2.6.24 regression: deadlock on coredump of big process

PATCH In review Tony Battersby 4 participants 15 messages Apr 28, 2008 – May 7, 2008

Subsystems (2)

Memory management MM core

lore

Share & tools

lore

Tony Battersby

author

2008-04-28 15:11 UTC #0 lore raw

Here is a program that can deadlock any kernel from 2.6.24-rc1 to current (2.6.25-git11). The deadlock happens due to oom during a coredump of a large process with multiple threads.

git-bisect reveals the following patch as the culprit:

commit 557ed1fa2620dc119adb86b34c614e152a629a80

Author: Nick Piggin <[email protected]>
Date:   Tue Oct 16 01:24:40 2007 -0700

remove ZERO_PAGE

The commit b5810039a54e5babf428e9a1e89fc1940fabff11 contains the note

A last caveat: the ZERO_PAGE is now refcounted and managed with rmap (and thus mapcounted and count towards shared rss). These writes to the struct page could cause excessive cacheline bouncing on big systems. There are a number of ways this could be addressed if it is an issue.

And indeed this cacheline bouncing has shown up on large SGI systems. There was a situation where an Altix system was essentially livelocked tearing down ZERO_PAGE pagetables when an HPC app aborted during startup. This situation can be avoided in userspace, but it does highlight the potential scalability problem with refcounting ZERO_PAGE, and corner cases where it can really hurt (we don't want the system to livelock!).

There are several broad ways to fix this problem:
1. add back some special casing to avoid refcounting ZERO_PAGE
2. per-node or per-cpu ZERO_PAGES
3. remove the ZERO_PAGE completely

I will argue for 3. The others should also fix the problem, but they result in more complex code than does 3, with little or no real benefit that I can see.

Why? Inserting a ZERO_PAGE for anonymous read faults appears to be a false optimisation: if an application is performance critical, it would not be doing many read faults of new memory, or at least it could be expected to write to that memory soon afterwards. If cache or memory use is critical, it should not be working with a significant number of ZERO_PAGEs anyway (a more compact representation of zeroes should be used).

As a sanity check -- mesuring on my desktop system, there are never many mappings to the ZERO_PAGE (eg. 2 or 3), thus memory usage here should not increase much without it.

When running a make -j4 kernel compile on my dual core system, there are about 1,000 mappings to the ZERO_PAGE created per second, but about 1,000 ZERO_PAGE COW faults per second (less than 1 ZERO_PAGE mapping per second is torn down without being COWed). So removing ZERO_PAGE will save 1,000 page faults per second when running kbuild, while keeping it only saves less than 1 page clearing operation per second. 1 page clear is cheaper than a thousand faults, presumably, so there isn't an obvious loss.

Neither the logical argument nor these basic tests give a guarantee of no regressions. However, this is a reasonable opportunity to try to remove the ZERO_PAGE from the pagefault path. If it is found to cause regressions, we can reintroduce it and just avoid refcounting it.

The /dev/zero ZERO_PAGE usage and TLB tricks also get nuked. I don't see much use to them except on benchmarks. All other users of ZERO_PAGE are converted just to use ZERO_PAGE(0) for simplicity. We can look at replacing them all and maybe ripping out ZERO_PAGE completely when we are more satisfied with this solution.

Signed-off-by Nick Piggin <[email protected]>

Signed-off-by Andrew Morton <[email protected]>

Signed-off-by Linus "snif" Torvalds <[email protected]>

I have verified that 2.6.24.5 with the above patch reverted coredumps successfully instead of deadlocking. The patch doesn't revert cleanly on 2.6.25, so I didn't test that.

Before finding the above patch with git-bisect, I tested Daniel Phillips' bio throttling patch
(http://zumastor.googlecode.com/svn/trunk/ddsnap/patches/2.6.24.2/bio.throttle.patch),
but it didn't prevent the deadlock.

I am testing on a simple 32-bit x86 system with Pentium 4 CPU, 256 MB DRAM, a IDE hard drive, and an ext3 filesystem. The software is a bare- bones embedded environment; I am not running X, device-mapper, RAID, or anything fancy. I am not using any network file systems. The system has no swap space, and in fact swap support is disabled in the kernel configuration.

When the kernel is deadlocked, I can switch VTs using Alt-<function key>. When typing characters on the keyboard, the characters are printed to the screen if I am on the VT that the core-dumping program was using, but keypresses on other VTs do not show up. The system is basically unusable.

If I let the kernel write the core file to disk directly (the default behavior), then pressing Alt-SysRq-I to kill all tasks and free up some memory will un-deadlock the coredump for a short while, but then it deadlocks again. If I pipe the core file to a program
(via /proc/sys/kernel/core_pattern) which doesn't write it to disk (e.g. cat > /dev/null), then the kernel still deadlocks, but Alt-SysRq-I kills the program and breaks the pipe, which un-deadlocks the system and allows me to log back in.

Below is the program that triggers the deadlock; compile with -D_REENTRANT -lpthread.

