 include/qemu/osdep.h: Don't include qapi/error.h
Commit 57cb38b included qapi/error.h into qemu/osdep.h to get the
Error typedef. Since then, we've moved to include qemu/osdep.h
everywhere. Its file comment explains: "To avoid getting into
possible circular include dependencies, this file should not include
any other QEMU headers, with the exceptions of config-host.h,
compiler.h, os-posix.h and os-win32.h, all of which are doing a
similar job to this file and are under similar constraints."
qapi/error.h doesn't do a similar job, and it doesn't adhere to
similar constraints: it includes qapi-types.h. That's in excess of
100KiB of crap most .c files don't actually need.
Add the typedef to qemu/typedefs.h, and include that instead of
qapi/error.h. Include qapi/error.h in .c files that need it and don't
get it now. Include qapi-types.h in qom/object.h for uint16List.
Update scripts/clean-includes accordingly. Update it further to match
reality: replace config.h by config-target.h, add sysemu/os-posix.h,
sysemu/os-win32.h. Update the list of includes in the qemu/osdep.h
comment quoted above similarly.
This reduces the number of objects depending on qapi/error.h from "all
of them" to less than a third. Unfortunately, the number depending on
qapi-types.h shrinks only a little. More work is needed for that one.
Signed-off-by: Markus Armbruster <armbru@redhat.com>
[Fix compilation without the spice devel packages. - Paolo]
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years ago block: Perform copy-on-read in loop
Improve our braindead copy-on-read implementation. Pre-patch,
we have multiple issues:
- we create a bounce buffer and perform a write for the entire
request, even if the active image already has 99% of the
clusters occupied, and really only needs to copy-on-read the
remaining 1% of the clusters
- our bounce buffer was as large as the read request, and can
needlessly exhaust our memory by using double the memory of
the request size (the original request plus our bounce buffer),
rather than a capped maximum overhead beyond the original
- if a driver has a max_transfer limit, we are bypassing the
normal code in bdrv_aligned_preadv() that fragments to that
limit, and instead attempt to read the entire buffer from the
driver in one go, which some drivers may assert on
- a client can request a large request of nearly 2G such that
rounding the request out to cluster boundaries results in a
byte count larger than 2G. While this cannot exceed 32 bits,
it DOES have some follow-on problems:
-- the call to bdrv_driver_pread() can assert for exceeding
BDRV_REQUEST_MAX_BYTES, if the driver is old and lacks
.bdrv_co_preadv
-- if the buffer is all zeroes, the subsequent call to
bdrv_co_do_pwrite_zeroes is a no-op due to a negative size,
which means we did not actually copy on read
Fix all of these issues by breaking up the action into a loop,
where each iteration is capped to sane limits. Also, querying
the allocation status allows us to optimize: when data is
already present in the active layer, we don't need to bounce.
Note that the code has a telling comment that copy-on-read
should probably be a filter driver rather than a bolt-on hack
in io.c; but that remains a task for another day.
CC: qemu-stable@nongnu.org
Signed-off-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
2 years ago block: Support byte-based aio callbacks
We are gradually moving away from sector-based interfaces, towards
byte-based. Add new sector-based aio callbacks for read and write,
to match the fact that bdrv_aio_pdiscard is already byte-based.
Ideally, drivers should be converted to use coroutine callbacks
rather than aio; but that is not quite as trivial (and if we were
to do that conversion, the null-aio driver would disappear), so for
the short term, converting the signature but keeping things with
aio is easier. However, we CAN declare that a driver that uses
the byte-based aio interfaces now defaults to byte-based
operations, and must explicitly provide a refresh_limits override
to stick with larger alignments (making the alignment issues more
obvious directly in the drivers touched in the next few patches).
Once all drivers are converted, the sector-based aio callbacks will
be removed; in the meantime, a FIXME comment is added due to a
slight inefficiency that will be touched up as part of that later
cleanup.
Simplify some instances of 'bs->drv' into 'drv' while touching this,
since the local variable already exists to reduce typing.
Signed-off-by: Eric Blake <eblake@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
1 year ago block: align bounce buffers to page
The following sequence
int fd = open(argv[1], O_RDWR | O_CREAT | O_DIRECT, 0644);
for (i = 0; i < 100000; i++)
write(fd, buf, 4096);
performs 5% better if buf is aligned to 4096 bytes.
The difference is quite reliable.
On the other hand we do not want at the moment to enforce bounce
buffering if guest request is aligned to 512 bytes.
The patch changes default bounce buffer optimal alignment to
MAX(page size, 4k). 4k is chosen as maximal known sector size on real
HDD.
The justification of the performance improve is quite interesting.
From the kernel point of view each request to the disk was split
by two. This could be seen by blktrace like this:
9,0 11 1 0.000000000 11151 Q WS 312737792 + 1023 [qemu-img]
9,0 11 2 0.000007938 11151 Q WS 312738815 + 8 [qemu-img]
9,0 11 3 0.000030735 11151 Q WS 312738823 + 1016 [qemu-img]
9,0 11 4 0.000032482 11151 Q WS 312739839 + 8 [qemu-img]
9,0 11 5 0.000041379 11151 Q WS 312739847 + 1016 [qemu-img]
9,0 11 6 0.000042818 11151 Q WS 312740863 + 8 [qemu-img]
9,0 11 7 0.000051236 11151 Q WS 312740871 + 1017 [qemu-img]
9,0 5 1 0.169071519 11151 Q WS 312741888 + 1023 [qemu-img]
After the patch the pattern becomes normal:
9,0 6 1 0.000000000 12422 Q WS 314834944 + 1024 [qemu-img]
9,0 6 2 0.000038527 12422 Q WS 314835968 + 1024 [qemu-img]
9,0 6 3 0.000072849 12422 Q WS 314836992 + 1024 [qemu-img]
9,0 6 4 0.000106276 12422 Q WS 314838016 + 1024 [qemu-img]
and the amount of requests sent to disk (could be calculated counting
number of lines in the output of blktrace) is reduced about 2 times.
