On Thu, Sep 28, 2023 at 12:43:57PM -0700, Zach O'Keefe wrote:
> Hey Ryan,
>
> Thanks for bringing this up.
>
> On Thu, Sep 28, 2023 at 4:59 AM Ryan Roberts <[email protected]> wrote:
> >
> > On 28/09/2023 11:54, Bagas Sanjaya wrote:
> > > On Thu, Sep 28, 2023 at 10:55:17AM +0100, Ryan Roberts wrote:
> > >> Hi all,
> > >>
> > >> I've just noticed that when applied to a file mapping for a file on xfs, MADV_COLLAPSE returns EINVAL. The same test case works fine if the file is on ext4.
> > >>
> > >> I think the root cause is that the implementation bails out if it finds a (non-PMD-sized) large folio in the page cache for any part of the file covered by the region. XFS does readahead into large folios so we hit this issue. See khugepaged.h:collapse_file():
> > >>
> > >> if (PageTransCompound(page)) {
> > >> struct page *head = compound_head(page);
> > >>
> > >> result = compound_order(head) == HPAGE_PMD_ORDER &&
> > >> head->index == start
> > >> /* Maybe PMD-mapped */
> > >> ? SCAN_PTE_MAPPED_HUGEPAGE
> > >> : SCAN_PAGE_COMPOUND;
> > >> goto out_unlock;
> > >> }
> > >
>
> Ya, non-PMD-sized THPs were just barely visible in my peripherals when
> writing this, and I'm still woefully behind on your work on them now
> (sorry!).
>
> I'd like to eventually make collapse (not just MADV_COLLAPSE, but
> khugepaged too) support arbitrary-sized large folios in general, but
> I'm very pressed for time right now. I think M. Wilcox is also
> interested in this, given he left the TODO to support it :P
Is the point of MADV_COLLAPSE to replace base pages with PMD-sized pages
in the pagecache for faster lookups? Or merely to replace them with
something larger, even if it's not PMD sized?
As of 6.6, XFS asks for folios of size min(read/readahead/write_len,
ondisk_mapping_length), so in theory the folio size should roughly
follow the access patterns. If the goal is merely larger folios, then
we are done here and can move on to some other part of the collapse.
OTOH if the goal is TLB savings, then I suppose you'd actually /want/ to
select a large (but not PMD) folio for collapsing into a PMD sized
folio, right?
e.g.
if (PageTransCompound(page)) {
struct page *head = compound_head(page);
if (head->index != start) {
/* not sure what _COMPOUND means here... */
result = SCAN_PAGE_COMPOUND;
goto out_unlock;
}
if (compound_order(head) == HPAGE_PMD_ORDER) {
result = SCAN_PTE_MAPPED_HUGEPAGE;
goto out_unlock;
}
/* result is still SCAN_SUCCEED, keep going */
}
I /think/ that would work? If the largefolio is dirty or not fully
uptodate then collapse won't touch it; and I think fs/iomap handles this
in a compatible way because it won't mark the folio uptodate until all
the blocks have been read, and it marks the folio dirty if any of the
blocks are dirty.
(says me, who doesn't really understand this part of the code.)
--D
> Thank you for the reproducer though! I haven't run it, but I'll
> probably come back here to steal it when the time comes.
>
> > > I don't see any hint to -EINVAL above. Am I missing something?
> >
> > The SCAN_PAGE_COMPOUND result ends up back at madvise_collapse() where it
> > eventually gets converted to -EINVAL by madvise_collapse_errno().
> >
> > >
> > >>
> > >> I'm not sure if this is already a known issue? I don't have time to work on a fix for this right now, so thought I would highlight it at least. I might get around to it at some point in the future if nobody else tackles it.
>
> My guess is Q1 2024 is when I'd be able to look into this, at the
> current level of urgency. It doesn't sound like it's blocking anything
> for your work right now -- lmk if that changes though!
>
> Thanks,
> Zach
>
>
>
> > >>
> > >> Thanks,
> > >> Ryan
> > >>
> > >>
> > >> Test case I've been using:
> > >>
> > >> -->8--
> > >>
> > >> #include <stdio.h>
> > >> #include <stdlib.h>
> > >> #include <sys/mman.h>
> > >> #include <sys/types.h>
> > >> #include <sys/stat.h>
> > >> #include <fcntl.h>
> > >> #include <unistd.h>
> > >>
> > >> #ifndef MADV_COLLAPSE
> > >> #define MADV_COLLAPSE 25
> > >> #endif
> > >>
> > >> #define handle_error(msg) do { perror(msg); exit(EXIT_FAILURE); } while (0)
> > >>
> > >> #define SZ_1K 1024
> > >> #define SZ_1M (SZ_1K * SZ_1K)
> > >> #define ALIGN(val, align) (((val) + ((align) - 1)) & ~((align) - 1))
> > >>
> > >> #if 1
> > >> // ext4
> > >> #define DATA_FILE "/home/ubuntu/data.txt"
> > >> #else
> > >> // xfs
> > >> #define DATA_FILE "/boot/data.txt"
> > >> #endif
> > >>
> > >> int main(void)
> > >> {
> > >> int fd;
> > >> char *mem;
> > >> int ret;
> > >>
> > >> fd = open(DATA_FILE, O_RDONLY);
> > >> if (fd == -1)
> > >> handle_error("open");
> > >>
> > >> mem = mmap(NULL, SZ_1M * 4, PROT_READ | PROT_EXEC, MAP_PRIVATE, fd, 0);
> > >> close(fd);
> > >> if (mem == MAP_FAILED)
> > >> handle_error("mmap");
> > >>
> > >> printf("1: pid=%d, mem=%p\n", getpid(), mem);
> > >> getchar();
> > >>
> > >> mem = (char *)ALIGN((unsigned long)mem, SZ_1M * 2);
> > >> ret = madvise(mem, SZ_1M * 2, MADV_COLLAPSE);
> > >> if (ret)
> > >> handle_error("madvise");
> > >>
> > >> printf("2: pid=%d, mem=%p\n", getpid(), mem);
> > >> getchar();
> > >>
> > >> return 0;
> > >> }
> > >>
> > >> -->8--
> > >>
> > >
> > > Confused...
