On Fri 20-07-12 15:11:08, Mel Gorman wrote:
> Sorry for the resend, I did not properly refresh Cong Wang's suggested
> fix. This V2 is still the mmap_sem approach that fixes a potential deadlock
> problem pointed out by Michal.
>
> Changelog since V1
> o Correct cut&paste error in race description (hugh)
> o Handle potential deadlock during fork (mhocko)
> o Reorder unlocking (wangcong)
>
> If a process creates a large hugetlbfs mapping that is eligible for page
> table sharing and forks heavily with children some of whom fault and
> others which destroy the mapping then it is possible for page tables to
> get corrupted. Some teardowns of the mapping encounter a "bad pmd" and
> output a message to the kernel log. The final teardown will trigger a
> BUG_ON in mm/filemap.c.
>
> This was reproduced in 3.4 but is known to have existed for a long time
> and goes back at least as far as 2.6.37. It was probably was introduced in
> 2.6.20 by [39dde65c: shared page table for hugetlb page]. The messages
> look like this;
>
> [ ..........] Lots of bad pmd messages followed by this
> [ 127.164256] mm/memory.c:391: bad pmd ffff880412e04fe8(80000003de4000e7).
> [ 127.164257] mm/memory.c:391: bad pmd ffff880412e04ff0(80000003de6000e7).
> [ 127.164258] mm/memory.c:391: bad pmd ffff880412e04ff8(80000003de0000e7).
> [ 127.186778] ------------[ cut here ]------------
> [ 127.186781] kernel BUG at mm/filemap.c:134!
> [ 127.186782] invalid opcode: 0000 [#1] SMP
> [ 127.186783] CPU 7
> [ 127.186784] Modules linked in: af_packet cpufreq_conservative
> cpufreq_userspace cpufreq_powersave acpi_cpufreq mperf ext3 jbd dm_mod
> coretemp crc32c_intel usb_storage ghash_clmulni_intel aesni_intel i2c_i801
> r8169 mii uas sr_mod cdrom sg iTCO_wdt iTCO_vendor_support shpchp serio_raw
> cryptd aes_x86_64 e1000e pci_hotplug dcdbas aes_generic container microcode
> ext4 mbcache jbd2 crc16 sd_mod crc_t10dif i915 drm_kms_helper drm
> i2c_algo_bit ehci_hcd ahci libahci usbcore rtc_cmos usb_common button
> i2c_core intel_agp video intel_gtt fan processor thermal thermal_sys hwmon
> ata_generic pata_atiixp libata scsi_mod
> [ 127.186801]
> [ 127.186802] Pid: 9017, comm: hugetlbfs-test Not tainted 3.4.0-autobuild
> #53 Dell Inc. OptiPlex 990/06D7TR
> [ 127.186804] RIP: 0010:[<ffffffff810ed6ce>] [<ffffffff810ed6ce>]
> __delete_from_page_cache+0x15e/0x160
> [ 127.186809] RSP: 0000:ffff8804144b5c08 EFLAGS: 00010002
> [ 127.186810] RAX: 0000000000000001 RBX: ffffea000a5c9000 RCX:
> 00000000ffffffc0
> [ 127.186811] RDX: 0000000000000000 RSI: 0000000000000009 RDI:
> ffff88042dfdad00
> [ 127.186812] RBP: ffff8804144b5c18 R08: 0000000000000009 R09:
> 0000000000000003
> [ 127.186813] R10: 0000000000000000 R11: 000000000000002d R12:
> ffff880412ff83d8
> [ 127.186814] R13: ffff880412ff83d8 R14: 0000000000000000 R15:
> ffff880412ff83d8
> [ 127.186815] FS: 00007fe18ed2c700(0000) GS:ffff88042dce0000(0000)
> knlGS:0000000000000000
> [ 127.186816] CS: 0010 DS: 0000 ES: 0000 CR0: 000000008005003b
> [ 127.186817] CR2: 00007fe340000503 CR3: 0000000417a14000 CR4:
> 00000000000407e0
> [ 127.186818] DR0: 0000000000000000 DR1: 0000000000000000 DR2:
> 0000000000000000
