ONCE(!srcu_read_lock_held(&vma_srcu));
+
+	do {
+		seq = read_seqbegin(&mm->mm_seq);
+		vma = __find_vma(mm, addr);
+	} while (read_seqretry(&mm->mm_seq, seq));
+
+	return vma;
+}
+
 /*
  * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
  */
@@ -2530,7 +2578,7 @@ detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
 	insertion_point = (prev ? &prev->vm_next : &mm->mmap);
 	vma->vm_prev = NULL;
 	do {
-		vma_rb_erase(vma, &mm->mm_rb);
+		vma_rb_erase(vma, mm);
 		mm->map_count--;
 		tail_vma = vma;
 		vma = vma->vm_next;
-- 
2.7.4

[toc] | [prev] | [next] | [standalone]

#1714462 — [PATCH v2 08/20] mm: Protect SPF handler against anon_vma changes

The speculative page fault handler must be protected against anon_vma
changes. This is because page_add_new_anon_rmap() is called during the
speculative path.

In addition, don't try speculative page fault if the VMA don't have an
anon_vma structure allocated because its allocation should be
protected by the mmap_sem.

In __vma_adjust() when importer->anon_vma is set, there is no need to
protect against speculative page faults since speculative page fault
is aborted if the vma->anon_vma is not set.

When calling page_add_new_anon_rmap() vma->anon_vma is necessarily
valid since we checked for it when locking the pte and the anon_vma is
removed once the pte is unlocked. So even if the speculative page
fault handler is running concurrently with do_unmap(), as the pte is
locked in unmap_region() - through unmap_vmas() - and the anon_vma
unlinked later, because we check for the vma sequence counter which is
updated in unmap_page_range() before locking the pte, and then in
free_pgtables() so when locking the pte the change will be detected.

Signed-off-by: Laurent Dufour <ldufour@linux.vnet.ibm.com>
---
 mm/memory.c | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/mm/memory.c b/mm/memory.c
index da3bd07bb052..68e4fdcce692 100644
--- a/mm/memory.c
+++ b/mm/memory.c
@@ -615,7 +615,9 @@ void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
 		 * Hide vma from rmap and truncate_pagecache before freeing
 		 * pgtables
 		 */
+		write_seqcount_begin(&vma->vm_sequence);
 		unlink_anon_vmas(vma);
+		write_seqcount_end(&vma->vm_sequence);
 		unlink_file_vma(vma);
 
 		if (is_vm_hugetlb_page(vma)) {
@@ -629,7 +631,9 @@ void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
 			       && !is_vm_hugetlb_page(next)) {
 				vma = next;
 				next = vma->vm_next;
+				write_seqcount_begin(&vma->vm_sequence);
 				unlink_anon_vmas(vma);
+				write_seqcount_end(&vma->vm_sequence);
 				unlink_file_vma(vma);
 			}
 			free_pgd_range(tlb, addr, vma->vm_end,
-- 
2.7.4

[toc] | [prev] | [next] | [standalone]

#1716071

Hello,

On (08/18/17 00:04), Laurent Dufour wrote:
> This is a port on kernel 4.13 of the work done by Peter Zijlstra to
> handle page fault without holding the mm semaphore [1].
> 
> The idea is to try to handle user space page faults without holding the
> mmap_sem. This should allow better concurrency for massively threaded
> process since the page fault handler will not wait for other threads memory
> layout change to be done, assuming that this change is done in another part
> of the process's memory space. This type page fault is named speculative
> page fault. If the speculative page fault fails because of a concurrency is
> detected or because underlying PMD or PTE tables are not yet allocating, it
> is failing its processing and a classic page fault is then tried.
> 
> The speculative page fault (SPF) has to look for the VMA matching the fault
> address without holding the mmap_sem, so the VMA list is now managed using
> SRCU allowing lockless walking. The only impact would be the deferred file
> derefencing in the case of a file mapping, since the file pointer is
> released once the SRCU cleaning is done.  This patch relies on the change
> done recently by Paul McKenney in SRCU which now runs a callback per CPU
> instead of per SRCU structure [1].
> 
> The VMA's attributes checked during the speculative page fault processing
> have to be protected against parallel changes. This is done by using a per
> VMA sequence lock. This sequence lock allows the speculative page fault
> handler to fast check for parallel changes in progress and to abort the
> speculative page fault in that case.
> 
> Once the VMA is found, the speculative page fault handler would check for
> the VMA's attributes to verify that the page fault has to be handled
> correctly or not. Thus the VMA is protected through a sequence lock which
> allows fast detection of concurrent VMA changes. If such a change is
> detected, the speculative page fault is aborted and a *classic* page fault
> is tried.  VMA sequence locks are added when VMA attributes which are
> checked during the page fault are modified.
> 
> When the PTE is fetched, the VMA is checked to see if it has been changed,
> so once the page table is locked, the VMA is valid, so any other changes
> leading to touching this PTE will need to lock the page table, so no
> parallel change is possible at this time.