Tony Battersby
Cybernetics

---------------------------------------------------------------------

#include <sys/time.h>
#include <sys/resource.h>
#include <stdlib.h>
#include <stdio.h>
#include <string.h>
#include <pthread.h>
#include <unistd.h>
#include <assert.h>

static void allow_coredump(void)
{
struct rlimit rlim;

rlim.rlim_cur = RLIM_INFINITY;
rlim.rlim_max = RLIM_INFINITY;
if (setrlimit(RLIMIT_CORE, &rlim))

      {
      perror("setrlimit");
      exit(EXIT_FAILURE);
      }
}

static void *thread_func(void *arg)
{
for (;;)

      {
      sleep(100);
      }

return NULL;
}

static void spawn_threads(int n_threads)
{
pthread_attr_t thread_attr;
int i;

   pthread_attr_init(&thread_attr);
   printf("spawn %d threads\n", n_threads);
   for (i = 0; i < n_threads; i++)
      {
      pthread_t thread;

      if (pthread_create(&thread, &thread_attr, &thread_func, NULL))
         {
         perror("pthread_create");
         exit(EXIT_FAILURE);
         }
      }

sleep(1);
}

static size_t get_max_malloc_len(void)
{
size_t min = 1;
size_t max = ~((size_t) 0);

   do
      {
      size_t len = min + (max - min) / 2;
      void *ptr;

      ptr = malloc(len);
      if (ptr == NULL)
         {
         max = len - 1;
         }
      else
         {
         free(ptr);
         min = len + 1;
         }
      } while (min < max);

return min;
}

static void malloc_all_but_x_mb(unsigned free_mb)
{
size_t len = get_max_malloc_len();
void *ptr;

   assert(len > free_mb << 20);
   len -= free_mb << 20;

   printf("allocate %u MB\n", len >> 20);
   ptr = malloc(len);
   assert(ptr != NULL);

/* if this triggers the oom killer, then use a larger free_mb */ memset(ptr, 0xab, len);
}

static void trigger_segfault(void)
{
printf("trigger segfault\n");
*(int *) 0 = 0;
}

int main(int argc, char *argv[])
{
allow_coredump();
malloc_all_but_x_mb(16);
spawn_threads(10);
trigger_segfault();
return 0;
}

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

KAMEZAWA Hiroyuki

2008-04-29 00:56 UTC #1 lore raw

Quoted 3 lines · On Mon, 28 Apr 2008 11:11:46 -0400 Tony Battersby <[email protected]> wrote:

Below is the program that triggers the deadlock; compile with -D_REENTRANT -lpthread.

What happens if you changes size of stack (of pthreads) smaller ? (maybe ulimit -s will work also for threads.)

Thanks,
-Kame

Tony Battersby

author

2008-04-29 14:10 UTC #2 lore raw

Quoted 8 lines · KAMEZAWA Hiroyuki

Quoted 4 lines · On Mon, 28 Apr 2008 11:11:46 -0400 Tony Battersby <[email protected]> wrote:
Below is the program that triggers the deadlock; compile with -D_REENTRANT -lpthread.

Quoted 8 lines

What happens if you changes size of stack (of pthreads) smaller ? (maybe ulimit -s will work also for threads.)
Thanks,
-Kame

If I leave more memory free by changing the argument to malloc_all_but_x_mb(), then I have to increase the number of threads required to trigger the deadlock. Changing the thread stack size via setrlimit(RLIMIT_STACK) also changes the number of threads that are required to trigger the deadlock. For example, with malloc_all_but_x_mb(16) and the default stack size of 8 MB, <= 5 threads will coredump successfully, and >= 6 threads will deadlock. With malloc_all_but_x_mb(16) and a reduced stack size of 4096 bytes, <= 8 threads will coredump successfully, and >= 9 threads will deadlock.

Also note that the "free" command reports 10 MB free memory while the program is running before the segfault is triggered.

Tony

KAMEZAWA Hiroyuki

2008-04-30 04:21 UTC #3 lore raw

PATCH 1/1 more ZERO_PAGE handling ( was 2.6.24 regression: deadlock on coredump of big process)

Quoted 11 lines · On Tue, 29 Apr 2008 10:10:58 -0400 Tony Battersby <[email protected]> wrote:

If I leave more memory free by changing the argument to malloc_all_but_x_mb(), then I have to increase the number of threads required to trigger the deadlock. Changing the thread stack size via setrlimit(RLIMIT_STACK) also changes the number of threads that are required to trigger the deadlock. For example, with malloc_all_but_x_mb(16) and the default stack size of 8 MB, <= 5 threads will coredump successfully, and >= 6 threads will deadlock. With malloc_all_but_x_mb(16) and a reduced stack size of 4096 bytes, <= 8 threads will coredump successfully, and >= 9 threads will deadlock.

Quoted 3 lines

Also note that the "free" command reports 10 MB free memory while the program is running before the segfault is triggered.

Hmm, my idea is below.

Nick's remove ZERO_PAGE patch includes following change

==
@@ -2252,39 +2158,24 @@ static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,

        spinlock_t *ptl;

{
<snip>
- page_add_new_anon_rmap(page, vma, address);
- } else {

-               /* Map the ZERO_PAGE - vm_page_prot is readonly */
-               page = ZERO_PAGE(address);
-               page_cache_get(page);
-               entry = mk_pte(page, vma->vm_page_prot);

+ if (unlikely(anon_vma_prepare(vma)))
+ goto oom;
+ page = alloc_zeroed_user_highpage_movable(vma, address);
==

above change is for avoiding to use ZERO_PAGE at read-page-fault to anonymous vma. This is reasonable I think. But at coredump, tons of read-but-never-written pages can be allocated.
==
coredump
-> get_user_pages()

       -> follow_page() returns NULL
            -> handle mm fault
                 -> do_anonymous page.

==
follow_page() returns ZERO_PAGE only when page table is not avaiable.

So, making follow_page() return ZERO_PAGE can be a fix of extra memory consumpstion at core dump. (Maybe someone can think of other fix.)

how about this patch ? Could you try ?