> [ 127.186819] DR3: 0000000000000000 DR6: 00000000ffff0ff0 DR7:
> 0000000000000400
> [ 127.186820] Process hugetlbfs-test (pid: 9017, threadinfo
> ffff8804144b4000, task ffff880417f803c0)
> [ 127.186821] Stack:
> [ 127.186822] ffffea000a5c9000 0000000000000000 ffff8804144b5c48
> ffffffff810ed83b
> [ 127.186824] ffff8804144b5c48 000000000000138a 0000000000001387
> ffff8804144b5c98
> [ 127.186825] ffff8804144b5d48 ffffffff811bc925 ffff8804144b5cb8
> 0000000000000000
> [ 127.186827] Call Trace:
> [ 127.186829] [<ffffffff810ed83b>] delete_from_page_cache+0x3b/0x80
> [ 127.186832] [<ffffffff811bc925>] truncate_hugepages+0x115/0x220
> [ 127.186834] [<ffffffff811bca43>] hugetlbfs_evict_inode+0x13/0x30
> [ 127.186837] [<ffffffff811655c7>] evict+0xa7/0x1b0
> [ 127.186839] [<ffffffff811657a3>] iput_final+0xd3/0x1f0
> [ 127.186840] [<ffffffff811658f9>] iput+0x39/0x50
> [ 127.186842] [<ffffffff81162708>] d_kill+0xf8/0x130
> [ 127.186843] [<ffffffff81162812>] dput+0xd2/0x1a0
> [ 127.186845] [<ffffffff8114e2d0>] __fput+0x170/0x230
> [ 127.186848] [<ffffffff81236e0e>] ? rb_erase+0xce/0x150
> [ 127.186849] [<ffffffff8114e3ad>] fput+0x1d/0x30
> [ 127.186851] [<ffffffff81117db7>] remove_vma+0x37/0x80
> [ 127.186853] [<ffffffff81119182>] do_munmap+0x2d2/0x360
> [ 127.186855] [<ffffffff811cc639>] sys_shmdt+0xc9/0x170
> [ 127.186857] [<ffffffff81410a39>] system_call_fastpath+0x16/0x1b
> [ 127.186858] Code: 0f 1f 44 00 00 48 8b 43 08 48 8b 00 48 8b 40 28 8b b0 40
> 03 00 00 85 f6 0f 88 df fe ff ff 48 89 df e8 e7 cb 05 00 e9 d2 fe ff ff <0f>
> 0b 55 83 e2 fd 48 89 e5 48 83 ec 30 48 89 5d d8 4c 89 65 e0
> [ 127.186868] RIP [<ffffffff810ed6ce>] __delete_from_page_cache+0x15e/0x160
> [ 127.186870] RSP <ffff8804144b5c08>
> [ 127.186871] ---[ end trace 7cbac5d1db69f426 ]---
>
> The bug is a race and not always easy to reproduce. To reproduce it I was
> doing the following on a single socket I7-based machine with 16G of RAM.
>
> $ hugeadm --pool-pages-max DEFAULT:13G
> $ echo $((18*1048576*1024)) > /proc/sys/kernel/shmmax
> $ echo $((18*1048576*1024)) > /proc/sys/kernel/shmall
> $ for i in `seq 1 9000`; do ./hugetlbfs-test; done
>
> On my particular machine, it usually triggers within 10 minutes but enabling
> debug options can change the timing such that it never hits. Once the bug is
> triggered, the machine is in trouble and needs to be rebooted. The machine
> will respond but processes accessing proc like "ps aux" will hang due to
> the BUG_ON. shutdown will also hang and needs a hard reset or a sysrq-b.
>
> The test case was mostly written by Michal Hocko with a few minor changes
> by me to reproduce this bug. Michal did a lot of heavy lifting eliminating
> possible sources of the race and saved me the embarrassment of posting a
> completely broken patch yesterday. He did not see this patch before
> going to the lists so any critical flaws are all mine!
>
> The basic problem is a race between page table sharing and teardown. For
> the most part page table sharing depends on i_mmap_mutex. In some cases,
> it is also taking the mm->page_table_lock for the PTE updates but with
> shared page tables, it is the i_mmap_mutex that is more important.
