Another step towards killing off sector-based block APIs.
Signed-off-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Message-id: 1468624988-423-15-git-send-email-eblake@redhat.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Another step towards killing off sector-based block APIs.
While at it, call directly into nbd-client.c instead of having
a pointless trivial wrapper in nbd.c.
Signed-off-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Message-id: 1468624988-423-14-git-send-email-eblake@redhat.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Another step towards killing off sector-based block APIs.
Unlike write_zeroes, where we can be handed unaligned requests
and must fail gracefully with -ENOTSUP for a fallback, we are
guaranteed that discard requests are always aligned because the
block layer already ignored unaligned head/tail.
Signed-off-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Message-id: 1468624988-423-13-git-send-email-eblake@redhat.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Another step towards killing off sector-based block APIs.
Signed-off-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Message-id: 1468624988-423-12-git-send-email-eblake@redhat.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Another step towards killing off sector-based block APIs.
Signed-off-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Message-id: 1468624988-423-11-git-send-email-eblake@redhat.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
There's enough drivers with a sector-based callback that it will
be easier to switch one at a time. This patch adds a byte-based
callback, and then after all drivers are swapped, we'll drop the
sector-based callback.
[checkpatch doesn't like the space after coroutine_fn in
block_int.h, but it's consistent with the rest of the file]
Signed-off-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Message-id: 1468624988-423-10-git-send-email-eblake@redhat.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Another step towards byte-based interfaces everywhere. Replace
the sector-based driver callback .bdrv_aio_discard() with a new
byte-based .bdrv_aio_pdiscard(). Only raw-posix and RBD drivers
are affected, so it was not worth splitting into multiple patches.
Signed-off-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Message-id: 1468624988-423-9-git-send-email-eblake@redhat.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
The internal function converts to byte-based before calling into
RBD code; hoist the conversion to the callers so that callers
can then be switched to byte-based themselves.
Signed-off-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Message-id: 1468624988-423-8-git-send-email-eblake@redhat.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
The only remaining uses of paio_submit() were flush (with no
offset or count) and discard (which we are switching to byte-based);
furthermore, the similarly named paio_submit_co() is already
byte-based.
Signed-off-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Message-id: 1468624988-423-7-git-send-email-eblake@redhat.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Change sector-based blk_discard(), blk_co_discard(), and
blk_aio_discard() to instead be byte-based blk_pdiscard(),
blk_co_pdiscard(), and blk_aio_pdiscard(). NBD gets a lot
simpler now that ignoring the unaligned portion of a
byte-based discard request is handled under the hood by
the block layer.
Signed-off-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Message-id: 1468624988-423-6-git-send-email-eblake@redhat.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Another step towards byte-based interfaces everywhere. Replace
the sector-based bdrv_aio_discard() with a new byte-based
bdrv_aio_pdiscard(), which silently ignores any unaligned head
or tail. Driver callbacks will be converted in followup patches.
Signed-off-by: Eric Blake <eblake@redhat.com>
Message-id: 1468624988-423-5-git-send-email-eblake@redhat.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
BlockRequest is the internal struct used by bdrv_aio_*. At the
moment, all such calls were sector-based, but we will eventually
convert to byte-based; start by changing the internal variables
to be byte-based. No change to behavior, although the read and
write code can now go byte-based through more of the stack.
Signed-off-by: Eric Blake <eblake@redhat.com>
Message-id: 1468624988-423-4-git-send-email-eblake@redhat.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Another step towards byte-based interfaces everywhere. Replace
the sector-based bdrv_discard() with a new byte-based
bdrv_pdiscard(), which silently ignores any unaligned head
or tail.
Signed-off-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Message-id: 1468624988-423-3-git-send-email-eblake@redhat.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Another step towards byte-based interfaces everywhere. Replace
the sector-based bdrv_co_discard() with a new byte-based
bdrv_co_pdiscard(), which silently ignores any unaligned head
or tail. Driver callbacks will be converted in followup patches.
By calculating the alignment outside of the loop, and clamping
the max discard to an aligned value, we can simplify the actions
done within the loop.
