Message ID | 1c7c39f7-91a7-be85-5906-e55180a91a5f@sandeen.net (mailing list archive) |
---|---|
State | Accepted |
Headers | show |
Series | [V4] libxfs: use FALLOC_FL_ZERO_RANGE in libxfs_device_zero | expand |
On Tue, Feb 25, 2020 at 10:13:55AM -0800, Eric Sandeen wrote: > I had a request from someone who cared about mkfs speed over > a slower network block device to look into using faster zeroing > methods, particularly for the log, during mkfs. > > Using FALLOC_FL_ZERO_RANGE is faster in this case than writing > a bunch of zeros across a wire. > > Signed-off-by: Eric Sandeen <sandeen@redhat.com> Looks good: Reviewed-by: Christoph Hellwig <hch@lst.de>
On Tue, Feb 25, 2020 at 10:13:55AM -0800, Eric Sandeen wrote: > I had a request from someone who cared about mkfs speed over > a slower network block device to look into using faster zeroing > methods, particularly for the log, during mkfs. > > Using FALLOC_FL_ZERO_RANGE is faster in this case than writing > a bunch of zeros across a wire. > > Signed-off-by: Eric Sandeen <sandeen@redhat.com> > --- > > V2: Clean up all the nasty stuff I'd flung out there as a wild first > cut, thanks Dave. > > V3: make len_bytes a size_t; leave "end_offset" where it is for the loop > use. It's a bit odd but ... just don't mess with it for now, one patch > one change. > > V4: Use EOPNOTSUPP not EOPNOTSUP (same on linux anyway but meh) > I ignored(tm) darrick's suggestion to make libxfs_device_zero accept > a longer length, for now - no callers need anything bigger at this time. But... > diff --git a/include/builddefs.in b/include/builddefs.in > index 4700b527..1dd27f76 100644 > --- a/include/builddefs.in > +++ b/include/builddefs.in > @@ -144,6 +144,9 @@ endif > ifeq ($(HAVE_GETFSMAP),yes) > PCFLAGS+= -DHAVE_GETFSMAP > endif > +ifeq ($(HAVE_FALLOCATE),yes) > +PCFLAGS += -DHAVE_FALLOCATE > +endif > > LIBICU_LIBS = @libicu_LIBS@ > LIBICU_CFLAGS = @libicu_CFLAGS@ > diff --git a/include/linux.h b/include/linux.h > index 8f3c32b0..8d5c4584 100644 > --- a/include/linux.h > +++ b/include/linux.h > @@ -20,6 +20,10 @@ > #include <stdio.h> > #include <asm/types.h> > #include <mntent.h> > +#include <fcntl.h> > +#if defined(HAVE_FALLOCATE) > +#include <linux/falloc.h> > +#endif > #ifdef OVERRIDE_SYSTEM_FSXATTR > # define fsxattr sys_fsxattr > #endif > @@ -164,6 +168,24 @@ static inline void platform_mntent_close(struct mntent_cursor * cursor) > endmntent(cursor->mtabp); > } > > +#if defined(FALLOC_FL_ZERO_RANGE) > +static inline int > +platform_zero_range( > + int fd, > + xfs_off_t start, > + size_t len) > +{ > + int ret; > + > + ret = fallocate(fd, FALLOC_FL_ZERO_RANGE, start, len); > + if (!ret) > + return 0; > + return -errno; > +} > +#else > +#define platform_zero_range(fd, s, l) (-EOPNOTSUPP) > +#endif > + > /* > * Check whether we have to define FS_IOC_FS[GS]ETXATTR ourselves. These > * are a copy of the definitions moved to linux/uapi/fs.h in the 4.5 kernel, > diff --git a/libxfs/rdwr.c b/libxfs/rdwr.c > index 0d9d7202..e2d9d790 100644 > --- a/libxfs/rdwr.c > +++ b/libxfs/rdwr.c > @@ -61,8 +61,18 @@ libxfs_device_zero(struct xfs_buftarg *btp, xfs_daddr_t start, uint len) > { > xfs_off_t start_offset, end_offset, offset; > ssize_t zsize, bytes; > + size_t len_bytes; > char *z; > - int fd; > + int error, fd; > + > + fd = libxfs_device_to_fd(btp->dev); > + start_offset = LIBXFS_BBTOOFF64(start); > + > + /* try to use special zeroing methods, fall back to writes if needed */ > + len_bytes = LIBXFS_BBTOOFF64(len); ...but if the caller passes in (say) 2^23 daddrs on a 32-bit system, this conversion will try to stuff 2^32 into a size_t (which is 32-bit), causing an integer overflow. I grok that no callers currently try this, but this seems like leaving a logic bomb that could go off on what are becoming difficult-to-test architectures. Granted the added overflow checking and whatnot required to convert that last parameter of libxfs_device_zero to unsigned long long could very well justify a separate patch for fixing the 64-bitness of the whole api. --D > + error = platform_zero_range(fd, start_offset, len_bytes); > + if (!error) > + return 0; > > zsize = min(BDSTRAT_SIZE, BBTOB(len)); > if ((z = memalign(libxfs_device_alignment(), zsize)) == NULL) { > @@ -73,9 +83,6 @@ libxfs_device_zero(struct xfs_buftarg *btp, xfs_daddr_t start, uint len) > } > memset(z, 0, zsize); > > - fd = libxfs_device_to_fd(btp->dev); > - start_offset = LIBXFS_BBTOOFF64(start); > - > if ((lseek(fd, start_offset, SEEK_SET)) < 0) { > fprintf(stderr, _("%s: %s seek to offset %llu failed: %s\n"), > progname, __FUNCTION__, > >
On 2/25/20 11:16 AM, Darrick J. Wong wrote: > On Tue, Feb 25, 2020 at 10:13:55AM -0800, Eric Sandeen wrote: >> I had a request from someone who cared about mkfs speed over >> a slower network block device to look into using faster zeroing >> methods, particularly for the log, during mkfs. >> >> Using FALLOC_FL_ZERO_RANGE is faster in this case than writing >> a bunch of zeros across a wire. >> >> Signed-off-by: Eric Sandeen <sandeen@redhat.com> >> --- >> >> V2: Clean up all the nasty stuff I'd flung out there as a wild first >> cut, thanks Dave. >> >> V3: make len_bytes a size_t; leave "end_offset" where it is for the loop >> use. It's a bit odd but ... just don't mess with it for now, one patch >> one change. >> >> V4: Use EOPNOTSUPP not EOPNOTSUP (same on linux anyway but meh) >> I ignored(tm) darrick's suggestion to make libxfs_device_zero accept >> a longer length, for now - no callers need anything bigger at this time. > > But... > ... >> diff --git a/libxfs/rdwr.c b/libxfs/rdwr.c >> index 0d9d7202..e2d9d790 100644 >> --- a/libxfs/rdwr.c >> +++ b/libxfs/rdwr.c >> @@ -61,8 +61,18 @@ libxfs_device_zero(struct xfs_buftarg *btp, xfs_daddr_t start, uint len) >> { >> xfs_off_t start_offset, end_offset, offset; >> ssize_t zsize, bytes; >> + size_t len_bytes; >> char *z; >> - int fd; >> + int error, fd; >> + >> + fd = libxfs_device_to_fd(btp->dev); >> + start_offset = LIBXFS_BBTOOFF64(start); >> + >> + /* try to use special zeroing methods, fall back to writes if needed */ >> + len_bytes = LIBXFS_BBTOOFF64(len); > > ...but if the caller passes in (say) 2^23 daddrs on a 32-bit system, > this conversion will try to stuff 2^32 into a size_t (which is 32-bit), > causing an integer overflow. I grok that no callers currently try this, > but this seems like leaving a logic bomb that could go off on what are > becoming difficult-to-test architectures. That's true, but .... > Granted the added overflow checking and whatnot required to convert that > last parameter of libxfs_device_zero to unsigned long long could very > well justify a separate patch for fixing the 64-bitness of the whole > api. > > --D > >> + error = platform_zero_range(fd, start_offset, len_bytes); >> + if (!error) >> + return 0; >> >> zsize = min(BDSTRAT_SIZE, BBTOB(len)); it's an existing problem, right? One which this patch doesn't address. :) -Eric
