On Fri, Sep 09, 2022 at 12:21:32PM +0200, Vlastimil Babka wrote: > > On 9/6/22 17:11, Vlastimil Babka wrote: > > On 9/6/22 16:56, Hyeonggon Yoo wrote: > >> On Tue, Sep 06, 2022 at 03:51:01PM +0800, kernel test robot wrote: > >>> Greeting, > >>> > >>> FYI, we noticed the following commit (built with gcc-11): > >>> > >>> commit: 3c4cafa313d978b31a1d5dc17c323074b19a1d63 ("mm/sl[au]b: rearrange > >>> struct slab fields to allow larger rcu_head") > >>> git://git.kernel.org/cgit/linux/kernel/git/vbabka/slab.git > >>> for-6.1/fit_rcu_head > >>> > >>> in testcase: fio-basic > >>> version: fio-x86_64-3.15-1_20220903 > >>> with following parameters: > >>> > >>>     disk: 2pmem > >>>     fs: xfs > >>>     runtime: 200s > >>>     nr_task: 50% > >>>     time_based: tb > >>>     rw: randrw > >>>     bs: 2M > >>>     ioengine: mmap > >>>     test_size: 200G > >>>     cpufreq_governor: performance > >>> > >>> test-description: Fio is a tool that will spawn a number of threads or > >>> processes doing a particular type of I/O action as specified by the user. > >>> test-url:https://github.com/axboe/fio > >>> > >>> > >>> on test machine: 96 threads 2 sockets Intel(R) Xeon(R) Gold 6252 CPU @ > >>> 2.10GHz (Cascade Lake) with 512G memory > >>> > >>> caused below changes (please refer to attached dmesg/kmsg for entire > >>> log/backtrace): > >>> > >>> > >>> [  304.700893][   C40] perf: interrupt took too long (12747 > 12477), > >>> lowering kernel.perf_event_max_sample_rate to 15000 > >>> [  305.015834][   C40] perf: interrupt took too long (15947 > 15933), > >>> lowering kernel.perf_event_max_sample_rate to 12000 > >>> [  305.954702][   C40] perf: interrupt took too long (19968 > 19933), > >>> lowering kernel.perf_event_max_sample_rate to 10000 > >>> [  309.554949][   C31] perf: interrupt took too long (25118 > 24960), > >>> lowering kernel.perf_event_max_sample_rate to 7000 > >>> [  315.068744][   C95] sched: RT throttling activated > >>> [  317.121806][  T590] general protection fault, probably for > >>> non-canonical address 0xdead000000000120: 0000 [#1] SMP NOPTI > >>> [  317.133291][  T590] CPU: 61 PID: 590 Comm: kcompactd0 Tainted: G > >>> S                 6.0.0-rc2-00002-g3c4cafa313d9 #1 > >>> [  317.144084][  T590] Hardware name: Intel Corporation > >>> S2600WFT/S2600WFT, BIOS SE5C620.86B.02.01.0008.031920191559 03/19/2019 > >>> [ 317.155668][ T590] RIP: 0010:isolate_movable_page (mm/migrate.c:103) > >>> [ 317.162016][ T590] Code: ba 28 00 0f 82 88 00 00 00 48 89 ef e8 e2 3a > >>> f8 ff 84 c0 74 74 48 8b 45 00 a9 00 00 04 00 75 69 48 8b 45 18 44 89 e6 > >>> 48 89 ef <48> 8b 40 fe ff d0 0f 1f 00 84 c0 74 52 48 8b 45 00 a9 00 00 04 00 > >>> All code > >>> ======== > >>>     0:    ba 28 00 0f 82           mov    $0x820f0028,%edx > >>>     5:    88 00                    mov    %al,(%rax) > >>>     7:    00 00                    add    %al,(%rax) > >>>     9:    48 89 ef                 mov    %rbp,%rdi > >>>     c:    e8 e2 3a f8 ff           callq  0xfffffffffff83af3 > >>>    11:    84 c0                    test   %al,%al > >>>    13:    74 74                    je     0x89 > >>>    15:    48 8b 45 00              mov    0x0(%rbp),%rax > >>>    19:    a9 00 00 04 00           test   $0x40000,%eax > >>>    1e:    75 69                    jne    0x89 > >>>    20:    48 8b 45 18              mov    0x18(%rbp),%rax > >>>    24:    44 89 e6                 mov    %r12d,%esi > >>>    27:    48 89 ef                 mov    %rbp,%rdi > >>>    2a:*    48 8b 40 fe              mov    -0x2(%rax),%rax        <-- > >>> trapping instruction > >>>    2e:    ff d0                    callq  *%rax > >>>    30:    0f 1f 00                 nopl   (%rax) > >>>    33:    84 c0                    test   %al,%al > >>>    35:    74 52                    je     0x89 > >>>    37:    48 8b 45 00              mov    0x0(%rbp),%rax > >>>    3b:    a9 00 00 04 00           test   $0x40000,%eax > >>> > >>> Code starting with the faulting instruction > >>> =========================================== > >>>     