[ 2311.315400] ======================================================
[ 2311.315401] WARNING: possible circular locking dependency detected
[ 2311.315403] 4.13.0-rc5-next-20170817-dbg-00039-gaf11d7500492-dirty #1743 Not tainted
[ 2311.315404] ------------------------------------------------------
[ 2311.315406] khugepaged/43 is trying to acquire lock:
[ 2311.315407]  (&mapping->i_mmap_rwsem){++++}, at: [<ffffffff8111b339>] rmap_walk_file+0x5a/0x147
[ 2311.315415] 
               but task is already holding lock:
[ 2311.315416]  (fs_reclaim){+.+.}, at: [<ffffffff810ebd80>] fs_reclaim_acquire+0x12/0x35
[ 2311.315420] 
               which lock already depends on the new lock.

[ 2311.315422] 
               the existing dependency chain (in reverse order) is:
[ 2311.315423] 
               -> #3 (fs_reclaim){+.+.}:
[ 2311.315427]        fs_reclaim_acquire+0x32/0x35
[ 2311.315429]        __alloc_pages_nodemask+0x8d/0x217
[ 2311.315432]        pte_alloc_one+0x13/0x5e
[ 2311.315434]        __pte_alloc+0x1f/0x83
[ 2311.315436]        move_page_tables+0x2c9/0x5ac
[ 2311.315438]        move_vma.isra.25+0xff/0x2a2
[ 2311.315439]        SyS_mremap+0x41b/0x49e
[ 2311.315442]        entry_SYSCALL_64_fastpath+0x18/0xad
[ 2311.315443] 
               -> #2 (&vma->vm_sequence/1){+.+.}:
[ 2311.315449]        write_seqcount_begin_nested+0x1b/0x1d
[ 2311.315451]        __vma_adjust+0x1b7/0x5d6
[ 2311.315453]        __split_vma+0x142/0x1a3
[ 2311.315454]        do_munmap+0x128/0x2af
[ 2311.315455]        vm_munmap+0x5a/0x73
[ 2311.315458]        elf_map+0xb1/0xce
[ 2311.315459]        load_elf_binary+0x8e0/0x1348
[ 2311.315462]        search_binary_handler+0x70/0x1f3
[ 2311.315464]        load_script+0x1a6/0x1b5
[ 2311.315466]        search_binary_handler+0x70/0x1f3
[ 2311.315468]        do_execveat_common+0x461/0x691
[ 2311.315471]        kernel_init+0x5a/0xf0
[ 2311.315472]        ret_from_fork+0x27/0x40
[ 2311.315473] 
               -> #1 (&vma->vm_sequence){+.+.}:
[ 2311.315478]        write_seqcount_begin_nested+0x1b/0x1d
[ 2311.315480]        __vma_adjust+0x19c/0x5d6
[ 2311.315481]        __split_vma+0x142/0x1a3
[ 2311.315482]        do_munmap+0x128/0x2af
[ 2311.315484]        vm_munmap+0x5a/0x73
[ 2311.315485]        elf_map+0xb1/0xce
[ 2311.315487]        load_elf_binary+0x8e0/0x1348
[ 2311.315489]        search_binary_handler+0x70/0x1f3
[ 2311.315490]        load_script+0x1a6/0x1b5
[ 2311.315492]        search_binary_handler+0x70/0x1f3
[ 2311.315494]        do_execveat_common+0x461/0x691
[ 2311.315496]        kernel_init+0x5a/0xf0
[ 2311.315497]        ret_from_fork+0x27/0x40
[ 2311.315498] 
               -> #0 (&mapping->i_mmap_rwsem){++++}:
[ 2311.315503]        lock_acquire+0x176/0x19e
[ 2311.315505]        down_read+0x3b/0x55
[ 2311.315507]        rmap_walk_file+0x5a/0x147
[ 2311.315508]        page_referenced+0x11c/0x134
[ 2311.315511]        shrink_page_list+0x36b/0xb80
[ 2311.315512]        shrink_inactive_list+0x1d9/0x437
[ 2311.315514]        shrink_node_memcg.constprop.71+0x3e7/0x571
[ 2311.315515]        shrink_node+0x3f/0x149
[ 2311.315517]        try_to_free_pages+0x270/0x45f
[ 2311.315518]        __alloc_pages_slowpath+0x34a/0xaa2
[ 2311.315520]        __alloc_pages_nodemask+0x111/0x217
[ 2311.315523]        khugepaged_alloc_page+0x17/0x45
[ 2311.315524]        khugepaged+0xa29/0x16b5
[ 2311.315527]        kthread+0xfb/0x103
[ 2311.315529]        ret_from_fork+0x27/0x40
[ 2311.315530] 
               other info that might help us debug this:

[ 2311.315531] Chain exists of:
                 &mapping->i_mmap_rwsem --> &vma->vm_sequence/1 --> fs_reclaim

[ 2311.315537]  Possible unsafe locking scenario:

[ 2311.315538]        CPU0                    CPU1
[ 2311.315539]        ----                    ----
[ 2311.315540]   lock(fs_reclaim);
[ 2311.315542]                                lock(&vma->vm_sequence/1);
[ 2311.315545]                                lock(fs_reclaim);
[ 2311.315547]   lock(&mapping->i_mmap_rwsem);
[ 2311.315549] 
                *** DEADLOCK ***