Both qemu-img and qemu-io are affected while qemu-kvm is not. The guest
does his job well and real requests comes properly aligned (to page).
Signed-off-by: Denis V. Lunev <den@openvz.org>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Message-id: 1431441056-26198-3-git-send-email-den@openvz.org
CC: Paolo Bonzini <pbonzini@redhat.com>
CC: Kevin Wolf <kwolf@redhat.com>
CC: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
4 years ago block: Fix hangs in synchronous APIs with iothreads
In the block layer, synchronous APIs are often implemented by creating a
coroutine that calls the asynchronous coroutine-based implementation and
then waiting for completion with BDRV_POLL_WHILE().
For this to work with iothreads (more specifically, when the synchronous
API is called in a thread that is not the home thread of the block
device, so that the coroutine will run in a different thread), we must
make sure to call aio_wait_kick() at the end of the operation. Many
places are missing this, so that BDRV_POLL_WHILE() keeps hanging even if
the condition has long become false.
Note that bdrv_dec_in_flight() involves an aio_wait_kick() call. This
corresponds to the BDRV_POLL_WHILE() in the drain functions, but it is
generally not enough for most other operations because they haven't set
the return value in the coroutine entry stub yet. To avoid race
conditions there, we need to kick after setting the return value.
The race window is small enough that the problem doesn't usually surface
in the common path. However, it does surface and causes easily
reproducible hangs if the operation can return early before even calling
bdrv_inc/dec_in_flight, which many of them do (trivial error or no-op
success paths).
The bug in bdrv_truncate(), bdrv_check() and bdrv_invalidate_cache() is
slightly different: These functions even neglected to schedule the
coroutine in the home thread of the node. This avoids the hang, but is
obviously wrong, too. Fix those to schedule the coroutine in the right
AioContext in addition to adding aio_wait_kick() calls.
Cc: qemu-stable@nongnu.org
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
11 months ago coroutine: move entry argument to qemu_coroutine_create
In practice the entry argument is always known at creation time, and
it is confusing that sometimes qemu_coroutine_enter is used with a
non-NULL argument to re-enter a coroutine (this happens in
block/sheepdog.c and tests/test-coroutine.c). So pass the opaque value
at creation time, for consistency with e.g. aio_bh_new.
Mostly done with the following semantic patch:
@ entry1 @
expression entry, arg, co;
@@
- co = qemu_coroutine_create(entry);
+ co = qemu_coroutine_create(entry, arg);
...
- qemu_coroutine_enter(co, arg);
+ qemu_coroutine_enter(co);
@ entry2 @
expression entry, arg;
identifier co;
@@
- Coroutine *co = qemu_coroutine_create(entry);
+ Coroutine *co = qemu_coroutine_create(entry, arg);
...
- qemu_coroutine_enter(co, arg);
+ qemu_coroutine_enter(co);
@ entry3 @
expression entry, arg;
@@
- qemu_coroutine_enter(qemu_coroutine_create(entry), arg);
+ qemu_coroutine_enter(qemu_coroutine_create(entry, arg));
@ reentry @
expression co;
@@
- qemu_coroutine_enter(co, NULL);
+ qemu_coroutine_enter(co);
except for the aforementioned few places where the semantic patch
stumbled (as expected) and for test_co_queue, which would otherwise
produce an uninitialized variable warning.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Reviewed-by: Fam Zheng <famz@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
3 years ago block: Use bdrv_coroutine_enter to start I/O coroutines
BDRV_POLL_WHILE waits for the started I/O by releasing bs's ctx then polling
the main context, which relies on the yielded coroutine continuing on bs->ctx
before notifying qemu_aio_context with bdrv_wakeup().
Thus, using qemu_coroutine_enter to start I/O is wrong because if the coroutine
is entered from main loop, co->ctx will be qemu_aio_context, as a result of the
"release, poll, acquire" loop of BDRV_POLL_WHILE, race conditions happen when
both main thread and the iothread access the same BDS:
main loop iothread
-----------------------------------------------------------------------
blockdev_snapshot
aio_context_acquire(bs->ctx)
virtio_scsi_data_plane_handle_cmd
bdrv_drained_begin(bs->ctx)
bdrv_flush(bs)
bdrv_co_flush(bs) aio_context_acquire(bs->ctx).enter
...
qemu_coroutine_yield(co)
BDRV_POLL_WHILE()
aio_context_release(bs->ctx)
aio_context_acquire(bs->ctx).return
...
aio_co_wake(co)
aio_poll(qemu_aio_context) ...
co_schedule_bh_cb() ...
qemu_coroutine_enter(co) ...
/* (A) bdrv_co_flush(bs) /* (B) I/O on bs */
continues... */
aio_context_release(bs->ctx)
aio_context_acquire(bs->ctx)
Note that in above case, bdrv_drained_begin() doesn't do the "release,
poll, acquire" in BDRV_POLL_WHILE, because bs->in_flight == 0.
Fix this by using bdrv_coroutine_enter and enter coroutine in the right
context.
iotests 109 output is updated because the coroutine reenter flow during
mirror job complete is different (now through co_queue_wakeup, instead
of the unconditional qemu_coroutine_switch before), making the end job
len different.