0:    48 8b 40 fe              mov    -0x2(%rax),%rax > >>>     4:    ff d0                    callq  *%rax > >>>     6:    0f 1f 00                 nopl   (%rax) > >>>     9:    84 c0                    test   %al,%al > >>>     b:    74 52                    je     0x5f > >>>     d:    48 8b 45 00              mov    0x0(%rbp),%rax > >>>    11:    a9 00 00 04 00           test   $0x40000,%eax > >>> [  317.182354][  T590] RSP: 0018:ffffc9000e1d3c78 EFLAGS: 00010246 > >>> [  317.188668][  T590] RAX: dead000000000122 RBX: ffffea0004031034 RCX: > >>> 000000000000000c > >>> [  317.196890][  T590] RDX: dead000000000101 RSI: 000000000000000c RDI: > >>> ffffea0004031000 > >>> [  317.205273][  T590] RBP: ffffea0004031000 R08: 0000000004031000 R09: > >>> 0000000000000004 > >>> [  317.213752][  T590] R10: 00000000000066b6 R11: 0000000000000004 R12: > >>> 000000000000000c > >>> [  317.222384][  T590] R13: ffffea0004031000 R14: 0000000000100c40 R15: > >>> ffffc9000e1d3df0 > >>> [  317.230679][  T590] FS:  0000000000000000(0000) > >>> GS:ffff88c04ff40000(0000) knlGS:0000000000000000 > >>> [  317.239896][  T590] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033 > >>> [  317.247098][  T590] CR2: 0000000000451c00 CR3: 0000008064ca4002 CR4: > >>> 00000000007706e0 > >>> [  317.255788][  T590] DR0: 0000000000000000 DR1: 0000000000000000 DR2: > >>> 0000000000000000 > >>> [  317.264256][  T590] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: > >>> 0000000000000400 > >>> [  317.272772][  T590] PKRU: 55555554 > >>> [  317.276783][  T590] Call Trace: > >>> [  317.280932][  T590]  > >>> [ 317.284315][ T590] isolate_migratepages_block (mm/compaction.c:982) > >>> [ 317.290702][ T590] isolate_migratepages (mm/compaction.c:1960) > >>> [ 317.296278][ T590] compact_zone (mm/compaction.c:2393) > >>> [ 317.301202][ T590] proactive_compact_node (mm/compaction.c:2661 > >>> (discriminator 2)) > >> Hmm... Let's debug. > >> > >> FYI, simply echo 1 > /proc/sys/vm/compact_memory invokes same bug on my test > >> environment. > >> > >> the 'mops' is invalid address in mm/migrate.c:103. > >> > >> Hmm, why is this slab page confused as movable page? > >> -> Because page->'mapping' and slab->slabs field has same offset. > >> > >> I think this is invoked because lowest two bits of slab->slabs is not 0. > >> > >> Vlastimil, any thoughts? > > > > Yeah, slabs->slabs could do that, and the remedy would be to exchange it > > with the slab->next field. > > However the report points to the value dead000000000122 which is > > LIST_POISON2, which unfortunately contains the lower bit after 4c6080cd6f8b > > ("lib/list: tweak LIST_POISON2 for better code generation on x86_64") > > > > Probably the simplest fix would be to check for PageSlab() before > > __PageMovable(). > > So I've done with the patch below, that I added to the for-6.1/fit_rcu_head > branch in slab.git. It's not very nice though with all the new membarriers. > I hope it's at least correct... > > > But heads up for Joel - if your rcu_head debugging info series (didn't > > check) has something like a counter in the 3rd 64bit word, where bit 1 can > > thus be set, it can cause the same issue fooling the __PageMovable() check. > > ----8<---- > From d6f9fbb33b908eb8162cc1f6ce7f7c970d0f285f Mon Sep 17 00:00:00 2001 > From: Vlastimil Babka > Date: Fri, 9 Sep 2022 12:03:10 +0200 > Subject: [PATCH 2/3] mm/migrate: make isolate_movable_page() skip slab pages > > In the next commit we want to rearrange struct slab fields to allow a > larger rcu_head. Afterwards, the page->mapping field will overlap > with SLUB's "struct list_head slab_list", where the value of prev > pointer can become LIST_POISON2, which is 0x122 + POISON_POINTER_DELTA. > Unfortunately the bit 1 being set can confuse PageMovable() to be a > false positive and cause a GPF as reported by lkp [1]. > > To fix this, make isolate_movable_page() skip pages with the PageSlab > flag set. This is a bit tricky as we need to add memory barriers to SLAB > and SLUB's page