[PATCH v7 3/6] mm/gup: Introduce memfd_pin_folios() for pinning memfd folios (v7)
Kasireddy, Vivek
vivek.kasireddy at intel.com
Wed Dec 13 08:44:51 UTC 2023
Hi David,
>
> On 12.12.23 08:38, Vivek Kasireddy wrote:
> > For drivers that would like to longterm-pin the folios associated
> > with a memfd, the memfd_pin_folios() API provides an option to
> > not only pin the folios via FOLL_PIN but also to check and migrate
> > them if they reside in movable zone or CMA block. This API
> > currently works with memfds but it should work with any files
> > that belong to either shmemfs or hugetlbfs. Files belonging to
> > other filesystems are rejected for now.
> >
> > The folios need to be located first before pinning them via FOLL_PIN.
> > If they are found in the page cache, they can be immediately pinned.
> > Otherwise, they need to be allocated using the filesystem specific
> > APIs and then pinned.
> >
> > v2:
> > - Drop gup_flags and improve comments and commit message (David)
> > - Allocate a page if we cannot find in page cache for the hugetlbfs
> > case as well (David)
> > - Don't unpin pages if there is a migration related failure (David)
> > - Drop the unnecessary nr_pages <= 0 check (Jason)
> > - Have the caller of the API pass in file * instead of fd (Jason)
> >
> > v3: (David)
> > - Enclose the huge page allocation code with #ifdef
> CONFIG_HUGETLB_PAGE
> > (Build error reported by kernel test robot <lkp at intel.com>)
> > - Don't forget memalloc_pin_restore() on non-migration related errors
> > - Improve the readability of the cleanup code associated with
> > non-migration related errors
> > - Augment the comments by describing FOLL_LONGTERM like behavior
> > - Include the R-b tag from Jason
> >
> > v4:
> > - Remove the local variable "page" and instead use 3 return statements
> > in alloc_file_page() (David)
> > - Add the R-b tag from David
> >
> > v5: (David)
> > - For hugetlb case, ensure that we only obtain head pages from the
> > mapping by using __filemap_get_folio() instead of find_get_page_flags()
> > - Handle -EEXIST when two or more potential users try to simultaneously
> > add a huge page to the mapping by forcing them to retry on failure
> >
> > v6: (Christoph)
> > - Rename this API to memfd_pin_user_pages() to make it clear that it
> > is intended for memfds
> > - Move the memfd page allocation helper from gup.c to memfd.c
> > - Fix indentation errors in memfd_pin_user_pages()
> > - For contiguous ranges of folios, use a helper such as
> > filemap_get_folios_contig() to lookup the page cache in batches
> >
> > v7:
> > - Rename this API to memfd_pin_folios() and make it return folios
> > and offsets instead of pages (David)
> > - Don't continue processing the folios in the batch returned by
> > filemap_get_folios_contig() if they do not have correct next_idx
> > - Add the R-b tag from Christoph
> >
>
> Sorry, I'm still not happy about the current state, because (1) the
> folio vs. pages handling is still mixed (2) we're returning+pinning a
> large folio multiple times.
I can address (1) in a follow-up series and as far as (2) is concerned, my
understanding is that we need to increase the folio's refcount as and
when the folio's tail pages are used. Is this not the case? It appears
this is what unpin_user_pages() expects as well. Do you see any
concern with this?
>
> See below if there is an easy way to clean this up.
>
> >> @@ -5,6 +5,7 @@
> > #include <linux/spinlock.h>
> >
> > #include <linux/mm.h>
> > +#include <linux/memfd.h>
> > #include <linux/memremap.h>
> > #include <linux/pagemap.h>
> > #include <linux/rmap.h>
> > @@ -17,6 +18,7 @@
> > #include <linux/hugetlb.h>
> > #include <linux/migrate.h>
> > #include <linux/mm_inline.h>
> > +#include <linux/pagevec.h>
> > #include <linux/sched/mm.h>
> > #include <linux/shmem_fs.h>
> >
> > @@ -3410,3 +3412,156 @@ long pin_user_pages_unlocked(unsigned long
> start, unsigned long nr_pages,
> > &locked, gup_flags);
> > }
> > EXPORT_SYMBOL(pin_user_pages_unlocked);
> > +
> > +/**
> > + * memfd_pin_folios() - pin folios associated with a memfd
> > + * @memfd: the memfd whose folios are to be pinned
> > + * @start: starting memfd offset
> > + * @nr_pages: number of pages from start to pin
>
> We're not pinning pages. An inclusive range [start, end] would be clearer.