[ 2311.315551] 1 lock held by khugepaged/43:
[ 2311.315552]  #0:  (fs_reclaim){+.+.}, at: [<ffffffff810ebd80>] fs_reclaim_acquire+0x12/0x35
[ 2311.315556] 
               stack backtrace:
[ 2311.315559] CPU: 0 PID: 43 Comm: khugepaged Not tainted 4.13.0-rc5-next-20170817-dbg-00039-gaf11d7500492-dirty #1743
[ 2311.315560] Call Trace:
[ 2311.315564]  dump_stack+0x67/0x8e
[ 2311.315568]  print_circular_bug.isra.39+0x1c7/0x1d4
[ 2311.315570]  __lock_acquire+0xb1a/0xe06
[ 2311.315572]  ? graph_unlock+0x69/0x69
[ 2311.315575]  lock_acquire+0x176/0x19e
[ 2311.315577]  ? rmap_walk_file+0x5a/0x147
[ 2311.315579]  down_read+0x3b/0x55
[ 2311.315581]  ? rmap_walk_file+0x5a/0x147
[ 2311.315583]  rmap_walk_file+0x5a/0x147
[ 2311.315585]  page_referenced+0x11c/0x134
[ 2311.315587]  ? page_vma_mapped_walk_done.isra.15+0xb/0xb
[ 2311.315589]  ? page_get_anon_vma+0x6d/0x6d
[ 2311.315591]  shrink_page_list+0x36b/0xb80
[ 2311.315593]  ? _raw_spin_unlock_irq+0x29/0x46
[ 2311.315595]  shrink_inactive_list+0x1d9/0x437
[ 2311.315597]  shrink_node_memcg.constprop.71+0x3e7/0x571
[ 2311.315600]  shrink_node+0x3f/0x149
[ 2311.315602]  try_to_free_pages+0x270/0x45f
[ 2311.315604]  __alloc_pages_slowpath+0x34a/0xaa2
[ 2311.315608]  ? ___might_sleep+0xd5/0x234
[ 2311.315609]  __alloc_pages_nodemask+0x111/0x217
[ 2311.315612]  khugepaged_alloc_page+0x17/0x45
[ 2311.315613]  khugepaged+0xa29/0x16b5
[ 2311.315616]  ? remove_wait_queue+0x47/0x47
[ 2311.315618]  ? collapse_shmem.isra.43+0x882/0x882
[ 2311.315620]  kthread+0xfb/0x103
[ 2311.315622]  ? __list_del_entry+0x1d/0x1d
[ 2311.315624]  ret_from_fork+0x27/0x40

	-ss

[toc] | [prev] | [next] | [standalone]

#1716110

On 08/18/2017 03:34 AM, Laurent Dufour wrote:
> This is a port on kernel 4.13 of the work done by Peter Zijlstra to
> handle page fault without holding the mm semaphore [1].
> 
> The idea is to try to handle user space page faults without holding the
> mmap_sem. This should allow better concurrency for massively threaded
> process since the page fault handler will not wait for other threads memory
> layout change to be done, assuming that this change is done in another part
> of the process's memory space. This type page fault is named speculative
> page fault. If the speculative page fault fails because of a concurrency is
> detected or because underlying PMD or PTE tables are not yet allocating, it
> is failing its processing and a classic page fault is then tried.
> 
> The speculative page fault (SPF) has to look for the VMA matching the fault
> address without holding the mmap_sem, so the VMA list is now managed using
> SRCU allowing lockless walking. The only impact would be the deferred file
> derefencing in the case of a file mapping, since the file pointer is
> released once the SRCU cleaning is done.  This patch relies on the change
> done recently by Paul McKenney in SRCU which now runs a callback per CPU
> instead of per SRCU structure [1].
> 
> The VMA's attributes checked during the speculative page fault processing
> have to be protected against parallel changes. This is done by using a per
> VMA sequence lock. This sequence lock allows the speculative page fault
> handler to fast check for parallel changes in progress and to abort the
> speculative page fault in that case.
> 
> Once the VMA is found, the speculative page fault handler would check for
> the VMA's attributes to verify that the page fault has to be handled
> correctly or not. Thus the VMA is protected through a sequence lock which
> allows fast detection of concurrent VMA changes. If such a change is
> detected, the speculative page fault is aborted and a *classic* page fault
> is tried.  VMA sequence locks are added when VMA attributes which are
> checked during the page fault are modified.
> 
> When the PTE is fetched, the VMA is checked to see if it has been changed,
> so once the page table is locked, the VMA is valid, so any other changes
> leading to touching this PTE will need to lock the page table, so no
> parallel change is possible at this time.
> 
> Compared to the Peter's initial work, this series introduces a spin_trylock
> when dealing with speculative page fault. This is required to avoid dead
> lock when handling a page fault while a TLB invalidate is requested by an
> other CPU holding the PTE. Another change due to a lock dependency issue
> with mapping->i_mmap_rwsem.
> 
> In addition some VMA field values which are used once the PTE is unlocked
> at the end the page fault path are saved into the vm_fault structure to
> used the values matching the VMA at the time the PTE was locked.
> 
> This series builds on top of v4.13-rc5 and is functional on x86 and
> PowerPC.
> 
> Tests have been made using a large commercial in-memory database on a
> PowerPC system with 752 CPU using RFC v5. The results are very encouraging
> since the loading of the 2TB database was faster by 14% with the
> speculative page fault.
> 

You specifically mention loading as most of the page faults will
happen at that time and then the working set will settle down with
very less page faults there after ? That means unless there is
another wave of page faults we wont notice performance improvement
during the runtime.