Signed-off-by: Fam Zheng <famz@redhat.com>
Acked-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
2 years ago block: Honor BDRV_REQ_FUA during write_zeroes
The block layer has a couple of cases where it can lose
Force Unit Access semantics when writing a large block of
zeroes, such that the request returns before the zeroes
have been guaranteed to land on underlying media.
SCSI does not support FUA during WRITESAME(10/16); FUA is only
supported if it falls back to WRITE(10/16). But where the
underlying device is new enough to not need a fallback, it
means that any upper layer request with FUA semantics was
silently ignoring BDRV_REQ_FUA.
Conversely, NBD has situations where it can support FUA but not
ZERO_WRITE; when that happens, the generic block layer fallback
to bdrv_driver_pwritev() (or the older bdrv_co_writev() in qemu
2.6) was losing the FUA flag.
The problem of losing flags unrelated to ZERO_WRITE has been
latent in bdrv_co_do_write_zeroes() since commit aa7bfbff, but
back then, it did not matter because there was no FUA flag. It
became observable when commit 93f5e6d8 paved the way for flags
that can impact correctness, when we should have been using
bdrv_co_writev_flags() with modified flags. Compare to commit
9eeb6dd, which got flag manipulation right in
bdrv_co_do_zero_pwritev().
Symptoms: I tested with qemu-io with default writethrough cache
(which is supposed to use FUA semantics on every write), and
targetted an NBD client connected to a server that intentionally
did not advertise NBD_FLAG_SEND_FUA. When doing 'write 0 512',
the NBD client sent two operations (NBD_CMD_WRITE then
NBD_CMD_FLUSH) to get the fallback FUA semantics; but when doing
'write -z 0 512', the NBD client sent only NBD_CMD_WRITE.
The fix is do to a cleanup bdrv_co_flush() at the end of the
operation if any step in the middle relied on a BDS that does
not natively support FUA for that step (note that we don't
need to flush after every operation, if the operation is broken
into chunks based on bounce-buffer sizing). Each BDS gains a
new flag .supported_zero_flags, which parallels the use of
.supported_write_flags but only when accessing a zero write
operation (the flags MUST be different, because of SCSI having
different semantics based on WRITE vs. WRITESAME; and also
because BDRV_REQ_MAY_UNMAP only makes sense on zero writes).
Also fix some documentation to describe -ENOTSUP semantics,
particularly since iscsi depends on those semantics.
Down the road, we may want to add a driver where its
.bdrv_co_pwritev() honors all three of BDRV_REQ_FUA,
BDRV_REQ_ZERO_WRITE, and BDRV_REQ_MAY_UNMAP, and advertise
this via bs->supported_write_flags for blocks opened by that
driver; such a driver should NOT supply .bdrv_co_write_zeroes
nor .supported_zero_flags. But none of the drivers touched
in this patch want to do that (the act of writing zeroes is
different enough from normal writes to deserve a second
callback).
Signed-off-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Fam Zheng <famz@redhat.com>
Acked-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
3 years ago block: Convert bdrv_get_block_status() to bytes
We are gradually moving away from sector-based interfaces, towards
byte-based. In the common case, allocation is unlikely to ever use
values that are not naturally sector-aligned, but it is possible
that byte-based values will let us be more precise about allocation
at the end of an unaligned file that can do byte-based access.
Changing the name of the function from bdrv_get_block_status() to
bdrv_block_status() ensures that the compiler enforces that all
callers are updated. For now, the io.c layer still assert()s that
all callers are sector-aligned, but that can be relaxed when a later
patch implements byte-based block status in the drivers.
There was an inherent limitation in returning the offset via the
return value: we only have room for BDRV_BLOCK_OFFSET_MASK bits, which
means an offset can only be mapped for sector-aligned queries (or,
if we declare that non-aligned input is at the same relative position
modulo 512 of the answer), so the new interface also changes things to
return the offset via output through a parameter by reference rather
than mashed into the return value. We'll have some glue code that
munges between the two styles until we finish converting all uses.
For the most part this patch is just the addition of scaling at the
callers followed by inverse scaling at bdrv_block_status(), coupled
with the tweak in calling convention. But some code, particularly
bdrv_is_allocated(), gets a lot simpler because it no longer has to
mess with sectors.
For ease of review, bdrv_get_block_status_above() will be tackled
separately.
Signed-off-by: Eric Blake <eblake@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
2 years ago block: Convert bdrv_get_block_status() to bytes
We are gradually moving away from sector-based interfaces, towards
byte-based. In the common case, allocation is unlikely to ever use
values that are not naturally sector-aligned, but it is possible
that byte-based values will let us be more precise about allocation
at the end of an unaligned file that can do byte-based access.
Changing the name of the function from bdrv_get_block_status() to
bdrv_block_status() ensures that the compiler enforces that all
callers are updated. For now, the io.c layer still assert()s that
all callers are sector-aligned, but that can be relaxed when a later
patch implements byte-based block status in the drivers.
There was an inherent limitation in returning the offset via the
return value: we only have room for BDRV_BLOCK_OFFSET_MASK bits, which
means an offset can only be mapped for sector-aligned queries (or,
if we declare that non-aligned input is at the same relative position
modulo 512 of the answer), so the new interface also changes things to
return the offset via output through a parameter by reference rather
than mashed into the return value. We'll have some glue code that
munges between the two styles until we finish converting all uses.
For the most part this patch is just the addition of scaling at the
callers followed by inverse scaling at bdrv_block_status(), coupled
with the tweak in calling convention. But some code, particularly
bdrv_is_allocated(), gets a lot simpler because it no longer has to
mess with sectors.
