linux-kernel.vger.kernel.org archive mirror
 help / color / mirror / Atom feed
From: Yang Shi <shy828301@gmail.com>
To: Hugh Dickins <hughd@google.com>
Cc: Andrew Morton <akpm@linux-foundation.org>,
	Shakeel Butt <shakeelb@google.com>,
	"Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>,
	Miaohe Lin <linmiaohe@huawei.com>,
	Mike Kravetz <mike.kravetz@oracle.com>,
	Michal Hocko <mhocko@suse.com>, Rik van Riel <riel@surriel.com>,
	Christoph Hellwig <hch@infradead.org>,
	Matthew Wilcox <willy@infradead.org>,
	"Eric W. Biederman" <ebiederm@xmission.com>,
	Alexey Gladkov <legion@kernel.org>,
	Chris Wilson <chris@chris-wilson.co.uk>,
	Matthew Auld <matthew.auld@intel.com>,
	Linux FS-devel Mailing List <linux-fsdevel@vger.kernel.org>,
	Linux Kernel Mailing List <linux-kernel@vger.kernel.org>,
	linux-api@vger.kernel.org, Linux MM <linux-mm@kvack.org>
Subject: Re: [PATCH 06/16] huge tmpfs: shmem_is_huge(vma, inode, index)
Date: Wed, 4 Aug 2021 12:01:17 -0700	[thread overview]
Message-ID: <CAHbLzkrvOCCbN3EcDeKwfqWrtU6kH0+7fuSv7aahyjpKtsHn3g@mail.gmail.com> (raw)
In-Reply-To: <8baad8b2-8f7a-2589-ce21-4135a59c5dc6@google.com>

On Wed, Aug 4, 2021 at 1:28 AM Hugh Dickins <hughd@google.com> wrote:
>
> On Mon, 2 Aug 2021, Yang Shi wrote:
> > On Sat, Jul 31, 2021 at 10:22 PM Hugh Dickins <hughd@google.com> wrote:
> > > On Fri, 30 Jul 2021, Yang Shi wrote:
> > > > On Fri, Jul 30, 2021 at 12:42 AM Hugh Dickins <hughd@google.com> wrote:
> > > > >
> > > > > Extend shmem_huge_enabled(vma) to shmem_is_huge(vma, inode, index), so
> > > > > that a consistent set of checks can be applied, even when the inode is
> > > > > accessed through read/write syscalls (with NULL vma) instead of mmaps
> > > > > (the index argument is seldom of interest, but required by mount option
> > > > > "huge=within_size").  Clean up and rearrange the checks a little.
> > > > >
> > > > > This then replaces the checks which shmem_fault() and shmem_getpage_gfp()
> > > > > were making, and eliminates the SGP_HUGE and SGP_NOHUGE modes: while it's
> > > > > still true that khugepaged's collapse_file() at that point wants a small
> > > > > page, the race that might allocate it a huge page is too unlikely to be
> > > > > worth optimizing against (we are there *because* there was at least one
> > > > > small page in the way), and handled by a later PageTransCompound check.
> > > >
> > > > Yes, it seems too unlikely. But if it happens the PageTransCompound
> > > > check may be not good enough since the page allocated by
> > > > shmem_getpage() may be charged to wrong memcg (root memcg). And it
> > > > won't be replaced by a newly allocated huge page so the wrong charge
> > > > can't be undone.
> > >
> > > Good point on the memcg charge: I hadn't thought of that.  Of course
> > > it's not specific to SGP_CACHE versus SGP_NOHUGE (this patch), but I
> > > admit that a huge mischarge is hugely worse than a small mischarge.
> >
> > The small page could be collapsed to a huge page sooner or later, so
> > the mischarge may be transient. But huge page can't be replaced.
>
> You're right, if all goes well, the mischarged small page could be
> collapsed to a correctly charged huge page sooner or later (but all
> may not go well), whereas the mischarged huge page is stuck there.
>
> >
> > >
> > > We could fix it by making shmem_getpage_gfp() non-static, and pointing
> > > to the vma (hence its mm, hence its memcg) here, couldn't we?  Easily
> > > done, but I don't really want to make shmem_getpage_gfp() public just
> > > for this, for two reasons.
> > >
> > > One is that the huge race it just so unlikely; and a mischarge to root
> > > is not the end of the world, so long as it's not reproducible.  It can
> > > only happen on the very first page of the huge extent, and the prior
> >
> > OK, if so the mischarge is not as bad as what I thought in the first place.
> >
> > > "Stop if extent has been truncated" check makes sure there was one
> > > entry in the extent at that point: so the race with hole-punch can only
> > > occur after we xas_unlock_irq(&xas) immediately before shmem_getpage()
> > > looks up the page in the tree (and I say hole-punch not truncate,
> > > because shmem_getpage()'s i_size check will reject when truncated).
> > > I don't doubt that it could happen, but stand by not optimizing against.
> >
> > I agree the race is so unlikely and it may be not worth optimizing
> > against it right now, but a note or a comment may be worth.
>
> Thanks, but despite us agreeing that the race is too unlikely to be worth
> optimizing against, it does still nag at me ever since you questioned it:
> silly, but I can't quite be convinced by my own dismissals.
>
> I do still want to get rid of SGP_HUGE and SGP_NOHUGE, clearing up those
> huge allocation decisions remains the intention; but now think to add
> SGP_NOALLOC for collapse_file() in place of SGP_NOHUGE or SGP_CACHE -
> to rule out that possibility of mischarge after racing hole-punch,
> no matter whether it's huge or small.  If any such race occurs,
> collapse_file() should just give up.
>
> This being the "Stupid me" SGP_READ idea, except that of course would
> not work: because half the point of that block in collapse_file() is
> to initialize the !Uptodate pages, whereas SGP_READ avoids doing so.
>
> There is, of course, the danger that in fixing this unlikely mischarge,
> I've got the code wrong and am introducing a bug: here's what a 17/16
> would look like, though it will be better inserted early.  I got sick
> of all the "if (page "s, and was glad of the opportunity to fix that
> outdated "bring it back from swap" comment - swap got done above.
>
> What do you think? Should I add this in or leave it out?

