Detected by a hang in the libnbd testsuite. If a client requests
multiple meta contexts (both base:allocation and qemu:dirty-bitmap:x)
at the same time, our attempt to silence a false-positive warning
about a potential uninitialized variable introduced botched logic: we
were short-circuiting the second context, and never sending the
NBD_REPLY_FLAG_DONE. Combining two 'if' into one 'if/else' in
bdf200a55 was wrong (I'm a bit embarrassed that such a change was my
initial suggestion after the v1 patch, then I did not review the v2
patch that actually got committed). Revert that, and instead silence
the false positive warning by replacing 'return ret' with 'return 0'
(the value it always has at that point in the code, even though it
eluded the deduction abilities of the robot that reported the false
positive).
Fixes: bdf200a553
Signed-off-by: Eric Blake <eblake@redhat.com>
Message-Id: <20200206173832.130004-1-eblake@redhat.com>
Reviewed-by: Laurent Vivier <laurent@vivier.eu>
Qemu as server currently won't accept export names larger than 256
bytes, nor create dirty bitmap names longer than 1023 bytes, so most
uses of qemu as client or server have no reason to get anywhere near
the NBD spec maximum of a 4k limit per string.
However, we weren't actually enforcing things, ignoring when the
remote side violates the protocol on input, and also having several
code paths where we send oversize strings on output (for example,
qemu-nbd --description could easily send more than 4k). Tighten
things up as follows:
client:
- Perform bounds check on export name and dirty bitmap request prior
to handing it to server
- Validate that copied server replies are not too long (ignoring
NBD_INFO_* replies that are not copied is not too bad)
server:
- Perform bounds check on export name and description prior to
advertising it to client
- Reject client name or metadata query that is too long
- Adjust things to allow full 4k name limit rather than previous
256 byte limit
Signed-off-by: Eric Blake <eblake@redhat.com>
Message-Id: <20191114024635.11363-4-eblake@redhat.com>
Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>
Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
As long as we limit NBD names to 256 bytes (the bare minimum permitted
by the standard), stack-allocation works for parsing a name received
from the client. But as mentioned in a comment, we eventually want to
permit up to the 4k maximum of the NBD standard, which is too large
for stack allocation; so switch everything in the server to use heap
allocation. For now, there is no change in actually supported name
length.
Signed-off-by: Eric Blake <eblake@redhat.com>
Message-Id: <20191114024635.11363-2-eblake@redhat.com>
[eblake: fix uninit variable compile failure]
Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>
Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
When iothreads are in use, the failure to grab the aio context results
in an assertion failure when trying to unlock things during blk_unref,
when trying to unlock a mutex that was not locked. In short, all
calls to nbd_export_put need to done while within the correct aio
context. But since nbd_export_put can recursively reach itself via
nbd_export_close, and recursively grabbing the context would deadlock,
we can't do the context grab directly in those functions, but must do
so in their callers.
Hoist the use of the correct aio_context from nbd_export_new() to its
caller qmp_nbd_server_add(). Then tweak qmp_nbd_server_remove(),
nbd_eject_notifier(), and nbd_esport_close_all() to grab the right
context, so that all callers during qemu now own the context before
nbd_export_put() can call blk_unref().
Remaining uses in qemu-nbd don't matter (since that use case does not
support iothreads).
Suggested-by: Kevin Wolf <kwolf@redhat.com>
Signed-off-by: Eric Blake <eblake@redhat.com>
Message-Id: <20190917023917.32226-1-eblake@redhat.com>
Reviewed-by: Sergio Lopez <slp@redhat.com>
On creation, the export's AioContext is set to the same one as the
BlockBackend, while the AioContext in the client QIOChannel is left
untouched.
As a result, when using data-plane, nbd_client_receive_next_request()
schedules coroutines in the IOThread AioContext, while the client's
QIOChannel is serviced from the main_loop, potentially triggering the
assertion at qio_channel_restart_[read|write].
To fix this, as soon we have the export corresponding to the client,
we call qio_channel_attach_aio_context() to attach the QIOChannel
context to the export's AioContext. This matches with the logic at
blk_aio_attached().
RHBZ: https://bugzilla.redhat.com/show_bug.cgi?id=1748253
Signed-off-by: Sergio Lopez <slp@redhat.com>
Message-Id: <20190912110032.26395-1-slp@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Signed-off-by: Eric Blake <eblake@redhat.com>
The server side is fairly straightforward: we can always advertise
support for detection of fast zero, and implement it by mapping the
request to the block layer BDRV_REQ_NO_FALLBACK.