Signed-off-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Message-id: 1468624988-423-2-git-send-email-eblake@redhat.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Now that the block layer honors max_request, we don't need to
bother with an EINVAL on overlarge requests, but can instead
assert that requests are well-behaved.
Signed-off-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Fam Zheng <famz@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Acked-by: Paolo Bonzini <pbonzini@redhat.com>
Message-id: 1468607524-19021-7-git-send-email-eblake@redhat.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Now that NBD relies on the block layer to fragment things, we no
longer need to track an offset argument for which fragment of
a request we are actually servicing.
While at it, use true and false instead of 0 and 1 for a bool
parameter.
Signed-off-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Fam Zheng <famz@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Acked-by: Paolo Bonzini <pbonzini@redhat.com>
Message-id: 1468607524-19021-6-git-send-email-eblake@redhat.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Now that the block layer will honor max_transfer, we can simplify
our code to rely on that guarantee.
The readv code can call directly into nbd-client, just as the
writev code has done since commit 52a4650.
Interestingly enough, while qemu-io 'w 0 40m' splits into a 32M
and 8M transaction, 'w -z 0 40m' splits into two 16M and an 8M,
because the block layer caps the bounce buffer for writing zeroes
at 16M. When we later introduce support for NBD_CMD_WRITE_ZEROES,
we can get a full 32M zero write (or larger, if the client and
server negotiate that write zeroes can use a larger size than
ordinary writes).
Signed-off-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Fam Zheng <famz@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Acked-by: Paolo Bonzini <pbonzini@redhat.com>
Message-id: 1468607524-19021-5-git-send-email-eblake@redhat.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Drivers should be able to rely on the block layer honoring the
max transfer length, rather than needing to return -EINVAL
(iscsi) or manually fragment things (nbd). We already fragment
write zeroes at the block layer; this patch adds the fragmentation
for normal writes, after requests have been aligned (fragmenting
before alignment would lead to multiple unaligned requests, rather
than just the head and tail).
When fragmenting a large request where FUA was requested, but
where we know that FUA is implemented by flushing all requests
rather than the given request, then we can still get by with
only one flush. Note, however, that we need a followup patch
to the raw format driver to avoid a regression in the number of
flushes actually issued.
The return value was previously nebulous on success (sometimes
zero, sometimes the length written); since we never have a short
write, and since fragmenting may store yet another positive
value in 'ret', change the function to always return 0 on success,
matching what we do in bdrv_aligned_preadv().
Signed-off-by: Eric Blake <eblake@redhat.com>
Message-id: 1468607524-19021-4-git-send-email-eblake@redhat.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
The raw format layer supports all flags via passthrough - but
it only makes sense to pass through flags that the lower layer
actually supports.
The next patch gives stronger reasoning for why this is correct.
At the moment, the raw format layer ignores the max_transfer
limit of its protocol layer, and an attempt to do the qemu-io
'w -f 0 40m' to an NBD server that lacks FUA will pass the entire
40m request to the NBD driver, which then fragments the request
itself into a 32m write, 8m write, and flush. But once the block
layer starts honoring limits and fragmenting packets, the raw
driver will hand the NBD driver two separate requests; if both
requests have BDRV_REQ_FUA set, then this would result in a 32m
write, flush, 8m write, and second flush. By having the raw
layer no longer advertise FUA support when the protocol layer
lacks it, we are back to a single flush at the block layer for
the overall 40m request.
Note that 'w -f -z 0 40m' does not currently exhibit the same
problem, because there, the fragmentation does not occur until
at the NBD layer (the raw layer has .bdrv_co_pwrite_zeroes, and
the NBD layer doesn't advertise max_pwrite_zeroes to constrain
things at the raw layer) - but the problem is latent and we
would again have too many flushes without this patch once the
NBD layer implements support for the new NBD_CMD_WRITE_ZEROES
command, if it sets max_pwrite_zeroes to the same 32m limit as
recommended by the NBD protocol.