> >
> > This is a user space test case that shows the problem; data.txt needs to be at
> > least 4MB and on a mounted ext4 and xfs filesystem. By toggling the '#if 1' to
> > 0, you can see the different behaviours for ext4 and xfs -
> > handle_error("madvise") fires with EINVAL in the xfs case. The getchar()s are
> > leftovers from me looking at the smaps file.
> >
On Thu, Sep 28, 2023 at 2:04 PM Darrick J. Wong <[email protected]> wrote:
>
> On Thu, Sep 28, 2023 at 12:43:57PM -0700, Zach O'Keefe wrote:
> > Hey Ryan,
> >
> > Thanks for bringing this up.
> >
> > On Thu, Sep 28, 2023 at 4:59 AM Ryan Roberts <[email protected]> wrote:
> > >
> > > On 28/09/2023 11:54, Bagas Sanjaya wrote:
> > > > On Thu, Sep 28, 2023 at 10:55:17AM +0100, Ryan Roberts wrote:
> > > >> Hi all,
> > > >>
> > > >> I've just noticed that when applied to a file mapping for a file on xfs, MADV_COLLAPSE returns EINVAL. The same test case works fine if the file is on ext4.
> > > >>
> > > >> I think the root cause is that the implementation bails out if it finds a (non-PMD-sized) large folio in the page cache for any part of the file covered by the region. XFS does readahead into large folios so we hit this issue. See khugepaged.h:collapse_file():
> > > >>
> > > >> if (PageTransCompound(page)) {
> > > >> struct page *head = compound_head(page);
> > > >>
> > > >> result = compound_order(head) == HPAGE_PMD_ORDER &&
> > > >> head->index == start
> > > >> /* Maybe PMD-mapped */
> > > >> ? SCAN_PTE_MAPPED_HUGEPAGE
> > > >> : SCAN_PAGE_COMPOUND;
> > > >> goto out_unlock;
> > > >> }
> > > >
> >
> > Ya, non-PMD-sized THPs were just barely visible in my peripherals when
> > writing this, and I'm still woefully behind on your work on them now
> > (sorry!).
> >
> > I'd like to eventually make collapse (not just MADV_COLLAPSE, but
> > khugepaged too) support arbitrary-sized large folios in general, but
> > I'm very pressed for time right now. I think M. Wilcox is also
> > interested in this, given he left the TODO to support it :P
>
> Is the point of MADV_COLLAPSE to replace base pages with PMD-sized pages
> in the pagecache for faster lookups? Or merely to replace them with
> something larger, even if it's not PMD sized?
Might depend on who you ask, but IMHO, the principle purpose of
collapse is saving TLB entries, with TLB coalescing complicating
things a little in terms of PMD-sized things or not. M. Wilcox's work
with descriptor-izing folios might make a nice case for memory savings
as well, down the line.
> As of 6.6, XFS asks for folios of size min(read/readahead/write_len,
> ondisk_mapping_length), so in theory the folio size should roughly
> follow the access patterns. If the goal is merely larger folios, then
> we are done here and can move on to some other part of the collapse.
>
> OTOH if the goal is TLB savings, then I suppose you'd actually /want/ to
> select a large (but not PMD) folio for collapsing into a PMD sized
> folio, right?
I suppose it might make some operations easier / faster during
collapse if we have less folios to process.
> e.g.
>
> if (PageTransCompound(page)) {
> struct page *head = compound_head(page);
>
> if (head->index != start) {
> /* not sure what _COMPOUND means here... */
> result = SCAN_PAGE_COMPOUND;
> goto out_unlock;
> }
>
> if (compound_order(head) == HPAGE_PMD_ORDER) {
> result = SCAN_PTE_MAPPED_HUGEPAGE;
> goto out_unlock;
> }
>
> /* result is still SCAN_SUCCEED, keep going */
> }
>
> I /think/ that would work? If the largefolio is dirty or not fully
> uptodate then collapse won't touch it; and I think fs/iomap handles this
> in a compatible way because it won't mark the folio uptodate until all
> the blocks have been read, and it marks the folio dirty if any of the
> blocks are dirty.
