| Message ID | 1362327978-30423-3-git-send-email-wenqing.lz@taobao.com |
|---|---|
| State | Accepted, archived |
| Headers | show |
On Mon, Mar 04, 2013 at 12:26:18AM +0800, Zheng Liu wrote: > From: Zheng Liu <wenqing.lz@taobao.com> > > This commit fixes a bug in filefrag that filefrag miss calculates > contiguous extent when physical blocks are contiguous but logical blocks > aren't. This case can be created by xfstests #218 or the following > script. > > for I in `seq 0 2 31`; do > dd if=/dev/zero of=$testfile bs=4k count=1 conv=notrunc \ > seek=$I oflag=sync &>/dev/null > done > > Meanwhile this commit prints expected logical block. Hmm, this (and your previous patch) fundamentally raises the question of what do we call an "extent", doesn't it? Ignoring for now the question of what xfstests #218 is expecting (if we disagree with what's "best", we should have a discussion with the other fs maintainers, and in the worst case, make our own version of the test), the question is how should defragmentation handle sparse files? In general, sparse files imply that random access workload, so whether or not the file is contiguous doesn't really matter much. If we want to optimize the time to copy said sparse file, and if we assume that by the time we are defragging said sparse file, we are done doing writes which will allocate new blocks, then having defrag optimize the file so that when the extents are sorted by logical block number, the physical block numbers are contiguous, then that's probably the best "figure of merit" to use. And I'll note that right now that's what filefrag is reporting, and what I think e4defrag should be changed to use when deciding whether the donor file is "better" than the original file. But that's not necessarily the only way to measure extents, and the current e4defrag code is clearly of the opinion that if the file is using a contiguous region of blocks, even if the blocks were allocated "backwards", that there's no point defragging the file, since after all, if the file was written in such a random order with respect to logical block numbers, it will probably be read in a random order, so keeping the file blocks used contiguous to minimize free block fragmentation is the best thing to shoot for. It's not clear that there's one right answer, but things will be a lot less confusing if we can agree amongst ourselves what answer we want to use --- and then if we need to either change the xfstests, or maybe create an option to filefrag to calculate the number of fragments that the xfstest is expecting. But we should first decide what is the right thing, and then we can see whether or not what it matches what the test is demanding. - Ted -- To unsubscribe from this list: send the line "unsubscribe linux-ext4" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Wed, Mar 13, 2013 at 04:16:11PM -0400, Theodore Ts'o wrote: > On Mon, Mar 04, 2013 at 12:26:18AM +0800, Zheng Liu wrote: > > From: Zheng Liu <wenqing.lz@taobao.com> > > > > This commit fixes a bug in filefrag that filefrag miss calculates > > contiguous extent when physical blocks are contiguous but logical blocks > > aren't. This case can be created by xfstests #218 or the following > > script. > > > > for I in `seq 0 2 31`; do > > dd if=/dev/zero of=$testfile bs=4k count=1 conv=notrunc \ > > seek=$I oflag=sync &>/dev/null > > done > > > > Meanwhile this commit prints expected logical block. > > Hmm, this (and your previous patch) fundamentally raises the question > of what do we call an "extent", doesn't it? > > Ignoring for now the question of what xfstests #218 is expecting (if > we disagree with what's "best", we should have a discussion with the > other fs maintainers, and in the worst case, make our own version of > the test), the question is how should defragmentation handle sparse > files? In general, sparse files imply that random access workload, so > whether or not the file is contiguous doesn't really matter much. > > If we want to optimize the time to copy said sparse file, and if we > assume that by the time we are defragging said sparse file, we are > done doing writes which will allocate new blocks, then having defrag > optimize the file so that when the extents are sorted by logical block > number, the physical block numbers are contiguous, then that's > probably the best "figure of merit" to use. And I'll note that right > now that's what filefrag is reporting, and what I think e4defrag > should be changed to use when deciding whether the donor file is > "better" than the original file. > > But that's not necessarily the only way to measure extents, and the > current e4defrag code is clearly of the opinion that if the file is > using a contiguous region of blocks, even if the blocks were allocated > "backwards", that there's no point defragging the file, since after > all, if the file was written in such a random order with respect to > logical block numbers, it will probably be read in a random order, so > keeping the file blocks used contiguous to minimize free block > fragmentation is the best thing to shoot for. > > It's not clear that there's one right answer, but things will be a lot > less confusing if we can agree amongst ourselves what answer we want > to use --- and then if we need to either change the xfstests, or maybe > create an option to filefrag to calculate the number of fragments that > the xfstest is expecting. But we should first decide what is the > right