(I'm sorry but I'll not be active for a week because my servers are powered off.)

-Kame

==
follow_page() returns ZERO_PAGE if page table is not available. but returns NULL pte is not presentl.

Signed-off-by KAMEZAWA Hiroyuki <[email protected]>

mm/memory.c +8 -6

Index: linux-2.6.25/mm/memory.c
===================================================================
@@ -926,15 +926,15 @@ struct page *follow_page(struct vm_area_
 	page = NULL;
 	pgd = pgd_offset(mm, address);
 	if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
-		goto no_page_table;
+		goto null_or_zeropage;
 
 	pud = pud_offset(pgd, address);
 	if (pud_none(*pud) || unlikely(pud_bad(*pud)))
-		goto no_page_table;
+		goto null_or_zeropage;
 	
 	pmd = pmd_offset(pud, address);
 	if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
-		goto no_page_table;
+		goto null_or_zeropage;
 
 	if (pmd_huge(*pmd)) {
 		BUG_ON(flags & FOLL_GET);

@@ -947,8 +947,10 @@ struct page *follow_page(struct vm_area_
 		goto out;
 
 	pte = *ptep;
-	if (!pte_present(pte))
-		goto unlock;
+	if (!(flags & FOLL_WRITE) && !pte_present(pte)) {
+		pte_unmap_unlock(ptep, ptl);
+		goto null_or_zeropage;
+	}
 	if ((flags & FOLL_WRITE) && !pte_write(pte))
 		goto unlock;
 	page = vm_normal_page(vma, address, pte);

@@ -968,7 +970,7 @@ unlock:
 out:
 	return page;
 
-no_page_table:
+null_or_zeropage:
 	/*
 	 * When core dumping an enormous anonymous area that nobody
 	 * has touched so far, we don't want to allocate page tables.

Nick Piggin

2008-04-30 04:46 UTC #4 lore raw

Re: [PATCH] more ZERO_PAGE handling ( was 2.6.24 regression: deadlock on coredump of big process)

Quoted 45 lines · KAMEZAWA Hiroyuki

Quoted 11 lines · On Tue, 29 Apr 2008 10:10:58 -0400 Tony Battersby <[email protected]> wrote:
If I leave more memory free by changing the argument to malloc_all_but_x_mb(), then I have to increase the number of threads required to trigger the deadlock. Changing the thread stack size via setrlimit(RLIMIT_STACK) also changes the number of threads that are required to trigger the deadlock. For example, with malloc_all_but_x_mb(16) and the default stack size of 8 MB, <= 5 threads will coredump successfully, and >= 6 threads will deadlock. With malloc_all_but_x_mb(16) and a reduced stack size of 4096 bytes, <= 8 threads will coredump successfully, and >= 9 threads will deadlock.
Quoted 3 lines
Also note that the "free" command reports 10 MB free memory while the program is running before the segfault is triggered.
Hmm, my idea is below.
Nick's remove ZERO_PAGE patch includes following change
==
(quoted diff) +3 -6
@@ -2252,39 +2158,24 @@ static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
        spinlock_t *ptl;
 {
<snip>
-               page_add_new_anon_rmap(page, vma, address);
-       } else {
-               /* Map the ZERO_PAGE - vm_page_prot is readonly */
-               page = ZERO_PAGE(address);
-               page_cache_get(page);
-               entry = mk_pte(page, vma->vm_page_prot);
+       if (unlikely(anon_vma_prepare(vma)))
+               goto oom;
+       page = alloc_zeroed_user_highpage_movable(vma, address);
==
above change is for avoiding to use ZERO_PAGE at read-page-fault to anonymous vma. This is reasonable I think. But at coredump, tons of read-but-never-written pages can be allocated.
==
coredump
-> get_user_pages()
       -> follow_page() returns NULL
            -> handle mm fault
                 -> do_anonymous page.

Quoted 8 lines

==
follow_page() returns ZERO_PAGE only when page table is not avaiable.
So, making follow_page() return ZERO_PAGE can be a fix of extra memory consumpstion at core dump. (Maybe someone can think of other fix.)
how about this patch ? Could you try ?

Ah, yes I stupidly missed this detail of follow_page. Definitely your patch is a good idea, and I think it would be a good idea even when we still had ZERO_PAGE, because it would prevent pagetable clearing from having to do extra teardown work here.

Good catch, and I agree with your patch. Thanks

Quoted 47 lines

(I'm sorry but I'll not be active for a week because my servers are powered off.)

-Kame

==
follow_page() returns ZERO_PAGE if page table is not available. but returns NULL pte is not presentl.

Signed-off-by KAMEZAWA Hiroyuki <[email protected]>

Index: linux-2.6.25/mm/memory.c
===================================================================

mm/memory.c +7 -5

@@ -926,15 +926,15 @@ struct page *follow_page(struct vm_area_
 	page = NULL;
 	pgd = pgd_offset(mm, address);
 	if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
-		goto no_page_table;
+		goto null_or_zeropage;
 
 	pud = pud_offset(pgd, address);
 	if (pud_none(*pud) || unlikely(pud_bad(*pud)))
-		goto no_page_table;
+		goto null_or_zeropage;
 
 	pmd = pmd_offset(pud, address);
 	if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
-		goto no_page_table;
+		goto null_or_zeropage;
 
 	if (pmd_huge(*pmd)) {
 		BUG_ON(flags & FOLL_GET);

