On Thu, Jun 18, 2020 at 6:15 PM Roman Gushchin <guro@xxxxxx> wrote:
>
> On Thu, Jun 18, 2020 at 05:46:24PM -0700, Yang Shi wrote:
> > On Thu, Jun 18, 2020 at 5:19 PM Roman Gushchin <guro@xxxxxx> wrote:
> > >
> > > Hi!
> > >
> > > I was consistently hitting a VM_BUG_ON_PAGE() in split_huge_page_to_list()
> > > when running vanilla 5.8-rc1 on my desktop. It was happening on every boot
> > > during the system start. I haven't seen this issue on 5.7.
> > >
> > > It looks like split_huge_page() expects the page to be locked,
> > > but it hasn't been changed from 5.7. I do not see any suspicious
> > > commits around the call side either.
> > >
> > > I've tried the following patch:
> > >
> > > --
> > > From 4af38fbf06a9354fadf22a78f1a42dfbb24fbc3a Mon Sep 17 00:00:00 2001
> > > From: Roman Gushchin <guro@xxxxxx>
> > > Date: Thu, 18 Jun 2020 16:33:47 -0700
> > > Subject: [PATCH] iommu/dma: lock page before calling split_huge_page()
> > >
> > > split_huge_page() expects a locked page. The following stacktrace
> > > is generated if debug is on. Fix this by locking the page before
> > > passing it to split_huge_page().
> > >
> > > [ 24.861385] page:ffffef044fb1fa00 refcount:1 mapcount:0 mapping:0000000000000000 index:0x0 head:ffffef044fb1fa00 order:2 compound_mapcount:0 compound_pincount:0
> > > [ 24.861389] flags: 0x17ffffc0010000(head)
> > > [ 24.861393] raw: 0017ffffc0010000 dead000000000100 dead000000000122 0000000000000000
> > > [ 24.861395] raw: 0000000000000000 0000000000000000 00000001ffffffff 0000000000000000
> > > [ 24.861396] page dumped because: VM_BUG_ON_PAGE(!PageLocked(head))
> > > [ 24.861411] ------------[ cut here ]------------
> > > [ 24.861413] kernel BUG at mm/huge_memory.c:2613!
> > > [ 24.861428] invalid opcode: 0000 [#1] SMP NOPTI
> > > [ 24.861432] CPU: 10 PID: 1505 Comm: pulseaudio Not tainted 5.8.0-rc1+ #689
> > > [ 24.861433] Hardware name: Gigabyte Technology Co., Ltd. AB350-Gaming/AB350-Gaming-CF, BIOS F25 01/16/2019
> > > [ 24.861441] RIP: 0010:split_huge_page_to_list+0x731/0xae0
> > > [ 24.861444] Code: 44 00 00 8b 47 34 85 c0 0f 84 b4 02 00 00 f0 ff 4f 34 75 c2 e8 e0 12 f7 ff eb bb 48 c7 c6 d0 16 39 ba 4c 89 c7 e8 ef 85 f9 ff <0f> 0b 48 c7 44 24 10 ff ff ff ff 31 db e9 bb fa ff ff 48 8b 7c 24
> > > [ 24.861446] RSP: 0018:ffffc1030254bb50 EFLAGS: 00010286
> > > [ 24.861449] RAX: 0000000000000000 RBX: 0000000000000002 RCX: ffff9b54cee98d08
> > > [ 24.861451] RDX: 00000000ffffffd8 RSI: 0000000000000000 RDI: ffff9b54cee98d00
> > > [ 24.861452] RBP: ffffef044fb1fa00 R08: 0000000000000547 R09: 0000000000000003
> > > [ 24.861454] R10: 0000000000000000 R11: 0000000000000001 R12: ffff9b54df37f188
> > > [ 24.861455] R13: ffff9b54df355000 R14: ffffef044fb1fa00 R15: ffffef044fb1fa00
> > > [ 24.861458] FS: 00007fd2dc132880(0000) GS:ffff9b54cee80000(0000) knlGS:0000000000000000
> > > [ 24.861460] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
> > > [ 24.861461] CR2: 00007fd2cb100000 CR3: 00000003feb16000 CR4: 00000000003406e0
> > > [ 24.861464] Call Trace:
> > > [ 24.861473] ? __mod_lruvec_state+0x41/0xf0
> > > [ 24.861478] ? __alloc_pages_nodemask+0x15c/0x320
