| Message ID | 20220827083607.2345453-6-jhubbard@nvidia.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | convert most filesystems to pin_user_pages_fast() | expand |
On Sat, Aug 27, 2022 at 01:36:06AM -0700, John Hubbard wrote: > Convert the NFS Direct IO layer to use pin_user_pages_fast() and > unpin_user_page(), instead of get_user_pages_fast() and put_page(). Again, this stuff can be hit with ITER_BVEC iterators > - result = iov_iter_get_pages_alloc2(iter, &pagevec, > + result = dio_w_iov_iter_pin_pages_alloc(iter, &pagevec, > rsize, &pgbase); and this will break on those.
On 8/27/22 15:48, Al Viro wrote: > On Sat, Aug 27, 2022 at 01:36:06AM -0700, John Hubbard wrote: >> Convert the NFS Direct IO layer to use pin_user_pages_fast() and >> unpin_user_page(), instead of get_user_pages_fast() and put_page(). > > Again, this stuff can be hit with ITER_BVEC iterators > >> - result = iov_iter_get_pages_alloc2(iter, &pagevec, >> + result = dio_w_iov_iter_pin_pages_alloc(iter, &pagevec, >> rsize, &pgbase); > > and this will break on those. If anyone has an example handy, of a user space program that leads to this situation (O_DIRECT with ITER_BVEC), it would really help me reach enlightenment a lot quicker in this area. :) thanks,
On Sat, Aug 27, 2022 at 04:55:18PM -0700, John Hubbard wrote: > On 8/27/22 15:48, Al Viro wrote: > > On Sat, Aug 27, 2022 at 01:36:06AM -0700, John Hubbard wrote: > >> Convert the NFS Direct IO layer to use pin_user_pages_fast() and > >> unpin_user_page(), instead of get_user_pages_fast() and put_page(). > > > > Again, this stuff can be hit with ITER_BVEC iterators > > > >> - result = iov_iter_get_pages_alloc2(iter, &pagevec, > >> + result = dio_w_iov_iter_pin_pages_alloc(iter, &pagevec, > >> rsize, &pgbase); > > > > and this will break on those. > > If anyone has an example handy, of a user space program that leads > to this situation (O_DIRECT with ITER_BVEC), it would really help > me reach enlightenment a lot quicker in this area. :) Er... splice(2) to O_DIRECT-opened file on e.g. ext4? Or sendfile(2) to the same, for that matter...
On Sun, Aug 28, 2022 at 01:38:57AM +0100, Al Viro wrote: > On Sat, Aug 27, 2022 at 04:55:18PM -0700, John Hubbard wrote: > > On 8/27/22 15:48, Al Viro wrote: > > > On Sat, Aug 27, 2022 at 01:36:06AM -0700, John Hubbard wrote: > > >> Convert the NFS Direct IO layer to use pin_user_pages_fast() and > > >> unpin_user_page(), instead of get_user_pages_fast() and put_page(). > > > > > > Again, this stuff can be hit with ITER_BVEC iterators > > > > > >> - result = iov_iter_get_pages_alloc2(iter, &pagevec, > > >> + result = dio_w_iov_iter_pin_pages_alloc(iter, &pagevec, > > >> rsize, &pgbase); > > > > > > and this will break on those. > > > > If anyone has an example handy, of a user space program that leads > > to this situation (O_DIRECT with ITER_BVEC), it would really help > > me reach enlightenment a lot quicker in this area. :) > > Er... splice(2) to O_DIRECT-opened file on e.g. ext4? Or > sendfile(2) to the same, for that matter... s/ext4/nfs/ to hit this particular codepath, obviously.