Thanks for keeping investigating this. The patch looks good to me. I
think we could go this way. Just a nit below.

>
> Thanks,
> Hugh
>
> --- a/include/linux/shmem_fs.h
> +++ b/include/linux/shmem_fs.h
> @@ -108,6 +108,7 @@ extern unsigned long shmem_partial_swap_usage(struct address_space *mapping,
>  /* Flag allocation requirements to shmem_getpage */
>  enum sgp_type {
>         SGP_READ,       /* don't exceed i_size, don't allocate page */
> +       SGP_NOALLOC,    /* like SGP_READ, but do use fallocated page */

The comment looks misleading, it seems SGP_NOALLOC does clear the
Uptodate flag but SGP_READ doesn't. Or it is fine not to distinguish
this difference?

>         SGP_CACHE,      /* don't exceed i_size, may allocate page */
>         SGP_WRITE,      /* may exceed i_size, may allocate !Uptodate page */
>         SGP_FALLOC,     /* like SGP_WRITE, but make existing page Uptodate */
> --- a/mm/khugepaged.c
> +++ b/mm/khugepaged.c
> @@ -1721,7 +1721,7 @@ static void collapse_file(struct mm_struct *mm,
>                                 xas_unlock_irq(&xas);
>                                 /* swap in or instantiate fallocated page */
>                                 if (shmem_getpage(mapping->host, index, &page,
> -                                                 SGP_CACHE)) {
> +                                                 SGP_NOALLOC)) {
>                                         result = SCAN_FAIL;
>                                         goto xa_unlocked;
>                                 }
> --- a/mm/shmem.c
> +++ b/mm/shmem.c
> @@ -1903,26 +1903,27 @@ static int shmem_getpage_gfp(struct inode *inode, pgoff_t index,
>                 return error;
>         }
>
> -       if (page)
> +       if (page) {
>                 hindex = page->index;
> -       if (page && sgp == SGP_WRITE)
> -               mark_page_accessed(page);
> -
> -       /* fallocated page? */
> -       if (page && !PageUptodate(page)) {
> +               if (sgp == SGP_WRITE)
> +                       mark_page_accessed(page);
> +               if (PageUptodate(page))
> +                       goto out;
> +               /* fallocated page */
>                 if (sgp != SGP_READ)
>                         goto clear;
>                 unlock_page(page);
>                 put_page(page);
> -               page = NULL;
> -               hindex = index;
>         }
> -       if (page || sgp == SGP_READ)
> -               goto out;
> +
> +       *pagep = NULL;
> +       if (sgp == SGP_READ)
> +               return 0;
> +       if (sgp == SGP_NOALLOC)
> +               return -ENOENT;
>
>         /*
> -        * Fast cache lookup did not find it:
> -        * bring it back from swap or allocate.
> +        * Fast cache lookup and swap lookup did not find it: allocate.
>          */
>
>         if (vma && userfaultfd_missing(vma)) {

  reply	other threads:[~2021-08-04 19:01 UTC|newest]