> Using ebizzy test [3], which spreads a lot of threads, the result are good
> when running on both a large or a small system. When using kernbench, the

The performance improvements are greater as there is a lot of creation
and destruction of anon mappings which generates constant flow of page
faults to be handled.

> result are quite similar which expected as not so much multi threaded
> processes are involved. But there is no performance degradation neither
> which is good.

If we compile with 'make -j N' there would be a lot of threads but I
guess the problem is SPF does not support handling file mapping IIUC
which limits the performance improvement for some workloads.

> 
> ------------------
> Benchmarks results
> 
> Note these test have been made on top of 4.13-rc3 with the following patch
> from Paul McKenney applied: 
>  "srcu: Provide ordering for CPU not involved in grace period" [5]

Is this patch an improvement for SRCU which we are using for walking VMAs.

> 
> Ebizzy:
> -------
> The test is counting the number of records per second it can manage, the
> higher is the best. I run it like this 'ebizzy -mTRp'. To get consistent
> result I repeated the test 100 times and measure the average result, mean
> deviation, max and min.
> 
> - 16 CPUs x86 VM
> Records/s	4.13-rc5	4.13-rc5-spf
> Average		11350.29	21760.36
> Mean deviation	396.56		881.40
> Max		13773		26194
> Min		10567		19223
> 
> - 80 CPUs Power 8 node:
> Records/s	4.13-rc5	4.13-rc5-spf
> Average		33904.67	58847.91
> Mean deviation	789.40		1753.19
> Max		36703		68958
> Min		31759		55125
> 

Can you also mention % improvement or degradation in a new column.

> The number of record per second is far better with the speculative page
> fault.
> The mean deviation is higher with the speculative page fault, may be
> because sometime the fault are not handled in a speculative way leading to
> more variation.

we need to analyze that. Why speculative page faults failed on those
occasions for exact same workload.

> 
> 
> Kernbench:
> ----------
> This test is building a 4.12 kernel using platform default config. The
> build has been run 5 times each time.
> 
> - 16 CPUs x86 VM
> Average Half load -j 8 Run (std deviation)
>  		 4.13.0-rc5		4.13.0-rc5-spf
> Elapsed Time     166.574 (0.340779)	145.754 (0.776325)		
> User Time        1080.77 (2.05871)	999.272 (4.12142)		
> System Time      204.594 (1.02449)	116.362 (1.22974)		
> Percent CPU 	 771.2 (1.30384)	765 (0.707107)
> Context Switches 46590.6 (935.591)	66316.4 (744.64)
> Sleeps           84421.2 (596.612)	85186 (523.041)		


> 
> Average Optimal load -j 16 Run (std deviation)
>  		 4.13.0-rc5		4.13.0-rc5-spf
> Elapsed Time     85.422 (0.42293)	74.81 (0.419345)
> User Time        1031.79 (51.6557)	954.912 (46.8439)
> System Time      186.528 (19.0575)	107.514 (9.36902)
> Percent CPU 	 1059.2 (303.607)	1056.8 (307.624)
> Context Switches 67240.3 (21788.9)	89360.6 (24299.9)
> Sleeps           89607.8 (5511.22)	90372.5 (5490.16)
> 
> The elapsed time is a bit shorter in the case of the SPF release, but the
> impact less important since there are less multithreaded processes involved
> here. 
> 
> - 80 CPUs Power 8 node:
> Average Half load -j 40 Run (std deviation)
>  		 4.13.0-rc5		4.13.0-rc5-spf
> Elapsed Time     117.176 (0.824093)	116.792 (0.695392)
> User Time        4412.34 (24.29)	4396.02 (24.4819)
> System Time      131.106 (1.28343)	133.452 (0.708851)
> Percent CPU      3876.8 (18.1439)	3877.6 (21.9955)
> Context Switches 72470.2 (466.181)	72971 (673.624)
> Sleeps           161294 (2284.85)	161946 (2217.9)
> 
> Average Optimal load -j 80 Run (std deviation)
>  		 4.13.0-rc5		4.13.0-rc5-spf
> Elapsed Time     111.176 (1.11123)	111.242 (0.801542)
> User Time        5930.03 (1600.07)	5929.89 (1617)
> System Time      166.258 (37.0662)	169.337 (37.8419)
> Percent CPU      5378.5 (1584.16)	5385.6 (1590.24)
> Context Switches 117389 (47350.1)	130132 (60256.3)
> Sleeps           163354 (4153.9)	163219 (2251.27)
> 

Can you also mention % improvement or degradation in a new column.

> Here the elapsed time is a bit shorter using the spf release, but we
> remain in the error margin. It has to be noted that this system is not
> correctly balanced on the NUMA point of view as all the available memory is
> attached to one core.

Why different NUMA configuration would have changed the outcome ?

> 
> ------------------------
> Changes since v1:
>  - Remove PERF_COUNT_SW_SPF_FAILED perf event.
>  - Add tracing events to details speculative page fault failures.
>  - Cache VMA fields values which are used once the PTE is unlocked at the
>  end of the page fault events.

Why is this required ?