On 8/27/22 17:39, Al Viro wrote: > On Sun, Aug 28, 2022 at 01:38:57AM +0100, Al Viro wrote: >> On Sat, Aug 27, 2022 at 04:55:18PM -0700, John Hubbard wrote: >>> On 8/27/22 15:48, Al Viro wrote: >>>> On Sat, Aug 27, 2022 at 01:36:06AM -0700, John Hubbard wrote: >>>>> Convert the NFS Direct IO layer to use pin_user_pages_fast() and >>>>> unpin_user_page(), instead of get_user_pages_fast() and put_page(). >>>> >>>> Again, this stuff can be hit with ITER_BVEC iterators >>>> >>>>> - result = iov_iter_get_pages_alloc2(iter, &pagevec, >>>>> + result = dio_w_iov_iter_pin_pages_alloc(iter, &pagevec, >>>>> rsize, &pgbase); >>>> >>>> and this will break on those. >>> >>> If anyone has an example handy, of a user space program that leads >>> to this situation (O_DIRECT with ITER_BVEC), it would really help >>> me reach enlightenment a lot quicker in this area. :) >> >> Er... splice(2) to O_DIRECT-opened file on e.g. ext4? Or >> sendfile(2) to the same, for that matter... > > s/ext4/nfs/ to hit this particular codepath, obviously. aha, thanks. I do remember that you alerted me earlier to splice(2) problems, and I thought I'd neatly sidestepped those this time, by staying with user_backed_iter(i) cases. But I see that it's going to take something more, after all. thanks,
On 8/27/22 17:39, Al Viro wrote: > On Sun, Aug 28, 2022 at 01:38:57AM +0100, Al Viro wrote: >> On Sat, Aug 27, 2022 at 04:55:18PM -0700, John Hubbard wrote: >>> On 8/27/22 15:48, Al Viro wrote: >>>> On Sat, Aug 27, 2022 at 01:36:06AM -0700, John Hubbard wrote: >>>>> Convert the NFS Direct IO layer to use pin_user_pages_fast() and >>>>> unpin_user_page(), instead of get_user_pages_fast() and put_page(). >>>> >>>> Again, this stuff can be hit with ITER_BVEC iterators >>>> >>>>> - result = iov_iter_get_pages_alloc2(iter, &pagevec, >>>>> + result = dio_w_iov_iter_pin_pages_alloc(iter, &pagevec, >>>>> rsize, &pgbase); >>>> >>>> and this will break on those. >>> >>> If anyone has an example handy, of a user space program that leads >>> to this situation (O_DIRECT with ITER_BVEC), it would really help >>> me reach enlightenment a lot quicker in this area. :) >> >> Er... splice(2) to O_DIRECT-opened file on e.g. ext4? Or >> sendfile(2) to the same, for that matter... > > s/ext4/nfs/ to hit this particular codepath, obviously. OK, I have a solution to this that's pretty easy: 1) Get rid of the user_backed_iter(i) check in dio_w_iov_iter_pin_pages() and dio_w_iov_iter_pin_pages_alloc(), and 2) At the call sites, match up the unpin calls appropriately. ...and apply a similar fix for the fuse conversion patch. However, the core block/bio conversion in patch 4 still does depend upon a key assumption, which I got from a 2019 email discussion with Christoph Hellwig and others here [1], which says: "All pages released by bio_release_pages should come from get_get_user_pages...". I really hope that still holds true. Otherwise this whole thing is in trouble. [1] https://lore.kernel.org/kvm/20190724053053.GA18330@infradead.org/ thanks,
On Sun 28-08-22 21:59:49, John Hubbard wrote: > On 8/27/22 17:39, Al Viro wrote: > > On Sun, Aug 28, 2022 at 01:38:57AM +0100, Al Viro wrote: > >> On Sat, Aug 27, 2022 at 04:55:18PM -0700, John Hubbard wrote: > >>> On 8/27/22 15:48, Al Viro wrote: > >>>> On Sat, Aug 27, 2022 at 01:36:06AM -0700, John Hubbard wrote: > >>>>> Convert the NFS Direct IO layer to use pin_user_pages_fast() and > >>>>> unpin_user_page(), instead of get_user_pages_fast() and put_page(). > >>>> > >>>> Again, this stuff can be hit with ITER_BVEC iterators > >>>> > >>>>> - result = iov_iter_get_pages_alloc2(iter, &pagevec, > >>>>> + result = dio_w_iov_iter_pin_pages_alloc(iter, &pagevec, > >>>>> rsize, &pgbase); > >>>> > >>>> and this will break on those. > >>> > >>> If anyone has an example handy, of a user space program that leads > >>> to this situation (O_DIRECT with ITER_BVEC), it would really help > >>> me reach enlightenment a lot quicker in this area. :) > >> > >> Er... splice(2) to O_DIRECT-opened file on e.g. ext4? Or > >> sendfile(2) to the same, for that matter... > > > > s/ext4/nfs/ to hit this particular codepath, obviously. > > OK, I have a solution to this that's pretty easy: > > 1) Get rid of the user_backed_iter(i) check in > dio_w_iov_iter_pin_pages() and dio_w_iov_iter_pin_pages_alloc(), and > > 2) At the call sites, match up the unpin calls appropriately. > > ...and apply a similar fix for the fuse conversion patch. > > However, the core block/bio conversion in patch 4 still does depend upon > a key assumption, which I got from a 2019 email discussion with > Christoph Hellwig and others here [1], which says: > > "All pages released by bio_release_pages should come from > get_get_user_pages...". > > I really hope that still holds true. Otherwise this whole thing is in > trouble. > > [1] https://lore.kernel.org/kvm/20190724053053.GA18330@infradead.org/ Well as far as I've checked that discussion, Christoph was aware of pipe pages etc. (i.e., bvecs) entering direct IO code. But he had some patches [2] which enabled GUP to work for bvecs as well (using the kernel mapping under the hood AFAICT from a quick glance at the series). I suppose we could also handle this in __iov_iter_get_pages_alloc() by grabbing pin reference instead of plain get_page() for the case of bvec iter. That way we should have only pinned pages in bio_release_pages() even for the bvec case. [2] http://git.infradead.org/users/hch/misc.git/shortlog/refs/heads/gup-bvec Honza
On 8/29/22 09:08, Jan Kara wrote: >> However, the core block/bio conversion in patch 4 still does depend upon >> a key assumption, which I got from a 2019 email discussion with >> Christoph Hellwig and others here [1], which says: >> >> "All pages released by bio_release_pages should come from >> get_get_user_pages...". >> >> I really hope that still holds true. Otherwise this whole thing is in >> trouble. >> >> [1] https://lore.kernel.org/kvm/20190724053053.GA18330@infradead.org/ > > Well as far as I've checked that discussion, Christoph was aware of pipe > pages etc. (i.e., bvecs) entering direct IO code. But he had some patches > [2] which enabled GUP to work for bvecs as well (using the kernel mapping > under the hood AFAICT from a quick glance at the series). I suppose we > could also handle this in __iov_iter_get_pages_alloc() by grabbing pin > reference instead of plain get_page() for the case of bvec iter. That way > we should have only pinned pages in bio_release_pages() even for the bvec > case. OK, thanks, that looks viable. So, that approach assumes that the remaining two cases in __iov_iter_get_pages_alloc() will never end up being released via bio_release_pages(): iov_iter_is_pipe(i) iov_iter_is_xarray(i) I'm actually a little worried about ITER_XARRAY, which is a recent addition. It seems to be used in ways that are similar to ITER_BVEC, and cephfs is using it. It's probably OK for now, for this series, which doesn't yet convert cephfs. > > [2] http://git.infradead.org/users/hch/misc.git/shortlog/refs/heads/gup-bvec Yes, I had looked through that again before sending this. The problem for me was that that it didn't have to deal with releasing pages differently (and therefore, differentiating between FOLL_PIN and FOLL_GET pages). But it did enable GUP to handle bvecs, so with that applied, one could then make the original claim about bio_release_pages() and GUP, yes. thanks,