@@ -947,8 +947,10 @@ struct page *follow_page(struct vm_area_
 		goto out;
 
 	pte = *ptep;
-	if (!pte_present(pte))
-		goto unlock;
+	if (!(flags & FOLL_WRITE) && !pte_present(pte)) {
+		pte_unmap_unlock(ptep, ptl);
+		goto null_or_zeropage;
+	}
 	if ((flags & FOLL_WRITE) && !pte_write(pte))
 		goto unlock;
 	page = vm_normal_page(vma, address, pte);

mm/memory.c unlock: · line 976

Quoted 8 lines

mm/memory.c +1 -1

@@ -968,7 +970,7 @@ unlock:
 out:
 	return page;
 
-no_page_table:
+null_or_zeropage:
 	/*
 	 * When core dumping an enormous anonymous area that nobody
 	 * has touched so far, we don't want to allocate page tables.

Mika Penttilä

2008-04-30 05:03 UTC #5 lore raw

Re: [PATCH] more ZERO_PAGE handling ( was 2.6.24 regression: deadlock on coredump of big process)

In reply to KAMEZAWA Hiroyuki

Quoted 48 lines · KAMEZAWA Hiroyuki

Quoted 10 lines · On Tue, 29 Apr 2008 10:10:58 -0400 Tony Battersby <[email protected]> wrote:
If I leave more memory free by changing the argument to malloc_all_but_x_mb(), then I have to increase the number of threads required to trigger the deadlock. Changing the thread stack size via setrlimit(RLIMIT_STACK) also changes the number of threads that are required to trigger the deadlock. For example, with malloc_all_but_x_mb(16) and the default stack size of 8 MB, <= 5 threads will coredump successfully, and >= 6 threads will deadlock. With malloc_all_but_x_mb(16) and a reduced stack size of 4096 bytes, <= 8 threads will coredump successfully, and >= 9 threads will deadlock.
Quoted 4 lines
Also note that the "free" command reports 10 MB free memory while the program is running before the segfault is triggered.
Hmm, my idea is below.
Nick's remove ZERO_PAGE patch includes following change
==
(quoted diff) +3 -6
@@ -2252,39 +2158,24 @@ static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
        spinlock_t *ptl;
 {
<snip>
-               page_add_new_anon_rmap(page, vma, address);
-       } else {
-               /* Map the ZERO_PAGE - vm_page_prot is readonly */
-               page = ZERO_PAGE(address);
-               page_cache_get(page);
-               entry = mk_pte(page, vma->vm_page_prot);
+       if (unlikely(anon_vma_prepare(vma)))
+               goto oom;
+       page = alloc_zeroed_user_highpage_movable(vma, address);
==
above change is for avoiding to use ZERO_PAGE at read-page-fault to anonymous
vma. This is reasonable I think. But at coredump, tons of read-but-never-written pages can be allocated. ==
coredump
-> get_user_pages()
       -> follow_page() returns NULL
            -> handle mm fault
                 -> do_anonymous page.
==
follow_page() returns ZERO_PAGE only when page table is not avaiable.

Quoted 10 lines

So, making follow_page() return ZERO_PAGE can be a fix of extra memory consumpstion at core dump. (Maybe someone can think of other fix.)
how about this patch ? Could you try ?
(I'm sorry but I'll not be active for a week because my servers are powered off.)
-Kame

But sure we still have to handle the fault for instance swapped pages, for other uses of get_user_pages();

--Mika

Quoted 53 lines

==
follow_page() returns ZERO_PAGE if page table is not available. but returns NULL pte is not presentl.

Signed-off-by KAMEZAWA Hiroyuki <[email protected]>

Index: linux-2.6.25/mm/memory.c
===================================================================

mm/memory.c +8 -7

@@ -926,15 +926,15 @@ struct page *follow_page(struct vm_area_
 	page = NULL;
 	pgd = pgd_offset(mm, address);
 	if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
-		goto no_page_table;
+		goto null_or_zeropage;
 
 	pud = pud_offset(pgd, address);
 	if (pud_none(*pud) || unlikely(pud_bad(*pud)))
-		goto no_page_table;
+		goto null_or_zeropage;
 
 	pmd = pmd_offset(pud, address);
 	if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
-		goto no_page_table;
+		goto null_or_zeropage;
 
 	if (pmd_huge(*pmd)) {
 		BUG_ON(flags & FOLL_GET);

@@ -947,8 +947,10 @@ struct page *follow_page(struct vm_area_
 		goto out;
 
 	pte = *ptep;
-	if (!pte_present(pte))
-		goto unlock;
+	if (!(flags & FOLL_WRITE) && !pte_present(pte)) {
+		pte_unmap_unlock(ptep, ptl);
+		goto null_or_zeropage;
+	}
 	if ((flags & FOLL_WRITE) && !pte_write(pte))
 		goto unlock;
 	page = vm_normal_page(vma, address, pte);

@@ -968,7 +970,7 @@ unlock:
 out:
 	return page;
 
-no_page_table:
+null_or_zeropage:
 	/*
 	 * When core dumping an enormous anonymous area that nobody
 	 * has touched so far, we don't want to allocate page tables.
 
