login
Search
No further response to our patches yet, so we are sending them again,=20 re-diffed against 2.6.23-rc5 Hi, recently we discovered writing to a nfs-exported lustre filesystem is rathe= r=20 slow (20-40 MB/s writing, but over 200 MB/s reading). As I already explained on the nfs mailing list, this happens since there is= an=20 offset on the very first page due to the nfs header. http://sourceforge.net/mailarchive/forum.php?thread_name=3D200708312003.304= 46.bernd-schubert%40gmx.de&forum_name=3Dnfs While this especially effects lustre, Olaf Kirch also noticed it on another= =20 filesystem before and wrote a nfs patch for it. This patch has two=20 disadvantages - it requires to move all data within the pages, IMHO rather= =20 cpu time consuming, furthermore, it presently causes data corruption when=20 more than one nfs thread is running. After thinking it over and over again we (Goswin and I) believe it would be= =20 best to improve generic_file_buffered_write(). If there is sufficient data now, as it is usual for aio writes,=20 generic_file_buffered_write() will now fill each page as much as possible a= nd=20 only then prepare/commit it. Before generic_file_buffered_write() commited= =20 chunks of pages even though there were still more data. Some statistics: num_writes =3D 4669440, bytes_total =3D 20231249633, segs_total =3D 5738644= ,=20 commit_loops =3D 7697604, commits_total =3D 6628750 commit_loops is the number commits without the patch and commits_total the= =20 number of commits we actually have now. This shows a saving of nearly 14% o= f=20 prepare, commit, cond_sched calls. < 1: Write size =3D 0, Num segs =3D 0 < 2: Write size =3D 20244, Num segs =3D 4455583 < 4: Write size =3D 6722, Num segs =3D 24 < 8: Write size =3D 19653, Num segs =3D 213842 < 16: Write size =3D 31778, Num segs =3D 0 < 32: Write size =3D 73395, Num segs =3D 0 < 64: Write size =3D 148840, Num segs =3D 0 < 128: Write size =3D 310178, Num segs =3D 0 < 256: Write size =3D 89027, Num segs =3D 0 < 512: Write size =3D 111903, Num segs =3D 0 < 1024: Write size =3D 140509, Num segs =3D 0 < 2048: Write size =3D 244052, Num segs =3D 0 < 4096: Write size =3D 217164, Num segs =3D 0 < 8192: Write size =3D 2784875, Num segs =3D 0 < 16384: Write size =3D 433506, Num segs =3D 0 < 32768: Write size =3D 11742, Num segs =3D 0 < 65536: Write size =3D 15783, Num segs =3D 0 < 131072: Write size =3D 6851, Num segs =3D 0 < 262144: Write size =3D 1562, Num segs =3D 0 < 524288: Write size =3D 755, Num segs =3D 0 < 1048576: Write size =3D 531, Num segs =3D 0 < 2097152: Write size =3D 272, Num segs =3D 0 < 4194304: Write size =3D 107, Num segs =3D 0 < 8388608: Write size =3D 0, Num segs =3D 0 Write size shows the number of writes with the total size smaller than deno= ted=20 in the first column. Num segs shows the number of writes with less segments= =20 than denoted in the first column. Most writes (~95%) only have one segment.