>
> (says me, who doesn't really understand this part of the code.)
I think there's a couple issues with this -- for example, the
head->index != start case is going to be the common-case for
non-PMD-sized large folios. Regardless, there is some more code in
hpage_collapse_scan_file() and her in collapse_file() that would need
to be updated. I'm taking a cursory look, and naively it doesn't look
too bad -- most things "should just work" in file/shmem collapse path.
ac492b9c70cac ("mm/khugepaged: skip shmem with userfaultfd" was merged
since the last I looked carefully at this path, so I would need to
spend more time understanding some changes there. So, from correctness
POV, maybe there's not anything drastic to be done for file/shmem.
Maybe this is a good place to start.
For anon, things are different, as we are coming at the pages from a
different angle, and are operating over the pmd directly. I'm not
immediately sure if it makes things easier or harder. Probably harder.
Can we even get non-PMD-sized large anon folios right now, without
Ryan's work?
From a khugepaged policy POV, there are some questions to be
answered.. but I think these mostly boil down to: scale the
present/swap/none checks by (1 << order).
Anyways, this isn't to be taken with much weight as a thorough audit
is required to understand any subtleties lurking around.
Thanks,
Zach
> --D
>
> > Thank you for the reproducer though! I haven't run it, but I'll
> > probably come back here to steal it when the time comes.
> >
> > > > I don't see any hint to -EINVAL above. Am I missing something?
> > >
> > > The SCAN_PAGE_COMPOUND result ends up back at madvise_collapse() where it
> > > eventually gets converted to -EINVAL by madvise_collapse_errno().
> > >
> > > >
> > > >>
> > > >> I'm not sure if this is already a known issue? I don't have time to work on a fix for this right now, so thought I would highlight it at least. I might get around to it at some point in the future if nobody else tackles it.
> >
> > My guess is Q1 2024 is when I'd be able to look into this, at the
> > current level of urgency. It doesn't sound like it's blocking anything
> > for your work right now -- lmk if that changes though!
> >
> > Thanks,
> > Zach
> >
> >
> >
> > > >>
> > > >> Thanks,
> > > >> Ryan
> > > >>
> > > >>
> > > >> Test case I've been using:
> > > >>
> > > >> -->8--
> > > >>
> > > >> #include <stdio.h>
> > > >> #include <stdlib.h>
> > > >> #include <sys/mman.h>
> > > >> #include <sys/types.h>
> > > >> #include <sys/stat.h>
> > > >> #include <fcntl.h>
> > > >> #include <unistd.h>
> > > >>
> > > >> #ifndef MADV_COLLAPSE
> > > >> #define MADV_COLLAPSE 25
> > > >> #endif
> > > >>
> > > >> #define handle_error(msg) do { perror(msg); exit(EXIT_FAILURE); } while (0)
> > > >>
> > > >> #define SZ_1K 1024
> > > >> #define SZ_1M (SZ_1K * SZ_1K)
> > > >> #define ALIGN(val, align) (((val) + ((align) - 1)) & ~((align) - 1))
> > > >>
> > > >> #if 1
> > > >> // ext4
> > > >> #define DATA_FILE "/home/ubuntu/data.txt"
> > > >> #else
> > > >> // xfs
> > > >> #define DATA_FILE "/boot/data.txt"
> > > >> #endif
> > > >>
> > > >> int main(void)
> > > >> {
> > > >> int fd;
> > > >> char *mem;
> > > >> int ret;
> > > >>
> > > >> fd = open(DATA_FILE, O_RDONLY);
> > > >> if (fd == -1)
> > > >> handle_error("open");
> > > >>
> > > >> mem = mmap(NULL, SZ_1M * 4, PROT_READ | PROT_EXEC, MAP_PRIVATE, fd, 0);
> > > >> close(fd);
> > > >> if (mem == MAP_FAILED)
> > > >> handle_error("mmap");
> > > >>
> > > >> printf("1: pid=%d, mem=%p\n", getpid(), mem);
> > > >> getchar();
> > > >>
> > > >> mem = (char *)ALIGN((unsigned long)mem, SZ_1M * 2);
> > > >> ret = madvise(mem, SZ_1M * 2, MADV_COLLAPSE);
> > > >> if (ret)
> > > >> handle_error("madvise");
> > > >>
> > > >> printf("2: pid=%d, mem=%p\n", getpid(), mem);
> > > >> getchar();
> > > >>
> > > >> return 0;
> > > >> }
> > > >>
> > > >> -->8--
> > > >>
> > > >
> > > > Confused...
> > >
> > > This is a user space test case that shows the problem; data.txt needs to be at
> > > least 4MB and on a mounted ext4 and xfs filesystem. By toggling the '#if 1' to
> > > 0, you can see the different behaviours for ext4 and xfs -
> > > handle_error("madvise") fires with EINVAL in the xfs case. The getchar()s are
> > > leftovers from me looking at the smaps file.
> > >