On Mon 29-08-22 12:59:26, John Hubbard wrote: > On 8/29/22 09:08, Jan Kara wrote: > >> However, the core block/bio conversion in patch 4 still does depend upon > >> a key assumption, which I got from a 2019 email discussion with > >> Christoph Hellwig and others here [1], which says: > >> > >> "All pages released by bio_release_pages should come from > >> get_get_user_pages...". > >> > >> I really hope that still holds true. Otherwise this whole thing is in > >> trouble. > >> > >> [1] https://lore.kernel.org/kvm/20190724053053.GA18330@infradead.org/ > > > > Well as far as I've checked that discussion, Christoph was aware of pipe > > pages etc. (i.e., bvecs) entering direct IO code. But he had some patches > > [2] which enabled GUP to work for bvecs as well (using the kernel mapping > > under the hood AFAICT from a quick glance at the series). I suppose we > > could also handle this in __iov_iter_get_pages_alloc() by grabbing pin > > reference instead of plain get_page() for the case of bvec iter. That way > > we should have only pinned pages in bio_release_pages() even for the bvec > > case. > > OK, thanks, that looks viable. So, that approach assumes that the > remaining two cases in __iov_iter_get_pages_alloc() will never end up > being released via bio_release_pages(): > > iov_iter_is_pipe(i) > iov_iter_is_xarray(i) > > I'm actually a little worried about ITER_XARRAY, which is a recent addition. > It seems to be used in ways that are similar to ITER_BVEC, and cephfs is > using it. It's probably OK for now, for this series, which doesn't yet > convert cephfs. So after looking into that a bit more, I think a clean approach would be to provide iov_iter_pin_pages2() and iov_iter_pages_alloc2(), under the hood in __iov_iter_get_pages_alloc() make sure we use pin_user_page() instead of get_page() in all the cases (using this in pipe_get_pages() and iter_xarray_get_pages() is easy) and then make all bio handling use the pinning variants for iters. I think at least iov_iter_is_pipe() case needs to be handled as well because as I wrote above, pipe pages can enter direct IO code e.g. for splice(2). Also I think that all iov_iter_get_pages2() (or the _alloc2 variant) users actually do want the "pin page" semantics in the end (they are accessing page contents) so eventually we should convert them all to iov_iter_pin_pages2() and remove iov_iter_get_pages2() altogether. But this will take some more conversion work with networking etc. so I'd start with converting bios only. Honza
On 8/31/22 02:43, Jan Kara wrote: >> OK, thanks, that looks viable. So, that approach assumes that the >> remaining two cases in __iov_iter_get_pages_alloc() will never end up >> being released via bio_release_pages(): >> >> iov_iter_is_pipe(i) >> iov_iter_is_xarray(i) >> >> I'm actually a little worried about ITER_XARRAY, which is a recent addition. >> It seems to be used in ways that are similar to ITER_BVEC, and cephfs is >> using it. It's probably OK for now, for this series, which doesn't yet >> convert cephfs. > > So after looking into that a bit more, I think a clean approach would be to > provide iov_iter_pin_pages2() and iov_iter_pages_alloc2(), under the hood > in __iov_iter_get_pages_alloc() make sure we use pin_user_page() instead of > get_page() in all the cases (using this in pipe_get_pages() and > iter_xarray_get_pages() is easy) and then make all bio handling use the > pinning variants for iters. I think at least iov_iter_is_pipe() case needs > to be handled as well because as I wrote above, pipe pages can enter direct > IO code e.g. for splice(2). > OK, yes. > Also I think that all iov_iter_get_pages2() (or the _alloc2 variant) users > actually do want the "pin page" semantics in the end (they are accessing > page contents) so eventually we should convert them all to > iov_iter_pin_pages2() and remove iov_iter_get_pages2() altogether. But this > will take some more conversion work with networking etc. so I'd start with > converting bios only. > > Honza I'll give this approach a spin, thanks very much for the guidance! thanks,