[toc] | [prev] | [next] | [standalone]

#1716950

On Mon, Aug 21, 2017 at 11:58:03AM +0530, Anshuman Khandual wrote:
> On 08/18/2017 03:34 AM, Laurent Dufour wrote:
> > This is a port on kernel 4.13 of the work done by Peter Zijlstra to
> > handle page fault without holding the mm semaphore [1].
> > 
> > The idea is to try to handle user space page faults without holding the
> > mmap_sem. This should allow better concurrency for massively threaded
> > process since the page fault handler will not wait for other threads memory
> > layout change to be done, assuming that this change is done in another part
> > of the process's memory space. This type page fault is named speculative
> > page fault. If the speculative page fault fails because of a concurrency is
> > detected or because underlying PMD or PTE tables are not yet allocating, it
> > is failing its processing and a classic page fault is then tried.
> > 
> > The speculative page fault (SPF) has to look for the VMA matching the fault
> > address without holding the mmap_sem, so the VMA list is now managed using
> > SRCU allowing lockless walking. The only impact would be the deferred file
> > derefencing in the case of a file mapping, since the file pointer is
> > released once the SRCU cleaning is done.  This patch relies on the change
> > done recently by Paul McKenney in SRCU which now runs a callback per CPU
> > instead of per SRCU structure [1].
> > 
> > The VMA's attributes checked during the speculative page fault processing
> > have to be protected against parallel changes. This is done by using a per
> > VMA sequence lock. This sequence lock allows the speculative page fault
> > handler to fast check for parallel changes in progress and to abort the
> > speculative page fault in that case.
> > 
> > Once the VMA is found, the speculative page fault handler would check for
> > the VMA's attributes to verify that the page fault has to be handled
> > correctly or not. Thus the VMA is protected through a sequence lock which
> > allows fast detection of concurrent VMA changes. If such a change is
> > detected, the speculative page fault is aborted and a *classic* page fault
> > is tried.  VMA sequence locks are added when VMA attributes which are
> > checked during the page fault are modified.
> > 
> > When the PTE is fetched, the VMA is checked to see if it has been changed,
> > so once the page table is locked, the VMA is valid, so any other changes
> > leading to touching this PTE will need to lock the page table, so no
> > parallel change is possible at this time.
> > 
> > Compared to the Peter's initial work, this series introduces a spin_trylock
> > when dealing with speculative page fault. This is required to avoid dead
> > lock when handling a page fault while a TLB invalidate is requested by an
> > other CPU holding the PTE. Another change due to a lock dependency issue
> > with mapping->i_mmap_rwsem.
> > 
> > In addition some VMA field values which are used once the PTE is unlocked
> > at the end the page fault path are saved into the vm_fault structure to
> > used the values matching the VMA at the time the PTE was locked.
> > 
> > This series builds on top of v4.13-rc5 and is functional on x86 and
> > PowerPC.
> > 
> > Tests have been made using a large commercial in-memory database on a
> > PowerPC system with 752 CPU using RFC v5. The results are very encouraging
> > since the loading of the 2TB database was faster by 14% with the
> > speculative page fault.
> > 
> 
> You specifically mention loading as most of the page faults will
> happen at that time and then the working set will settle down with
> very less page faults there after ? That means unless there is
> another wave of page faults we wont notice performance improvement
> during the runtime.
> 
> > Using ebizzy test [3], which spreads a lot of threads, the result are good
> > when running on both a large or a small system. When using kernbench, the
> 
> The performance improvements are greater as there is a lot of creation
> and destruction of anon mappings which generates constant flow of page
> faults to be handled.
> 
> > result are quite similar which expected as not so much multi threaded
> > processes are involved. But there is no performance degradation neither
> > which is good.
> 
> If we compile with 'make -j N' there would be a lot of threads but I
> guess the problem is SPF does not support handling file mapping IIUC
> which limits the performance improvement for some workloads.
> 
> > 
> > ------------------
> > Benchmarks results
> > 
> > Note these test have been made on top of 4.13-rc3 with the following patch
> > from Paul McKenney applied: 
> >  "srcu: Provide ordering for CPU not involved in grace period" [5]
> 
> Is this patch an improvement for SRCU which we are using for walking VMAs.

It is a tweak to an earlier patch that parallelizes SRCU callback
handling.