thing, and then we can see whether or not what it matches what > the test is demanding. Thanks for the explanation. Indeed the key problem is how to define an extent. I agree with you that we shouldn't only match what the test expects, and just let test case pass. Regards, - Zheng -- To unsubscribe from this list: send the line "unsubscribe linux-ext4" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Thu, Mar 14, 2013 at 08:54:42PM +0800, Zheng Liu wrote: > On Wed, Mar 13, 2013 at 04:16:11PM -0400, Theodore Ts'o wrote: > > On Mon, Mar 04, 2013 at 12:26:18AM +0800, Zheng Liu wrote: > > > From: Zheng Liu <wenqing.lz@taobao.com> > > > > > > This commit fixes a bug in filefrag that filefrag miss calculates > > > contiguous extent when physical blocks are contiguous but logical blocks > > > aren't. This case can be created by xfstests #218 or the following > > > script. > > > > > > for I in `seq 0 2 31`; do > > > dd if=/dev/zero of=$testfile bs=4k count=1 conv=notrunc \ > > > seek=$I oflag=sync &>/dev/null > > > done > > > > > > Meanwhile this commit prints expected logical block. > > > > Hmm, this (and your previous patch) fundamentally raises the question > > of what do we call an "extent", doesn't it? > > > > Ignoring for now the question of what xfstests #218 is expecting (if > > we disagree with what's "best", we should have a discussion with the > > other fs maintainers, and in the worst case, make our own version of > > the test), the question is how should defragmentation handle sparse > > files? In general, sparse files imply that random access workload, so > > whether or not the file is contiguous doesn't really matter much. > > > > If we want to optimize the time to copy said sparse file, and if we > > assume that by the time we are defragging said sparse file, we are > > done doing writes which will allocate new blocks, then having defrag > > optimize the file so that when the extents are sorted by logical block > > number, the physical block numbers are contiguous, then that's > > probably the best "figure of merit" to use. And I'll note that right > > now that's what filefrag is reporting, and what I think e4defrag > > should be changed to use when deciding whether the donor file is > > "better" than the original file. > > > > But that's not necessarily the only way to measure extents, and the > > current e4defrag code is clearly of the opinion that if the file is > > using a contiguous region of blocks, even if the blocks were allocated > > "backwards", that there's no point defragging the file, since after > > all, if the file was written in such a random order with respect to > > logical block numbers, it will probably be read in a random order, so > > keeping the file blocks used contiguous to minimize free block > > fragmentation is the best thing to shoot for. > > > > It's not clear that there's one right answer, but things will be a lot > > less confusing if we can agree amongst ourselves what answer we want > > to use --- and then if we need to either change the xfstests, or maybe > > create an option to filefrag to calculate the number of fragments that > > the xfstest is expecting. But we should first decide what is the > > right thing, and then we can see whether or not what it matches what > > the test is demanding. > > Thanks for the explanation. Indeed the key problem is how to define an > extent. I agree with you that we shouldn't only match what the test > expects, and just let test case pass. AFAIC, there is no ambiguity into what defines an extent. An extent defines a single logical offset to physical block relationship. As soon as either logical offset or physical block are no longer able to be described by the relationship, you need to define a new relationship. That's the definition every filesystem uses for recording extents on disk, and that gives a pretty solid basis for what everyone expects filefrag to report to users as an extent. IOWs, a sparse file is simply a bunch of extent records that have discontiguous logical offsets. Similarly, a fragmented file is simply a bunch of extent records that are physically discontiguous. Therefore, the definition of "sparse" and "fragmented" is not dependent on the definition of "extent", rather it is dependent on the relationship between extents and how the filesystem itself defines that relationship. So, back to test 218. If you have 10 extents that are logically contiguous, but physically discontiguous, is that a fragmented file? The question 218 is asking is not whether those 10 extents were allocated backwards (that's just an underhanded trick to make the XFS allocator create a fragmented file), but rather whether the defragmenter can recognise files that match the above definition of a fragmented file and hence defragment them down to a single extent. Similarly, 218 then creates sparse files and determines whether the defragmenter behaves as expected for a sparse file. That is, you cannot physically merge the extents because they are discontiguous, and hence the file should not be touched by the defragmenter. Hence I think you are reading way too much into the test case. It's testing a basic premise of defragmenter behaviour - recognising sparse and/or fragmented files and dealing with them appropriately. If you want to change how e2defrag defines "fragmented" and "sparse" to be something more complex and different to what 218 expects, then you'll need a new e2defrag specific test. Cheers, Dave.