Signed-off-by: Eric Blake <eblake@redhat.com>
Message-Id: <20190823143726.27062-5-eblake@redhat.com>
Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
[eblake: update iotests 223, 233]
Commit fe0480d6 and friends added BDRV_REQ_NO_FALLBACK as a way to
avoid wasting time on a preliminary write-zero request that will later
be rewritten by actual data, if it is known that the write-zero
request will use a slow fallback; but in doing so, could not optimize
for NBD. The NBD specification is now considering an extension that
will allow passing on those semantics; this patch updates the new
protocol bits and 'qemu-nbd --list' output to recognize the bit, as
well as the new errno value possible when using the new flag; while
upcoming patches will improve the client to use the feature when
present, and the server to advertise support for it.
The NBD spec recommends (but not requires) that ENOTSUP be avoided for
all but failures of a fast zero (the only time it is mandatory to
avoid an ENOTSUP failure is when fast zero is supported but not
requested during write zeroes; the questionable use is for ENOTSUP to
other actions like a normal write request). However, clients that get
an unexpected ENOTSUP will either already be treating it the same as
EINVAL, or may appreciate the extra bit of information. We were
equally loose for returning EOVERFLOW in more situations than
recommended by the spec, so if it turns out to be a problem in
practice, a later patch can tighten handling for both error codes.
Signed-off-by: Eric Blake <eblake@redhat.com>
Message-Id: <20190823143726.27062-3-eblake@redhat.com>
Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
[eblake: tweak commit message, also handle EOPNOTSUPP]
When creating a read-only image, we are still advertising support for
TRIM and WRITE_ZEROES to the client, even though the client should not
be issuing those commands. But seeing this requires looking across
multiple functions:
All callers to nbd_export_new() passed a single flag based solely on
whether the export allows writes. Later, we then pass a constant set
of flags to nbd_negotiate_options() (namely, the set of flags which we
always support, at least for writable images), which is then further
dynamically modified with NBD_FLAG_SEND_DF based on client requests
for structured options. Finally, when processing NBD_OPT_EXPORT_NAME
or NBD_OPT_EXPORT_GO we bitwise-or the original caller's flag with the
runtime set of flags we've built up over several functions.
Let's refactor things to instead compute a baseline of flags as soon
as possible which gets shared between multiple clients, in
nbd_export_new(), and changing the signature for the callers to pass
in a simpler bool rather than having to figure out flags. We can then
get rid of the 'myflags' parameter to various functions, and instead
refer to client for everything we need (we still have to perform a
bitwise-OR for NBD_FLAG_SEND_DF during NBD_OPT_EXPORT_NAME and
NBD_OPT_EXPORT_GO, but it's easier to see what is being computed).
This lets us quit advertising senseless flags for read-only images, as
well as making the next patch for exposing FAST_ZERO support easier to
write.
Signed-off-by: Eric Blake <eblake@redhat.com>
Message-Id: <20190823143726.27062-2-eblake@redhat.com>
Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
[eblake: improve commit message, update iotest 223]
The NBD specification defines NBD_FLAG_CAN_MULTI_CONN, which can be
advertised when the server promises cache consistency between
simultaneous clients (basically, rules that determine what FUA and
flush from one client are able to guarantee for reads from another
client). When we don't permit simultaneous clients (such as qemu-nbd
without -e), the bit makes no sense; and for writable images, we
probably have a lot more work before we can declare that actions from
one client are cache-consistent with actions from another. But for
read-only images, where flush isn't changing any data, we might as
well advertise multi-conn support. What's more, advertisement of the
bit makes it easier for clients to determine if 'qemu-nbd -e' was in
use, where a second connection will succeed rather than hang until the
first client goes away.
This patch affects qemu as server in advertising the bit. We may want
to consider patches to qemu as client to attempt parallel connections
for higher throughput by spreading the load over those connections
when a server advertises multi-conn, but for now sticking to one
connection per nbd:// BDS is okay.
See also: https://bugzilla.redhat.com/1708300
Signed-off-by: Eric Blake <eblake@redhat.com>
Message-Id: <20190815185024.7010-1-eblake@redhat.com>
[eblake: tweak blockdev-nbd.c to not request shared when writable,
fix iotest 233]
Reviewed-by: John Snow <jsnow@redhat.com>
Add a public interface for get. While we're at it,
rename "bdrv_get_dirty_bitmap_locked" to "bdrv_dirty_bitmap_get_locked".
(There are more functions to rename to the bdrv_dirty_bitmap_VERB form,
but they will wait until the conclusion of this series.)
