[mmotm,v2] tmpfs: do not allocate pages on read

Commit Message

Hugh Dickins March 6, 2022, 10:59 p.m. UTC

Mikulas asked in
https://lore.kernel.org/linux-mm/[e-mail]/
Do we still need a0ee5ec520ed ("tmpfs: allocate on read when stacked")?

Lukas noticed this unusual behavior of loop device backed by tmpfs in
https://lore.kernel.org/linux-mm/20211126075100.gd64odg2bcptiqeb@work/

Normally, shmem_file_read_iter() copies the ZERO_PAGE when reading holes;
but if it looks like it might be a read for "a stacking filesystem", it
allocates actual pages to the page cache, and even marks them as dirty.
And reads from the loop device do satisfy the test that is used.

This oddity was added for an old version of unionfs, to help to limit
its usage to the limited size of the tmpfs mount involved; but about
the same time as the tmpfs mod went in (2.6.25), unionfs was reworked
to proceed differently; and the mod kept just in case others needed it.

Do we still need it? I cannot answer with more certainty than "Probably
not". It's nasty enough that we really should try to delete it; but if
a regression is reported somewhere, then we might have to revert later.

It's not quite as simple as just removing the test (as Mikulas did):
xfstests generic/013 hung because splice from tmpfs failed on page not
up-to-date and page mapping unset.  That can be fixed just by marking
the ZERO_PAGE as Uptodate, which of course it is: do so in
pagecache_init() - it might be useful to others than tmpfs.

My intention, though, was to stop using the ZERO_PAGE here altogether:
surely iov_iter_zero() is better for this case?  Sadly not: it relies
on clear_user(), and the x86 clear_user() is slower than its copy_user():
https://lore.kernel.org/lkml/[e-mail]/

But while we are still using the ZERO_PAGE, let's stop dirtying its
struct page cacheline with unnecessary get_page() and put_page().

Reported-by: Mikulas Patocka <[e-mail]>
Reported-by: Lukas Czerner <[e-mail]>
Signed-off-by: Hugh Dickins <[e-mail]>
---
v2: Set ZERO_PAGE uptodate during init, per hch.

 mm/filemap.c |  6 ++++++
 mm/shmem.c   | 20 ++++++--------------
 2 files changed, 12 insertions(+), 14 deletions(-)

Comments

Christoph Hellwig March 7, 2022, 6:44 a.m. UTC | #1

On Sun, Mar 06, 2022 at 02:59:05PM -0800, Hugh Dickins wrote:
> Mikulas asked in
> https://lore.kernel.org/linux-mm/[e-mail]/
> Do we still need a0ee5ec520ed ("tmpfs: allocate on read when stacked")?
> 
> Lukas noticed this unusual behavior of loop device backed by tmpfs in
> https://lore.kernel.org/linux-mm/20211126075100.gd64odg2bcptiqeb@work/
> 
> Normally, shmem_file_read_iter() copies the ZERO_PAGE when reading holes;
> but if it looks like it might be a read for "a stacking filesystem", it
> allocates actual pages to the page cache, and even marks them as dirty.
> And reads from the loop device do satisfy the test that is used.
> 
> This oddity was added for an old version of unionfs, to help to limit
> its usage to the limited size of the tmpfs mount involved; but about
> the same time as the tmpfs mod went in (2.6.25), unionfs was reworked
> to proceed differently; and the mod kept just in case others needed it.
> 
> Do we still need it? I cannot answer with more certainty than "Probably
> not". It's nasty enough that we really should try to delete it; but if
> a regression is reported somewhere, then we might have to revert later.
> 
> It's not quite as simple as just removing the test (as Mikulas did):
> xfstests generic/013 hung because splice from tmpfs failed on page not
> up-to-date and page mapping unset.  That can be fixed just by marking
> the ZERO_PAGE as Uptodate, which of course it is: do so in
> pagecache_init() - it might be useful to others than tmpfs.
> 
> My intention, though, was to stop using the ZERO_PAGE here altogether:
> surely iov_iter_zero() is better for this case?  Sadly not: it relies
> on clear_user(), and the x86 clear_user() is slower than its copy_user():
> https://lore.kernel.org/lkml/[e-mail]/
> 
> But while we are still using the ZERO_PAGE, let's stop dirtying its
> struct page cacheline with unnecessary get_page() and put_page().
> 
> Reported-by: Mikulas Patocka <[e-mail]>
> Reported-by: Lukas Czerner <[e-mail]>
> Signed-off-by: Hugh Dickins <[e-mail]>

I would have split the uptodate setting of ZERO_PAGE into a separate,
clearly documented patch, but otherwise this looks good:

