block/file-posix.c: Use pwritev2() with RWF_DSYNC for FUA
PATCH v6 1/1 block/file-posix.c: Use pwritev2() with RWF_DSYNC for FUA
Full Unit Access (FUA) is an optimization where a disk write with the flag set will be persisted to disk immediately instead of potentially remaining in the disk's write cache.
This commit address the todo task
for using pwritev2() with RWF_DSYNC in the thread pool section of raw_co_prw(), if pwritev2() with RWF_DSYNC is available in the host, which is always the case for Linux kernel >= 4.7.
The intent for FUA is indicated with the BDRV_REQ_FUA flag. The old code paths are preserved in case BDRV_REQ_FUA is off or pwritev2() with RWF_DSYNC is not available.
Support for disk writes with FUA is handled in qemu_pwritev_fua(), which uses pwritev2() with RWF_DSYNC if available, otherwise falls back to pwritev2() with no flags followed by flush using handle_aiocb_flush().
If pwritev2() is not implemented, then disk write in the linear FUA will fallback to pwrite() + handle_aiocb_flush().
---
v5:
- Use pwritev for unsupported OSes
v4:
- Add fallback when qemu_pwritev_fua() returns ENOSYS
- Similar fallback was not added for handle_aiocb_rw_vector() since there is a preadv_present check in handle_aiocb_rw()
v3:
- Changed signature to add fd, iov, nr_iov
- Return -ENOSYS for non-Linux hosts
v2:
- Moved handle_aiocb_flush() into qemu_pwritev_fua()
- In handle_aiocb_rw_linear(), iovec with iovcnt=1 is created based on the assumption that there will be only one buffer
--- block/file-posix.c | 68 ++++++++++++++++++++++++++++++++++++++-------- 1 file changed, 56 insertions(+), 12 deletions(-)
block/file-posix.c +56 -12
@@ -229,6 +229,7 @@ typedef struct RawPosixAIOData {unsigned long op; } zone_mgmt; }; +BdrvRequestFlags flags; } RawPosixAIOData;#if defined(__FreeBSD__) || defined(__FreeBSD_kernel__)
@@ -1674,6 +1675,20 @@ qemu_pwritev(int fd, const struct iovec *iov, int nr_iov, off_t offset) return pwritev(fd, iov, nr_iov, offset); } +static ssize_t +qemu_pwritev_fua (int fd,struct iovec *iov,int nr_iov,off_t offset,RawPosixAIOData *aiocb) +{ +#ifdef RWF_DSYNC +return pwritev2 (fd, iov, nr_iov, offset,RWF_DSYNC ); +#else +ssize_t len =pwritev (fd, iov, nr_iov, offset); +if (len ==0 ) { + len =handle_aiocb_flush (aiocb); + } +return len; +#endif +} +#else static bool preadv_present = false;
@@ -1690,6 +1705,11 @@ qemu_pwritev(int fd, const struct iovec *iov, int nr_iov, off_t offset) return -ENOSYS; } +static ssize_t +qemu_pwritev_fua (int fd,struct iovec *iov,int nr_iov,off_t offset,const RawPosixAIOData *aiocb) +{ +return -ENOSYS ; +}#endif static ssize_t handle_aiocb_rw_vector (RawPosixAIOData *aiocb)
@@ -1698,10 +1718,16 @@ static ssize_t handle_aiocb_rw_vector(RawPosixAIOData *aiocb) len =RETRY_ON_EINTR ( (aiocb->aio_type & (QEMU_AIO_WRITE |QEMU_AIO_ZONE_APPEND )) ? -qemu_pwritev (aiocb->aio_fildes, - aiocb->io.iov, - aiocb->io.niov, - aiocb->aio_offset) : + (aiocb->flags &BDRV_REQ_FUA ) ? +qemu_pwritev_fua (aiocb->aio_fildes, + aiocb->io.iov, + aiocb->io.niov, + aiocb->aio_offset, + aiocb) : +qemu_pwritev (aiocb->aio_fildes, + aiocb->io.iov, + aiocb->io.niov, + aiocb->aio_offset) : qemu_preadv(aiocb->aio_fildes, aiocb->io.iov, aiocb->io.niov,
