c/mirror_zfs
1
0
Fork 0
mirror of https://git.proxmox.com/git/mirror_zfs.git synced 2024-12-25 18:59:33 +03:00
Code Issues Packages Projects Releases Wiki Activity

abd_iter_page: rework to handle multipage scatterlists

Browse Source 
Previously, abd_iter_page() would assume that every scatterlist would
contain a single page (compound or no), because that's all we ever
create in abd_alloc_chunks(). However, scatterlists can contain multiple
pages of arbitrary provenance, and if we get one of those, we'd get all
the math wrong.

This reworks things to handle multiple pages in a scatterlist, by
properly finding the right page within it for the given offset, and
understanding better where the end of the page is and not crossing it.

Sponsored-by: Klara, Inc.
Sponsored-by: Wasabi Technology, Inc.
Reported-by: Brian Atkinson <batkinson@lanl.gov>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Brian Atkinson <batkinson@lanl.gov>
Signed-off-by: Rob Norris <rob.norris@klarasystems.com>
Closes #16108
This commit is contained in:
Rob N 2024-04-20 09:41:31 +10:00 committed by GitHub
parent 9f83eec039
commit f4f156157d
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 74 additions and 46 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

120
module/os/linux/zfs/abd_os.c
View File

@ -1015,10 +1015,50 @@ abd_cache_reap_now(void)
}
#if defined(_KERNEL)
/*
* Yield the next page struct and data offset and size within it, without
* This is abd_iter_page(), the function underneath abd_iterate_page_func().
* It yields the next page struct and data offset and size within it, without
* mapping it into the address space.
*/
/*
* "Compound pages" are a group of pages that can be referenced from a single
* struct page *. Its organised as a "head" page, followed by a series of
* "tail" pages.
*
* In OpenZFS, compound pages are allocated using the __GFP_COMP flag, which we
* get from scatter ABDs and SPL vmalloc slabs (ie >16K allocations). So a
* great many of the IO buffers we get are going to be of this type.
*
* The tail pages are just regular PAGESIZE pages, and can be safely used
* as-is. However, the head page has length covering itself and all the tail
* pages. If the ABD chunk spans multiple pages, then we can use the head page
* and a >PAGESIZE length, which is far more efficient.
*
* Before kernel 4.5 however, compound page heads were refcounted separately
* from tail pages, such that moving back to the head page would require us to
* take a reference to it and releasing it once we're completely finished with
* it. In practice, that means when our caller is done with the ABD, which we
* have no insight into from here. Rather than contort this API to track head
* page references on such ancient kernels, we disable this special compound
* page handling on 4.5, instead just using treating each page within it as a
* regular PAGESIZE page (which it is). This is slightly less efficient, but
* makes everything far simpler.
*
* The below test sets/clears ABD_ITER_COMPOUND_PAGES to enable/disable the
* special handling, and also defines the ABD_ITER_PAGE_SIZE(page) macro to
* understand compound pages, or not, as required.
*/
#if LINUX_VERSION_CODE >= KERNEL_VERSION(4, 5, 0)
#define ABD_ITER_COMPOUND_PAGES 1
#define ABD_ITER_PAGE_SIZE(page) \
(PageCompound(page) ? page_size(page) : PAGESIZE)
#else
#undef ABD_ITER_COMPOUND_PAGES
#define ABD_ITER_PAGE_SIZE(page) (PAGESIZE)
#endif
void
abd_iter_page(struct abd_iter *aiter)
{
@ -1032,6 +1072,12 @@ abd_iter_page(struct abd_iter *aiter)
struct page *page;
size_t doff, dsize;
/*
* Find the page, and the start of the data within it. This is computed
* differently for linear and scatter ABDs; linear is referenced by
* virtual memory location, while scatter is referenced by page
* pointer.
*/
if (abd_is_linear(aiter->iter_abd)) {
ASSERT3U(aiter->iter_pos, ==, aiter->iter_offset);
@ -1044,57 +1090,24 @@ abd_iter_page(struct abd_iter *aiter)
/* offset of address within the page */
doff = offset_in_page(paddr);
/* total data remaining in abd from this position */
dsize = aiter->iter_abd->abd_size - aiter->iter_offset;
} else {
ASSERT(!abd_is_gang(aiter->iter_abd));
/* current scatter page */
page = sg_page(aiter->iter_sg);
page = nth_page(sg_page(aiter->iter_sg),
aiter->iter_offset >> PAGE_SHIFT);
/* position within page */
doff = aiter->iter_offset;
/* remaining data in scatterlist */
dsize = MIN(aiter->iter_sg->length - aiter->iter_offset,
aiter->iter_abd->abd_size - aiter->iter_pos);
doff = aiter->iter_offset & (PAGESIZE - 1);
}
ASSERT(page);
#if LINUX_VERSION_CODE >= KERNEL_VERSION(4, 5, 0)
#ifdef ABD_ITER_COMPOUND_PAGES
if (PageTail(page)) {
/*
* This page is part of a "compound page", which is a group of
* pages that can be referenced from a single struct page *.
* Its organised as a "head" page, followed by a series of
* "tail" pages.
*
* In OpenZFS, compound pages are allocated using the
* __GFP_COMP flag, which we get from scatter ABDs and SPL
* vmalloc slabs (ie >16K allocations). So a great many of the
* IO buffers we get are going to be of this type.
*
* The tail pages are just regular PAGE_SIZE pages, and can be
* safely used as-is. However, the head page has length
* covering itself and all the tail pages. If this ABD chunk
* spans multiple pages, then we can use the head page and a
* >PAGE_SIZE length, which is far more efficient.
*
* To do this, we need to adjust the offset to be counted from
* the head page. struct page for compound pages are stored
* contiguously, so we can just adjust by a simple offset.
*
* Before kernel 4.5, compound page heads were refcounted
* separately, such that moving back to the head page would
* require us to take a reference to it and releasing it once
* we're completely finished with it. In practice, that means
* when our caller is done with the ABD, which we have no
* insight into from here. Rather than contort this API to
* track head page references on such ancient kernels, we just
* compile this block out and use the tail pages directly. This
* is slightly less efficient, but makes everything far
* simpler.
* If this is a compound tail page, move back to the head, and
* adjust the offset to match. This may let us yield a much
* larger amount of data from a single logical page, and so
* leave our caller with fewer pages to process.
*/
struct page *head = compound_head(page);
doff += ((page - head) * PAGESIZE);
@ -1102,12 +1115,27 @@ abd_iter_page(struct abd_iter *aiter)
}
#endif
/* final page and position within it */
ASSERT(page);
/*
* Compute the maximum amount of data we can take from this page. This
* is the smaller of:
* - the remaining space in the page
* - the remaining space in this scatterlist entry (which may not cover
* the entire page)
* - the remaining space in the abd (which may not cover the entire
* scatterlist entry)
*/
dsize = MIN(ABD_ITER_PAGE_SIZE(page) - doff,
aiter->iter_abd->abd_size - aiter->iter_pos);
if (!abd_is_linear(aiter->iter_abd))
dsize = MIN(dsize, aiter->iter_sg->length - aiter->iter_offset);
ASSERT3U(dsize, >, 0);
/* final iterator outputs */
aiter->iter_page = page;
aiter->iter_page_doff = doff;
/* amount of data in the chunk, up to the end of the page */
aiter->iter_page_dsize = MIN(dsize, page_size(page) - doff);
aiter->iter_page_dsize = dsize;
}
/*