Signed-off-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Fam Zheng <famz@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Message-id: 1468607524-19021-3-git-send-email-eblake@redhat.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Drivers should be able to rely on the block layer honoring the
max transfer length, rather than needing to return -EINVAL
(iscsi) or manually fragment things (nbd). This patch adds
the fragmentation in the block layer, after requests have been
aligned (fragmenting before alignment would lead to multiple
unaligned requests, rather than just the head and tail).
The return value was previously nebulous on success on whether
it was zero or the length read; and fragmenting may introduce
yet other non-zero values if we use the last length read. But
as at least some callers are sloppy and expect only zero on
success, it is easiest to just guarantee 0.
[Fix uninitialized ret local variable in bdrv_aligned_preadv().
--Stefan]
Signed-off-by: Eric Blake <eblake@redhat.com>
Message-id: 1468607524-19021-2-git-send-email-eblake@redhat.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
until now the allocation map was used only as a hint if a cluster
is allocated or not. If a block was not allocated (or Qemu had
no info about the allocation status) a get_block_status call was
issued to check the allocation status and possibly avoid
a subsequent read of unallocated sectors. If a block known to be
allocated the get_block_status call was omitted. In the other case
a get_block_status call was issued before every read to avoid
the necessity for a consistent allocation map. To avoid the
potential overhead of calling get_block_status for each and
every read request this took only place for the bigger requests.
This patch enhances this mechanism to cache the allocation
status and avoid calling get_block_status for blocks where
the allocation status has been queried before. This allows
for bypassing the read request even for smaller requests and
additionally omits calling get_block_status for known to be
unallocated blocks.
Signed-off-by: Peter Lieven <pl@kamp.de>
Message-Id: <1468831940-15556-3-git-send-email-pl@kamp.de>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
when setting clusters as alloacted the boundaries have
to be expanded. As Paolo pointed out the calculation of
the number of clusters is wrong:
Suppose cluster_sectors is 2, sector_num = 1, nb_sectors = 6:
In the "mark allocated" case, you want to set 0..8, i.e.
cluster_num=0, nb_clusters=4.
0--.--2--.--4--.--6--.--8
<--|_________________|--> (<--> = expanded)
Instead you are setting nb_clusters=3, so that 6..8 is not marked.
0--.--2--.--4--.--6--.--8
<--|______________|!!! (! = wrong)
Cc: qemu-stable@nongnu.org
Reported-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Peter Lieven <pl@kamp.de>
Message-Id: <1468831940-15556-2-git-send-email-pl@kamp.de>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Some guests (win2008 server for example) do a lot of unnecessary
flushing when underlying media has not changed. This adds additional
overhead on host when calling fsync/fdatasync.
This change introduces a write generation scheme in BlockDriverState.
Current write generation is checked against last flushed generation to
avoid unnessesary flushes.
The problem with excessive flushing was found by a performance test
which does parallel directory tree creation (from 2 processes).
Results improved from 0.424 loops/sec to 0.432 loops/sec.
Each loop creates 10^3 directories with 10 files in each.
This affected some blkdebug testcases that were expecting error logs from
failure-injected flushes which are now skipped entirely
(tests 026 071 089).
This also affects the performance of block jobs and thus BLOCK_JOB_READY
events for driver-mirror and active block-commit commands now arrives
faster, before QMP send successfully returns to caller (tests 141 144).
Signed-off-by: Evgeny Yakovlev <eyakovlev@virtuozzo.com>
Signed-off-by: Denis V. Lunev <den@openvz.org>
Reviewed-by: Paolo Bonzini <pbonzini@redhat.com>
Message-id: 1468870792-7411-5-git-send-email-den@openvz.org
CC: Kevin Wolf <kwolf@redhat.com>
CC: Max Reitz <mreitz@redhat.com>
CC: Stefan Hajnoczi <stefanha@redhat.com>
CC: Fam Zheng <famz@redhat.com>
CC: John Snow <jsnow@redhat.com>
Signed-off-by: John Snow <jsnow@redhat.com>
Invoking io_setup(MAX_EVENTS) we ask kernel to create ring buffer for us
with specified number of events. But kernel ring buffer allocation logic
is a bit tricky (ring buffer is page size aligned + some percpu allocation
are required) so eventually more than requested events number is allocated.