On Wed, Aug 31, 2022 at 11:43:49AM +0200, Jan Kara wrote: > So after looking into that a bit more, I think a clean approach would be to > provide iov_iter_pin_pages2() and iov_iter_pages_alloc2(), under the hood > in __iov_iter_get_pages_alloc() make sure we use pin_user_page() instead of > get_page() in all the cases (using this in pipe_get_pages() and > iter_xarray_get_pages() is easy) and then make all bio handling use the > pinning variants for iters. I think at least iov_iter_is_pipe() case needs > to be handled as well because as I wrote above, pipe pages can enter direct > IO code e.g. for splice(2). > > Also I think that all iov_iter_get_pages2() (or the _alloc2 variant) users > actually do want the "pin page" semantics in the end (they are accessing > page contents) so eventually we should convert them all to > iov_iter_pin_pages2() and remove iov_iter_get_pages2() altogether. But this > will take some more conversion work with networking etc. so I'd start with > converting bios only. Not sure, TBH... FWIW, quite a few of the callers of iov_iter_get_pages2() do *NOT* need to grab any references for BVEC/XARRAY/PIPE cases. What's more, it would be bloody useful to have a variant that doesn't grab references for !iter->user_backed case - that could be usable for KVEC as well, simplifying several callers. Requirements: * recepients of those struct page * should have a way to make dropping the page refs conditional (obviously); bio machinery can be told to do so. * callers should *NOT* do something like "set an ITER_BVEC iter, with page references grabbed and stashed in bio_vec array, call async read_iter() and drop the references in array - the refs we grab in dio will serve" Note that for sync IO that pattern is fine whether we grab/drop anything inside read_iter(); for async we could take depopulating the bio_vec array to the IO completion or downstream of that. * the code dealing with the references returned by iov_iter_..._pages should *NOT* play silly buggers with refcounts - something like "I'll grab a reference, start DMA and report success; page will stay around until I get around to dropping the ref and callers don't need to wait for that" deep in the bowels of infinibad stack (or something equally tasteful) is seriously asking for trouble. Future plans from the last cycle included iov_iter_find_pages{,_alloc}() that would *not* grab references on anything other than IOVEC and UBUF (would advance the iterator, same as iov_iter_get_pages2(), though). Then iov_iter_get_...() would become a wrapper for that. After that - look into switching the users of ..._get_... to ..._find_.... Hadn't done much in that direction yet, though - need to redo the analysis first. That primitive might very well do FOLL_PIN instead of FOLL_GET for IOVEC and UBUF...
On Thu 01-09-22 01:38:21, Al Viro wrote: > On Wed, Aug 31, 2022 at 11:43:49AM +0200, Jan Kara wrote: > > > So after looking into that a bit more, I think a clean approach would be to > > provide iov_iter_pin_pages2() and iov_iter_pages_alloc2(), under the hood > > in __iov_iter_get_pages_alloc() make sure we use pin_user_page() instead of > > get_page() in all the cases (using this in pipe_get_pages() and > > iter_xarray_get_pages() is easy) and then make all bio handling use the > > pinning variants for iters. I think at least iov_iter_is_pipe() case needs > > to be handled as well because as I wrote above, pipe pages can enter direct > > IO code e.g. for splice(2). > > > > Also I think that all iov_iter_get_pages2() (or the _alloc2 variant) users > > actually do want the "pin page" semantics in the end (they are accessing > > page contents) so eventually we should convert them all to > > iov_iter_pin_pages2() and remove iov_iter_get_pages2() altogether. But this > > will take some more conversion work with networking etc. so I'd start with > > converting bios only. > > Not sure, TBH... > > FWIW, quite a few of the callers of iov_iter_get_pages2() do *NOT* need to > grab any references for BVEC/XARRAY/PIPE cases. What's more, it would be > bloody useful to have a variant that doesn't grab references for > !iter->user_backed