Reviewed-by: Christoph Hellwig <[e-mail]>

Darrick J. Wong March 8, 2022, 5:27 p.m. UTC | #2

On Mon, Mar 07, 2022 at 07:44:34AM +0100, Christoph Hellwig wrote:
> On Sun, Mar 06, 2022 at 02:59:05PM -0800, Hugh Dickins wrote:
> > Mikulas asked in
> > https://lore.kernel.org/linux-mm/[e-mail]/
> > Do we still need a0ee5ec520ed ("tmpfs: allocate on read when stacked")?
> > 
> > Lukas noticed this unusual behavior of loop device backed by tmpfs in
> > https://lore.kernel.org/linux-mm/20211126075100.gd64odg2bcptiqeb@work/
> > 
> > Normally, shmem_file_read_iter() copies the ZERO_PAGE when reading holes;
> > but if it looks like it might be a read for "a stacking filesystem", it
> > allocates actual pages to the page cache, and even marks them as dirty.
> > And reads from the loop device do satisfy the test that is used.
> > 
> > This oddity was added for an old version of unionfs, to help to limit
> > its usage to the limited size of the tmpfs mount involved; but about
> > the same time as the tmpfs mod went in (2.6.25), unionfs was reworked
> > to proceed differently; and the mod kept just in case others needed it.
> > 
> > Do we still need it? I cannot answer with more certainty than "Probably
> > not". It's nasty enough that we really should try to delete it; but if
> > a regression is reported somewhere, then we might have to revert later.
> > 
> > It's not quite as simple as just removing the test (as Mikulas did):
> > xfstests generic/013 hung because splice from tmpfs failed on page not
> > up-to-date and page mapping unset.  That can be fixed just by marking
> > the ZERO_PAGE as Uptodate, which of course it is: do so in
> > pagecache_init() - it might be useful to others than tmpfs.
> > 
> > My intention, though, was to stop using the ZERO_PAGE here altogether:
> > surely iov_iter_zero() is better for this case?  Sadly not: it relies
> > on clear_user(), and the x86 clear_user() is slower than its copy_user():
> > https://lore.kernel.org/lkml/[e-mail]/
> > 
> > But while we are still using the ZERO_PAGE, let's stop dirtying its
> > struct page cacheline with unnecessary get_page() and put_page().
> > 
> > Reported-by: Mikulas Patocka <[e-mail]>
> > Reported-by: Lukas Czerner <[e-mail]>
> > Signed-off-by: Hugh Dickins <[e-mail]>
> 
> I would have split the uptodate setting of ZERO_PAGE into a separate,
> clearly documented patch, but otherwise this looks good:
> 
> Reviewed-by: Christoph Hellwig <[e-mail]>

I've long wondered (for my own nefarious purposes) why tmpfs files
didn't just grab the zero page, so:

Acked-by: Darrick J. Wong <[e-mail]>

--D

Hugh Dickins March 8, 2022, 9:46 p.m. UTC | #3

On Tue, 8 Mar 2022, Darrick J. Wong wrote:
> 
> I've long wondered (for my own nefarious purposes) why tmpfs files
> didn't just grab the zero page,

(tmpfs files have been using the zero page for reads for many years:
it was just this odd internal "could it be for a stacking filesystem?"
case, which /dev/loop also fell into, which was doing allocation on read.

I wonder what your nefarious purposes are ;) Maybe related to pages
faulted into an mmap: those pages tmpfs has always allocated for, then
they're freed up later by page reclaim if still undirtied. We may change
that in future, and use the zero page even there: there are advantages
of course, but some care and code needed - never been a priority.)

> so:
> 
> Acked-by: Darrick J. Wong <[e-mail]>

Thanks,
Hugh

Patch

--- a/mm/filemap.c
+++ b/mm/filemap.c
@@ -1063,6 +1063,12 @@  void __init pagecache_init(void)
 		init_waitqueue_head(&folio_wait_table[i]);
 
 	page_writeback_init();
+
+	/*
+	 * tmpfs uses the ZERO_PAGE for reading holes: it is up-to-date,
+	 * and splice's page_cache_pipe_buf_confirm() needs to see that.
+	 */
+	SetPageUptodate(ZERO_PAGE(0));
 }
 
 /*
diff --git a/mm/shmem.c b/mm/shmem.c
index 3c346f2e557f..659bd599d731 100644
--- a/mm/shmem.c
+++ b/mm/shmem.c
@@ -2501,19 +2501,10 @@  static ssize_t shmem_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
 	struct address_space *mapping = inode->i_mapping;
 	pgoff_t index;
 	unsigned long offset;
-	enum sgp_type sgp = SGP_READ;
 	int error = 0;
 	ssize_t retval = 0;
 	loff_t *ppos = &iocb->ki_pos;
 
-	/*
-	 * Might this read be for a stacking filesystem?  Then when reading
-	 * holes of a sparse file, we actually need to allocate those pages,
-	 * and even mark them dirty, so it cannot exceed the max_blocks limit.
-	 */
-	if (!iter_is_iovec(to))
-		sgp = SGP_CACHE;
-
 	index = *ppos >> PAGE_SHIFT;
 	offset = *ppos & ~PAGE_MASK;
 
@@ -2522,6 +2513,7 @@  static ssize_t shmem_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
 		pgoff_t end_index;
 		unsigned long nr, ret;
 		loff_t i_size = i_size_read(inode);
+		bool got_page;
 
 		end_index = i_size >> PAGE_SHIFT;
 		if (index > end_index)
@@ -2532,15 +2524,13 @@  static ssize_t shmem_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
 				break;
 		}
 
-		error = shmem_getpage(inode, index, &page, sgp);
+		error = shmem_getpage(inode, index, &page, SGP_READ);
 		if (error) {
 			if (error == -EINVAL)
 				error = 0;
 			break;
 		}
 		if (page) {
-			if (sgp == SGP_CACHE)
-				set_page_dirty(page);
 			unlock_page(page);
 
 			if (PageHWPoison(page)) {
@@ -2580,9 +2570,10 @@  static ssize_t shmem_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
 			 */
 			if (!offset)
 				mark_page_accessed(page);
+			got_page = true;
 		} else {
 			page = ZERO_PAGE(0);
-			get_page(page);
+			got_page = false;
 		}
 
 		/*
@@ -2595,7 +2586,8 @@  static ssize_t shmem_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
 		index += offset >> PAGE_SHIFT;
 		offset &= ~PAGE_MASK;
 
-		put_page(page);
+		if (got_page)
+			put_page(page);
 		if (!iov_iter_count(to))
 			break;
 		if (ret < nr) {