From a userspace side we have to follow the convention and should not try
to io_submit() more or logic, which consumes completed events, should be
changed accordingly. The pitfall is in the following sequence:
MAX_EVENTS = 128
io_setup(MAX_EVENTS)
io_submit(MAX_EVENTS)
io_submit(MAX_EVENTS)
/* now 256 events are in-flight */
io_getevents(MAX_EVENTS) = 128
/* we can handle only 128 events at once, to be sure
* that nothing is pended the io_getevents(MAX_EVENTS)
* call must be invoked once more or hang will happen. */
To prevent the hang or reiteration of io_getevents() call this patch
restricts the number of in-flights, which is now limited to MAX_EVENTS.
Signed-off-by: Roman Pen <roman.penyaev@profitbricks.com>
Reviewed-by: Fam Zheng <famz@redhat.com>
Reviewed-by: Paolo Bonzini <pbonzini@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Message-id: 1468415004-31755-1-git-send-email-roman.penyaev@profitbricks.com
Cc: Stefan Hajnoczi <stefanha@redhat.com>
Cc: qemu-devel@nongnu.org
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
This has better performance because it executes fewer system calls
and does not use a bottom half per disk.
Originally proposed by Ming Lei.
[Changed #include "raw-aio.h" to "block/raw-aio.h" in win32-aio.c to fix
build error as reported by Peter Maydell <peter.maydell@linaro.org>.
--Stefan]
Acked-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-id: 1467650000-51385-1-git-send-email-pbonzini@redhat.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
squash! linux-aio: share one LinuxAioState within an AioContext
First, bdrv_open_child() expects all options for the child to be
prefixed by the child's name (and a separating dot). Second,
bdrv_open_child() does not take ownership of the QDict passed to it but
only extracts all options for the child, so if a QDict is created for
the sole purpose of passing it to bdrv_open_child(), it needs to be
freed afterwards.
This patch makes vvfat adhere to both of these rules.
Signed-off-by: Max Reitz <mreitz@redhat.com>
Message-id: 20160711135452.11304-1-mreitz@redhat.com
Reviewed-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Signed-off-by: Max Reitz <mreitz@redhat.com>
Commit "cdeaf1f vmdk: add bdrv_co_write_zeroes" causes a regression on
writes. It writes metadata after every write instead of doing it only once
for each cluster.
vmdk_pwritev() writes metadata whenever m_data is set as valid so this patch
sets m_data as valid only when we have a new cluster which hasn't been
allocated before or a zero grain.
Signed-off-by: Reda Sallahi <fullmanet@gmail.com>
Message-id: 20160707084249.29084-1-fullmanet@gmail.com
Reviewed-by: Fam Zheng <famz@redhat.com>
Signed-off-by: Max Reitz <mreitz@redhat.com>
ratelimit_calculate_delay() previously reset the accounting every time
slice, no matter how much data had been processed before. This had (at
least) two consequences:
1. The minimum speed is rather large, e.g. 5 MiB/s for commit and stream.
Not sure if there are real-world use cases where this would be a
problem. Mirroring and backup over a slow link (e.g. DSL) would
come to mind, though.
2. Tests for block job operations (e.g. cancel) were rather racy
All block jobs currently use a time slice of 100ms. That's a
reasonable value to get smooth output during regular
operation. However this also meant that the state of block jobs
changed every 100ms, no matter how low the configured limit was. On
busy hosts, qemu often transferred additional chunks until the test
case had a chance to cancel the job.
Fix the block job rate limit code to delay for more than one time
slice to address the above issues. To make it easier to handle
oversized chunks we switch the semantics from returning a delay
_before_ the current request to a delay _after_ the current
request. If necessary, this delay consists of multiple time slice
units.
Since the mirror job sends multiple chunks in one go even if the rate
limit was exceeded in between, we need to keep track of the start of
the current time slice so we can correctly re-compute the delay for
the updated amount of data.