= =20 However, no nfs activity has been done, which is actually the case we made= =20 the patches for. size\num 1 2 3 4 5 6 7+ =20 < 1: 0 24 0 0 0 0 0 =20 < 2: 20244 0 0 0 0 641526 0 =20 < 4: 6722 0 0 0 0 0 0 =20 < 8: 19653 0 0 0 0 213842 0 =20 < 16: 31778 0 0 0 0 213856 0 =20 < 32: 73395 0 0 0 0 590 0 =20 < 64: 147730 0 0 0 0 93626 0 =20 < 128: 100888 0 0 0 0 119597 0 =20 < 256: 85588 0 0 0 0 12 0 =20 < 512: 111900 0 0 0 0 3 0 =20 < 1024: 140509 0 0 0 0 0 0 =20 < 2048: 244052 0 0 0 0 0 0 =20 < 4096: 217160 4 0 0 0 0 0 =20 < 8192: 2784855 20 0 0 0 0 0 =20 < 16384: 433506 0 0 0 0 0 0 =20 < 32768: 11742 0 0 0 0 0 0 =20 < 65536: 15783 0 0 0 0 0 0 =20 < 131072: 6851 0 0 0 0 0 0 =20 < 262144: 1562 0 0 0 0 0 0 =20 < 524288: 755 0 0 0 0 0 0 =20 < 1048576: 531 0 0 0 0 0 0 =20 < 2097152: 272 0 0 0 0 0 0 =20 < 4194304: 107 0 0 0 0 0 0 =20 < 8388608: 0 0 0 0 0 0 0 =20 This table shows the number of segments smaller than denoted in the first=20 column within a vector of sizes 1 to 7 segments. This shows that without nf= s=20 only 1, 2 and 6 segment vectors appear to happen. Depending on wsize nfs wi= ll=20 have a lot more segments (129 for 512kiB wsize). Btw, whats the cause of the 24 vectors with 2 segments, one of them being s= ize=20 zero? Signed-off-by: Bernd Schubert <bs@q-leap.de> Signed-off-by: Goswin von Brederlow <goswin@vonbrederlow.de> Cheers, Goswin and Bernd mm/filemap.c | 139 +++++++++++++++++++++++++++++++++---------------- 1 file changed, 95 insertions(+), 44 deletions(-) =20 Index: linux-2.6.23-rc5/mm/filemap.c =3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D= =3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D= =3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D =2D-- linux-2.6.23-rc5.orig/mm/filemap.c 2007-09-06 18:33:11.000000000 +0200 +++ linux-2.6.23-rc5/mm/filemap.c 2007-09-06 18:33:15.000000000 +0200 @@ -1834,6 +1834,21 @@ } EXPORT_SYMBOL(generic_file_direct_write); =20 +/** + * generic_file_buffered_write - handle iov'ectors + * @iob: file operations + * @iov: vector of data to write + * @nr_segs: number of iov segments + * @pos: position in the file + * @ppos: position in the file after this function + * @count: number of bytes to write + * written: offset in iov->base (data to skip on write) + * + * This function will do 3 main tasks for each iov: + * - prepare a write + * - copy the data from iov into a new page + * - commit this page + */ ssize_t generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov, unsigned long nr_segs, loff_t pos, loff_t *ppos, @@ -1851,6 +1866,11 @@ const struct iovec *cur_iov =3D iov; /* current iovec */ size_t iov_base =3D 0; /* offset in the current iovec */ char __user *buf; + unsigned long data_start =3D (pos & (PAGE_CACHE_SIZE -1)); /* Within page= */ + loff_t wpos =3D pos; /* the position in the file we will return */ + + /* position in file as index of pages */ + unsigned long index =3D pos >> PAGE_CACHE_SHIFT; =20 pagevec_init(&lru_pvec, 0); =20 @@ -1864,9 +1884,15 @@ buf =3D cur_iov->iov_base + iov_base; } =20 + page =3D __grab_cache_page(mapping, index, &cached_page, &lru_pvec); + if (!page) { + status =3D -ENOMEM; + goto out; + } + do { =2D unsigned long index; unsigned long offset; + unsigned long data_end; /* end of data within the page */ size_t copied; =20 offset =3D (pos & (PAGE_CACHE_SIZE -1)); /* Within page */ @@ -1897,11 +1923,6 @@ */ fault_in_pages_readable(buf, bytes); } =2D page =3D __grab_cache_page(mapping,index,&cached_page,&lru_pvec); =2D if (!page) { =2D status =3D -ENOMEM; =2D break; =2D } =20 if (unlikely(bytes =3D=3D 0)) { status =3D 0; @@ -1909,22 +1930,26 @@ goto zero_length_segment; } =20 =2D status =3D a_ops->prepare_write(file, page, offset, offset+bytes); =2D if (unlikely(status)) { =2D loff_t isize =3D i_size_read(inode); + data_end =3D offset + bytes; =20 =2D if (status !