For ease of review, bdrv_get_block_status_above() will be tackled
separately.
Signed-off-by: Eric Blake <eblake@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
2 years ago block: Convert bdrv_get_block_status() to bytes
We are gradually moving away from sector-based interfaces, towards
byte-based. In the common case, allocation is unlikely to ever use
values that are not naturally sector-aligned, but it is possible
that byte-based values will let us be more precise about allocation
at the end of an unaligned file that can do byte-based access.
Changing the name of the function from bdrv_get_block_status() to
bdrv_block_status() ensures that the compiler enforces that all
callers are updated. For now, the io.c layer still assert()s that
all callers are sector-aligned, but that can be relaxed when a later
patch implements byte-based block status in the drivers.
There was an inherent limitation in returning the offset via the
return value: we only have room for BDRV_BLOCK_OFFSET_MASK bits, which
means an offset can only be mapped for sector-aligned queries (or,
if we declare that non-aligned input is at the same relative position
modulo 512 of the answer), so the new interface also changes things to
return the offset via output through a parameter by reference rather
than mashed into the return value. We'll have some glue code that
munges between the two styles until we finish converting all uses.
For the most part this patch is just the addition of scaling at the
callers followed by inverse scaling at bdrv_block_status(), coupled
with the tweak in calling convention. But some code, particularly
bdrv_is_allocated(), gets a lot simpler because it no longer has to
mess with sectors.
For ease of review, bdrv_get_block_status_above() will be tackled
separately.
Signed-off-by: Eric Blake <eblake@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
2 years ago block: Convert bdrv_get_block_status() to bytes
We are gradually moving away from sector-based interfaces, towards
byte-based. In the common case, allocation is unlikely to ever use
values that are not naturally sector-aligned, but it is possible
that byte-based values will let us be more precise about allocation
at the end of an unaligned file that can do byte-based access.
Changing the name of the function from bdrv_get_block_status() to
bdrv_block_status() ensures that the compiler enforces that all
callers are updated. For now, the io.c layer still assert()s that
all callers are sector-aligned, but that can be relaxed when a later
patch implements byte-based block status in the drivers.
There was an inherent limitation in returning the offset via the
return value: we only have room for BDRV_BLOCK_OFFSET_MASK bits, which
means an offset can only be mapped for sector-aligned queries (or,
if we declare that non-aligned input is at the same relative position
modulo 512 of the answer), so the new interface also changes things to
return the offset via output through a parameter by reference rather
than mashed into the return value. We'll have some glue code that
munges between the two styles until we finish converting all uses.
For the most part this patch is just the addition of scaling at the
callers followed by inverse scaling at bdrv_block_status(), coupled
with the tweak in calling convention. But some code, particularly
bdrv_is_allocated(), gets a lot simpler because it no longer has to
mess with sectors.
For ease of review, bdrv_get_block_status_above() will be tackled
separately.
Signed-off-by: Eric Blake <eblake@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
2 years ago block: Convert bdrv_get_block_status() to bytes
We are gradually moving away from sector-based interfaces, towards
byte-based. In the common case, allocation is unlikely to ever use
values that are not naturally sector-aligned, but it is possible
that byte-based values will let us be more precise about allocation
at the end of an unaligned file that can do byte-based access.
Changing the name of the function from bdrv_get_block_status() to
bdrv_block_status() ensures that the compiler enforces that all
callers are updated. For now, the io.c layer still assert()s that
all callers are sector-aligned, but that can be relaxed when a later
patch implements byte-based block status in the drivers.
There was an inherent limitation in returning the offset via the
return value: we only have room for BDRV_BLOCK_OFFSET_MASK bits, which
means an offset can only be mapped for sector-aligned queries (or,
if we declare that non-aligned input is at the same relative position
modulo 512 of the answer), so the new interface also changes things to
return the offset via output through a parameter by reference rather
than mashed into the return value. We'll have some glue code that
munges between the two styles until we finish converting all uses.
For the most part this patch is just the addition of scaling at the
callers followed by inverse scaling at bdrv_block_status(), coupled
with the tweak in calling convention. But some code, particularly
bdrv_is_allocated(), gets a lot simpler because it no longer has to
mess with sectors.
For ease of review, bdrv_get_block_status_above() will be tackled
separately.
Signed-off-by: Eric Blake <eblake@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
2 years ago block: Perform copy-on-read in loop
Improve our braindead copy-on-read implementation. Pre-patch,
we have multiple issues:
- we create a bounce buffer and perform a write for the entire
request, even if the active image already has 99% of the
clusters occupied, and really only needs to copy-on-read the
remaining 1% of the clusters
- our bounce buffer was as large as the read request, and can
needlessly exhaust our memory by using double the memory of
the request size (the original request plus our bounce buffer),
rather than a capped maximum overhead beyond the original
- if a driver has a max_transfer limit, we are bypassing the
normal code in bdrv_aligned_preadv() that fragments to that
limit, and instead attempt to read the entire buffer from the
driver in one go, which some drivers may assert on
- a client can request a large request of nearly 2G such that
rounding the request out to cluster boundaries results in a
byte count larger than 2G. While this cannot exceed 32 bits,
it DOES have some follow-on problems:
-- the call to bdrv_driver_pread() can assert for exceeding
BDRV_REQUEST_MAX_BYTES, if the driver is old and lacks
.bdrv_co_preadv
-- if the buffer is all zeroes, the subsequent call to
bdrv_co_do_pwrite_zeroes is a no-op due to a negative size,
which means we did not actually copy on read
Fix all of these issues by breaking up the action into a loop,
where each iteration is capped to sane limits. Also, querying
the allocation status allows us to optimize: when data is
already present in the active layer, we don't need to bounce.