--

Quoted 6 lines

To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

Nick Piggin

2008-04-30 05:09 UTC #6 lore raw

Re: [PATCH] more ZERO_PAGE handling ( was 2.6.24 regression: deadlock on coredump of big process)

Quoted 60 lines · Mika Penttila

Quoted 59 lines · KAMEZAWA Hiroyuki
Quoted 10 lines · On Tue, 29 Apr 2008 10:10:58 -0400 Tony Battersby <[email protected]> wrote:
If I leave more memory free by changing the argument to malloc_all_but_x_mb(), then I have to increase the number of threads required to trigger the deadlock. Changing the thread stack size via setrlimit(RLIMIT_STACK) also changes the number of threads that are required to trigger the deadlock. For example, with malloc_all_but_x_mb(16) and the default stack size of 8 MB, <= 5 threads will coredump successfully, and >= 6 threads will deadlock. With malloc_all_but_x_mb(16) and a reduced stack size of 4096 bytes, <= 8 threads will coredump successfully, and >= 9 threads will deadlock.
Quoted 4 lines
Also note that the "free" command reports 10 MB free memory while the program is running before the segfault is triggered.
Hmm, my idea is below.
Nick's remove ZERO_PAGE patch includes following change
==
(quoted diff) +3 -6
@@ -2252,39 +2158,24 @@ static int do_anonymous_page(struct mm_struct *mm, 
 struct vm_area_struct *vma,
       spinlock_t *ptl;
 {
<snip>
-               page_add_new_anon_rmap(page, vma, address);
-       } else {
-               /* Map the ZERO_PAGE - vm_page_prot is readonly */
-               page = ZERO_PAGE(address);
-               page_cache_get(page);
-               entry = mk_pte(page, vma->vm_page_prot);
+       if (unlikely(anon_vma_prepare(vma)))
+               goto oom;
+       page = alloc_zeroed_user_highpage_movable(vma, address);
==
above change is for avoiding to use ZERO_PAGE at read-page-fault to anonymous
vma. This is reasonable I think. But at coredump, tons of read-but-never-written pages can be allocated.
==
coredump
-> get_user_pages()
      -> follow_page() returns NULL
           -> handle mm fault
                -> do_anonymous page.
==
follow_page() returns ZERO_PAGE only when page table is not avaiable.
So, making follow_page() return ZERO_PAGE can be a fix of extra memory consumpstion at core dump. (Maybe someone can think of other fix.)
how about this patch ? Could you try ?
(I'm sorry but I'll not be active for a week because my servers are powered off.)

Quoted 7 lines

Quoted 3 lines
-Kame
But sure we still have to handle the fault for instance swapped pages, for other uses of get_user_pages();

Yeah, it does need to test for pte_none.

KAMEZAWA Hiroyuki

2008-04-30 05:13 UTC #7 lore raw

Re: [PATCH] more ZERO_PAGE handling ( was 2.6.24 regression: deadlock on coredump of big process)

In reply to Mika Penttilä

Quoted 35 lines · On Wed, 30 Apr 2008 08:03:33 +0300 Mika Penttilä <[email protected]>…

Quoted 35 lines
==
(quoted diff) +3 -6
@@ -2252,39 +2158,24 @@ static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
        spinlock_t *ptl;
 {
<snip>
-               page_add_new_anon_rmap(page, vma, address);
-       } else {
-               /* Map the ZERO_PAGE - vm_page_prot is readonly */
-               page = ZERO_PAGE(address);
-               page_cache_get(page);
-               entry = mk_pte(page, vma->vm_page_prot);
+       if (unlikely(anon_vma_prepare(vma)))
+               goto oom;
+       page = alloc_zeroed_user_highpage_movable(vma, address);
==
above change is for avoiding to use ZERO_PAGE at read-page-fault to anonymous vma. This is reasonable I think. But at coredump, tons of read-but-never-written pages can be allocated.
==
coredump
-> get_user_pages()
       -> follow_page() returns NULL
            -> handle mm fault
                 -> do_anonymous page.
==
follow_page() returns ZERO_PAGE only when page table is not avaiable.
So, making follow_page() return ZERO_PAGE can be a fix of extra memory consumpstion at core dump. (Maybe someone can think of other fix.)
how about this patch ? Could you try ?
(I'm sorry but I'll not be active for a week because my servers are powered off.)

Quoted 8 lines

Quoted 3 lines
-Kame
But sure we still have to handle the fault for instance swapped pages, for other uses of get_user_pages();

Ah, my bad.....how about this ? I changed !pte_present() to pte_none().

-Kame
==
follow_page() returns ZERO_PAGE if a page table is not available. but returns NULL if a page table exists. If NULL, handle_mm_fault() allocates a new page.

This behavior increases page consumption at coredump, which tend to do read-once-but-never-written page fault. This patch is for avoiding this.

Changelog:
- fixed to check pte_none() not !pte_present().

Signed-off-by KAMEZAWA Hiroyuki <[email protected]>

mm/memory.c +8 -6

Index: linux-2.6.25/mm/memory.c
===================================================================
@@ -926,15 +926,15 @@ struct page *follow_page(struct vm_area_
 	page = NULL;
 	pgd = pgd_offset(mm, address);
 	if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
-		goto no_page_table;
+		goto null_or_zeropage;
 
 	pud = pud_offset(pgd, address);
 	if (pud_none(*pud) || unlikely(pud_bad(*pud)))
-		goto no_page_table;
+		goto null_or_zeropage;
 	
 	pmd = pmd_offset(pud, address);
 	if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
-		goto no_page_table;
+		goto null_or_zeropage;
 
 	if (pmd_huge(*pmd)) {
 		BUG_ON(flags & FOLL_GET);

@@ -947,8 +947,10 @@ struct page *follow_page(struct vm_area_
 		goto out;
 
 	pte = *ptep;
-	if (!pte_present(pte))
-		goto unlock;
+	if (!(flags & FOLL_WRITE) && pte_none(pte)) {
+		pte_unmap_unlock(ptep, ptl);
+		goto null_or_zeropage;
+	}
 	if ((flags & FOLL_WRITE) && !pte_write(pte))
 		goto unlock;
 	page = vm_normal_page(vma, address, pte);

@@ -968,7 +970,7 @@ unlock:
 out:
 	return page;
 
-no_page_table:
+null_or_zeropage:
 	/*
 	 * When core dumping an enormous anonymous area that nobody
 	 * has touched so far, we don't want to allocate page tables.