The minimum bandwidth now is 1 data unit per time slice. The block
jobs are currently passing the amount of data transferred in sectors
and using 100ms time slices, so this translates to 5120
bytes/second. With chunk sizes usually being O(512KiB), tests have
plenty of time (O(100s)) to operate on block jobs. The chance of a
race condition now is fairly remote, except possibly on insanely
loaded systems.
Signed-off-by: Sascha Silbe <silbe@linux.vnet.ibm.com>
Message-id: 1467127721-9564-2-git-send-email-silbe@linux.vnet.ibm.com
Reviewed-by: Max Reitz <mreitz@redhat.com>
Signed-off-by: Max Reitz <mreitz@redhat.com>
Recently, qcow2_get_cluster_offset() has been changed to work with bytes
instead of sectors. This invalidated some assertions and introduced a
possible integer multiplication overflow.
This could be reproduced using e.g.
$ qemu-img create -f qcow2 -o cluster_size=1M blub.qcow2 8G
Formatting 'foo.qcow2', fmt=qcow2 size=8589934592 encryption=off
cluster_size=1048576 lazy_refcounts=off refcount_bits=16
$ qemu-io -c map blub.qcow2
qemu-io: qemu/block/qcow2-cluster.c:504: qcow2_get_cluster_offset:
Assertion `bytes_needed <= INT_MAX' failed.
[1] 20775 abort (core dumped) qemu-io -c map foo.qcow2
This patch removes the now wrong assertion, adding comments and more
assertions to prove its correctness (and fixing the overflow which would
become apparent with the original assertion removed).
Signed-off-by: Max Reitz <mreitz@redhat.com>
Message-id: 20160620142623.24471-3-mreitz@redhat.com
Reviewed-by: Eric Blake <eblake@redhat.com>
Signed-off-by: Max Reitz <mreitz@redhat.com>
We refuse to open images whose L1 table we deem "too big". Consequently,
we should not produce such images ourselves.
Cc: qemu-stable@nongnu.org
Signed-off-by: Max Reitz <mreitz@redhat.com>
Message-id: 20160615153630.2116-3-mreitz@redhat.com
Reviewed-by: Eric Blake <eblake@redhat.com>
[mreitz: Added QEMU_BUILD_BUG_ON()]
Signed-off-by: Max Reitz <mreitz@redhat.com>
The rerror/werror policies are implemented in the devices, so that's
where they should be configured. In comparison to the old options in
-drive, the qdev properties are only added to those devices that
actually support them.
If the option isn't given (or "auto" is specified), the setting of the
BlockBackend is used for compatibility with the old options. For block
jobs, "auto" is the same as "enospc".
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
Commit implemented the 'enospc' policy as 'ignore' if the error was not
ENOSPC. The QAPI documentation promises that it's treated as 'stop'.
Using the common block job error handling function fixes this and also
adds the missing QMP event.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
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>
Signed-off-by: Fam Zheng <famz@redhat.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: John Snow <jsnow@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
This patch adds a new optional 'job-id' parameter to 'block-commit',
allowing the user to specify the ID of the block job to be created.
Signed-off-by: Alberto Garcia <berto@igalia.com>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
This patch adds a new optional 'job-id' parameter to 'block-stream',
allowing the user to specify the ID of the block job to be created.
The HMP 'block_stream' command remains unchanged.
Signed-off-by: Alberto Garcia <berto@igalia.com>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
This patch adds a new optional 'job-id' parameter to 'blockdev-backup'
and 'drive-backup', allowing the user to specify the ID of the block
job to be created.
The HMP 'drive_backup' command remains unchanged.
Signed-off-by: Alberto Garcia <berto@igalia.com>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
This patch adds a new optional 'job-id' parameter to 'blockdev-mirror'
and 'drive-mirror', allowing the user to specify the ID of the block
job to be created.
The HMP 'drive_mirror' command remains unchanged.
Signed-off-by: Alberto Garcia <berto@igalia.com>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
When a new job is created, the job ID is taken from the device name of
the BDS. This patch adds a new 'job_id' parameter to let the caller
provide one instead.
This patch also verifies that the ID is always unique and well-formed.
This causes problems in a couple of places where no ID is being set,
because the BDS does not have a device name.