Note that the code has a telling comment that copy-on-read
should probably be a filter driver rather than a bolt-on hack
in io.c; but that remains a task for another day.
CC: qemu-stable@nongnu.org
Signed-off-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
2 years ago block: Perform copy-on-read in loop
Improve our braindead copy-on-read implementation. Pre-patch,
we have multiple issues:
- we create a bounce buffer and perform a write for the entire
request, even if the active image already has 99% of the
clusters occupied, and really only needs to copy-on-read the
remaining 1% of the clusters
- our bounce buffer was as large as the read request, and can
needlessly exhaust our memory by using double the memory of
the request size (the original request plus our bounce buffer),
rather than a capped maximum overhead beyond the original
- if a driver has a max_transfer limit, we are bypassing the
normal code in bdrv_aligned_preadv() that fragments to that
limit, and instead attempt to read the entire buffer from the
driver in one go, which some drivers may assert on
- a client can request a large request of nearly 2G such that
rounding the request out to cluster boundaries results in a
byte count larger than 2G. While this cannot exceed 32 bits,
it DOES have some follow-on problems:
-- the call to bdrv_driver_pread() can assert for exceeding
BDRV_REQUEST_MAX_BYTES, if the driver is old and lacks
.bdrv_co_preadv
-- if the buffer is all zeroes, the subsequent call to
bdrv_co_do_pwrite_zeroes is a no-op due to a negative size,
which means we did not actually copy on read
Fix all of these issues by breaking up the action into a loop,
where each iteration is capped to sane limits. Also, querying
the allocation status allows us to optimize: when data is
already present in the active layer, we don't need to bounce.
Note that the code has a telling comment that copy-on-read
should probably be a filter driver rather than a bolt-on hack
in io.c; but that remains a task for another day.
CC: qemu-stable@nongnu.org
Signed-off-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
2 years ago block: Perform copy-on-read in loop
Improve our braindead copy-on-read implementation. Pre-patch,
we have multiple issues:
- we create a bounce buffer and perform a write for the entire
request, even if the active image already has 99% of the
clusters occupied, and really only needs to copy-on-read the
remaining 1% of the clusters
- our bounce buffer was as large as the read request, and can
needlessly exhaust our memory by using double the memory of
the request size (the original request plus our bounce buffer),
rather than a capped maximum overhead beyond the original
- if a driver has a max_transfer limit, we are bypassing the
normal code in bdrv_aligned_preadv() that fragments to that
limit, and instead attempt to read the entire buffer from the
driver in one go, which some drivers may assert on
- a client can request a large request of nearly 2G such that
rounding the request out to cluster boundaries results in a
byte count larger than 2G. While this cannot exceed 32 bits,
it DOES have some follow-on problems:
-- the call to bdrv_driver_pread() can assert for exceeding
BDRV_REQUEST_MAX_BYTES, if the driver is old and lacks
.bdrv_co_preadv
-- if the buffer is all zeroes, the subsequent call to
bdrv_co_do_pwrite_zeroes is a no-op due to a negative size,
which means we did not actually copy on read
Fix all of these issues by breaking up the action into a loop,
where each iteration is capped to sane limits. Also, querying
the allocation status allows us to optimize: when data is
already present in the active layer, we don't need to bounce.
Note that the code has a telling comment that copy-on-read
should probably be a filter driver rather than a bolt-on hack
in io.c; but that remains a task for another day.
CC: qemu-stable@nongnu.org
Signed-off-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
2 years ago block: Perform copy-on-read in loop
Improve our braindead copy-on-read implementation. Pre-patch,
we have multiple issues:
- we create a bounce buffer and perform a write for the entire
request, even if the active image already has 99% of the
clusters occupied, and really only needs to copy-on-read the
remaining 1% of the clusters
- our bounce buffer was as large as the read request, and can
needlessly exhaust our memory by using double the memory of
the request size (the original request plus our bounce buffer),
rather than a capped maximum overhead beyond the original
- if a driver has a max_transfer limit, we are bypassing the
normal code in bdrv_aligned_preadv() that fragments to that
limit, and instead attempt to read the entire buffer from the
driver in one go, which some drivers may assert on
- a client can request a large request of nearly 2G such that
rounding the request out to cluster boundaries results in a
byte count larger than 2G. While this cannot exceed 32 bits,
it DOES have some follow-on problems:
-- the call to bdrv_driver_pread() can assert for exceeding
BDRV_REQUEST_MAX_BYTES, if the driver is old and lacks
.bdrv_co_preadv
-- if the buffer is all zeroes, the subsequent call to
bdrv_co_do_pwrite_zeroes is a no-op due to a negative size,
which means we did not actually copy on read
Fix all of these issues by breaking up the action into a loop,
where each iteration is capped to sane limits. Also, querying
the allocation status allows us to optimize: when data is
already present in the active layer, we don't need to bounce.
Note that the code has a telling comment that copy-on-read
should probably be a filter driver rather than a bolt-on hack
in io.c; but that remains a task for another day.