Ok, I'll make this change in the next version.
>
> > + * @folios: array that receives pointers to the folios pinned.
> > + * Should be at-least nr_pages long.
> > + * @offsets: array that receives offsets of pages in their folios.
> > + * Should be at-least nr_pages long.
>
> See below, I'm wondering if this is really required once we return each folio
> only once.
The offsets can be calculated by the caller (udmabuf) as well but doing so
in this interface would prevent special handling in the caller for the hugetlb
case. Please look at patch 5 in this series (udmabuf: Pin the pages using
memfd_pin_folios() API (v5)) for more details as to what I mean.
>
> > + *
> > + * Attempt to pin folios associated with a memfd; given that a memfd is
> > + * either backed by shmem or hugetlb, the folios can either be found in
> > + * the page cache or need to be allocated if necessary. Once the folios
> > + * are located, they are all pinned via FOLL_PIN and the @offsets array
> > + * is populated with offsets of the pages in their respective folios.
> > + * Therefore, for each page the caller requested, there will be a
> > + * corresponding entry in both @folios and @offsets. And, eventually,
> > + * these pinned folios need to be released either using
> unpin_user_pages()
> > + * or unpin_user_page().
>
> Oh, we don't have a folio function yet? Should be easy to add, and we'd
> really
> add them.
Sure, I'll add them in a follow-up series.
>
> > + *
> > + * It must be noted that the folios may be pinned for an indefinite amount
> > + * of time. And, in most cases, the duration of time they may stay pinned
> > + * would be controlled by the userspace. This behavior is effectively the
> > + * same as using FOLL_LONGTERM with other GUP APIs.
> > + *
> > + * Returns number of folios pinned. This would be equal to the number of
> > + * pages requested. If no folios were pinned, it returns -errno.
> > + */
> > +long memfd_pin_folios(struct file *memfd, unsigned long start,
> > + unsigned long nr_pages, struct folio **folios,
> > + pgoff_t *offsets)
> > +{
> > + unsigned long end = start + (nr_pages << PAGE_SHIFT) - 1;
> > + unsigned int max_pgs, pgoff, pgshift = PAGE_SHIFT;
> > + pgoff_t start_idx, end_idx, next_idx;
> > + unsigned int flags, nr_folios, i, j;
> > + struct folio *folio = NULL;
> > + struct folio_batch fbatch;
> > + struct page **pages;
> > + struct hstate *h;
> > + long ret;
> > +
> > + if (!nr_pages)
> > + return -EINVAL;
> > +
> > + if (!memfd)
> > + return -EINVAL;
> > +
> > + if (!shmem_file(memfd) && !is_file_hugepages(memfd))
> > + return -EINVAL;
> > +
> > + pages = kmalloc_array(nr_pages, sizeof(*pages), GFP_KERNEL);
> > + if (!pages)
> > + return -ENOMEM;
> > +
> > + if (is_file_hugepages(memfd)) {
> > + h = hstate_file(memfd);
> > + pgshift = huge_page_shift(h);
> > + }
> > +
> > + flags = memalloc_pin_save();
> > + do {
> > + i = 0;
> > + start_idx = start >> pgshift;
> > + end_idx = end >> pgshift;
> > + if (is_file_hugepages(memfd)) {
> > + start_idx <<= huge_page_order(h);
> > + end_idx <<= huge_page_order(h);
> > + }
> > +
> > + folio_batch_init(&fbatch);
> > + while (start_idx <= end_idx) {
> > + /*
> > + * In most cases, we should be able to find the folios
> > + * in the page cache. If we cannot find them for some
> > + * reason, we try to allocate them and add them to
> the
> > + * page cache.