In the case of test_block_job_start() (from test-blockjob-txn.c) we
can simply use this new 'job_id' parameter to set the missing ID.
In the case of img_commit() (from qemu-img.c) we still don't have the
API to make commit_active_start() set the job ID, so we solve it by
setting a default value. We'll get rid of this as soon as we extend
the API.
Signed-off-by: Alberto Garcia <berto@igalia.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
The 'id' field of the BlockJob structure will be able to hold any ID,
not only a device name. This patch updates the description of that
field and the error messages where it is being used.
Soon we'll add the ability to set an arbitrary ID when creating a
block job.
Signed-off-by: Alberto Garcia <berto@igalia.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Tracked down with an ugly, brittle and probably buggy Perl script.
Also move includes converted to <...> up so they get included before
ours where that's obviously okay.
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Tested-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Richard Henderson <rth@twiddle.net>
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1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=BgpV
-----END PGP SIGNATURE-----
Merge remote-tracking branch 'remotes/armbru/tags/pull-qapi-2016-07-06' into staging
QAPI patches for 2016-07-06
# gpg: Signature made Wed 06 Jul 2016 10:00:51 BST
# gpg: using RSA key 0x3870B400EB918653
# gpg: Good signature from "Markus Armbruster <armbru@redhat.com>"
# gpg: aka "Markus Armbruster <armbru@pond.sub.org>"
# Primary key fingerprint: 354B C8B3 D7EB 2A6B 6867 4E5F 3870 B400 EB91 8653
* remotes/armbru/tags/pull-qapi-2016-07-06:
replay: Use new QAPI cloning
sockets: Use new QAPI cloning
qapi: Add new clone visitor
qapi: Add new visit_complete() function
tests: Factor out common code in qapi output tests
tests: Clean up test-string-output-visitor
qmp-output-visitor: Favor new visit_free() function
string-output-visitor: Favor new visit_free() function
qmp-input-visitor: Favor new visit_free() function
string-input-visitor: Favor new visit_free() function
opts-visitor: Favor new visit_free() function
qapi: Add new visit_free() function
qapi: Add parameter to visit_end_*
qemu-img: Don't leak errors when outputting JSON
qapi: Improve use of qmp/types.h
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Making each output visitor provide its own output collection
function was the only remaining reason for exposing visitor
sub-types to the rest of the code base. Add a polymorphic
visit_complete() function which is a no-op for input visitors,
and which populates an opaque pointer for output visitors. For
maximum type-safety, also add a parameter to the output visitor
constructors with a type-correct version of the output pointer,
and assert that the two uses match.
This approach was considered superior to either passing the
output parameter only during construction (action at a distance
during visit_free() feels awkward) or only during visit_complete()
(defeating type safety makes it easier to use incorrectly).
Most callers were function-local, and therefore a mechanical
conversion; the testsuite was a bit trickier, but the previous
cleanup patch minimized the churn here.
The visit_complete() function may be called at most once; doing
so lets us use transfer semantics rather than duplication or
ref-count semantics to get the just-built output back to the
caller, even though it means our behavior is not idempotent.
