On Wed, Nov 13, 2024 at 5:47 PM Jingbo Xu <jefflexu@xxxxxxxxxxxxxxxxx> wrote: > > > > On 11/14/24 8:39 AM, Joanne Koong wrote: > > On Tue, Nov 12, 2024 at 1:25 AM Jingbo Xu <jefflexu@xxxxxxxxxxxxxxxxx> wrote: > >> > >> Hi Joanne, > >> > >> On 11/8/24 7:56 AM, Joanne Koong wrote: > >>> Currently, we allocate and copy data to a temporary folio when > >>> handling writeback in order to mitigate the following deadlock scenario > >>> that may arise if reclaim waits on writeback to complete: > >>> * single-threaded FUSE server is in the middle of handling a request > >>> that needs a memory allocation > >>> * memory allocation triggers direct reclaim > >>> * direct reclaim waits on a folio under writeback > >>> * the FUSE server can't write back the folio since it's stuck in > >>> direct reclaim > >>> > >>> To work around this, we allocate a temporary folio and copy over the > >>> original folio to the temporary folio so that writeback can be > >>> immediately cleared on the original folio. This additionally requires us > >>> to maintain an internal rb tree to keep track of writeback state on the > >>> temporary folios. > >>> > >>> A recent change prevents reclaim logic from waiting on writeback for > >>> folios whose mappings have the AS_WRITEBACK_MAY_BLOCK flag set in it. > >>> This commit sets AS_WRITEBACK_MAY_BLOCK on FUSE inode mappings (which > >>> will prevent FUSE folios from running into the reclaim deadlock described > >>> above) and removes the temporary folio + extra copying and the internal > >>> rb tree. > >>> > >>> fio benchmarks -- > >>> (using averages observed from 10 runs, throwing away outliers) > >>> > >>> Setup: > >>> sudo mount -t tmpfs -o size=30G tmpfs ~/tmp_mount > >>> ./libfuse/build/example/passthrough_ll -o writeback -o max_threads=4 -o source=~/tmp_mount ~/fuse_mount > >>> > >>> fio --name=writeback --ioengine=sync --rw=write --bs={1k,4k,1M} --size=2G > >>> --numjobs=2 --ramp_time=30 --group_reporting=1 --directory=/root/fuse_mount > >>> > >>> bs = 1k 4k 1M > >>> Before 351 MiB/s 1818 MiB/s 1851 MiB/s > >>> After 341 MiB/s 2246 MiB/s 2685 MiB/s > >>> % diff -3% 23% 45% > >>> > >>> Signed-off-by: Joanne Koong <joannelkoong@xxxxxxxxx> > >> > >> I think there are some places checking or waiting for writeback could be > >> reconsidered if they are still needed or not after we dropping the temp > >> page design. > >> > >> As they are inherited from the original implementation, at least they > >> are harmless. I think they could be remained in this patch, and could > >> be cleaned up later if really needed. > >> > > > > Thank you for the thorough inspection! > > > >> > >>> @@ -891,7 +813,7 @@ static int fuse_do_readfolio(struct file *file, struct folio *folio) > >>> * have writeback that extends beyond the lifetime of the folio. So > >>> * make sure we read a properly synced folio. > >>> */ > >>> - fuse_wait_on_folio_writeback(inode, folio); > >>> + folio_wait_writeback(folio); > >> > >> I doubt if wait-on-writeback is needed here, as now page cache won't be > >> freed until the writeback IO completes. > >> > >> The routine attempts to free page cache, e.g. invalidate_inode_pages2() > >> (generally called by distributed filesystems when the file content has > >> been modified from remote) or truncate_inode_pages() (called from > >> truncate(2) or inode eviction) will wait for writeback completion (if > >> any) before freeing page. > >> > >> Thus I don't think there's any possibility that .read_folio() or > >> .readahead() will be called when the writeback has not completed. > >> > > > > Great point. I'll remove this line and the comment above it too. > > > >> > >>> @@ -1172,7 +1093,7 @@ static ssize_t fuse_send_write_pages(struct