CC: qemu-stable@nongnu.org
Signed-off-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
2 years ago block: Perform copy-on-read in loop
Improve our braindead copy-on-read implementation. Pre-patch,
we have multiple issues:
- we create a bounce buffer and perform a write for the entire
request, even if the active image already has 99% of the
clusters occupied, and really only needs to copy-on-read the
remaining 1% of the clusters
- our bounce buffer was as large as the read request, and can
needlessly exhaust our memory by using double the memory of
the request size (the original request plus our bounce buffer),
rather than a capped maximum overhead beyond the original
- if a driver has a max_transfer limit, we are bypassing the
normal code in bdrv_aligned_preadv() that fragments to that
limit, and instead attempt to read the entire buffer from the
driver in one go, which some drivers may assert on
- a client can request a large request of nearly 2G such that
rounding the request out to cluster boundaries results in a
byte count larger than 2G. While this cannot exceed 32 bits,
it DOES have some follow-on problems:
-- the call to bdrv_driver_pread() can assert for exceeding
BDRV_REQUEST_MAX_BYTES, if the driver is old and lacks
.bdrv_co_preadv
-- if the buffer is all zeroes, the subsequent call to
bdrv_co_do_pwrite_zeroes is a no-op due to a negative size,
which means we did not actually copy on read
Fix all of these issues by breaking up the action into a loop,
where each iteration is capped to sane limits. Also, querying
the allocation status allows us to optimize: when data is
already present in the active layer, we don't need to bounce.
Note that the code has a telling comment that copy-on-read
should probably be a filter driver rather than a bolt-on hack
in io.c; but that remains a task for another day.
CC: qemu-stable@nongnu.org
Signed-off-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
2 years ago block: Perform copy-on-read in loop
Improve our braindead copy-on-read implementation. Pre-patch,
we have multiple issues:
- we create a bounce buffer and perform a write for the entire
request, even if the active image already has 99% of the
clusters occupied, and really only needs to copy-on-read the
remaining 1% of the clusters
- our bounce buffer was as large as the read request, and can
needlessly exhaust our memory by using double the memory of
the request size (the original request plus our bounce buffer),
rather than a capped maximum overhead beyond the original
- if a driver has a max_transfer limit, we are bypassing the
normal code in bdrv_aligned_preadv() that fragments to that
limit, and instead attempt to read the entire buffer from the
driver in one go, which some drivers may assert on
- a client can request a large request of nearly 2G such that
rounding the request out to cluster boundaries results in a
byte count larger than 2G. While this cannot exceed 32 bits,
it DOES have some follow-on problems:
-- the call to bdrv_driver_pread() can assert for exceeding
BDRV_REQUEST_MAX_BYTES, if the driver is old and lacks
.bdrv_co_preadv
-- if the buffer is all zeroes, the subsequent call to
bdrv_co_do_pwrite_zeroes is a no-op due to a negative size,
which means we did not actually copy on read
Fix all of these issues by breaking up the action into a loop,
where each iteration is capped to sane limits. Also, querying
the allocation status allows us to optimize: when data is
already present in the active layer, we don't need to bounce.
Note that the code has a telling comment that copy-on-read
should probably be a filter driver rather than a bolt-on hack
in io.c; but that remains a task for another day.
CC: qemu-stable@nongnu.org
Signed-off-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
2 years ago block: Perform copy-on-read in loop
Improve our braindead copy-on-read implementation. Pre-patch,
we have multiple issues:
- we create a bounce buffer and perform a write for the entire
request, even if the active image already has 99% of the
clusters occupied, and really only needs to copy-on-read the
remaining 1% of the clusters
- our bounce buffer was as large as the read request, and can
needlessly exhaust our memory by using double the memory of
the request size (the original request plus our bounce buffer),
rather than a capped maximum overhead beyond the original
- if a driver has a max_transfer limit, we are bypassing the
normal code in bdrv_aligned_preadv() that fragments to that
limit, and instead attempt to read the entire buffer from the
driver in one go, which some drivers may assert on
- a client can request a large request of nearly 2G such that
rounding the request out to cluster boundaries results in a
byte count larger than 2G. While this cannot exceed 32 bits,
it DOES have some follow-on problems:
-- the call to bdrv_driver_pread() can assert for exceeding
BDRV_REQUEST_MAX_BYTES, if the driver is old and lacks
.bdrv_co_preadv
-- if the buffer is all zeroes, the subsequent call to
bdrv_co_do_pwrite_zeroes is a no-op due to a negative size,
which means we did not actually copy on read
Fix all of these issues by breaking up the action into a loop,
where each iteration is capped to sane limits. Also, querying
the allocation status allows us to optimize: when data is
already present in the active layer, we don't need to bounce.
Note that the code has a telling comment that copy-on-read
should probably be a filter driver rather than a bolt-on hack
in io.c; but that remains a task for another day.
CC: qemu-stable@nongnu.org
Signed-off-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
2 years ago block: Perform copy-on-read in loop
Improve our braindead copy-on-read implementation. Pre-patch,
we have multiple issues:
- we create a bounce buffer and perform a write for the entire
request, even if the active image already has 99% of the
clusters occupied, and really only needs to copy-on-read the
remaining 1% of the clusters
- our bounce buffer was as large as the read request, and can
needlessly exhaust our memory by using double the memory of
the request size (the original request plus our bounce buffer),
rather than a capped maximum overhead beyond the original
- if a driver has a max_transfer limit, we are bypassing the
normal code in bdrv_aligned_preadv() that fragments to that
limit, and instead attempt to read the entire buffer from the
driver in one go, which some drivers may assert on
- a client can request a large request of nearly 2G such that
rounding the request out to cluster boundaries results in a
byte count larger than 2G. While this cannot exceed 32 bits,
it DOES have some follow-on problems:
-- the call to bdrv_driver_pread() can assert for exceeding
BDRV_REQUEST_MAX_BYTES, if the driver is old and lacks
.bdrv_co_preadv
-- if the buffer is all zeroes, the subsequent call to
bdrv_co_do_pwrite_zeroes is a no-op due to a negative size,
which means we did not actually copy on read
Fix all of these issues by breaking up the action into a loop,
where each iteration is capped to sane limits. Also, querying
the allocation status allows us to optimize: when data is
already present in the active layer, we don't need to bounce.