Thread overview: 45+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2021-07-30  7:22 [PATCH 00/16] tmpfs: HUGEPAGE and MEM_LOCK fcntls and memfds Hugh Dickins
2021-07-30  7:25 ` [PATCH 01/16] huge tmpfs: fix fallocate(vanilla) advance over huge pages Hugh Dickins
2021-07-30 21:36   ` Yang Shi
2021-08-01  3:38     ` Hugh Dickins
2021-08-02 20:36       ` Yang Shi
2021-07-30  7:28 ` [PATCH 02/16] huge tmpfs: fix split_huge_page() after FALLOC_FL_KEEP_SIZE Hugh Dickins
2021-07-30 23:48   ` Yang Shi
2021-07-30  7:30 ` [PATCH 03/16] huge tmpfs: remove shrinklist addition from shmem_setattr() Hugh Dickins
2021-07-30 21:50   ` Yang Shi
2021-07-30  7:36 ` [PATCH 04/16] huge tmpfs: revert shmem's use of transhuge_vma_enabled() Hugh Dickins
2021-07-30 21:56   ` Yang Shi
2021-08-01  4:01     ` Hugh Dickins
2021-08-02 20:39       ` Yang Shi
2021-07-30  7:39 ` [PATCH 05/16] huge tmpfs: move shmem_huge_enabled() upwards Hugh Dickins
2021-07-30 21:57   ` Yang Shi
2021-07-30  7:42 ` [PATCH 06/16] huge tmpfs: shmem_is_huge(vma, inode, index) Hugh Dickins
2021-07-30 23:34   ` Yang Shi
2021-08-01  5:22     ` Hugh Dickins
2021-08-01  5:37       ` Hugh Dickins
2021-08-02 21:14       ` Yang Shi
2021-08-04  8:28         ` Hugh Dickins
2021-08-04 19:01           ` Yang Shi [this message]
2021-08-06  5:21             ` Hugh Dickins
2021-08-06 17:41               ` Yang Shi
2021-08-05 23:04         ` Yang Shi
2021-08-06  5:43           ` Hugh Dickins
2021-08-06 17:57             ` Yang Shi
2021-08-12 18:19               ` Yang Shi
2021-07-30  7:45 ` [PATCH 07/16] memfd: memfd_create(name, MFD_HUGEPAGE) for shmem huge pages Hugh Dickins
2021-08-04 14:03   ` Kirill A. Shutemov
2021-08-06  3:33     ` Hugh Dickins
2021-07-30  7:48 ` [PATCH 08/16] huge tmpfs: fcntl(fd, F_HUGEPAGE) and fcntl(fd, F_NOHUGEPAGE) Hugh Dickins
2021-08-04 14:08   ` Kirill A. Shutemov
2021-08-06  4:34     ` Hugh Dickins
2021-07-30  7:51 ` [PATCH 09/16] huge tmpfs: decide stat.st_blksize by shmem_is_huge() Hugh Dickins
2021-07-30 23:40   ` Yang Shi
2021-07-30  7:55 ` [PATCH 10/16] tmpfs: fcntl(fd, F_MEM_LOCK) to memlock a tmpfs file Hugh Dickins
2021-08-03  1:38   ` Matthew Wilcox
2021-08-04  9:15     ` Hugh Dickins
2021-07-30  7:57 ` [PATCH 11/16] tmpfs: fcntl(fd, F_MEM_LOCKED) to test if memlocked Hugh Dickins
2021-07-30  8:00 ` [PATCH 12/16] tmpfs: refuse memlock when fallocated beyond i_size Hugh Dickins
2021-07-30  8:03 ` [PATCH 13/16] mm: bool user_shm_lock(loff_t size, struct ucounts *) Hugh Dickins
2021-07-30  8:06 ` [PATCH 14/16] mm: user_shm_lock(,,getuc) and user_shm_unlock(,,putuc) Hugh Dickins
2021-07-30  8:09 ` [PATCH 15/16] tmpfs: permit changing size of memlocked file Hugh Dickins
2021-07-30  8:13 ` [PATCH 16/16] memfd: memfd_create(name, MFD_MEM_LOCK) for memlocked shmem Hugh Dickins

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=CAHbLzkrvOCCbN3EcDeKwfqWrtU6kH0+7fuSv7aahyjpKtsHn3g@mail.gmail.com \
    --to=shy828301@gmail.com \
    --cc=akpm@linux-foundation.org \
    --cc=chris@chris-wilson.co.uk \
    --cc=ebiederm@xmission.com \
    --cc=hch@infradead.org \
    --cc=hughd@google.com \
    --cc=kirill.shutemov@linux.intel.com \
    --cc=legion@kernel.org \
    --cc=linmiaohe@huawei.com \
    --cc=linux-api@vger.kernel.org \
    --cc=linux-fsdevel@vger.kernel.org \
    --cc=linux-kernel@vger.kernel.org \
    --cc=linux-mm@kvack.org \
    --cc=matthew.auld@intel.com \
    --cc=mhocko@suse.com \
    --cc=mike.kravetz@oracle.com \
    --cc=riel@surriel.com \
    --cc=shakeelb@google.com \
    --cc=willy@infradead.org \
    --subject='Re: [PATCH 06/16] huge tmpfs: shmem_is_huge(vma, inode, index)' \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).