> > + */
> > + nr_folios = filemap_get_folios_contig(memfd-
> >f_mapping,
> > + &start_idx,
> > + end_idx,
> > + &fbatch);
> > + if (folio) {
> > + folio_put(folio);
> > + folio = NULL;
> > + }
> > +
> > + next_idx = 0;
> > + for (j = 0; j < nr_folios; j++) {
> > + if (next_idx &&
> > + next_idx != folio_index(fbatch.folios[j]))
> > + continue;
> > +
> > + folio = try_grab_folio(&fbatch.folios[j]->page,
> > + 1, FOLL_PIN);
> > + if (!folio) {
> > + folio_batch_release(&fbatch);
> > + kfree(pages);
> > + goto err;
> > + }
> > +
> > + max_pgs = folio_nr_pages(folio);
> > + if (i == 0) {
> > + pgoff = offset_in_folio(folio, start);
> > + pgoff >>= PAGE_SHIFT;
> > + }
> > +
> > + do {
> > + folios[i] = folio;
> > + offsets[i] = pgoff << PAGE_SHIFT;
> > + pages[i] = folio_page(folio, 0);
> > + folio_add_pin(folio);
> > +
> > + pgoff++;
> > + i++;
> > + } while (pgoff < max_pgs && i < nr_pages);
> > +
> > + pgoff = 0;
> > + next_idx = folio_next_index(folio);
> > + gup_put_folio(folio, 1, FOLL_PIN);
> > + }
> > +
> > + folio = NULL;
> > + folio_batch_release(&fbatch);
> > + if (!nr_folios) {
> > + folio = memfd_alloc_folio(memfd, start_idx);
> > + if (IS_ERR(folio)) {
> > + ret = PTR_ERR(folio);
> > + if (ret != -EEXIST) {
> > + kfree(pages);
> > + goto err;
> > + }
> > + }
> > + }
> > + }
> > +
> > + ret = check_and_migrate_movable_pages(nr_pages, pages);
>
> Having a folio variant would avoid having to mess with pages here at all.
> Further, we're now returning+pinning the same folio multiple times, instead
> of
> just once like the folio batching variant would.
It should be possible to pin the folio only once but I don't see any problem with
pinning it multiple times -- once per each subpage used -- as long as it is unpinned
correctly the same number of times. Is this not ok?
>
> I'm wondering if the following wouldn't make more sense, assuming we add
> check_and_migrate_movable_folios(), which should be pretty easy to add.
>
> Obviously untested, just to express what I have in mind:
Thank you for taking the time to do this!
>
>
>
> /**
> * memfd_pin_folios() - pin folios associated with a memfd
> * @memfd: the memfd whose folios are to be pinned
> * @start: the starting memfd offset
> * @end: the final memfd offset (inclusive)
> * @folios: array that receives pointers to the folios pinned
> * @max_folios: the number of entries in the array for folios
> * @offsets: the offset into the first folio
Given that my goal is to do the following in udmabuf driver:
ret = sg_alloc_table(sg, ubuf->pagecount, GFP_KERNEL);
for_each_sg(sg->sgl, sgl, ubuf->pagecount, i)
sg_set_folio(sgl, ubuf->folios[i], PAGE_SIZE, ubuf->offsets[i]);
ret = dma_map_sgtable(dev, sg, direction, 0);
That is, populate a scatterlist with ubuf->pagecount number of entries,
where each segment if of size PAGE_SIZE, in order to be consistent and
support a wide variety of DMA importers that may not probably handle
segments that are larger than PAGE_SIZE.
Therefore, in the hugetlb case, there would be multiple entries pointing to
the same folio with different offsets. The question really is whether these
entries associated with @folios and @offsets would need to be populated
by the caller (udmabuf) or the API (memfd_pin_folios). I have tried both of
these approaches in the earlier versions and they all work fine but I think
populating the entries in memfd_pin_folios() seems to be cleaner as the
caller does not need to do any special handling (hugetlb vs shmem).
> *
> * Attempt to pin folios associated with a memfd; given that a memfd is
> * either backed by shmem or hugetlb, the folios can either be found in
> * the page cache or need to be allocated if necessary. Once the folios
> * are located, they are all pinned via FOLL_PIN and @offset is populated
> * with the offset into the first folio.
> *
> * Pinned folios must be released using unpin_folio() or unpin_folios().
> *
> * It must be noted that the folios may be pinned for an indefinite amount
> * of time. And, in most cases, the duration of time they may stay pinned
> * would be controlled by the userspace. This behavior is effectively the
> * same as using FOLL_LONGTERM with other GUP APIs.
> *
> * Returns number of folios pinned, which might be less than @max_folios
> * only if the whole range was pinned. If no folios were pinned, it returns
> * -errno.