Note that the code has a telling comment that copy-on-read
should probably be a filter driver rather than a bolt-on hack
in io.c; but that remains a task for another day.
CC: qemu-stable@nongnu.org
Signed-off-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
2 years ago block: Perform copy-on-read in loop
Improve our braindead copy-on-read implementation. Pre-patch,
we have multiple issues:
- we create a bounce buffer and perform a write for the entire
request, even if the active image already has 99% of the
clusters occupied, and really only needs to copy-on-read the
remaining 1% of the clusters
- our bounce buffer was as large as the read request, and can
needlessly exhaust our memory by using double the memory of
the request size (the original request plus our bounce buffer),
rather than a capped maximum overhead beyond the original
- if a driver has a max_transfer limit, we are bypassing the
normal code in bdrv_aligned_preadv() that fragments to that
limit, and instead attempt to read the entire buffer from the
driver in one go, which some drivers may assert on
- a client can request a large request of nearly 2G such that
rounding the request out to cluster boundaries results in a
byte count larger than 2G. While this cannot exceed 32 bits,
it DOES have some follow-on problems:
-- the call to bdrv_driver_pread() can assert for exceeding
BDRV_REQUEST_MAX_BYTES, if the driver is old and lacks
.bdrv_co_preadv
-- if the buffer is all zeroes, the subsequent call to
bdrv_co_do_pwrite_zeroes is a no-op due to a negative size,
which means we did not actually copy on read
Fix all of these issues by breaking up the action into a loop,
where each iteration is capped to sane limits. Also, querying
the allocation status allows us to optimize: when data is
already present in the active layer, we don't need to bounce.
Note that the code has a telling comment that copy-on-read
should probably be a filter driver rather than a bolt-on hack
in io.c; but that remains a task for another day.
CC: qemu-stable@nongnu.org
Signed-off-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
2 years ago block: Perform copy-on-read in loop
Improve our braindead copy-on-read implementation. Pre-patch,
we have multiple issues:
- we create a bounce buffer and perform a write for the entire
request, even if the active image already has 99% of the
clusters occupied, and really only needs to copy-on-read the
remaining 1% of the clusters
- our bounce buffer was as large as the read request, and can
needlessly exhaust our memory by using double the memory of
the request size (the original request plus our bounce buffer),
rather than a capped maximum overhead beyond the original
- if a driver has a max_transfer limit, we are bypassing the
normal code in bdrv_aligned_preadv() that fragments to that
limit, and instead attempt to read the entire buffer from the
driver in one go, which some drivers may assert on
- a client can request a large request of nearly 2G such that
rounding the request out to cluster boundaries results in a
byte count larger than 2G. While this cannot exceed 32 bits,
it DOES have some follow-on problems:
-- the call to bdrv_driver_pread() can assert for exceeding
BDRV_REQUEST_MAX_BYTES, if the driver is old and lacks
.bdrv_co_preadv
-- if the buffer is all zeroes, the subsequent call to
bdrv_co_do_pwrite_zeroes is a no-op due to a negative size,
which means we did not actually copy on read
Fix all of these issues by breaking up the action into a loop,
where each iteration is capped to sane limits. Also, querying
the allocation status allows us to optimize: when data is
already present in the active layer, we don't need to bounce.
Note that the code has a telling comment that copy-on-read
should probably be a filter driver rather than a bolt-on hack
in io.c; but that remains a task for another day.
CC: qemu-stable@nongnu.org
Signed-off-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
2 years ago block: Perform copy-on-read in loop
Improve our braindead copy-on-read implementation. Pre-patch,
we have multiple issues:
- we create a bounce buffer and perform a write for the entire
request, even if the active image already has 99% of the
clusters occupied, and really only needs to copy-on-read the
remaining 1% of the clusters
- our bounce buffer was as large as the read request, and can
needlessly exhaust our memory by using double the memory of
the request size (the original request plus our bounce buffer),
rather than a capped maximum overhead beyond the original
- if a driver has a max_transfer limit, we are bypassing the
normal code in bdrv_aligned_preadv() that fragments to that
limit, and instead attempt to read the entire buffer from the
driver in one go, which some drivers may assert on
- a client can request a large request of nearly 2G such that
rounding the request out to cluster boundaries results in a
byte count larger than 2G. While this cannot exceed 32 bits,
it DOES have some follow-on problems:
-- the call to bdrv_driver_pread() can assert for exceeding
BDRV_REQUEST_MAX_BYTES, if the driver is old and lacks
.bdrv_co_preadv
-- if the buffer is all zeroes, the subsequent call to
bdrv_co_do_pwrite_zeroes is a no-op due to a negative size,
which means we did not actually copy on read
Fix all of these issues by breaking up the action into a loop,
where each iteration is capped to sane limits. Also, querying
the allocation status allows us to optimize: when data is
already present in the active layer, we don't need to bounce.
Note that the code has a telling comment that copy-on-read
should probably be a filter driver rather than a bolt-on hack
in io.c; but that remains a task for another day.