>
> Unfortunately it appears to be also insufficient. Consider the following
> situation
>
> Process A Process B
> --------- ---------
> hugetlb_fault shmdt
> LockWrite(mmap_sem)
> do_munmap
> unmap_region
> unmap_vmas
> unmap_single_vma
> unmap_hugepage_range
>
> Lock(i_mmap_mutex)
>
> Lock(mm->page_table_lock)
>
> huge_pmd_unshare/unmap tables <--- (1)
>
> Unlock(mm->page_table_lock)
>
> Unlock(i_mmap_mutex)
> huge_pte_alloc ...
> Lock(i_mmap_mutex) ...
> vma_prio_walk, find svma, spte ...
> Lock(mm->page_table_lock) ...
> share spte ...
> Unlock(mm->page_table_lock) ...
> Unlock(i_mmap_mutex) ...
> hugetlb_no_page
> <--- (2)
> free_pgtables
> unlink_file_vma
> hugetlb_free_pgd_range
> remove_vma_list
>
> In this scenario, it is possible for Process A to share page tables with
> Process B that is trying to tear them down. The i_mmap_mutex on its own
> does not prevent Process A walking Process B's page tables. At (1) above,
> the page tables are not shared yet so it unmaps the PMDs. Process A sets
> up page table sharing and at (2) faults a new entry. Process B then trips
> up on it in free_pgtables.
>
> This patch takes the mmap_sem for read and then the page_table_lock of
> address spaces being considered for page table sharing. I verified that
> page table sharing still occurs using the awesome technology of printk
> to spit out a message when huge_pmd_share is successful. libhugetlbfs
> regression test suite passed.
>
> I strongly suggest this be treated as a -stable candidate if it is merged.
>
> Test program is as follows.
>
> ==== CUT HERE ====
>
> static size_t huge_page_size = (2UL << 20);
> static size_t nr_huge_page_A = 512;
> static size_t nr_huge_page_B = 5632;
>
> unsigned int get_random(unsigned int max)
> {
> struct timeval tv;
>
> gettimeofday(&tv, NULL);
> srandom(tv.tv_usec);
> return random() % max;
> }
>
> static void play(void *addr, size_t size)
> {
> unsigned char *start = addr,
> *end = start + size,
> *a;
> start += get_random(size/2);
>
> /* we could itterate on huge pages but let's give it more time. */
> for (a = start; a < end; a += 4096)
> *a = 0;
> }
>
> int main(int argc, char **argv)
> {
> key_t key = IPC_PRIVATE;
> size_t sizeA = nr_huge_page_A * huge_page_size;
> size_t sizeB = nr_huge_page_B * huge_page_size;
> int shmidA, shmidB;
> void *addrA = NULL, *addrB = NULL;
> int nr_children = 300, n = 0;
>
> if ((shmidA = shmget(key, sizeA, IPC_CREAT|SHM_HUGETLB|0660)) == -1) {
> perror("shmget:");
> return 1;
> }
>
> if ((addrA = shmat(shmidA, addrA, SHM_R|SHM_W)) == (void *)-1UL) {
> perror("shmat");
> return 1;
> }
> if ((shmidB = shmget(key, sizeB, IPC_CREAT|SHM_HUGETLB|0660)) == -1) {
> perror("shmget:");
> return 1;
> }
>
> if ((addrB = shmat(shmidB, addrB, SHM_R|SHM_W)) == (void *)-1UL) {
> perror("shmat");
> return 1;
> }
>
> fork_child:
> switch(fork()) {
> case 0:
> switch (n%3) {
> case 0:
> play(addrA, sizeA);
> break;
> case 1:
> play(addrB, sizeB);
> break;
> case 2:
> break;
> }
> break;
> case -1:
> perror("fork:");
> break;
> default:
> if (++n < nr_children)
> goto fork_child;
> play(addrA, sizeA);
> break;
> }
> shmdt(addrA);
> shmdt(addrB);
> do {
> wait(NULL);
> } while (--n > 0);
> shmctl(shmidA, IPC_RMID, NULL);
> shmctl(shmidB, IPC_RMID, NULL);
> return 0;
> }
>
> Signed-off-by: Mel Gorman <[email protected]>
Yes this looks correct. mmap_sem will make sure that unmap_vmas and
free_pgtables are executed atomicaly wrt. huge_pmd_share so it doesn't
see non-NULL spte on the way out. I am just wondering whether we need
the page_table_lock as well. It is not harmful but I guess we can drop
it because both exit_mmap and shmdt are not taking it and mmap_sem is
sufficient for them.