@@ -1727,10 +1753,31 @@ static ssize_t handle_aiocb_rw_linear(RawPosixAIOData *aiocb, char *buf)while (offset < aiocb->aio_nbytes) {if (aiocb->aio_type & (QEMU_AIO_WRITE |QEMU_AIO_ZONE_APPEND )) { - len =pwrite (aiocb->aio_fildes, - (const char *)buf + offset, - aiocb->aio_nbytes - offset, - aiocb->aio_offset + offset); +if (aiocb->flags &BDRV_REQ_FUA ) { +struct iovec iov = { + .iov_base = buf + offset, + .iov_len = aiocb->aio_nbytes - offset, + }; + len =qemu_pwritev_fua (aiocb->aio_fildes, + &iov, +1 , + aiocb->aio_offset + offset, + aiocb); +if (len == -ENOSYS ) { + len =pwrite (aiocb->aio_fildes, + (const char *)buf + offset, + aiocb->aio_nbytes - offset, + aiocb->aio_offset + offset); +if (len ==0 ) { + len =handle_aiocb_flush (aiocb); + } + } + }else { + len =pwrite (aiocb->aio_fildes, + (const char *)buf + offset, + aiocb->aio_nbytes - offset, + aiocb->aio_offset + offset); + } }else { len = pread(aiocb->aio_fildes, buf + offset,
@@ -2539,14 +2586,11 @@ static int coroutine_fn raw_co_prw(BlockDriverState *bs, int64_t *offset_ptr, .iov = qiov->iov, .niov = qiov->niov, }, + .flags = flags, };assert (qiov->size == bytes); ret =raw_thread_pool_submit (handle_aiocb_rw, &acb); -if (ret ==0 && (flags &BDRV_REQ_FUA )) { -/* TODO Use pwritev2() instead if it's available */ - ret =raw_co_flush_to_disk (bs); - } goto out;/* Avoid the compiler err of unused label */ out:
Signature · 1 line
2.43.0
Quoted 37 lines · Pinku Deb Nath
Full Unit Access (FUA) is an optimization where a disk write with the flag set will be persisted to disk immediately instead of potentially remaining in the disk's write cache.
This commit address the todo task
for usingpwritev2()withRWF_DSYNCin the thread pool section ofraw_co_prw(), ifpwritev2()withRWF_DSYNCis available in the host, which is always the case for Linux kernel >= 4.7.The intent for FUA is indicated with the
BDRV_REQ_FUAflag. The old code paths are preserved in caseBDRV_REQ_FUAis off orpwritev2()withRWF_DSYNCis not available.Support for disk writes with FUA is handled in
qemu_pwritev_fua(), which usespwritev2()withRWF_DSYNCif available, otherwise falls back topwritev2()with no flags followed by flush usinghandle_aiocb_flush().If
pwritev2()is not implemented, then disk write in the linear FUA will fallback topwrite()+handle_aiocb_flush().Signed-off-by Pinku Deb Nath <[email protected]>---
v5:
- Use pwritev for unsupported OSesv4:
- Add fallback whenqemu_pwritev_fua()returns ENOSYS
- Similar fallback was not added forhandle_aiocb_rw_vector()since there is apreadv_presentcheck inhandle_aiocb_rw()v3:
- Changed signature to add fd, iov,nr_iov
- Return-ENOSYSfor non-Linux hosts
Quoted 7 lines
v2:
- Movedhandle_aiocb_flush()intoqemu_pwritev_fua()
- Inhandle_aiocb_rw_linear(), iovec with iovcnt=1 is created based on the assumption that there will be only one buffer--- block/file-posix.c | 68 ++++++++++++++++++++++++++++++++++++++-------- 1 file changed, 56 insertions(+), 12 deletions(-)
Yay! 😊
Quoted 48 lines · On Tue, Apr 8, 2025, 8:43 a.m. Stefan Hajnoczi <[email protected]> wrote:
Quoted 37 lines · Pinku Deb Nath
Full Unit Access (FUA) is an optimization where a disk write with the flag set will be persisted to disk immediately instead of potentially remaining in the disk's write cache.