> */
> long memfd_pin_folios(struct file *memfd, unsigned long start,
> unsigned long end, struct folio **folios,
> unsigned int max_folios, unsigned long *offset)
> {
> unsigned int pgshift = PAGE_SHIFT;
> unsigned int flags, nr_folios, cur_folios, i;
> pgoff_t start_idx, end_idx;
> struct folio_batch fbatch;
> struct folio *folio;
> struct hstate *h;
> long ret;
>
> if (start > end || !max_folios)
> return -EINVAL;
>
> if (!memfd)
> return -EINVAL;
>
> if (!shmem_file(memfd) && !is_file_hugepages(memfd))
> return -EINVAL;
>
> if (is_file_hugepages(memfd)) {
> h = hstate_file(memfd);
> pgshift = huge_page_shift(h);
> }
>
> flags = memalloc_pin_save();
> folio_batch_init(&fbatch);
> do {
> nr_folios = 0;
> start_idx = start >> pgshift;
> end_idx = end >> pgshift;
> if (is_file_hugepages(memfd)) {
> start_idx <<= huge_page_order(h);
> end_idx <<= huge_page_order(h);
> }
>
> while (start_idx <= end_idx) {
> /*
> * In most cases, we should be able to find the folios
> * in the page cache. If we cannot find them for some
> * reason, we try to allocate them and add them to
> the
> * page cache.
> */
> folio_batch_release(&fbatch);
> cur_folios = filemap_get_folios_contig(memfd-
> >f_mapping,
> &start_idx,
> end_idx,
> &fbatch);
> if (!cur_folios) {
> folio = memfd_alloc_folio(memfd, start_idx);
> if (IS_ERR(folio)) {
> ret = PTR_ERR(folio);
> if (ret != -EEXIST)
> goto err;
> }
> folio_put(folio);
> continue;
> }
>
> /* Let's pin each folio, which shouldn't really fail. */
> for (i = 0; i < cur_folios; i++) {
> folio = try_grab_folio(&fbatch.folios[i]->page,
> 1, FOLL_PIN);
> if (!folio)
> goto err;
>
> if (!nr_folios)
> *offset = offset_in_folio(folio, start);
> folios[nr_folios++] = folio;
>
> if (max_folios == nr_folios)
> break;
> }
> if (max_folios == nr_folios)
> break;
> }
> folio_batch_release(&fbatch);
>
> ret = check_and_migrate_movable_folios(nr_folios, folios);
> } while (ret == -EAGAIN);
>
> memalloc_pin_restore(flags);
> return ret ? ret : nr_folios;
> err:
> folio_batch_release(&fbatch);
> memalloc_pin_restore(flags);
> while (i-- > 0)
> if (folios[i])
> gup_put_folio(folios[i], 1, FOLL_PIN);
>
> return ret;
> }
> EXPORT_SYMBOL_GPL(memfd_pin_folios);
>
>
>
> I'm still wondering about the offset handling, though. Could it happen that
> why we are
> repeatedly calling filemap_get_folios_contig(), that we would need offset!=0
> on any of
> the other folios besides the first one? My current understanding (and looking
> at
> filemap_get_folios_contig()) is: no.
I am not entirely sure but while testing this series with Qemu master + kernel
snapshot of drm-tip which is 6.7 RC1, I noticed strange behavior of
filemap_get_folios_contig() and the batches it returns particularly for the
hugetlb folios. Assuming we have order-9 folios in the memfd (my test-case),
and if the range [start, end] cuts across more than one folio: lets say start is
at subpage 490 (in folio-0) and end is at subpage 520 (in folio-1), then start_idx
would be 0 and end_idx would be 512. In this case, I would have expected
filemap_get_folios_contig() to return two entries in the batch that included
folio-0 and folio-1. However, it returned a batch with 15 entries (max batch size)
with all the entries pointing to folio-0. This is why I added the check:
if (next_idx &&
next_idx != folio_index(fbatch.folios[j]))
continue;
Anyway, based on the code you wrote, I have realized that we both have a
different view on how many entries need to be there in the @folios array
for a given range [start, end] in the hugetlb case.
I have assumed that it is highly desirable to have a segment length of
PAGE_SIZE for consistency and interoperability reasons but I guess it might
be ok to do:
sg_set_folio(sgl, ubuf->folios[i], nr_tails * PAGE_SIZE, ubuf->offsets[i]);
I'll run some experiments to see if this would work in most cases or not.
>
> I'm primarily concerned about concurrent fallocate(PUNCH_HOLE) and THP
> collapse/splitting.
Could you please elaborate on what the issue would be in this case?
Thanks,
Vivek
>
> --
> Cheers,
>
> David / dhildenb
More information about the dri-devel
mailing list