Signed-off-by: John Snow <jsnow@redhat.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
Message-id: 20190709232550.10724-11-jsnow@redhat.com
Signed-off-by: John Snow <jsnow@redhat.com>
Commit 3d068aff (3.0) introduced NBD_MAX_BITMAP_EXTENTS as a limit on
how large we would allow a reply to NBD_CMD_BLOCK_STATUS to grow when
it is visiting a qemu:dirty-bitmap: context. Later, commit fb7afc79
(3.1) reused the constant to limit base:allocation context replies,
although the name is now less appropriate in that situation.
Rename things, and improve the macro to use units.h for better
legibility. Then reformat the comment to comply with checkpatch rules
added in the meantime. No semantic change.
Signed-off-by: Eric Blake <eblake@redhat.com>
Message-Id: <20190510151735.29687-1-eblake@redhat.com>
Reviewed-by: Stefano Garzarella <sgarzare@redhat.com>
Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
This adds a new parameter to blk_new() which requires its callers to
declare from which AioContext this BlockBackend is going to be used (or
the locks of which AioContext need to be taken anyway).
The given context is only stored and kept up to date when changing
AioContexts. Actually applying the stored AioContext to the root node
is saved for another commit.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
The NBD server uses an AioContext notifier, so it can tolerate that its
BlockBackend is switched to a different AioContext. Before we start
actually calling bdrv_try_set_aio_context(), which checks for
consistency, outside of test cases, we need to make sure that the NBD
server actually allows this.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
In commit 0c1d50bd, I added a couple of TODO comments about whether we
consult bl.request_alignment when responding to NBD_OPT_INFO. At the
time, qemu as server was hard-coding an advertised alignment of 512 to
clients that promised to obey constraints, and there was no function
for getting at a device's preferred alignment. But in hindsight,
advertising 512 when the block device prefers 1 caused other
compliance problems, and commit b0245d64 changed one of the two TODO
comments to advertise a more accurate alignment. Time to fix the other
TODO. Doesn't really impact qemu as client (our normal client doesn't
use NBD_OPT_INFO, and qemu-nbd --list promises to obey block sizes),
but it might prove useful to other clients.
Fixes: b0245d64
Signed-off-by: Eric Blake <eblake@redhat.com>
Message-Id: <20190403030526.12258-4-eblake@redhat.com>
Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
We've recently added traces for clients to flag server non-compliance;
let's do the same for servers to flag client non-compliance. According
to the spec, if the client requests NBD_INFO_BLOCK_SIZE, it is
promising to send all requests aligned to those boundaries. Of
course, if the client does not request NBD_INFO_BLOCK_SIZE, then it
made no promises so we shouldn't flag anything; and because we are
willing to handle clients that made no promises (the spec allows us to
use NBD_REP_ERR_BLOCK_SIZE_REQD if we had been unwilling), we already
have to handle unaligned requests (which the block layer already does
on our behalf). So even though the spec allows us to return EINVAL
for clients that promised to behave, it's easier to always answer
unaligned requests. Still, flagging non-compliance can be useful in
debugging a client that is trying to be maximally portable.
Qemu as client used to have one spot where it sent non-compliant
requests: if the server sends an unaligned reply to
NBD_CMD_BLOCK_STATUS, and the client was iterating over the entire
disk, the next request would start at that unaligned point; this was
fixed in commit a39286dd when the client was taught to work around
server non-compliance; but is equally fixed if the server is patched
to not send unaligned replies in the first place (yes, qemu 4.0 as
server still has few such bugs, although they will be patched in
4.1). Fortunately, I did not find any more spots where qemu as client
was non-compliant. I was able to test the patch by using the following
hack to convince qemu-io to run various unaligned commands, coupled
with serving 512-byte alignment by intentionally omitting '-f raw' on
the server while viewing server traces.
| diff --git i/nbd/client.c w/nbd/client.c
| index 427980bdd22..1858b2aac35 100644
| --- i/nbd/client.c
| +++ w/nbd/client.c
| @@ -449,6 +449,7 @@ static int nbd_opt_info_or_go(QIOChannel *ioc, uint32_t opt,
| nbd_send_opt_abort(ioc);
| return -1;
| }
| + info->min_block = 1;//hack
| if (!is_power_of_2(info->min_block)) {
| error_setg(errp, "server minimum block size %" PRIu32
| " is not a power of two", info->min_block);
Signed-off-by: Eric Blake <eblake@redhat.com>
Message-Id: <20190403030526.12258-3-eblake@redhat.com>
[eblake: address minor review nits]
Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Don't increment remaining_bytes until we know that we will actually be
including the current block status extent in the reply; otherwise, the
value traced will include a bytes value that is oversized by the
length of the next block status extent which did not get sent because
it instead ended the loop.