One more nit bellow.
I will send my version of the fix which took another path as a reply to
this email to have something to compare with.
[...]
> diff --git a/arch/x86/mm/hugetlbpage.c b/arch/x86/mm/hugetlbpage.c
> index f6679a7..944b2df 100644
> --- a/arch/x86/mm/hugetlbpage.c
> +++ b/arch/x86/mm/hugetlbpage.c
> @@ -58,7 +58,8 @@ static int vma_shareable(struct vm_area_struct *vma,
> unsigned long addr)
> /*
> * search for a shareable pmd page for hugetlb.
> */
> -static void huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t
> *pud)
> +static void huge_pmd_share(struct mm_struct *mm, struct mm_struct *locked_mm,
> + unsigned long addr, pud_t *pud)
> {
> struct vm_area_struct *vma = find_vma(mm, addr);
> struct address_space *mapping = vma->vm_file->f_mapping;
> @@ -68,14 +69,40 @@ static void huge_pmd_share(struct mm_struct *mm, unsigned
> long addr, pud_t *pud)
> struct vm_area_struct *svma;
> unsigned long saddr;
> pte_t *spte = NULL;
> + spinlock_t *spage_table_lock = NULL;
> + struct rw_semaphore *smmap_sem = NULL;
>
> if (!vma_shareable(vma, addr))
> return;
>
> +retry:
> mutex_lock(&mapping->i_mmap_mutex);
> vma_prio_tree_foreach(svma, &iter, &mapping->i_mmap, idx, idx) {
> if (svma == vma)
> continue;
> + if (svma->vm_mm == vma->vm_mm)
> + continue;
> +
> + /*
> + * The target mm could be in the process of tearing down
> + * its page tables and the i_mmap_mutex on its own is
> + * not sufficient. To prevent races against teardown and
> + * pagetable updates, we acquire the mmap_sem and pagetable
> + * lock of the remote address space. down_read_trylock()
> + * is necessary as the other process could also be trying
> + * to share pagetables with the current mm. In the fork
> + * case, we are already both mm's so check for that
> + */
> + if (locked_mm != svma->vm_mm) {
> + if (!down_read_trylock(&svma->vm_mm->mmap_sem)) {
> + mutex_unlock(&mapping->i_mmap_mutex);
> + goto retry;
> + }
> + smmap_sem = &svma->vm_mm->mmap_sem;
> + }
> +
> + spage_table_lock = &svma->vm_mm->page_table_lock;
> + spin_lock_nested(spage_table_lock, SINGLE_DEPTH_NESTING);
>
> saddr = page_table_shareable(svma, vma, addr, idx);
> if (saddr) {
[...]
> diff --git a/mm/hugetlb.c b/mm/hugetlb.c
> index ae8f708..4832277 100644
> --- a/mm/hugetlb.c
> +++ b/mm/hugetlb.c
> @@ -2244,7 +2244,7 @@ int copy_hugetlb_page_range(struct mm_struct *dst,
> struct mm_struct *src,
> src_pte = huge_pte_offset(src, addr);
> if (!src_pte)
> continue;
> - dst_pte = huge_pte_alloc(dst, addr, sz);
> + dst_pte = huge_pte_alloc(dst, src, addr, sz);
> if (!dst_pte)
> goto nomem;
>
> @@ -2745,7 +2745,7 @@ int hugetlb_fault(struct mm_struct *mm, struct
> vm_area_struct *vma,
> VM_FAULT_SET_HINDEX(h - hstates);
> }
>
> - ptep = huge_pte_alloc(mm, address, huge_page_size(h));
> + ptep = huge_pte_alloc(mm, NULL, address, huge_page_size(h));
strictly speaking we should provide current->mm here because we are in
the page fault path and mmap_sem is held for reading. This doesn't
matter here though because huge_pmd_share will take it for reading so
nesting wouldn't hurt. Maybe a small comment that this is intentional
and correct would be nice.
> if (!ptep)
> return VM_FAULT_OOM;
>
>
> --
> Mel Gorman
> SUSE Labs
--
Michal Hocko
SUSE Labs
SUSE LINUX s.r.o.
Lihovarska 1060/12
190 00 Praha 9
Czech Republic
--
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/