Nick Piggin

2008-04-30 05:19 UTC #8 lore raw

Re: [PATCH] more ZERO_PAGE handling ( was 2.6.24 regression: deadlock on coredump of big process)

Quoted 54 lines · KAMEZAWA Hiroyuki

Quoted 35 lines · On Wed, 30 Apr 2008 08:03:33 +0300 Mika Penttila <[email protected]>…
Quoted 35 lines
==
(quoted diff) +3 -6
@@ -2252,39 +2158,24 @@ static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
        spinlock_t *ptl;
 {
<snip>
-               page_add_new_anon_rmap(page, vma, address);
-       } else {
-               /* Map the ZERO_PAGE - vm_page_prot is readonly */
-               page = ZERO_PAGE(address);
-               page_cache_get(page);
-               entry = mk_pte(page, vma->vm_page_prot);
+       if (unlikely(anon_vma_prepare(vma)))
+               goto oom;
+       page = alloc_zeroed_user_highpage_movable(vma, address);
==
above change is for avoiding to use ZERO_PAGE at read-page-fault to anonymous vma. This is reasonable I think. But at coredump, tons of read-but-never-written pages can be allocated.
==
coredump
-> get_user_pages()
       -> follow_page() returns NULL
            -> handle mm fault
                 -> do_anonymous page.
==
follow_page() returns ZERO_PAGE only when page table is not avaiable.
So, making follow_page() return ZERO_PAGE can be a fix of extra memory consumpstion at core dump. (Maybe someone can think of other fix.)
how about this patch ? Could you try ?
(I'm sorry but I'll not be active for a week because my servers are powered off.)
Quoted 8 lines
Quoted 3 lines
-Kame
But sure we still have to handle the fault for instance swapped pages, for other uses of get_user_pages();
Ah, my bad.....how about this ? I changed !pte_present() to pte_none().
-Kame
==
follow_page() returns ZERO_PAGE if a page table is not available. but returns NULL if a page table exists. If NULL, handle_mm_fault() allocates a new page.

Quoted 3 lines

This behavior increases page consumption at coredump, which tend to do read-once-but-never-written page fault. This patch is for avoiding this.

I think you still need the pte_present test too, otherwise !present and !none ptes can slip through and be treated as present.

Something like this should do:

if (!pte_present(pte)) {
	if (pte_none(pte)) {
		pte_unmap_unlock
		goto null_or_zeropage;
	}
	goto unlock;
}

Quoted 46 lines

Changelog:
- fixed to check pte_none() not !pte_present().

Signed-off-by KAMEZAWA Hiroyuki <[email protected]>

Index: linux-2.6.25/mm/memory.c
===================================================================

mm/memory.c +7 -5

@@ -926,15 +926,15 @@ struct page *follow_page(struct vm_area_
 	page = NULL;
 	pgd = pgd_offset(mm, address);
 	if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
-		goto no_page_table;
+		goto null_or_zeropage;
 
 	pud = pud_offset(pgd, address);
 	if (pud_none(*pud) || unlikely(pud_bad(*pud)))
-		goto no_page_table;
+		goto null_or_zeropage;
 
 	pmd = pmd_offset(pud, address);
 	if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
-		goto no_page_table;
+		goto null_or_zeropage;
 
 	if (pmd_huge(*pmd)) {
 		BUG_ON(flags & FOLL_GET);

@@ -947,8 +947,10 @@ struct page *follow_page(struct vm_area_
 		goto out;
 
 	pte = *ptep;
-	if (!pte_present(pte))
-		goto unlock;
+	if (!(flags & FOLL_WRITE) && pte_none(pte)) {
+		pte_unmap_unlock(ptep, ptl);
+		goto null_or_zeropage;
+	}
 	if ((flags & FOLL_WRITE) && !pte_write(pte))
 		goto unlock;
 	page = vm_normal_page(vma, address, pte);

mm/memory.c unlock: · line 976

Quoted 8 lines

mm/memory.c +1 -1

@@ -968,7 +970,7 @@ unlock:
 out:
 	return page;
 
-no_page_table:
+null_or_zeropage:
 	/*
 	 * When core dumping an enormous anonymous area that nobody
 	 * has touched so far, we don't want to allocate page tables.

KAMEZAWA Hiroyuki

2008-04-30 05:31 UTC #9 lore raw

Re: [PATCH] more ZERO_PAGE handling ( was 2.6.24 regression: deadlock on coredump of big process)

Quoted 10 lines · On Wed, 30 Apr 2008 07:19:32 +0200 Nick Piggin <[email protected]> wrote:

Something like this should do:

if (!pte_present(pte)) {
	if (pte_none(pte)) {
		pte_unmap_unlock
		goto null_or_zeropage;
	}
	goto unlock;
}

Sorry for broken work and thank you for advice.
updated.