diff --git a/include/builddefs.in b/include/builddefs.in index 4700b527..1dd27f76 100644 --- a/include/builddefs.in +++ b/include/builddefs.in @@ -144,6 +144,9 @@ endif ifeq ($(HAVE_GETFSMAP),yes) PCFLAGS+= -DHAVE_GETFSMAP endif +ifeq ($(HAVE_FALLOCATE),yes) +PCFLAGS += -DHAVE_FALLOCATE +endif LIBICU_LIBS = @libicu_LIBS@ LIBICU_CFLAGS = @libicu_CFLAGS@ diff --git a/include/linux.h b/include/linux.h index 8f3c32b0..8d5c4584 100644 --- a/include/linux.h +++ b/include/linux.h @@ -20,6 +20,10 @@ #include <stdio.h> #include <asm/types.h> #include <mntent.h> +#include <fcntl.h> +#if defined(HAVE_FALLOCATE) +#include <linux/falloc.h> +#endif #ifdef OVERRIDE_SYSTEM_FSXATTR # define fsxattr sys_fsxattr #endif @@ -164,6 +168,24 @@ static inline void platform_mntent_close(struct mntent_cursor * cursor) endmntent(cursor->mtabp); } +#if defined(FALLOC_FL_ZERO_RANGE) +static inline int +platform_zero_range( + int fd, + xfs_off_t start, + size_t len) +{ + int ret; + + ret = fallocate(fd, FALLOC_FL_ZERO_RANGE, start, len); + if (!ret) + return 0; + return -errno; +} +#else +#define platform_zero_range(fd, s, l) (-EOPNOTSUPP) +#endif + /* * Check whether we have to define FS_IOC_FS[GS]ETXATTR ourselves. These * are a copy of the definitions moved to linux/uapi/fs.h in the 4.5 kernel, diff --git a/libxfs/rdwr.c b/libxfs/rdwr.c index 0d9d7202..e2d9d790 100644 --- a/libxfs/rdwr.c +++ b/libxfs/rdwr.c @@ -61,8 +61,18 @@ libxfs_device_zero(struct xfs_buftarg *btp, xfs_daddr_t start, uint len) { xfs_off_t start_offset, end_offset, offset; ssize_t zsize, bytes; + size_t len_bytes; char *z; - int fd; + int error, fd; + + fd = libxfs_device_to_fd(btp->dev); + start_offset = LIBXFS_BBTOOFF64(start); + + /* try to use special zeroing methods, fall back to writes if needed */ + len_bytes = LIBXFS_BBTOOFF64(len); + error = platform_zero_range(fd, start_offset, len_bytes); + if (!error) + return 0; zsize = min(BDSTRAT_SIZE, BBTOB(len)); if ((z = memalign(libxfs_device_alignment(), zsize)) == NULL) { @@ -73,9 +83,6 @@ libxfs_device_zero(struct xfs_buftarg *btp, xfs_daddr_t start, uint len) } memset(z, 0, zsize); - fd = libxfs_device_to_fd(btp->dev); - start_offset = LIBXFS_BBTOOFF64(start); - if ((lseek(fd, start_offset, SEEK_SET)) < 0) { fprintf(stderr, _("%s: %s seek to offset %llu failed: %s\n"), progname, __FUNCTION__,
I had a request from someone who cared about mkfs speed over a slower network block device to look into using faster zeroing methods, particularly for the log, during mkfs. Using FALLOC_FL_ZERO_RANGE is faster in this case than writing a bunch of zeros across a wire. Signed-off-by: Eric Sandeen <sandeen@redhat.com> --- V2: Clean up all the nasty stuff I'd flung out there as a wild first cut, thanks Dave. V3: make len_bytes a size_t; leave "end_offset" where it is for the loop use. It's a bit odd but ... just don't mess with it for now, one patch one change. V4: Use EOPNOTSUPP not EOPNOTSUP (same on linux anyway but meh) I ignored(tm) darrick's suggestion to make libxfs_device_zero accept a longer length, for now - no callers need anything bigger at this time.