This commit address the todo task
for usingpwritev2()withRWF_DSYNCin the thread pool section ofraw_co_prw(), ifpwritev2()withRWF_DSYNCis available in the host, which is always the case for Linux kernel >= 4.7.The intent for FUA is indicated with the
BDRV_REQ_FUAflag. The old code paths are preserved in caseBDRV_REQ_FUAis off orpwritev2()withRWF_DSYNCis not available.Support for disk writes with FUA is handled in
qemu_pwritev_fua(), which usespwritev2()withRWF_DSYNCif available, otherwise falls back topwritev2()with no flags followed by flush usinghandle_aiocb_flush().If
pwritev2()is not implemented, then disk write in the linear FUA will fallback topwrite()+handle_aiocb_flush().Signed-off-by Pinku Deb Nath <[email protected]>---
v5:
- Use pwritev for unsupported OSesv4:
- Add fallback whenqemu_pwritev_fua()returns ENOSYS
- Similar fallback was not added forhandle_aiocb_rw_vector()since there is apreadv_presentcheck inhandle_aiocb_rw()v3:
- Changed signature to add fd, iov,nr_iov
- Return-ENOSYSfor non-Linux hostsQuoted 7 lines
v2:
- Movedhandle_aiocb_flush()intoqemu_pwritev_fua()
- Inhandle_aiocb_rw_linear(), iovec with iovcnt=1 is created based on the assumption that there will be only one buffer--- block/file-posix.c | 68 ++++++++++++++++++++++++++++++++++++++-------- 1 file changed, 56 insertions(+), 12 deletions(-)Reviewed-by Stefan Hajnoczi <[email protected]>
Am 07.04.2025 um 17:47 hat Pinku Deb Nath geschrieben:
Quoted 58 lines
Full Unit Access (FUA) is an optimization where a disk write with the flag set will be persisted to disk immediately instead of potentially remaining in the disk's write cache.
This commit address the todo task
for usingpwritev2()withRWF_DSYNCin the thread pool section ofraw_co_prw(), ifpwritev2()withRWF_DSYNCis available in the host, which is always the case for Linux kernel >= 4.7.The intent for FUA is indicated with the
BDRV_REQ_FUAflag. The old code paths are preserved in caseBDRV_REQ_FUAis off orpwritev2()withRWF_DSYNCis not available.Support for disk writes with FUA is handled in
qemu_pwritev_fua(), which usespwritev2()withRWF_DSYNCif available, otherwise falls back topwritev2()with no flags followed by flush usinghandle_aiocb_flush().If
pwritev2()is not implemented, then disk write in the linear FUA will fallback topwrite()+handle_aiocb_flush().Signed-off-by Pinku Deb Nath <[email protected]>---
v5:
- Use pwritev for unsupported OSesv4:
- Add fallback whenqemu_pwritev_fua()returns ENOSYS
- Similar fallback was not added forhandle_aiocb_rw_vector()since there is apreadv_presentcheck inhandle_aiocb_rw()v3:
- Changed signature to add fd, iov,nr_iov
- Return-ENOSYSfor non-Linux hostsv2:
- Movedhandle_aiocb_flush()intoqemu_pwritev_fua()
- Inhandle_aiocb_rw_linear(), iovec with iovcnt=1 is created based on the assumption that there will be only one buffer--- block/file-posix.c | 68 ++++++++++++++++++++++++++++++++++++++-------- 1 file changed, 56 insertions(+), 12 deletions(-)block/file-posix.c +1
-0@@ -229,6 +229,7 @@ typedef struct RawPosixAIOData {unsigned long op; } zone_mgmt; }; +BdrvRequestFlags flags; } RawPosixAIOData;#if defined(__FreeBSD__) || defined(__FreeBSD_kernel__)
block/file-posix.c qemu_pwritev(int fd, const struct iovec *iov, int nr_iov, off_t offset) · line 1682
Quoted 8 lines
block/file-posix.c +5
-0@@ -1674,6 +1675,20 @@ qemu_pwritev(int fd, const struct iovec *iov, int nr_iov, off_t offset)return pwritev (fd, iov, nr_iov, offset); } +static ssize_t +qemu_pwritev_fua (int fd,struct iovec *iov,int nr_iov,off_t offset,RawPosixAIOData *aiocb) +{ +#ifdef RWF_DSYNC +return pwritev2 (fd, iov, nr_iov, offset,RWF_DSYNC );
If pwritev2() fails with ENOSYS, we'll disable even pwritev(). I think we should only try pwritev2(), and if it returns ENOSYS, we can execute the code below with pwritev() + handle_aiocb_flush(), which would then be outside the #ifdef.