CC: qemu-stable@nongnu.org
Signed-off-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
2 years ago block: Perform copy-on-read in loop
Improve our braindead copy-on-read implementation. Pre-patch,
we have multiple issues:
- we create a bounce buffer and perform a write for the entire
request, even if the active image already has 99% of the
clusters occupied, and really only needs to copy-on-read the
remaining 1% of the clusters
- our bounce buffer was as large as the read request, and can
needlessly exhaust our memory by using double the memory of
the request size (the original request plus our bounce buffer),
rather than a capped maximum overhead beyond the original
- if a driver has a max_transfer limit, we are bypassing the
normal code in bdrv_aligned_preadv() that fragments to that
limit, and instead attempt to read the entire buffer from the
driver in one go, which some drivers may assert on
- a client can request a large request of nearly 2G such that
rounding the request out to cluster boundaries results in a
byte count larger than 2G. While this cannot exceed 32 bits,
it DOES have some follow-on problems:
-- the call to bdrv_driver_pread() can assert for exceeding
BDRV_REQUEST_MAX_BYTES, if the driver is old and lacks
.bdrv_co_preadv
-- if the buffer is all zeroes, the subsequent call to
bdrv_co_do_pwrite_zeroes is a no-op due to a negative size,
which means we did not actually copy on read
Fix all of these issues by breaking up the action into a loop,
where each iteration is capped to sane limits. Also, querying
the allocation status allows us to optimize: when data is
already present in the active layer, we don't need to bounce.
Note that the code has a telling comment that copy-on-read
should probably be a filter driver rather than a bolt-on hack
in io.c; but that remains a task for another day.
CC: qemu-stable@nongnu.org
Signed-off-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
2 years ago block: Make bdrv_is_allocated() byte-based
We are gradually moving away from sector-based interfaces, towards
byte-based. In the common case, allocation is unlikely to ever use
values that are not naturally sector-aligned, but it is possible
that byte-based values will let us be more precise about allocation
at the end of an unaligned file that can do byte-based access.
Changing the signature of the function to use int64_t *pnum ensures
that the compiler enforces that all callers are updated. For now,
the io.c layer still assert()s that all callers are sector-aligned
on input and that *pnum is sector-aligned on return to the caller,
but that can be relaxed when a later patch implements byte-based
block status. Therefore, this code adds usages like
DIV_ROUND_UP(,BDRV_SECTOR_SIZE) to callers that still want aligned
values, where the call might reasonbly give non-aligned results
in the future; on the other hand, no rounding is needed for callers
that should just continue to work with byte alignment.
For the most part this patch is just the addition of scaling at the
callers followed by inverse scaling at bdrv_is_allocated(). But
some code, particularly bdrv_commit(), gets a lot simpler because it
no longer has to mess with sectors; also, it is now possible to pass
NULL if the caller does not care how much of the image is allocated
beyond the initial offset. Leave comments where we can further
simplify once a later patch eliminates the need for sector-aligned
requests through bdrv_is_allocated().
For ease of review, bdrv_is_allocated_above() will be tackled
separately.
Signed-off-by: Eric Blake <eblake@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
2 years ago block: Honor BDRV_REQ_FUA during write_zeroes
The block layer has a couple of cases where it can lose
Force Unit Access semantics when writing a large block of
zeroes, such that the request returns before the zeroes
have been guaranteed to land on underlying media.
SCSI does not support FUA during WRITESAME(10/16); FUA is only
supported if it falls back to WRITE(10/16). But where the
underlying device is new enough to not need a fallback, it
means that any upper layer request with FUA semantics was
silently ignoring BDRV_REQ_FUA.
Conversely, NBD has situations where it can support FUA but not
ZERO_WRITE; when that happens, the generic block layer fallback
to bdrv_driver_pwritev() (or the older bdrv_co_writev() in qemu
2.6) was losing the FUA flag.
The problem of losing flags unrelated to ZERO_WRITE has been
latent in bdrv_co_do_write_zeroes() since commit aa7bfbff, but
back then, it did not matter because there was no FUA flag. It
became observable when commit 93f5e6d8 paved the way for flags
that can impact correctness, when we should have been using
bdrv_co_writev_flags() with modified flags. Compare to commit
9eeb6dd, which got flag manipulation right in
bdrv_co_do_zero_pwritev().
Symptoms: I tested with qemu-io with default writethrough cache
(which is supposed to use FUA semantics on every write), and
targetted an NBD client connected to a server that intentionally
did not advertise NBD_FLAG_SEND_FUA. When doing 'write 0 512',
the NBD client sent two operations (NBD_CMD_WRITE then
NBD_CMD_FLUSH) to get the fallback FUA semantics; but when doing
'write -z 0 512', the NBD client sent only NBD_CMD_WRITE.
The fix is do to a cleanup bdrv_co_flush() at the end of the
operation if any step in the middle relied on a BDS that does
not natively support FUA for that step (note that we don't
need to flush after every operation, if the operation is broken
into chunks based on bounce-buffer sizing). Each BDS gains a
new flag .supported_zero_flags, which parallels the use of
.supported_write_flags but only when accessing a zero write
operation (the flags MUST be different, because of SCSI having
different semantics based on WRITE vs. WRITESAME; and also
because BDRV_REQ_MAY_UNMAP only makes sense on zero writes).
Also fix some documentation to describe -ENOTSUP semantics,
particularly since iscsi depends on those semantics.
Down the road, we may want to add a driver where its
.bdrv_co_pwritev() honors all three of BDRV_REQ_FUA,
BDRV_REQ_ZERO_WRITE, and BDRV_REQ_MAY_UNMAP, and advertise
this via bs->supported_write_flags for blocks opened by that
driver; such a driver should NOT supply .bdrv_co_write_zeroes
nor .supported_zero_flags. But none of the drivers touched
in this patch want to do that (the act of writing zeroes is
different enough from normal writes to deserve a second
callback).
Signed-off-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Fam Zheng <famz@redhat.com>
Acked-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
3 years ago |