Regards,
-Kame
==
follow_page() returns ZERO_PAGE if a page table is not available. but returns NULL if a page table exists. If NULL, handle_mm_fault() allocates a new page.

This behavior increases page consumption at coredump, which tend to do read-once-but-never-written page fault. This patch is for avoiding this.

Changelog:
- fixed to check pte_present()/pte_none() in proper way.

Signed-off-by KAMEZAWA Hiroyuki <[email protected]>

mm/memory.c +10 -5

Index: linux-2.6.25/mm/memory.c
===================================================================
@@ -926,15 +926,15 @@ struct page *follow_page(struct vm_area_
 	page = NULL;
 	pgd = pgd_offset(mm, address);
 	if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
-		goto no_page_table;
+		goto null_or_zeropage;
 
 	pud = pud_offset(pgd, address);
 	if (pud_none(*pud) || unlikely(pud_bad(*pud)))
-		goto no_page_table;
+		goto null_or_zeropage;
 	
 	pmd = pmd_offset(pud, address);
 	if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
-		goto no_page_table;
+		goto null_or_zeropage;
 
 	if (pmd_huge(*pmd)) {
 		BUG_ON(flags & FOLL_GET);

@@ -947,8 +947,13 @@ struct page *follow_page(struct vm_area_
 		goto out;
 
 	pte = *ptep;
-	if (!pte_present(pte))
+	if (!pte_present(pte)) {
+		if (!(flags & FOLL_WRITE) && pte_none(pte)) {
+			pte_unmap_unlock(ptep, ptl);
+			goto null_or_zeropage;
+		}
 		goto unlock;
+	}
 	if ((flags & FOLL_WRITE) && !pte_write(pte))
 		goto unlock;
 	page = vm_normal_page(vma, address, pte);

@@ -968,7 +973,7 @@ unlock:
 out:
 	return page;
 
-no_page_table:
+null_or_zeropage:
 	/*
 	 * When core dumping an enormous anonymous area that nobody
 	 * has touched so far, we don't want to allocate page tables.

Nick Piggin

Acked-by

2008-04-30 06:11 UTC #10 lore raw

Re: [PATCH] more ZERO_PAGE handling ( was 2.6.24 regression: deadlock on coredump of big process)

Quoted 13 lines · KAMEZAWA Hiroyuki

Quoted 10 lines · On Wed, 30 Apr 2008 07:19:32 +0200 Nick Piggin <[email protected]> wrote:
Something like this should do:
if (!pte_present(pte)) {
	if (pte_none(pte)) {
		pte_unmap_unlock
		goto null_or_zeropage;
	}
	goto unlock;
}

Quoted 2 lines

Sorry for broken work and thank you for advice.
updated.

Don't be sorry. The most important thing is that you found this tricky problem. That's very good work IMO!

Quoted 43 lines

Regards,
-Kame
==
follow_page() returns ZERO_PAGE if a page table is not available. but returns NULL if a page table exists. If NULL, handle_mm_fault() allocates a new page.
This behavior increases page consumption at coredump, which tend to do read-once-but-never-written page fault. This patch is for avoiding this.
Changelog:
- fixed to check pte_present()/pte_none() in proper way.
Signed-off-by KAMEZAWA Hiroyuki <[email protected]>
Index: linux-2.6.25/mm/memory.c
===================================================================
mm/memory.c +3 -3
@@ -926,15 +926,15 @@ struct page *follow_page(struct vm_area_
 	page = NULL;
 	pgd = pgd_offset(mm, address);
 	if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
-		goto no_page_table;
+		goto null_or_zeropage;
 
 	pud = pud_offset(pgd, address);
 	if (pud_none(*pud) || unlikely(pud_bad(*pud)))
-		goto no_page_table;
+		goto null_or_zeropage;
 
 	pmd = pmd_offset(pud, address);
 	if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
-		goto no_page_table;
+		goto null_or_zeropage;
 
 	if (pmd_huge(*pmd)) {
 		BUG_ON(flags & FOLL_GET);
@@ -947,8 +947,13 @@ struct page *follow_page(struct vm_area_
 		goto out;

mm/memory.c struct page *follow_page(struct vm_area_ · line 956

Quoted 10 lines

mm/memory.c +6 -1

 
 	pte = *ptep;
-	if (!pte_present(pte))
+	if (!pte_present(pte)) {
+		if (!(flags & FOLL_WRITE) && pte_none(pte)) {
+			pte_unmap_unlock(ptep, ptl);
+			goto null_or_zeropage;
+		}
 		goto unlock;
+	}

Just a small nitpick: I guess you don't need this FOLL_WRITE test because null_or_zeropage will test FOLL_ANON which implies !FOLL_WRITE. It should give slightly smaller code.

Otherwise, looks good to me:

Acked-by Nick Piggin <[email protected]>

mm/memory.c unlock: · line 979

Quoted 12 lines

mm/memory.c +1 -1

 	if ((flags & FOLL_WRITE) && !pte_write(pte))
 		goto unlock;
 	page = vm_normal_page(vma, address, pte);

@@ -968,7 +973,7 @@ unlock:
 out:
 	return page;
 
-no_page_table:
+null_or_zeropage:
 	/*
 	 * When core dumping an enormous anonymous area that nobody
 	 * has touched so far, we don't want to allocate page tables.