fuse_io_args *ia, > >>> int err; > >>> > >>> for (i = 0; i < ap->num_folios; i++) > >>> - fuse_wait_on_folio_writeback(inode, ap->folios[i]); > >>> + folio_wait_writeback(ap->folios[i]); > >> > >> Ditto. > > Actually this is a typo and I originally meant that waiting for > writeback in fuse_send_readpages() is unneeded as page cache won't be > freed until the writeback IO completes. > > > - wait_event(fi->page_waitq, !fuse_range_is_writeback(inode, first, last)); > > + filemap_fdatawait_range(inode->i_mapping, first, last); > Gotcha and agreed. I'll remove this wait from readahead(). > > In fact the above waiting for writeback in fuse_send_write_pages() is > needed. The reason is as follows: > > >> > > > > Why did we need this fuse_wait_on_folio_writeback() even when we had > > the temp pages? If I'm understanding it correctly, > > fuse_send_write_pages() is only called for the writethrough case (by > > fuse_perform_write()), so these folios would never even be under > > writeback, no? > > I think mmap write could make the page dirty and the writeback could be > triggered then. > Ohhh I see, thanks for the explanation. > > > >> > >> > >>> static void fuse_writepage_args_page_fill(struct fuse_writepage_args *wpa, struct folio *folio, > >>> - struct folio *tmp_folio, uint32_t folio_index) > >>> + uint32_t folio_index) > >>> { > >>> struct inode *inode = folio->mapping->host; > >>> struct fuse_args_pages *ap = &wpa->ia.ap; > >>> > >>> - folio_copy(tmp_folio, folio); > >>> - > >>> - ap->folios[folio_index] = tmp_folio; > >>> + folio_get(folio); > >> > >> I still think this folio_get() here is harmless but redundant. > >> > >> Ditto page cache won't be freed before writeback completes. > >> > >> Besides, other .writepages() implementaions e.g. iomap_writepages() also > >> doen't get the refcount when constructing the writeback IO. > >> > > > > Point taken. I'll remove this then, since other .writepages() also > > don't obtain a refcount. > > > >> > >>> @@ -2481,7 +2200,7 @@ static int fuse_write_begin(struct file *file, struct address_space *mapping, > >>> if (IS_ERR(folio)) > >>> goto error; > >>> > >>> - fuse_wait_on_page_writeback(mapping->host, folio->index); > >>> + folio_wait_writeback(folio); > >> > >> I also doubt if wait_on_writeback() is needed here, as now there won't > >> be duplicate writeback IOs for the same page. > > > > What prevents there from being a duplicate writeback write for the > > same page? This is the path I'm looking at: > > > > ksys_write() > > vfs_write() > > new_sync_write() > > op->write_iter() > > fuse_file_write_iter() > > fuse_cache_write_iter() > > generic_file_write_iter() > > __generic_file_write_iter() > > generic_perform_write() > > op->write_begin() > > fuse_write_begin() > > > > but I'm not seeing where there's anything that prevents a duplicate > > write from happening. > > I mean there won't be duplicate *writeback* rather than *write* for the > same page. You could write the page cache and make it dirty at the time > when the writeback for the same page is still on going, as long as we > can ensure that even when the page is dirtied again, there won't be a > duplicate writeback IO for the same page when the previous writeback IO > has not completed yet. > I think we still need this folio_wait_writeback() since we're calling fuse_do_readfolio() and removing the folio_wait_writeback() from fuse_do_readfolio(). else we could read back stale data if the writeback hasn't gone through yet. I think we could probably move the folio_wait_writeback() here in fuse_write_begin() to be right before the fuse_do_readfolio() call and skip waiting on writeback if we hit the "success" gotos. Thanks, Joanne > > > -- > Thanks, > Jingbo