=3D AOP_TRUNCATED_PAGE) =2D unlock_page(page); =2D page_cache_release(page); =2D if (status =3D=3D AOP_TRUNCATED_PAGE) =2D continue; + if (data_end =3D=3D PAGE_CACHE_SIZE || count =3D=3D bytes) { /* =2D * prepare_write() may have instantiated a few blocks =2D * outside i_size. Trim these off again. + * Only prepare a write if its an entire page or if + * we don't have more data */ =2D if (pos + bytes > isize) =2D vmtruncate(inode, isize); =2D break; + status =3D a_ops->prepare_write(file, page, data_start, data_end); + if (unlikely(status)) { + loff_t isize =3D i_size_read(inode); + + if (status =3D=3D AOP_TRUNCATED_PAGE) + continue; + /* + * prepare_write() may have instantiated a few blocks + * outside i_size. Trim these off again. + */ + if (pos + bytes > isize) + vmtruncate(inode, isize); + } } if (likely(nr_segs =3D=3D 1)) copied =3D filemap_copy_from_user(page, offset, @@ -1932,60 +1957,86 @@ else copied =3D filemap_copy_from_user_iovec(page, offset, cur_iov, iov_base, bytes); =2D flush_dcache_page(page); =2D status =3D a_ops->commit_write(file, page, offset, offset+bytes); =2D if (status =3D=3D AOP_TRUNCATED_PAGE) { + + if (data_end =3D=3D PAGE_CACHE_SIZE || count =3D=3D bytes) { + /* + * Same as above, always try fill pages up to + * PAGE_CACHE_SIZE if possible (sufficient data) + */ + flush_dcache_page(page); + status =3D a_ops->commit_write(file, page, + data_start, data_end); + if (status =3D=3D AOP_TRUNCATED_PAGE) { + continue; + } + unlock_page(page); + mark_page_accessed(page); page_cache_release(page); =2D continue; + balance_dirty_pages_ratelimited(mapping); + cond_resched(); + if (bytes < count) { /* still more iov data to write */ + page =3D __grab_cache_page(mapping, index + 1, + &cached_page, &lru_pvec); + if (!page) { + status =3D -ENOMEM; + goto out; + } + } else { + page =3D NULL; + } + written +=3D data_end - data_start; + wpos +=3D data_end - data_start; + data_start =3D 0; /* correct would be data_end % PAGE_SIZE + * but if this would be !=3D 0, we don't + * have more data + */ } zero_length_segment: if (likely(copied >=3D 0)) { =2D if (!status) =2D status =3D copied; =2D =2D if (status >=3D 0) { =2D written +=3D status; =2D count -=3D status; =2D pos +=3D status; =2D buf +=3D status; + if (!status) { + count -=3D copied; + pos +=3D copied; + buf +=3D copied; if (unlikely(nr_segs > 1)) { filemap_set_next_iovec(&cur_iov, =2D &iov_base, status); + &iov_base, copied); if (count) buf =3D cur_iov->iov_base + iov_base; } else { =2D iov_base +=3D status; + iov_base +=3D copied; } } } if (unlikely(copied !=3D bytes)) =2D if (status >=3D 0) + if (!status) status =3D -EFAULT; =2D unlock_page(page); =2D mark_page_accessed(page); =2D page_cache_release(page); =2D if (status < 0) + if (status) break; =2D balance_dirty_pages_ratelimited(mapping); =2D cond_resched(); } while (count); =2D *ppos =3D pos; + +out: + *ppos =3D wpos; =20 if (cached_page) page_cache_release(cached_page); =20 + if (page) { + unlock_page(page); + page_cache_release(page); + } + /* * For now, when the user asks for O_SYNC, we'll actually give O_DSYNC */ =2D if (likely(status >=3D 0)) { + if (likely(!status)) { if (unlikely((file->f_flags & O_SYNC) || IS_SYNC(inode))) { if (!a_ops->writepage || !is_sync_kiocb(iocb)) status =3D generic_osync_inode(inode, mapping, OSYNC_METADATA|OSYNC_DATA); } =2D } =2D=09 + } + /* * If we get here for O_DIRECT writes then we must have fallen through * to buffered writes (block instantiation inside i_size). So we sync =2D-=20 Bernd Schubert Q-Leap Networks GmbH