-----BEGIN PGP SIGNED MESSAGE----- Hash: SHA1 On 3/13/2013 4:16 PM, Theodore Ts'o wrote: > If we want to optimize the time to copy said sparse file, and if > we assume that by the time we are defragging said sparse file, we > are done doing writes which will allocate new blocks, then having > defrag optimize the file so that when the extents are sorted by > logical block number, the physical block numbers are contiguous, > then that's probably the best "figure of merit" to use. And I'll > note that right now that's what filefrag is reporting, and what I > think e4defrag should be changed to use when deciding whether the > donor file is "better" than the original file. That's certainly exactly what I expect to get after defragging a file. > But that's not necessarily the only way to measure extents, and > the current e4defrag code is clearly of the opinion that if the > file is using a contiguous region of blocks, even if the blocks > were allocated "backwards", that there's no point defragging the > file, since after all, if the file was written in such a random > order with respect to logical block numbers, it will probably be > read in a random order, so keeping the file blocks used contiguous > to minimize free block fragmentation is the best thing to shoot > for. That seems wrong to me. It shouldn't matter what condition the file starts in; the end should always be a file with blocks allocated in the proper order. If there's no point in defragging a randomly accessed file, then the user shouldn't bother running defrag on it, and conversely, if they do, it should perform the task they requested. -----BEGIN PGP SIGNATURE----- Version: GnuPG v2.0.17 (MingW32) Comment: Using GnuPG with Thunderbird - http://www.enigmail.net/ iQEcBAEBAgAGBQJRSLIeAAoJEJrBOlT6nu7526UH/RTFmZWDXgrzGUwbbBJ1IN6t aylOR1Wg3SWKEp5Cs717iHCNPMosF31bHkHG+MbpoSvJpAfRPZ7svTDTbzfBZ17F dLKLOAYx3KRBeX8TSYIC7e5E/qTg1C6wmiS1Rlab5boeRkmUL/kP7WPpCgKHpydZ tgUtzO7npCWVxP3eG9NnOR9y9U0PHuS3B9Q5S2vhbeqpxQnEi4cphoSw1bOJVcCj RkddmEUuj+usDl8kHwxlyoiRhcwMUY59UuTqEWqgDkntQL5DTdX//AfrUdW3PBmB DKIC4dCaoivXoKarKRuLaHM04rldOvZdxNFFNnoKlnQMhh4i82RnEYSio26778c= =0Iea -----END PGP SIGNATURE----- -- To unsubscribe from this list: send the line "unsubscribe linux-ext4" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
diff --git a/misc/filefrag.c b/misc/filefrag.c index ee03a07..3af938f 100644 --- a/misc/filefrag.c +++ b/misc/filefrag.c @@ -116,16 +116,17 @@ static int get_bmap(int fd, unsigned long block, unsigned long *phy_blk) static void print_extent_header(void) { - printf(" ext: %*s %*s length: %*s flags:\n", + printf(" ext: %*s %*s length: %*s %*s flags:\n", logical_width * 2 + 3, "logical_offset:", physical_width * 2 + 3, "physical_offset:", - physical_width + 1, - "expected:"); + physical_width + 1, "log_expected:", + physical_width + 1, "phy_expected:"); } static void print_extent_info(struct fiemap_extent *fm_extent, int cur_ex, - unsigned long long expected, int blk_shift, + unsigned long long lexpected, + unsigned long long pexpected, int blk_shift, ext2fs_struct_stat *st) { unsigned long long physical_blk; @@ -133,6 +134,8 @@ static void print_extent_info(struct fiemap_extent *fm_extent, int cur_ex, unsigned