							Thanx, Paul

> > Ebizzy:
> > -------
> > The test is counting the number of records per second it can manage, the
> > higher is the best. I run it like this 'ebizzy -mTRp'. To get consistent
> > result I repeated the test 100 times and measure the average result, mean
> > deviation, max and min.
> > 
> > - 16 CPUs x86 VM
> > Records/s	4.13-rc5	4.13-rc5-spf
> > Average		11350.29	21760.36
> > Mean deviation	396.56		881.40
> > Max		13773		26194
> > Min		10567		19223
> > 
> > - 80 CPUs Power 8 node:
> > Records/s	4.13-rc5	4.13-rc5-spf
> > Average		33904.67	58847.91
> > Mean deviation	789.40		1753.19
> > Max		36703		68958
> > Min		31759		55125
> > 
> 
> Can you also mention % improvement or degradation in a new column.
> 
> > The number of record per second is far better with the speculative page
> > fault.
> > The mean deviation is higher with the speculative page fault, may be
> > because sometime the fault are not handled in a speculative way leading to
> > more variation.
> 
> we need to analyze that. Why speculative page faults failed on those
> occasions for exact same workload.
> 
> > 
> > 
> > Kernbench:
> > ----------
> > This test is building a 4.12 kernel using platform default config. The
> > build has been run 5 times each time.
> > 
> > - 16 CPUs x86 VM
> > Average Half load -j 8 Run (std deviation)
> >  		 4.13.0-rc5		4.13.0-rc5-spf
> > Elapsed Time     166.574 (0.340779)	145.754 (0.776325)		
> > User Time        1080.77 (2.05871)	999.272 (4.12142)		
> > System Time      204.594 (1.02449)	116.362 (1.22974)		
> > Percent CPU 	 771.2 (1.30384)	765 (0.707107)
> > Context Switches 46590.6 (935.591)	66316.4 (744.64)
> > Sleeps           84421.2 (596.612)	85186 (523.041)		
> 
> 
> > 
> > Average Optimal load -j 16 Run (std deviation)
> >  		 4.13.0-rc5		4.13.0-rc5-spf
> > Elapsed Time     85.422 (0.42293)	74.81 (0.419345)
> > User Time        1031.79 (51.6557)	954.912 (46.8439)
> > System Time      186.528 (19.0575)	107.514 (9.36902)
> > Percent CPU 	 1059.2 (303.607)	1056.8 (307.624)
> > Context Switches 67240.3 (21788.9)	89360.6 (24299.9)
> > Sleeps           89607.8 (5511.22)	90372.5 (5490.16)
> > 
> > The elapsed time is a bit shorter in the case of the SPF release, but the
> > impact less important since there are less multithreaded processes involved
> > here. 
> > 
> > - 80 CPUs Power 8 node:
> > Average Half load -j 40 Run (std deviation)
> >  		 4.13.0-rc5		4.13.0-rc5-spf
> > Elapsed Time     117.176 (0.824093)	116.792 (0.695392)
> > User Time        4412.34 (24.29)	4396.02 (24.4819)
> > System Time      131.106 (1.28343)	133.452 (0.708851)
> > Percent CPU      3876.8 (18.1439)	3877.6 (21.9955)
> > Context Switches 72470.2 (466.181)	72971 (673.624)
> > Sleeps           161294 (2284.85)	161946 (2217.9)
> > 
> > Average Optimal load -j 80 Run (std deviation)
> >  		 4.13.0-rc5		4.13.0-rc5-spf
> > Elapsed Time     111.176 (1.11123)	111.242 (0.801542)
> > User Time        5930.03 (1600.07)	5929.89 (1617)
> > System Time      166.258 (37.0662)	169.337 (37.8419)
> > Percent CPU      5378.5 (1584.16)	5385.6 (1590.24)
> > Context Switches 117389 (47350.1)	130132 (60256.3)
> > Sleeps           163354 (4153.9)	163219 (2251.27)
> > 
> 
> Can you also mention % improvement or degradation in a new column.
> 
> > Here the elapsed time is a bit shorter using the spf release, but we
> > remain in the error margin. It has to be noted that this system is not
> > correctly balanced on the NUMA point of view as all the available memory is
> > attached to one core.
> 
> Why different NUMA configuration would have changed the outcome ?
> 
> > 
> > ------------------------
> > Changes since v1:
> >  - Remove PERF_COUNT_SW_SPF_FAILED perf event.
> >  - Add tracing events to details speculative page fault failures.
> >  - Cache VMA fields values which are used once the PTE is unlocked at the
> >  end of the page fault events.
> 
> Why is this required ?

[toc] | [prev] | [next] | [standalone]

#1719887

On 21/08/2017 08:28, Anshuman Khandual wrote:
> On 08/18/2017 03:34 AM, Laurent Dufour wrote:
>> This is a port on kernel 4.13 of the work done by Peter Zijlstra to
>> handle page fault without holding the mm semaphore [1].
>>
>> The idea is to try to handle user space page faults without holding the
>> mmap_sem. This should allow better concurrency for massively threaded
>> process since the page fault handler will not wait for other threads memory
>> layout change to be done, assuming that this change is done in another part
>> of the process's memory space. This type page fault is named speculative
>> page fault. If the speculative page fault fails because of a concurrency is
>> detected or because underlying PMD or PTE tables are not yet allocating, it
>> is failing its processing and a classic page fault is then tried.
>>
>> The speculative page fault (SPF) has to look for the VMA matching the fault
>> address without holding the mmap_sem, so the VMA list is now managed using
>> SRCU allowing lockless walking. The only impact would be the deferred file
>> derefencing in the case of a file mapping, since the file pointer is
>> released once the SRCU cleaning is done.  This patch relies on the change
>> done recently by Paul McKenney in SRCU which now runs a callback per CPU
>> instead of per SRCU structure [1].
>>
>> The VMA's attributes checked during the speculative page fault processing
>> have to be protected against parallel changes. This is done by using a per
>> VMA sequence lock. This sequence lock allows the speculative page fault
>> handler to fast check for parallel changes in progress and to abort the
>> speculative page fault in that case.
>>
>> Once the VMA is found, the speculative page fault handler would check for
>> the VMA's attributes to verify that the page fault has to be handled
>> correctly or not. Thus the VMA is protected through a sequence lock which
>> allows fast detection of concurrent VMA changes. If such a change is
>> detected, the speculative page fault is aborted and a *classic* page fault
>> is tried.  VMA sequence locks are added when VMA attributes which are
>> checked during the page fault are modified.
>>
>> When the PTE is fetched, the VMA is checked to see if it has been changed,
>> so once the page table is locked, the VMA is valid, so any other changes
>> leading to touching this PTE will need to lock the page table, so no
>> parallel change is possible at this time.
>>
>> Compared to the Peter's initial work, this series introduces a spin_trylock
>> when dealing with speculative page fault. This is required to avoid dead
>> lock when handling a page fault while a TLB invalidate is requested by an
>> other CPU holding the PTE. Another change due to a lock dependency issue
>> with mapping->i_mmap_rwsem.
>>
>> In addition some VMA field values which are used once the PTE is unlocked
>> at the end the page fault path are saved into the vm_fault structure to
>> used the values matching the VMA at the time the PTE was locked.
>>
>> This series builds on top of v4.13-rc5 and is functional on x86 and
>> PowerPC.
>>
>> Tests have been made using a large commercial in-memory database on a
>> PowerPC system with 752 CPU using RFC v5. The results are very encouraging
>> since the loading of the 2TB database was faster by 14% with the
>> speculative page fault.
>>
> 
> You specifically mention loading as most of the page faults will
> happen at that time and then the working set will settle down with
> very less page faults there after ? That means unless there is
> another wave of page faults we wont notice performance improvement
> during the runtime.

