If an NBD server advertises a size that is not a multiple of a sector,
the block layer rounds up that size, even though we set info.size to
the exact byte value sent by the server. The block layer then proceeds
to let us read or query block status on the hole that it added past
EOF, which the NBD server is unlikely to be happy with. Fortunately,
qemu as a server never advertizes an unaligned size, so we generally
don't run into this problem; but the nbdkit server makes it easy to
test:
$ printf %1000d 1 > f1
$ ~/nbdkit/nbdkit -fv file f1 & pid=$!
$ qemu-img convert -f raw nbd://localhost:10809 f2
$ kill $pid
$ qemu-img compare f1 f2
Pre-patch, the server attempts a 1024-byte read, which nbdkit
rightfully rejects as going beyond its advertised 1000 byte size; the
conversion fails and the output files differ (not even the first
sector is copied, because qemu-img does not follow ddrescue's habit of
trying smaller reads to get as much information as possible in spite
of errors). Post-patch, the client's attempts to read (and query block
status, for new enough nbdkit) are properly truncated to the server's
length, with sane handling of the hole the block layer forced on
us. Although f2 ends up as a larger file (1024 bytes instead of 1000),
qemu-img compare shows the two images to have identical contents for
display to the guest.
I didn't add iotests coverage since I didn't want to add a dependency
on nbdkit in iotests. I also did NOT patch write, trim, or write
zeroes - these commands continue to fail (usually with ENOSPC, but
whatever the server chose), because we really can't write to the end
of the file, and because 'qemu-img convert' is the most common case
where we care about being tolerant (which is read-only). Perhaps we
could truncate the request if the client is writing zeros to the tail,
but that seems like more work, especially if the block layer is fixed
in 4.1 to track byte-accurate sizing (in which case this patch would
be reverted as unnecessary).
Signed-off-by: Eric Blake <eblake@redhat.com>
Message-Id: <20190329042750.14704-5-eblake@redhat.com>
Tested-by: Richard W.M. Jones <rjones@redhat.com>
It is desirable for 'qemu-img map' to have the same output for a file
whether it is served over file or nbd protocols. However, ever since
we implemented block status for NBD (2.12), the NBD protocol forgot to
inform the block layer that as the final layer in the chain, the
offset is valid; without an offset, the human-readable form of
qemu-img map gives up with the unhelpful:
$ nbdkit -U - data data="1" size=512 --run 'qemu-img map $nbd'
Offset Length Mapped to File
qemu-img: File contains external, encrypted or compressed clusters.
The --output=json form always works, because it is reporting the
lower-level bdrv_block_status results directly rather than trying to
filter out sparse ranges for human consumption - but now it also
shows the offset member.
With this patch, the human output changes to:
Offset Length Mapped to File
0 0x200 0 nbd+unix://?socket=/tmp/nbdkitOxeoLa/socket
This change is observable to several iotests.
Fixes: 78a33ab5
Reported-by: Richard W.M. Jones <rjones@redhat.com>
Signed-off-by: Eric Blake <eblake@redhat.com>
Message-Id: <20190329042750.14704-4-eblake@redhat.com>
Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
The NBD spec is clear that a server that advertises a minimum block
size should reply to NBD_CMD_BLOCK_STATUS with extents aligned
accordingly. However, we know that the qemu NBD server implementation
has had a corner-case bug where it is not compliant with the spec,
present since the introduction of NBD_CMD_BLOCK_STATUS in qemu 2.12
(and unlikely to be patched in time for 4.0). Namely, when qemu is
serving a file that is not a multiple of 512 bytes, it rounds the size
advertised over NBD up to the next sector boundary (someday, I'd like
to fix that to be byte-accurate, but it's a much bigger audit not
appropriate for this release); yet if the final sector contains data
prior to EOF, lseek(SEEK_HOLE) will point to the implicit hole
mid-sector which qemu then reported over NBD.
We are well within our rights to hang up on a server that can't follow
the spec, but it is more useful to try and keep the connection alive
in spite of the problem. Do so by tracing a message about the problem,
and then either truncating the request back to an aligned boundary (if
it covered more than the final sector) or widening it out to the full
boundary with a forced status of data (since truncating would result
in 0 bytes, but we have to make progress, and valid since data is a
default-safe answer). And in practice, since the problem only happens
on a sector that starts with data and ends with a hole, we are going
to want to read that full sector anyway (where qemu as the server
fills in the tail beyond EOF with appropriate NUL bytes).
Easy reproduction:
$ printf %1000d 1 > file
$ qemu-nbd -f raw -t file & pid=$!
$ qemu-img map --output=json -f raw nbd://localhost:10809
qemu-img: Could not read file metadata: Invalid argument
$ kill $pid
where the patched version instead succeeds with:
[{ "start": 0, "length": 1024, "depth": 0, "zero": false, "data": true}]
Signed-off-by: Eric Blake <eblake@redhat.com>
Message-Id: <20190326171317.4036-1-eblake@redhat.com>
Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
The NBD spec is clear that when structured replies are active, a
simple error reply is acceptable to any command except for
NBD_CMD_READ. However, we were mistakenly requiring structured errors
for NBD_CMD_BLOCK_STATUS, and hanging up on a server that gave a
simple error (since qemu does not behave as such a server, we didn't
notice the problem until now). Broken since its introduction in
commit 78a33ab5 (v2.12).