Similar to preadv_present, we can have a bool pwritev2_present to avoid trying pwritev2() every time when we already know it's not there.
block/file-posix.c qemu_pwritev(int fd, const struct iovec *iov, int nr_iov, off_t offset) · line 1685
Quoted 3 lines
block/file-posix.c +3
-0+#else +ssize_t len =pwritev (fd, iov, nr_iov, offset); +if (len ==0 ) {
pwritev() doesn't return 0, but the number of written bytes on success (you even called the result 'len'!), so with len == 0, we would only flush if someone tries to write at EOF of a block device or something.
Quoted 10 lines
block/file-posix.c +6
-0+ len = handle_aiocb_flush(aiocb); + } +return len; +#endif +} +#else static bool preadv_present = false;
block/file-posix.c qemu_pwritev(int fd, const struct iovec *iov, int nr_iov, off_t offset) · line 1712
Quoted 8 lines
block/file-posix.c +5
-0@@ -1690,6 +1705,11 @@ qemu_pwritev(int fd, const struct iovec *iov, int nr_iov, off_t offset)return -ENOSYS ; } +static ssize_t +qemu_pwritev_fua (int fd,struct iovec *iov,int nr_iov,off_t offset,const RawPosixAIOData *aiocb) +{ +return -ENOSYS ; +}
This is inconsistent with the interface of the function above. The function above returns -1 with errno set on error, this one directly returns a negative errno code. We need to pick one interface and stick to it.
The existing definitions for qemu_preadv() are misleading because they have the same problem - but they are never actually called because preadv_present is always false. But qemu_pwritev_fua() is called below in handle_aiocb_rw_linear() without checking preadv_present, so it doesn't have the same excuse.
If you check a new pwritev2_present in handle_aiocb_rw_linear() and initialise it as false in this #else block, then what we return here doesn't matter any more and you can keep it consistent with the other (misleading) functions.
(I think we should actually replace the existing return -ENOSYS with g_assert_not_reached(), but that is for a separate patch.)
block/file-posix.c static ssize_t handle_aiocb_rw_vector(RawPosixAIOData *aiocb) · line 1721
Quoted 12 lines
block/file-posix.c +1
-4#endif static ssize_t handle_aiocb_rw_vector (RawPosixAIOData *aiocb)@@ -1698,10 +1718,16 @@ static ssize_t handle_aiocb_rw_vector(RawPosixAIOData *aiocb) len = RETRY_ON_EINTR( (aiocb->aio_type & (QEMU_AIO_WRITE | QEMU_AIO_ZONE_APPEND)) ? - qemu_pwritev(aiocb->aio_fildes, - aiocb->io.iov, - aiocb->io.niov, - aiocb->aio_offset) : + (aiocb->flags & BDRV_REQ_FUA) ?
Double space.
Quoted 21 lines
block/file-posix.c +10
-4+ qemu_pwritev_fua(aiocb->aio_fildes, + aiocb->io.iov, + aiocb->io.niov, + aiocb->aio_offset, + aiocb) : + qemu_pwritev(aiocb->aio_fildes, + aiocb->io.iov, + aiocb->io.niov, + aiocb->aio_offset) : qemu_preadv(aiocb->aio_fildes, aiocb->io.iov, aiocb->io.niov,@@ -1727,10 +1753,31 @@ static ssize_t handle_aiocb_rw_linear(RawPosixAIOData *aiocb, char *buf)while (offset < aiocb->aio_nbytes) {if (aiocb->aio_type & (QEMU_AIO_WRITE |QEMU_AIO_ZONE_APPEND )) { - len =pwrite (aiocb->aio_fildes, - (const char *)buf + offset, - aiocb->aio_nbytes - offset, - aiocb->aio_offset + offset); +if (aiocb->flags &BDRV_REQ_FUA ) {
block/file-posix.c static ssize_t handle_aiocb_rw_linear(RawPosixAIOData *aiocb, char *buf) · line 1766
Quoted 10 lines
block/file-posix.c +10
-0+struct iovec iov = { + .iov_base = buf + offset, + .iov_len = aiocb->aio_nbytes - offset, + }; + len =qemu_pwritev_fua (aiocb->aio_fildes, + &iov, +1 , + aiocb->aio_offset + offset, + aiocb); +if (len == -ENOSYS ) {
This requires checking errno instead, like other instances in this function do.