long long ext_len; unsigned long long ext_blks; char flags[256] = ""; + char logex_flags[32] = ""; + char phyex_flags[32] = ""; /* For inline data all offsets should be in bytes, not blocks */ if (fm_extent->fe_flags & FIEMAP_EXTENT_DATA_INLINE) @@ -143,11 +146,20 @@ static void print_extent_info(struct fiemap_extent *fm_extent, int cur_ex, logical_blk = fm_extent->fe_logical >> blk_shift; physical_blk = fm_extent->fe_physical >> blk_shift; - if (expected) - sprintf(flags, ext_fmt == hex_fmt ? "%*llx: " : "%*llu: ", - physical_width, expected >> blk_shift); + if (lexpected) + sprintf(logex_flags, ext_fmt == hex_fmt ? "%*llx: " : "%*llu: ", + physical_width, lexpected >> blk_shift); else - sprintf(flags, "%.*s ", physical_width, " "); + sprintf(logex_flags, "%.*s ", physical_width, " "); + + if (pexpected) + sprintf(phyex_flags, ext_fmt == hex_fmt ? "%*llx: " : "%*llu: ", + physical_width, pexpected >> blk_shift); + else + sprintf(phyex_flags, "%.*s ", physical_width, " "); + + strcat(flags, logex_flags); + strcat(flags, phyex_flags); if (fm_extent->fe_flags & FIEMAP_EXTENT_UNKNOWN) strcat(flags, "unknown,"); @@ -188,7 +200,7 @@ static int filefrag_fiemap(int fd, int blk_shift, int *num_extents, struct fiemap_extent *fm_ext = &fiemap->fm_extents[0]; int count = (sizeof(buf) - sizeof(*fiemap)) / sizeof(struct fiemap_extent); - unsigned long long expected = 0; + unsigned long long lexpected = 0, pexpected = 0; unsigned long flags = 0; unsigned int i; static int fiemap_incompat_printed; @@ -230,18 +242,21 @@ static int filefrag_fiemap(int fd, int blk_shift, int *num_extents, for (i = 0; i < fiemap->fm_mapped_extents; i++) { if (fm_ext[i].fe_logical != 0 && - fm_ext[i].fe_physical != expected) { + (fm_ext[i].fe_logical != lexpected || + fm_ext[i].fe_physical != pexpected)) { tot_extents++; } else { - expected = 0; + lexpected = 0; + pexpected = 0; if (!tot_extents) tot_extents = 1; } if (verbose) - print_extent_info(&fm_ext[i], n, expected, - blk_shift, st); + print_extent_info(&fm_ext[i], n, lexpected, + pexpected, blk_shift, st); - expected = fm_ext[i].fe_physical + fm_ext[i].fe_length; + lexpected = fm_ext[i].fe_logical + fm_ext[i].fe_length; + pexpected = fm_ext[i].fe_physical + fm_ext[i].fe_length; if (fm_ext[i].fe_flags & FIEMAP_EXTENT_LAST) last = 1; n++; @@ -308,7 +323,7 @@ static int filefrag_fibmap(int fd, int blk_shift, int *num_extents, if (force_extent && last_block != 0 && (block != last_block + 1 || fm_ext.fe_logical + fm_ext.fe_length != logical)) { - print_extent_info(&fm_ext, *num_extents - 1, + print_extent_info(&fm_ext, *num_extents - 1, logical, (last_block + 1) * st->st_blksize, blk_shift, st); fm_ext.fe_logical = logical; @@ -325,7 +340,7 @@ static int filefrag_fibmap(int fd, int blk_shift, int *num_extents, } if (force_extent) - print_extent_info(&fm_ext, *num_extents - 1, + print_extent_info(&fm_ext, *num_extents - 1, logical, last_block * st->st_blksize, blk_shift, st); return count; @@ -339,7 +354,7 @@ static void frag_report(const char *filename) long fd; unsigned long numblocks; int data_blocks_per_cyl = 1; - int num_extents = 1, expected = ~0; + int num_extents = 1, expected = 0; int is_ext2 = 0; static dev_t last_device; unsigned int flags;