> > > [ 24.861483] iommu_dma_alloc+0x316/0x580
> > > [ 24.861496] snd_dma_alloc_pages+0xdf/0x160 [snd_pcm]
> > > [ 24.861508] snd_dma_alloc_pages_fallback+0x5d/0x80 [snd_pcm]
> > > [ 24.861516] snd_malloc_sgbuf_pages+0x166/0x380 [snd_pcm]
> > > [ 24.861523] ? snd_pcm_hw_refine+0x29d/0x310 [snd_pcm]
> > > [ 24.861529] ? _cond_resched+0x16/0x40
> > > [ 24.861535] snd_dma_alloc_pages+0x64/0x160 [snd_pcm]
> > > [ 24.861542] snd_pcm_lib_malloc_pages+0x136/0x1d0 [snd_pcm]
> > > [ 24.861550] ? snd_pcm_lib_ioctl+0x167/0x210 [snd_pcm]
> > > [ 24.861556] snd_pcm_hw_params+0x3c0/0x490 [snd_pcm]
> > > [ 24.861563] snd_pcm_common_ioctl+0x1c5/0x1110 [snd_pcm]
> > > [ 24.861571] ? snd_pcm_info_user+0x64/0x80 [snd_pcm]
> > > [ 24.861578] snd_pcm_ioctl+0x23/0x30 [snd_pcm]
> > > [ 24.861583] ksys_ioctl+0x82/0xc0
> > > [ 24.861587] __x64_sys_ioctl+0x16/0x20
> > > [ 24.861593] do_syscall_64+0x4d/0x90
> > > [ 24.861597] entry_SYSCALL_64_after_hwframe+0x44/0xa9
> > >
> > > Signed-off-by: Roman Gushchin <guro@xxxxxx>
> > > ---
> > > drivers/iommu/dma-iommu.c | 12 ++++++++++--
> > > 1 file changed, 10 insertions(+), 2 deletions(-)
> > >
> > > diff --git a/drivers/iommu/dma-iommu.c b/drivers/iommu/dma-iommu.c
> > > index 4959f5df21bd..31e4e305d8d5 100644
> > > --- a/drivers/iommu/dma-iommu.c
> > > +++ b/drivers/iommu/dma-iommu.c
> > > @@ -24,6 +24,7 @@
> > > #include <linux/scatterlist.h>
> > > #include <linux/vmalloc.h>
> > > #include <linux/crash_dump.h>
> > > +#include <linux/pagemap.h>
> > >
> > > struct iommu_dma_msi_page {
> > > struct list_head list;
> > > @@ -549,8 +550,15 @@ static struct page **__iommu_dma_alloc_pages(struct device *dev,
> > > if (!PageCompound(page)) {
> > > split_page(page, order);
> > > break;
> > > - } else if (!split_huge_page(page)) {
> > > - break;
> > > + } else {
> > > + int err;
> > > +
> > > + lock_page(page);
> > > + err = split_huge_page(page);
> > > + unlock_page(page);
> >
> > Yes, THP split does need the page locked, in addition it needs the
> > caller hold a pin on the page too (refcount bump).
> >
> > But, I don't get how the code could even really work by a quick look.
> > Actually split_huge_page() assumes the passed in THP is user THP (anon
> > or file cache) and the order is PMD order However, it looks the iommu
> > driver just wants to allocate a bunch of base pages by allocating a
> > huge page (could by any order if I read the code correctly) then split
> > them to base pages. I don't think this is the correct approach IMO.
> > Anyway I'm not iommu expert, if I miss anything please feel free to
> > let me know.
>
> I agree. The whole
>
> page = alloc_pages_node(nid, alloc_flags, order);
> if (!page)
> continue;
> if (!order)
> break;
> if (!PageCompound(page)) {
> split_page(page, order);
> break;
> } else if (!split_huge_page(page)) {
> break;
> }
>
> looks very suspicious to me.
> My wild guess is that gfp flags changed somewhere above, so we hit
> the branch which was never hit before.
It seems so. The page flag has PG_head set then hit the branch, but
the order is 2. I guess __GFP_COMP is set in gfp flag somewhere.