I just captured performance statistic during the database loading then
since the database was not stimulated, there was no page faults generated.
Further tests will be made while the database is running but I didn't have
the framework to do so right now.

> 
>> Using ebizzy test [3], which spreads a lot of threads, the result are good
>> when running on both a large or a small system. When using kernbench, the
> 
> The performance improvements are greater as there is a lot of creation
> and destruction of anon mappings which generates constant flow of page
> faults to be handled.
> 
>> result are quite similar which expected as not so much multi threaded
>> processes are involved. But there is no performance degradation neither
>> which is good.
> 
> If we compile with 'make -j N' there would be a lot of threads but I
> guess the problem is SPF does not support handling file mapping IIUC
> which limits the performance improvement for some workloads.

Yes but that test is showing that there is no performance degradation which
is good.

>>
>> ------------------
>> Benchmarks results
>>
>> Note these test have been made on top of 4.13-rc3 with the following patch
>> from Paul McKenney applied: 
>>  "srcu: Provide ordering for CPU not involved in grace period" [5]
> 
> Is this patch an improvement for SRCU which we are using for walking VMAs.
> 
>>
>> Ebizzy:
>> -------
>> The test is counting the number of records per second it can manage, the
>> higher is the best. I run it like this 'ebizzy -mTRp'. To get consistent
>> result I repeated the test 100 times and measure the average result, mean
>> deviation, max and min.
>>
>> - 16 CPUs x86 VM
>> Records/s	4.13-rc5	4.13-rc5-spf
>> Average		11350.29	21760.36
>> Mean deviation	396.56		881.40
>> Max		13773		26194
>> Min		10567		19223
>>
>> - 80 CPUs Power 8 node:
>> Records/s	4.13-rc5	4.13-rc5-spf
>> Average		33904.67	58847.91
>> Mean deviation	789.40		1753.19
>> Max		36703		68958
>> Min		31759		55125
>>
> 
> Can you also mention % improvement or degradation in a new column.

Fair enough:

- 16 CPUs x86 VM
Records/s	4.13-rc5	4.13-rc5-spf
Average		11350.29	21760.36	+92%
Mean deviation	396.56		881.40		+122%
Max		13773		26194		+90%
Min		10567		19223		+82%

- 80 CPUs Power 8 node:
Records/s	4.13-rc5	4.13-rc5-spf
Average		33904.67	58847.91	+74%
Mean deviation	789.40		1753.19		+122%
Max		36703		68958		+88%
Min		31759		55125		+74%


> 
>> The number of record per second is far better with the speculative page
>> fault.
>> The mean deviation is higher with the speculative page fault, may be
>> because sometime the fault are not handled in a speculative way leading to
>> more variation.
> 
> we need to analyze that. Why speculative page faults failed on those
> occasions for exact same workload.

That's not even clear that the mean deviation increasing is due to
speculative page fault failure. This will need to be study, but even if the
mean deviation is more important, the result are far better anyway.