KAMEZAWA Hiroyuki

2008-05-07 02:10 UTC #11 lore raw

Re: [PATCH] more ZERO_PAGE handling ( was 2.6.24 regression: deadlock on coredump of big process)

Quoted 14 lines · On Wed, 30 Apr 2008 08:11:25 +0200 Nick Piggin <[email protected]> wrote:

mm/memory.c struct page *follow_page(struct vm_area_ · line 956
Quoted 9 lines
mm/memory.c +6 -1
 	pte = *ptep;
-	if (!pte_present(pte))
+	if (!pte_present(pte)) {
+		if (!(flags & FOLL_WRITE) && pte_none(pte)) {
+			pte_unmap_unlock(ptep, ptl);
+			goto null_or_zeropage;
+		}
 		goto unlock;
+	}
Just a small nitpick: I guess you don't need this FOLL_WRITE test because null_or_zeropage will test FOLL_ANON which implies !FOLL_WRITE. It should give slightly smaller code.

Quoted 2 lines

Otherwise, looks good to me:

Hmm, but

do_execve()
-> copy_strings()

       -> get_arg_page()
            -> get_user_pages()

can do write-page-fault in ANON (and it's a valid ops.)

So, I think it's safe not to remove FOLL_WRITE check here.

Thanks,
-Kame

KAMEZAWA Hiroyuki

2008-05-07 02:23 UTC #12 lore raw

Re: [PATCH] more ZERO_PAGE handling ( was 2.6.24 regression: deadlock on coredump of big process)

Quoted 18 lines · On Wed, 7 May 2008 11:14:04 +0900 KAMEZAWA Hiroyuki…

Quoted 14 lines
mm/memory.c struct page *follow_page(struct vm_area_ · line 956
Quoted 9 lines
mm/memory.c +6 -1
 	pte = *ptep;
-	if (!pte_present(pte))
+	if (!pte_present(pte)) {
+		if (!(flags & FOLL_WRITE) && pte_none(pte)) {
+			pte_unmap_unlock(ptep, ptl);
+			goto null_or_zeropage;
+		}
 		goto unlock;
+	}
Just a small nitpick: I guess you don't need this FOLL_WRITE test because null_or_zeropage will test FOLL_ANON which implies !FOLL_WRITE. It should give slightly smaller code.
Quoted 2 lines
Otherwise, looks good to me:
Hmm, but

Quoted 9 lines

do_execve()
-> copy_strings()
       -> get_arg_page()
            -> get_user_pages()
can do write-page-fault in ANON (and it's a valid ops.)
So, I think it's safe not to remove FOLL_WRITE check here.

BTW, in above case, returning ZERO_PAGE() when pgd/pud/pmd is not available is safe ? (above path is expanding-stack at exec.)

Thanks,
-Kame

Tony Battersby

author Tested-by

2008-04-30 13:57 UTC #13 lore raw

Re: [PATCH] more ZERO_PAGE handling ( was 2.6.24 regression: deadlock on coredump of big process)

In reply to KAMEZAWA Hiroyuki

Quoted 56 lines · KAMEZAWA Hiroyuki

follow_page() returns ZERO_PAGE if a page table is not available. but returns NULL if a page table exists. If NULL, handle_mm_fault() allocates a new page.
This behavior increases page consumption at coredump, which tend to do read-once-but-never-written page fault. This patch is for avoiding this.
Changelog:
- fixed to check pte_present()/pte_none() in proper way.
Signed-off-by KAMEZAWA Hiroyuki <[email protected]>
Index: linux-2.6.25/mm/memory.c
===================================================================
mm/memory.c +9 -4
@@ -926,15 +926,15 @@ struct page *follow_page(struct vm_area_
 	page = NULL;
 	pgd = pgd_offset(mm, address);
 	if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
-		goto no_page_table;
+		goto null_or_zeropage;
 
 	pud = pud_offset(pgd, address);
 	if (pud_none(*pud) || unlikely(pud_bad(*pud)))
-		goto no_page_table;
+		goto null_or_zeropage;
 
 	pmd = pmd_offset(pud, address);
 	if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
-		goto no_page_table;
+		goto null_or_zeropage;
 
 	if (pmd_huge(*pmd)) {
 		BUG_ON(flags & FOLL_GET);
@@ -947,8 +947,13 @@ struct page *follow_page(struct vm_area_
 		goto out;
 
 	pte = *ptep;
-	if (!pte_present(pte))
+	if (!pte_present(pte)) {
+		if (!(flags & FOLL_WRITE) && pte_none(pte)) {
+			pte_unmap_unlock(ptep, ptl);
+			goto null_or_zeropage;
+		}
 		goto unlock;
+	}
 	if ((flags & FOLL_WRITE) && !pte_write(pte))
 		goto unlock;
 	page = vm_normal_page(vma, address, pte);

mm/memory.c unlock: · line 979

Quoted 8 lines

mm/memory.c +1 -1

@@ -968,7 +973,7 @@ unlock:
 out:
 	return page;
 
-no_page_table:
+null_or_zeropage:
 	/*
 	 * When core dumping an enormous anonymous area that nobody
 	 * has touched so far, we don't want to allocate page tables.

This patch fixes the deadlock. Tested on 2.6.24.5. Thanks!

Tested-by Tony Battersby <[email protected]>

Tony

[email protected]

2008-05-01 08:39 UTC #14 lore raw

Re: Re: [PATCH] more ZERO_PAGE handling ( was 2.6.24 regression: deadlock on coredump of big process)

Quoted 4 lines

This patch fixes the deadlock. Tested on 2.6.24.5. Thanks!
Tested-by Tony Battersby <[email protected]>

thank you for test. I'll post this again when I'm back.

Regards,
-Kame