Message ID | 20200831140127.657134-3-nsoffer@redhat.com (mailing list archive) |
---|---|
State | New, archived |
Headers | show |
Series | Replace posix_fallocate() with falloate() | expand |
On Mon 31 Aug 2020 04:01:27 PM CEST, Nir Soffer wrote: > If fallocate() is not supported, posix_fallocate() falls back to > inefficient allocation, writing one byte for every 4k bytes[1]. This is > very slow compared with writing zeros. In oVirt we measured ~400% > improvement in allocation time when replacing posix_fallocate() with > manually writing zeroes[2]. > > We also know that posix_fallocated() does not work well when using OFD > locks[3]. We don't know the reason yet for this issue yet. > > Change preallocate_falloc() to use fallocate() instead of > posix_falloate(), and fall back to full preallocation if not > supported. > case PREALLOC_MODE_FALLOC: > result = preallocate_falloc(fd, current_length, offset, errp); > - goto out; > + if (result != -ENOTSUP) > + goto out; > + /* If fallocate() is not supported, fallback to full preallocation. */ With the missing braces in this if statement, Reviewed-by: Alberto Garcia <berto@igalia.com> Berto
On Mon, Aug 31, 2020 at 05:01:27PM +0300, Nir Soffer wrote: > If fallocate() is not supported, posix_fallocate() falls back to > inefficient allocation, writing one byte for every 4k bytes[1]. This is > very slow compared with writing zeros. In oVirt we measured ~400% > improvement in allocation time when replacing posix_fallocate() with > manually writing zeroes[2]. > > We also know that posix_fallocated() does not work well when using OFD > locks[3]. We don't know the reason yet for this issue yet. > > Change preallocate_falloc() to use fallocate() instead of > posix_falloate(), and fall back to full preallocation if not supported. > > Here are quick test results with this change. > > Before (qemu-img-5.1.0-2.fc32.x86_64): > > $ time qemu-img create -f raw -o preallocation=falloc /tmp/nfs3/test.raw 6g > Formatting '/tmp/nfs3/test.raw', fmt=raw size=6442450944 preallocation=falloc > > real 0m42.100s > user 0m0.602s > sys 0m4.137s > > NFS stats: > calls retrans authrefrsh write > 1571583 0 1572205 1571321 > > After: > > $ time ./qemu-img create -f raw -o preallocation=falloc /tmp/nfs3/test.raw 6g > Formatting '/tmp/nfs3/test.raw', fmt=raw size=6442450944 preallocation=falloc > > real 0m15.551s > user 0m0.070s > sys 0m2.623s > > NFS stats: > calls retrans authrefrsh write > 24620 0 24624 24567 > > [1] https://code.woboq.org/userspace/glibc/sysdeps/posix/posix_fallocate.c.html#96 > [2] https://bugzilla.redhat.com/1850267#c25 > [3] https://bugzilla.redhat.com/1851097 This bug appears to be private to RH employees only, so rather than link to it, please summarize any important facts in it for benefit of nonn-RH QEMU contributors. > > Signed-off-by: Nir Soffer <nsoffer@redhat.com> > --- > block/file-posix.c | 32 +++++++++----------------- > docs/system/qemu-block-drivers.rst.inc | 11 +++++---- > docs/tools/qemu-img.rst | 11 +++++---- > qapi/block-core.json | 4 ++-- > 4 files changed, 25 insertions(+), 33 deletions(-) Regards, Daniel
On Mon, Sep 14, 2020 at 8:32 PM Daniel P. Berrangé <berrange@redhat.com> wrote: > > On Mon, Aug 31, 2020 at 05:01:27PM +0300, Nir Soffer wrote: > > If fallocate() is not supported, posix_fallocate() falls back to > > inefficient allocation, writing one byte for every 4k bytes[1]. This is > > very slow compared with writing zeros. In oVirt we measured ~400% > > improvement in allocation time when replacing posix_fallocate() with > > manually writing zeroes[2]. > > > > We also know that posix_fallocated() does not work well when using OFD > > locks[3]. We don't know the reason yet for this issue yet. > > > > Change preallocate_falloc() to use fallocate() instead of > > posix_falloate(), and fall back to full preallocation if not supported. > > > > Here are quick test results with this change. > > > > Before (qemu-img-5.1.0-2.fc32.x86_64): > > > > $ time qemu-img create -f raw -o preallocation=falloc /tmp/nfs3/test.raw 