Fixes: fb7afc79
Signed-off-by: Eric Blake <eblake@redhat.com>
Message-Id: <20190403030526.12258-2-eblake@redhat.com>
Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Both NBD_CMD_BLOCK_STATUS and structured NBD_CMD_READ will split their
reply according to bdrv_block_status() boundaries. If the block device
has a request_alignment smaller than 512, but we advertise a block
alignment of 512 to the client, then this can result in the server
reply violating client expectations by reporting a smaller region of
the export than what the client is permitted to address (although this
is less of an issue for qemu 4.0 clients, given recent client patches
to overlook our non-compliance at EOF). Since it's always better to
be strict in what we send, it is worth advertising the actual minimum
block limit rather than blindly rounding it up to 512.
Note that this patch is not foolproof - it is still possible to
provoke non-compliant server behavior using:
$ qemu-nbd --image-opts driver=blkdebug,align=512,image.driver=file,image.filename=/path/to/non-aligned-file
That is arguably a bug in the blkdebug driver (it should never pass
back block status smaller than its alignment, even if it has to make
multiple bdrv_get_status calls and determine the
least-common-denominator status among the group to return). It may
also be possible to observe issues with a backing layer with smaller
alignment than the active layer, although so far I have been unable to
write a reliable iotest for that scenario (but again, an issue like
that could be argued to be a bug in the block layer, or something
where we need a flag to bdrv_block_status() to state whether the
result must be aligned to the current layer's limits or can be
subdivided for accuracy when chasing backing files).
Anyways, as blkdebug is not normally used, and as this patch makes our
server more interoperable with qemu 3.1 clients, it is worth applying
now, even while we still work on a larger patch series for the 4.1
timeframe to have byte-accurate file lengths.
Note that the iotests output changes - for 223 and 233, we can see the
server's better granularity advertisement; and for 241, the three test
cases have the following effects:
- natural alignment: the server's smaller alignment is now advertised,
and the hole reported at EOF is now the right result; we've gotten rid
of the server's non-compliance
- forced server alignment: the server still advertises 512 bytes, but
still sends a mid-sector hole. This is still a server compliance bug,
which needs to be fixed in the block layer in a later patch; output
does not change because the client is already being tolerant of the
non-compliance
- forced client alignment: the server's smaller alignment means that
the client now sees the server's status change mid-sector without any
protocol violations, but the fact that the map shows an unaligned
mid-sector hole is evidence of the block layer problems with aligned
block status, to be fixed in a later patch
Signed-off-by: Eric Blake <eblake@redhat.com>
Message-Id: <20190329042750.14704-7-eblake@redhat.com>
Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
[eblake: rebase to enhanced iotest 241 coverage]
Instead of checking against busy, inconsistent, or read only directly,
use a check function with permissions bits that let us streamline the
checks without reproducing them in many places.
Included in this patch are permissions changes that simply add the
inconsistent check to existing permissions call spots, without
addressing existing bugs.
In general, this means that busy+readonly checks become BDRV_BITMAP_DEFAULT,
which checks against all three conditions. busy-only checks become
BDRV_BITMAP_ALLOW_RO.
Notably, remove allows inconsistent bitmaps, so it doesn't follow the pattern.
Signed-off-by: John Snow <jsnow@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Message-id: 20190301191545.8728-4-jsnow@redhat.com
Signed-off-by: John Snow <jsnow@redhat.com>
These mean the same thing now. Unify them and rename the merged call
bdrv_dirty_bitmap_busy to indicate semantically what we are describing,
as well as help disambiguate from the various _locked and _unlocked
versions of bitmap helpers that refer to mutex locks.
Signed-off-by: John Snow <jsnow@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Message-id: 20190223000614.13894-8-jsnow@redhat.com
Signed-off-by: John Snow <jsnow@redhat.com>
Check that the bitmap is not in use prior to it checking if it is
not enabled/recording guest writes. The bitmap being busy was likely
at the behest of the user, so this error has a greater chance of being
understood by the user.
Signed-off-by: John Snow <jsnow@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Message-id: 20190223000614.13894-6-jsnow@redhat.com
Signed-off-by: John Snow <jsnow@redhat.com>
Eric Blake
Pan NengyuanandLaurent Vivier
Sergio LopezandEric Blake
John Snow
Peter Maydell
Markus Armbruster
Vladimir Sementsov-OgievskiyandEric Blake
Kevin Wolf