case - that could be usable for KVEC as well, simplifying > several callers. > > Requirements: > * recepients of those struct page * should have a way to make > dropping the page refs conditional (obviously); bio machinery can be told > to do so. > * callers should *NOT* do something like > "set an ITER_BVEC iter, with page references grabbed and stashed in > bio_vec array, call async read_iter() and drop the references in array - the > refs we grab in dio will serve" > Note that for sync IO that pattern is fine whether we grab/drop anything > inside read_iter(); for async we could take depopulating the bio_vec > array to the IO completion or downstream of that. > * the code dealing with the references returned by iov_iter_..._pages > should *NOT* play silly buggers with refcounts - something like "I'll grab > a reference, start DMA and report success; page will stay around until I > get around to dropping the ref and callers don't need to wait for that" deep > in the bowels of infinibad stack (or something equally tasteful) is seriously > asking for trouble. I agree we could get away without grabbing references in some cases. But it is a performance vs robustness tradeoff I'd say. E.g. with XARRAY case I can see people feeding pages from struct address_space and it is unclear what else than page reference would protect them from being freed by reclaim. Furthermore if e.g. writeback can happen for such struct address_space (or other filesystem operations requiring stable data), we really need a full page pin and not just page reference to signal that the page may be under DMA and may be changing under your hands. So I'm not against having some special cases that avoid grabbing page reference / page pin but I think it is justified only in performance sensitive cases and when we can make sure filesystem backed page (because of above mentioned data stability issues) or anon page (because of cow handling) cannot possibly enter this path. Honza
diff --git a/fs/nfs/direct.c b/fs/nfs/direct.c index 1707f46b1335..f6e47329e092 100644 --- a/fs/nfs/direct.c +++ b/fs/nfs/direct.c @@ -142,13 +142,6 @@ int nfs_swap_rw(struct kiocb *iocb, struct iov_iter *iter) return 0; } -static void nfs_direct_release_pages(struct page **pages, unsigned int npages) -{ - unsigned int i; - for (i = 0; i < npages; i++) - put_page(pages[i]); -} - void nfs_init_cinfo_from_dreq(struct nfs_commit_info *cinfo, struct nfs_direct_req *dreq) { @@ -332,11 +325,11 @@ static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq, size_t pgbase; unsigned npages, i; - result = iov_iter_get_pages_alloc2(iter, &pagevec, + result = dio_w_iov_iter_pin_pages_alloc(iter, &pagevec, rsize, &pgbase); if (result < 0) break; - + bytes = result; npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE; for (i = 0; i < npages; i++) { @@ -362,7 +355,7 @@ static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq, pos += req_len; dreq->bytes_left -= req_len; } - nfs_direct_release_pages(pagevec, npages); + dio_w_unpin_user_pages(pagevec, npages); kvfree(pagevec); if (result < 0) break; @@ -791,8 +784,8 @@ static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq, size_t pgbase; unsigned npages, i; - result = iov_iter_get_pages_alloc2(iter, &pagevec, - wsize, &pgbase); + result = dio_w_iov_iter_pin_pages_alloc(iter, &pagevec, + wsize, &pgbase); if (result < 0) break; @@ -829,7 +822,7 @@ static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq, pos += req_len; dreq->bytes_left -= req_len; } - nfs_direct_release_pages(pagevec, npages); + dio_w_unpin_user_pages(pagevec, npages); kvfree(pagevec); if (result < 0) break;
Convert the NFS Direct IO layer to use pin_user_pages_fast() and unpin_user_page(), instead of get_user_pages_fast() and put_page(). Signed-off-by: John Hubbard <jhubbard@nvidia.com> --- fs/nfs/direct.c | 19 ++++++------------- 1 file changed, 6 insertions(+), 13 deletions(-)