allocation and freeing, and their counterparts to > isolate_movable_page(). Hello, I just took a quick grasp, Is this approach okay with folio_test_anon()? > > [1] https://lore.kernel.org/all/208c1757-5edd-fd42-67d4-1940cc43b50f(a)intel.com/ > > Reported-by: kernel test robot > Signed-off-by: Vlastimil Babka > --- > mm/compaction.c | 2 +- > mm/migrate.c | 12 +++++++++++- > mm/slab.c | 6 +++++- > mm/slub.c | 6 +++++- > 4 files changed, 22 insertions(+), 4 deletions(-) > > diff --git a/mm/compaction.c b/mm/compaction.c > index 640fa76228dd..b697c207beec 100644 > --- a/mm/compaction.c > +++ b/mm/compaction.c > @@ -972,7 +972,7 @@ isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn, > * __PageMovable can return false positive so we need > * to verify it under page_lock. > */ > - if (unlikely(__PageMovable(page)) && > + if (unlikely(!PageSlab(page) && __PageMovable(page)) && > !PageIsolated(page)) { > if (locked) { > unlock_page_lruvec_irqrestore(locked, flags); > diff --git a/mm/migrate.c b/mm/migrate.c > index 6a1597c92261..7f661b45d431 100644 > --- a/mm/migrate.c > +++ b/mm/migrate.c > @@ -78,7 +78,7 @@ int isolate_movable_page(struct page *page, isolate_mode_t mode) > * assumes anybody doesn't touch PG_lock of newly allocated page > * so unconditionally grabbing the lock ruins page's owner side. > */ > - if (unlikely(!__PageMovable(page))) > + if (unlikely(!__PageMovable(page) || PageSlab(page))) > goto out_putpage; > /* > * As movable pages are not isolated from LRU lists, concurrent > @@ -94,9 +94,19 @@ int isolate_movable_page(struct page *page, isolate_mode_t mode) > if (unlikely(!trylock_page(page))) > goto out_putpage; > > + if (unlikely(PageSlab(page))) > + goto out_no_isolated; > + /* Pairs with smp_wmb() in slab freeing, e.g. SLUB's __free_slab() */ > + smp_rmb(); > + > if (!PageMovable(page) || PageIsolated(page)) > goto out_no_isolated; > > + /* Pairs with smp_wmb() in slab allocation, e.g. SLUB's alloc_slab_page() */ > + smp_rmb(); > + if (unlikely(PageSlab(page))) > + goto out_no_isolated; > + > mops = page_movable_ops(page); > VM_BUG_ON_PAGE(!mops, page); > > diff --git a/mm/slab.c b/mm/slab.c > index 10e96137b44f..25e9a6ef4f74 100644 > --- a/mm/slab.c > +++ b/mm/slab.c > @@ -1370,6 +1370,8 @@ static struct slab *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, > > account_slab(slab, cachep->gfporder, cachep, flags); > __folio_set_slab(folio); > + /* Make the flag visible before any changes to folio->mapping */ > + smp_wmb(); > /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */ > if (sk_memalloc_socks() && page_is_pfmemalloc(folio_page(folio, 0))) > slab_set_pfmemalloc(slab); > @@ -1387,9 +1389,11 @@ static void kmem_freepages(struct kmem_cache *cachep, struct slab *slab) > > BUG_ON(!folio_test_slab(folio)); > __slab_clear_pfmemalloc(slab); > - __folio_clear_slab(folio); > page_mapcount_reset(folio_page(folio, 0)); > folio->mapping = NULL; > + /* Make the mapping reset visible before clearing the flag */ > + smp_wmb(); > + __folio_clear_slab(folio); > > if (current->reclaim_state) > current->reclaim_state->reclaimed_slab += 1 << order; > diff --git a/mm/slub.c b/mm/slub.c > index d86be1b0d09f..2f9cb6e67de3 100644 > --- a/mm/slub.c > +++ b/mm/slub.c > @@ -1830,6 +1830,8 @@ static inline struct slab *alloc_slab_page(gfp_t flags, int node, > > slab = folio_slab(folio); > __folio_set_slab(folio); > + /* Make the flag visible before any changes to folio->mapping */ > + smp_wmb(); > if (page_is_pfmemalloc(folio_page(folio, 0))) > slab_set_pfmemalloc(slab); > > @@ -2037,8 +2039,10 @@ static void __free_slab(struct kmem_cache *s, struct slab *slab) > int pages = 1 << order; > > __slab_clear_pfmemalloc(slab); > - __folio_clear_slab(folio); > folio->mapping = NULL; > + /* Make the mapping reset visible before clearing the flag */ > + smp_wmb(); > + __folio_clear_slab(folio); > if (current->reclaim_state) > current->reclaim_state->reclaimed_slab += pages; > unaccount_slab(slab, order, s); > -- > 2.37.3 > > > -- Thanks, Hyeonggon