>>
>>
>> Kernbench:
>> ----------
>> This test is building a 4.12 kernel using platform default config. The
>> build has been run 5 times each time.
>>
>> - 16 CPUs x86 VM
>> Average Half load -j 8 Run (std deviation)
>>  		 4.13.0-rc5		4.13.0-rc5-spf
>> Elapsed Time     166.574 (0.340779)	145.754 (0.776325)		
>> User Time        1080.77 (2.05871)	999.272 (4.12142)		
>> System Time      204.594 (1.02449)	116.362 (1.22974)		
>> Percent CPU 	 771.2 (1.30384)	765 (0.707107)
>> Context Switches 46590.6 (935.591)	66316.4 (744.64)
>> Sleeps           84421.2 (596.612)	85186 (523.041)		
> 
> 
>>
>> Average Optimal load -j 16 Run (std deviation)
>>  		 4.13.0-rc5		4.13.0-rc5-spf
>> Elapsed Time     85.422 (0.42293)	74.81 (0.419345)
>> User Time        1031.79 (51.6557)	954.912 (46.8439)
>> System Time      186.528 (19.0575)	107.514 (9.36902)
>> Percent CPU 	 1059.2 (303.607)	1056.8 (307.624)
>> Context Switches 67240.3 (21788.9)	89360.6 (24299.9)
>> Sleeps           89607.8 (5511.22)	90372.5 (5490.16)
>>
>> The elapsed time is a bit shorter in the case of the SPF release, but the
>> impact less important since there are less multithreaded processes involved
>> here. 
>>
>> - 80 CPUs Power 8 node:
>> Average Half load -j 40 Run (std deviation)
>>  		 4.13.0-rc5		4.13.0-rc5-spf
>> Elapsed Time     117.176 (0.824093)	116.792 (0.695392)
>> User Time        4412.34 (24.29)	4396.02 (24.4819)
>> System Time      131.106 (1.28343)	133.452 (0.708851)
>> Percent CPU      3876.8 (18.1439)	3877.6 (21.9955)
>> Context Switches 72470.2 (466.181)	72971 (673.624)
>> Sleeps           161294 (2284.85)	161946 (2217.9)
>>
>> Average Optimal load -j 80 Run (std deviation)
>>  		 4.13.0-rc5		4.13.0-rc5-spf
>> Elapsed Time     111.176 (1.11123)	111.242 (0.801542)
>> User Time        5930.03 (1600.07)	5929.89 (1617)
>> System Time      166.258 (37.0662)	169.337 (37.8419)
>> Percent CPU      5378.5 (1584.16)	5385.6 (1590.24)
>> Context Switches 117389 (47350.1)	130132 (60256.3)
>> Sleeps           163354 (4153.9)	163219 (2251.27)
>>
> 
> Can you also mention % improvement or degradation in a new column.

Fair enough:
- 16 CPUs x86 VM
Average Half load -j 8 Run (std deviation)
                 4.13.0-rc5             4.13.0-rc5-spf
Elapsed Time     166.574 (0.340779)     145.754 (0.776325)	-12.5%
User Time        1080.77 (2.05871)      999.272 (4.12142)	-7.54%
System Time      204.594 (1.02449)      116.362 (1.22974)	-43.13%
Percent CPU      771.2 (1.30384)        765 (0.707107)		-0.8%
Context Switches 46590.6 (935.591)      66316.4 (744.64)	+42.34%
Sleeps           84421.2 (596.612)      85186 (523.041)		+0.9%

Average Optimal load -j 16 Run (std deviation)
                 4.13.0-rc5             4.13.0-rc5-spf
Elapsed Time     85.422 (0.42293)       74.81 (0.419345)	-12.42%
User Time        1031.79 (51.6557)      954.912 (46.8439)	-7.45%
System Time      186.528 (19.0575)      107.514 (9.36902)	-42.36%
Percent CPU      1059.2 (303.607)       1056.8 (307.624)	-0.23%
Context Switches 67240.3 (21788.9)      89360.6 (24299.9)	+32.9%
Sleeps           89607.8 (5511.22)      90372.5 (5490.16)	+0.85%

- 80 CPUs Power 8 node:
Average Half load -j 40 Run (std deviation)
                 4.13.0-rc5             4.13.0-rc5-spf
Elapsed Time     117.176 (0.824093)     116.792 (0.695392)	-0.33%
User Time        4412.34 (24.29)        4396.02 (24.4819)	-0.37%
System Time      131.106 (1.28343)      133.452 (0.708851)	+1.79%
Percent CPU      3876.8 (18.1439)       3877.6 (21.9955)	+0.02%
Context Switches 72470.2 (466.181)      72971 (673.624)		+0.69%
Sleeps           161294 (2284.85)       161946 (2217.9)		+0.40%

Average Optimal load -j 80 Run (std deviation)
                 4.13.0-rc5             4.13.0-rc5-spf
Elapsed Time     111.176 (1.11123)      111.242 (0.801542)	+0.06%
User Time        5930.03 (1600.07)      5929.89 (1617)		+0%
System Time      166.258 (37.0662)      169.337 (37.8419)	+1.85%
Percent CPU      5378.5 (1584.16)       5385.6 (1590.24)	+0.13%
Context Switches 117389 (47350.1)       130132 (60256.3)	+10.86%
Sleeps           163354 (4153.9)        163219 (2251.27)	-0.08%


>> Here the elapsed time is a bit shorter using the spf release, but we
>> remain in the error margin. It has to be noted that this system is not
>> correctly balanced on the NUMA point of view as all the available memory is
>> attached to one core.
> 
> Why different NUMA configuration would have changed the outcome ?

I guess, process will have been scheduled nearest the memory, or spread in
a different way on the core if memory will be attached to.

>>
>> ------------------------
>> Changes since v1:
>>  - Remove PERF_COUNT_SW_SPF_FAILED perf event.
>>  - Add tracing events to details speculative page fault failures.
>>  - Cache VMA fields values which are used once the PTE is unlocked at the
>>  end of the page fault events.
> 
> Why is this required ?

Please see patch 07/20 for details.

Cheers,
Laurent.

[toc] | [prev] | [standalone]

Back to top | Article view | linux.kernel

csiph-web