6g > > Formatting '/tmp/nfs3/test.raw', fmt=raw size=6442450944 preallocation=falloc > > > > real 0m42.100s > > user 0m0.602s > > sys 0m4.137s > > > > NFS stats: > > calls retrans authrefrsh write > > 1571583 0 1572205 1571321 > > > > After: > > > > $ time ./qemu-img create -f raw -o preallocation=falloc /tmp/nfs3/test.raw 6g > > Formatting '/tmp/nfs3/test.raw', fmt=raw size=6442450944 preallocation=falloc > > > > real 0m15.551s > > user 0m0.070s > > sys 0m2.623s > > > > NFS stats: > > calls retrans authrefrsh write > > 24620 0 24624 24567 > > > > [1] https://code.woboq.org/userspace/glibc/sysdeps/posix/posix_fallocate.c.html#96 > > [2] https://bugzilla.redhat.com/1850267#c25 > > [3] https://bugzilla.redhat.com/1851097 > > This bug appears to be private to RH employees only, so rather than link > to it, please summarize any important facts in it for benefit of nonn-RH > QEMU contributors. Thanks, I missed that detail when linking to the bug. The bug is public now. > > Signed-off-by: Nir Soffer <nsoffer@redhat.com> > > --- > > block/file-posix.c | 32 +++++++++----------------- > > docs/system/qemu-block-drivers.rst.inc | 11 +++++---- > > docs/tools/qemu-img.rst | 11 +++++---- > > qapi/block-core.json | 4 ++-- > > 4 files changed, 25 insertions(+), 33 deletions(-) > > Regards, > Daniel > -- > |: https://berrange.com -o- https://www.flickr.com/photos/dberrange :| > |: https://libvirt.org -o- https://fstop138.berrange.com :| > |: https://entangle-photo.org -o- https://www.instagram.com/dberrange :| >
diff --git a/block/file-posix.c b/block/file-posix.c index 341ffb1cb4..eac3c0b412 100644 --- a/block/file-posix.c +++ b/block/file-posix.c @@ -1835,36 +1835,24 @@ static int allocate_first_block(int fd, size_t max_size) static int preallocate_falloc(int fd, int64_t current_length, int64_t offset, Error **errp) { -#ifdef CONFIG_POSIX_FALLOCATE +#ifdef CONFIG_FALLOCATE int result; if (offset == current_length) return 0; - /* - * Truncating before posix_fallocate() makes it about twice slower on - * file systems that do not support fallocate(), trying to check if a - * block is allocated before allocating it, so don't do that here. - */ - - result = -posix_fallocate(fd, current_length, - offset - current_length); + result = do_fallocate(fd, 0, current_length, offset - current_length); if (result != 0) { - /* posix_fallocate() doesn't set errno. */ - error_setg_errno(errp, -result, - "Could not preallocate new data"); + error_setg_errno(errp, -result, "Could not preallocate new data"); return result; } if (current_length == 0) { /* - * posix_fallocate() uses fallocate() if the filesystem supports - * it, or fallback to manually writing zeroes. If fallocate() - * was used, unaligned reads from the fallocated area in - * raw_probe_alignment() will succeed, hence we need to allocate - * the first block. + * Unaligned reads from the fallocated area in raw_probe_alignment() + * will succeed, hence we need to allocate the first block. * - * Optimize future alignment probing; ignore failures. + * Optimizes future alignment probing; ignore failures. */ allocate_first_block(fd, offset); } @@ -1973,10 +1961,12 @@ static int handle_aiocb_truncate(void *opaque) } switch (prealloc) { -#ifdef CONFIG_POSIX_FALLOCATE +#ifdef CONFIG_FALLOCATE case PREALLOC_MODE_FALLOC: result = preallocate_falloc(fd, current_length, offset, errp); - goto out; + if (result != -ENOTSUP) + goto out; + /* If fallocate() is not supported, fallback to full preallocation. */ #endif case PREALLOC_MODE_FULL: result = preallocate_full(fd, current_length, offset, errp); @@ -3080,7 +3070,7 @@ static QemuOptsList raw_create_opts = { .name = BLOCK_OPT_PREALLOC, .type = QEMU_OPT_STRING, .help = "Preallocation mode (allowed values: off" -#ifdef CONFIG_POSIX_FALLOCATE +#ifdef CONFIG_FALLOCATE ", falloc" #endif ", full)" diff --git a/docs/system/qemu-block-drivers.rst.inc b/docs/system/qemu-block-drivers.rst.inc index b052a6d14e..8e4acf397e 100644 --- a/docs/system/qemu-block-drivers.rst.inc +++ b/docs/system/qemu-block-drivers.rst.inc @@ -25,11 +25,12 @@ This section describes each format and the options that are supported for it. .. program:: raw .. option:: preallocation - Preallocation mode (allowed values: ``off``, ``falloc``, - ``full``). ``falloc`` mode preallocates space for image by - calling ``posix_fallocate()``. ``full`` mode preallocates space - for image by writing data to underlying storage. This data may or - may not be zero, depending on the storage location. + Preallocation mode (allowed values: ``off``, ``falloc``, ``full``). + ``falloc`` mode preallocates space for image by calling + ``fallocate()``, and falling back to ``full` mode if not supported. + ``full`` mode preallocates space for image by writing data to + underlying storage. This data may or may not be zero, depending on + the storage location. .. program:: image-formats .. option:: qcow2 diff --git a/docs/tools/qemu-img.rst b/docs/tools/qemu-img.rst index c35bd64822..a2089bd1b7 100644 --- a/docs/tools/qemu-img.rst +++ b/docs/tools/qemu-img.rst @@ -750,11 +750,12 @@ Supported image file formats: Supported options: ``preallocation`` - Preallocation mode (allowed values: ``off``, ``falloc``, - ``full``). ``falloc`` mode preallocates space for image by - calling ``posix_fallocate()``. ``full`` mode preallocates space - for image by writing data to underlying storage. This data may or - may not be zero, depending on the storage location. + Preallocation mode (allowed values: ``off``, ``falloc``, ``full``). + ``falloc`` mode preallocates space for image by calling + ``fallocate()``, and falling back to ``full` mode if not supported. + ``full`` mode preallocates space for image by writing data to + underlying storage. This data may or may not be zero, depending on + the storage location. ``qcow2`` diff --git a/qapi/block-core.json b/qapi/block-core.json index db08c58d78..681d79ec63 100644 --- a/qapi/block-core.json +++ b/qapi/block-core.json @@ -5021,8 +5021,8 @@ # # @off: no preallocation # @metadata: preallocate only for metadata -# @falloc: like @full preallocation but allocate disk space by -# posix_fallocate() rather than writing data. +# @falloc: try to allocate disk space by fallocate(), and fallback to +# @full preallocation if not supported. # @full: preallocate all data by writing it to the device to ensure # disk space is really available. This data may or may not be # zero, depending on the image format and storage.
If fallocate() is not supported, posix_fallocate() falls back to inefficient allocation, writing one byte for every 4k bytes[1]. This is very slow compared with writing zeros. In oVirt we measured ~400% improvement in allocation time when replacing posix_fallocate() with manually writing zeroes[2]. We also know that posix_fallocated() does not work well when using OFD locks[3]. We don't know the reason yet for this issue yet. Change preallocate_falloc() to use fallocate() instead of posix_falloate(), and fall back to full preallocation if not supported. Here are quick test results with this change. Before (qemu-img-5.1.0-2.fc32.x86_64): $ time qemu-img create -f raw -o preallocation=falloc /tmp/nfs3/test.raw 6g Formatting '/tmp/nfs3/test.raw', fmt=raw size=6442450944 preallocation=falloc real 0m42.100s user 0m0.602s sys 0m4.137s NFS stats: calls retrans authrefrsh write 1571583 0 1572205 1571321 After: $ time ./qemu-img create -f raw -o preallocation=falloc /tmp/nfs3/test.raw 6g Formatting '/tmp/nfs3/test.raw', fmt=raw size=6442450944 preallocation=falloc real 0m15.551s user 0m0.070s sys 0m2.623s NFS stats: calls retrans authrefrsh write 24620 0 24624 24567 [1] https://code.woboq.org/userspace/glibc/sysdeps/posix/posix_fallocate.c.html#96 [2] https://bugzilla.redhat.com/1850267#c25 [3] https://bugzilla.redhat.com/1851097 Signed-off-by: Nir Soffer <nsoffer@redhat.com> --- block/file-posix.c | 32 +++++++++----------------- docs/system/qemu-block-drivers.rst.inc | 11 +++++---- docs/tools/qemu-img.rst | 11 +++++---- qapi/block-core.json | 4 ++-- 4 files changed, 25 insertions(+), 33 deletions(-)