Re: [PATCH v4] page cache: Store only head pages in i_pages

From: Chris Wilson
Date: Sun Jun 02 2019 - 09:15:58 EST


Quoting Matthew Wilcox (2019-06-02 11:51:50)
> Thanks for the reports, Chris.
>
> I think they're both canaries; somehow the page cache / swap cache has
> got corrupted and contains entries that it shouldn't.
>
> This second one (with the VM_BUG_ON_PAGE in __delete_from_swap_cache)
> shows a regular (non-huge) page at index 1. There are two ways we might
> have got there; one is that we asked to delete a page at index 1 which is
> no longer in the cache. The other is that we asked to delete a huge page
> at index 0, but the page wasn't subsequently stored in indices 1-511.
>
> We dump the page that we found; not the page we're looking for, so I don't
> know which. If this one's easy to reproduce, you could add:
>
> for (i = 0; i < nr; i++) {
> void *entry = xas_store(&xas, NULL);
> + if (entry != page) {
> + printk("Oh dear %d %d\n", i, nr);
> + dump_page(page, "deleting page");
> + }

[ 113.423120] Oh dear 0 1
[ 113.423141] page:ffffea000911cdc0 refcount:0 mapcount:0 mapping:ffff88826aee7bb1 index:0x7fce6ff37
[ 113.423146] anon
[ 113.423150] flags: 0x8000000000080445(locked|uptodate|workingset|owner_priv_1|swapbacked)
[ 113.423161] raw: 8000000000080445 dead000000000100 dead000000000200 ffff88826aee7bb1
[ 113.423167] raw: 00000007fce6ff37 0000000000054537 00000000ffffffff 0000000000000000
[ 113.423171] page dumped because: deleting page
[ 113.423176] page:ffffea0009118000 refcount:1 mapcount:0 mapping:ffff88826aee7bb1 index:0x7fce6fe00
[ 113.423182] anon
[ 113.423183] flags: 0x8000000000080454(uptodate|lru|workingset|owner_priv_1|swapbacked)
[ 113.423191] raw: 8000000000080454 ffffea0009118048 ffffea000911ce08 ffff88826aee7bb1
[ 113.423198] raw: 00000007fce6fe00 0000000000054400 00000001ffffffff ffff8882693e5000
[ 113.423204] page dumped because: VM_BUG_ON_PAGE(entry != page)
[ 113.423209] page->mem_cgroup:ffff8882693e5000
[ 113.423222] ------------[ cut here ]------------
[ 113.423227] kernel BUG at mm/swap_state.c:174!
[ 113.423236] invalid opcode: 0000 [#1] PREEMPT SMP DEBUG_PAGEALLOC
[ 113.423243] CPU: 1 PID: 131 Comm: kswapd0 Tainted: G U 5.2.0-rc2+ #251
[ 113.423248] Hardware name: /NUC6CAYB, BIOS AYAPLCEL.86A.0029.2016.1124.1625 11/24/2016
[ 113.423260] RIP: 0010:__delete_from_swap_cache.cold.17+0x30/0x36
[ 113.423265] Code: 48 c7 c7 13 94 bf 81 e8 cd 7f f3 ff 48 89 df 48 c7 c6 24 94 bf 81 e8 95 6c fd ff 48 c7 c6 32 94 bf 81 4c 89 ff e8 86 6c fd ff <0f> 0b 90 90 90 90 48 8b 07 48 8b 16 48 c1 e8 3a 48 c1 ea 3a 29 d0
[ 113.423274] RSP: 0018:ffffc900008b3a80 EFLAGS: 00010046
[ 113.423280] RAX: 0000000000000000 RBX: ffffea000911cdc0 RCX: 0000000000000006
[ 113.423285] RDX: 0000000000000007 RSI: 0000000000000092 RDI: ffff888276c963c0
[ 113.423290] RBP: ffff888265a98d20 R08: 00000000000002ce R09: 0000000000000000
[ 113.423296] R10: 0000000272bc445c R11: 0000000000000000 R12: 0000000000000001
[ 113.423301] R13: 0000000000000000 R14: 0000000000000000 R15: ffffea0009118000
[ 113.423306] FS: 0000000000000000(0000) GS:ffff888276c80000(0000) knlGS:0000000000000000
[ 113.423313] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
[ 113.423317] CR2: 00007fce7c857000 CR3: 0000000002c09000 CR4: 00000000001406e0
[ 113.423323] Call Trace:
[ 113.423331] __remove_mapping+0x1c2/0x380
[ 113.423337] shrink_page_list+0x123c/0x1d00
[ 113.423343] shrink_inactive_list+0x130/0x300
[ 113.423348] shrink_node_memcg+0x20e/0x740
[ 113.423354] shrink_node+0xba/0x420
[ 113.423359] balance_pgdat+0x27d/0x4d0
[ 113.423365] kswapd+0x216/0x300
[ 113.423372] ? wait_woken+0x80/0x80
[ 113.423378] ? balance_pgdat+0x4d0/0x4d0
[ 113.423384] kthread+0x106/0x120
[ 113.423389] ? kthread_create_on_node+0x40/0x40
[ 113.423398] ret_from_fork+0x1f/0x30
[ 113.423405] Modules linked in: i915 intel_gtt drm_kms_helper
[ 113.423414] ---[ end trace 328930613dd77e06 ]---
[ 113.454546] RIP: 0010:__delete_from_swap_cache.cold.17+0x30/0x36

> VM_BUG_ON_PAGE(entry != page, entry);
> set_page_private(page + i, 0);
> xas_next(&xas);
> }
>
> I'll re-read the patch and see if I can figure out how the cache is getting
> screwed up. Given what you said, probably on the swap-in path.

It may be self-incriminating, but this only occurs when i915.ko is also
involved via shrink_slab.
-Chris