Generated code is simplified as follows for events:
|@@ -26,7 +26,7 @@ void qapi_event_send_acpi_device_ost(ACP
| QDict *qmp;
| Error *err = NULL;
| QMPEventFuncEmit emit;
|- QmpOutputVisitor *qov;
|+ QObject *obj;
| Visitor *v;
| q_obj_ACPI_DEVICE_OST_arg param = {
| info
|@@ -39,8 +39,7 @@ void qapi_event_send_acpi_device_ost(ACP
|
| qmp = qmp_event_build_dict("ACPI_DEVICE_OST");
|
|- qov = qmp_output_visitor_new();
|- v = qmp_output_get_visitor(qov);
|+ v = qmp_output_visitor_new(&obj);
|
| visit_start_struct(v, "ACPI_DEVICE_OST", NULL, 0, &err);
| if (err) {
|@@ -55,7 +54,8 @@ void qapi_event_send_acpi_device_ost(ACP
| goto out;
| }
|
|- qdict_put_obj(qmp, "data", qmp_output_get_qobject(qov));
|+ visit_complete(v, &obj);
|+ qdict_put_obj(qmp, "data", obj);
| emit(QAPI_EVENT_ACPI_DEVICE_OST, qmp, &err);
and for commands:
| {
| Error *err = NULL;
|- QmpOutputVisitor *qov = qmp_output_visitor_new();
| Visitor *v;
|
|- v = qmp_output_get_visitor(qov);
|+ v = qmp_output_visitor_new(ret_out);
| visit_type_AddfdInfo(v, "unused", &ret_in, &err);
|- if (err) {
|- goto out;
|+ if (!err) {
|+ visit_complete(v, ret_out);
| }
|- *ret_out = qmp_output_get_qobject(qov);
|-
|-out:
| error_propagate(errp, err);
Signed-off-by: Eric Blake <eblake@redhat.com>
Message-Id: <1465490926-28625-13-git-send-email-eblake@redhat.com>
Reviewed-by: Markus Armbruster <armbru@redhat.com>
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Now that we have a polymorphic visit_free(), we no longer need
qmp_output_visitor_cleanup(); however, we still need to
expose the subtype for qmp_output_get_qobject().
Signed-off-by: Eric Blake <eblake@redhat.com>
Message-Id: <1465490926-28625-10-git-send-email-eblake@redhat.com>
Reviewed-by: Markus Armbruster <armbru@redhat.com>
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Now that we have a polymorphic visit_free(), we no longer need
opts_visitor_cleanup(); which in turn means we no longer need
to return a subtype from opts_visitor_new() nor a public upcast
function.
Signed-off-by: Eric Blake <eblake@redhat.com>
Message-Id: <1465490926-28625-6-git-send-email-eblake@redhat.com>
Reviewed-by: Markus Armbruster <armbru@redhat.com>
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Rather than making the dealloc visitor track of stack of pointers
remembered during visit_start_* in order to free them during
visit_end_*, it's a lot easier to just make all callers pass the
same pointer to visit_end_*. The generated code has access to the
same pointer, while all other users are doing virtual walks and
can pass NULL. The dealloc visitor is then greatly simplified.
All three visit_end_*() functions intentionally take a void**,
even though the visit_start_*() functions differ between void**,
GenericList**, and GenericAlternate**. This is done for several
reasons: when doing a virtual walk, passing NULL doesn't care
what the type is, but when doing a generated walk, we already
have to cast the caller's specific FOO* to call visit_start,
while using void** lets us use visit_end without a cast. Also,
an upcoming patch will add a clone visitor that wants to use
the same implementation for all three visit_end callbacks,
which is made easier if all three share the same signature.
For visitors with already track per-object state (the QMP visitors
via a stack, and the string visitors which do not allow nesting),
add an assertion that the caller is indeed passing the same
pointer to paired calls.
Signed-off-by: Eric Blake <eblake@redhat.com>
Message-Id: <1465490926-28625-4-git-send-email-eblake@redhat.com>
Reviewed-by: Markus Armbruster <armbru@redhat.com>
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Don't use the cpu_to_*w() functions, which we are trying to deprecate.
Instead either just use cpu_to_*() to do the byteswap, or use
st*_be_p() if we need to do the store somewhere other than to a
variable that's already the correct type.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Message-id: 1466093177-17890-1-git-send-email-peter.maydell@linaro.org
Reviewed-by: Eric Blake <eblake@redhat.com>
Signed-off-by: Max Reitz <mreitz@redhat.com>
This is the final patch for converting the common I/O path to take
a BdrvChild parameter instead of BlockDriverState.
The completion of this conversion means that all users that perform I/O
on an image need to actually hold a reference (in the form of BdrvChild,
possible as part of a BlockBackend) to that image. This also protects
against inconsistent use of BlockBackend vs. BlockDriverState functions
because direct use of a BlockDriverState isn't possible any more and
blk->root is private for block-backends.c.
In addition, we can now distinguish different users in the I/O path,
and the future op blockers work is going to add assertions based on
permissions stored in BdrvChild.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Acked-by: Stefan Hajnoczi <stefanha@redhat.com>