block/file-posix.c static ssize_t handle_aiocb_rw_linear(RawPosixAIOData *aiocb, char *buf) · line 1771
Quoted 5 lines
block/file-posix.c +5
-0+ len =pwrite (aiocb->aio_fildes, + (const char *)buf + offset, + aiocb->aio_nbytes - offset, + aiocb->aio_offset + offset); +if (len ==0 ) {
Same problem as above with len == 0.
Quoted 18 lines
block/file-posix.c +10
-0+ len =handle_aiocb_flush (aiocb); + } + } + }else { + len =pwrite (aiocb->aio_fildes, + (const char *)buf + offset, + aiocb->aio_nbytes - offset, + aiocb->aio_offset + offset); + } }else { len = pread(aiocb->aio_fildes, buf + offset,@@ -2539,14 +2586,11 @@ static int coroutine_fn raw_co_prw(BlockDriverState *bs, int64_t *offset_ptr, .iov = qiov->iov, .niov = qiov->niov, }, + .flags = flags, };
block/file-posix.c static int coroutine_fn raw_co_prw(BlockDriverState *bs, int64_t *offset_ptr, · line 2596
Quoted 10 lines
block/file-posix.c +0
-4assert (qiov->size == bytes); ret =raw_thread_pool_submit (handle_aiocb_rw, &acb); -if (ret ==0 && (flags &BDRV_REQ_FUA )) { -/* TODO Use pwritev2() instead if it's available */ - ret =raw_co_flush_to_disk (bs); - } goto out;/* Avoid the compiler err of unused label */ out:
Kevin
Quoted 8 lines · Pinku Deb Nath
Full Unit Access (FUA) is an optimization where a disk write with the flag set will be persisted to disk immediately instead of potentially remaining in the disk's write cache.
This commit address the todo task
for usingpwritev2()withRWF_DSYNCin the thread pool section ofraw_co_prw(), ifpwritev2()withRWF_DSYNCis available in the host, which is always the case for Linux kernel >= 4.7.
[Side note - I should take a look at using this syscall in libnbd's file plugin]
Quoted 11 lines
block/file-posix.c +3
-4@@ -1698,10 +1718,16 @@ static ssize_t handle_aiocb_rw_vector(RawPosixAIOData *aiocb) len = RETRY_ON_EINTR( (aiocb->aio_type & (QEMU_AIO_WRITE | QEMU_AIO_ZONE_APPEND)) ? - qemu_pwritev(aiocb->aio_fildes, - aiocb->io.iov, - aiocb->io.niov, - aiocb->aio_offset) : + (aiocb->flags & BDRV_REQ_FUA) ? + qemu_pwritev_fua(aiocb->aio_fildes, + aiocb->io.iov,
block/file-posix.c static ssize_t handle_aiocb_rw_vector(RawPosixAIOData *aiocb) · line 1733
Quoted 10 lines
block/file-posix.c +7
-0+ aiocb->io.niov, + aiocb->aio_offset, + aiocb) : + qemu_pwritev(aiocb->aio_fildes, + aiocb->io.iov, + aiocb->io.niov, + aiocb->aio_offset) : qemu_preadv(aiocb->aio_fildes, aiocb->io.iov, aiocb->io.niov,
This ?: chain is getting ridiculous. Can we instead do:
if (aiocb->aio_type & (QEMU_AIO_WRITE | QEMU_AIO_ZONE_APPEND)) {
if (aiocb->flags & BDRV_REQ_FUA) {
fn = qemu_pwritev_fua;
} else {
fn = qemu_pwritev;
}
} else {
fn = qemu_preadv;
}len = RETRY_ON_EINTR(fn(aiocb->aio_fildes, ... ))
by having all of the various dispatch functions take the same signature and a suitably-defined fn function pointer variable?
Signature · 3 lines
Eric Blake, Principal Software Engineer Red Hat, Inc. Virtualization: qemu.org | libguestfs.org