Noticed while debugging a separate failure reported by nbdkit while
working out its initial implementation of BLOCK_STATUS, although it
turns out that nbdkit also chose to send structured error replies for
BLOCK_STATUS, so I had to manually provoke the situation by hacking
qemu's server to send a simple error reply:
| diff --git i/nbd/server.c w/nbd/server.c
| index fd013a2817a..833288d7c45 100644
| 00--- i/nbd/server.c
| +++ w/nbd/server.c
| @@ -2269,6 +2269,8 @@ static coroutine_fn int nbd_handle_request(NBDClient *client,
| "discard failed", errp);
|
| case NBD_CMD_BLOCK_STATUS:
| + return nbd_co_send_simple_reply(client, request->handle, ENOMEM,
| + NULL, 0, errp);
| if (!request->len) {
| return nbd_send_generic_reply(client, request->handle, -EINVAL,
| "need non-zero length", errp);
|
Signed-off-by: Eric Blake <eblake@redhat.com>
Acked-by: Richard W.M. Jones <rjones@redhat.com>
Message-Id: <20190325190104.30213-3-eblake@redhat.com>
Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
When the server replies with a (structured [*]) error to
NBD_CMD_BLOCK_STATUS, without any extent information sent first, the
client code was blindly throwing away the server's error code and
instead telling the caller that EIO occurred. This has been broken
since its introduction in 78a33ab5 (v2.12, where we should have called:
error_setg(&local_err, "Server did not reply with any status extents");
nbd_iter_error(&iter, false, -EIO, &local_err);
to declare the situation as a non-fatal error if no earlier error had
already been flagged, rather than just blindly slamming iter.err and
iter.ret), although it is more noticeable since commit 7f86068d, which
actually tries hard to preserve the server's code thanks to a separate
iter.request_ret.
[*] The spec is clear that the server is also permitted to reply with
a simple error, but that's a separate fix.
I was able to provoke this scenario with a hack to the server, then
seeing whether ENOMEM makes it back to the caller:
| diff --git a/nbd/server.c b/nbd/server.c
| index fd013a2817a..29c7995de02 100644
| --- a/nbd/server.c
| +++ b/nbd/server.c
| @@ -2269,6 +2269,8 @@ static coroutine_fn int nbd_handle_request(NBDClient *client,
| "discard failed", errp);
|
| case NBD_CMD_BLOCK_STATUS:
| + return nbd_send_generic_reply(client, request->handle, -ENOMEM,
| + "no status for you today", errp);
| if (!request->len) {
| return nbd_send_generic_reply(client, request->handle, -EINVAL,
| "need non-zero length", errp);
| --
Signed-off-by: Eric Blake <eblake@redhat.com>
Message-Id: <20190325190104.30213-2-eblake@redhat.com>
Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
The NBD spec states that NBD_CMD_FLAG_REQ_ONE (which we currently
always use) should not reply with an extent larger than our request,
and that the server's response should be exactly one extent. Right
now, that means that if a server sends more than one extent, we treat
the server as broken, fail the block status request, and disconnect,
which prevents all further use of the block device. But while good
software should be strict in what it sends, it should be tolerant in
what it receives.
While trying to implement NBD_CMD_BLOCK_STATUS in nbdkit, we
temporarily had a non-compliant server sending too many extents in
spite of REQ_ONE. Oddly enough, 'qemu-img convert' with qemu 3.1
failed with a somewhat useful message:
qemu-img: Protocol error: invalid payload for NBD_REPLY_TYPE_BLOCK_STATUS
which then disappeared with commit d8b4bad8, on the grounds that an
error message flagged only at the time of coroutine teardown is
pointless, and instead we should rely on the actual failed API to
report an error - in other words, the 3.1 behavior was masking the
fact that qemu-img was not reporting an error. That has since been
fixed in the previous patch, where qemu-img convert now fails with:
qemu-img: error while reading block status of sector 0: Invalid argument
But even that is harsh. Since we already partially relaxed things in
commit acfd8f7a to tolerate a server that exceeds the cap (although
that change was made prior to the NBD spec actually putting a cap on
the extent length during REQ_ONE - in fact, the NBD spec change was
BECAUSE of the qemu behavior prior to that commit), it's not that much
harder to argue that we should also tolerate a server that sends too
many extents. But at the same time, it's nice to trace when we are
being tolerant of server non-compliance, in order to help server
writers fix their implementations to be more portable (if they refer
to our traces, rather than just stderr).
Reported-by: Richard W.M. Jones <rjones@redhat.com>
Signed-off-by: Eric Blake <eblake@redhat.com>
Message-Id: <20190323212639.579-3-eblake@redhat.com>
Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
bdrv_drain() must not leave connection_co scheduled, so bs->in_flight
needs to be increased while the coroutine is waiting to be scheduled
in the new AioContext after nbd_client_attach_aio_context().
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Instead of using the convenience wrapper qio_channel_read_all_eof(), use
the lower level QIOChannel API. This means duplicating some code, but
we'll need this because this coroutine yield is special: We want it to
be interruptible so that nbd_client_attach_aio_context() can correctly
reenter the coroutine.
This moves the bdrv_dec/inc_in_flight() pair into nbd_read_eof(), so
that connection_co will always sit in this exact qio_channel_yield()
call when bdrv_drain() returns.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
nbd_client_attach_aio_context() schedules connection_co in the new
AioContext and this way reenters it in any arbitrary place that has
yielded. We can restrict this a bit to the function call where the
coroutine actually sits waiting when it's idle.
This doesn't solve any bug yet, but it shows where in the code we need
to support this random reentrance and where we don't have to care.
Add FIXME comments for the existing bugs that the rest of this series
will fix.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
This coroutine will serve nbd reconnects, so, rename it to be something
more generic.
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-Id: <20190201130138.94525-7-vsementsov@virtuozzo.com>
Signed-off-by: Eric Blake <eblake@redhat.com>
We have several paranoid checks for ioc != NULL. But ioc may become
NULL only on close, which should not happen during requests handling.
Also, we check ioc only sometimes, not after each yield, which is
inconsistent. Let's drop these checks. However, for safety, let's leave
asserts instead.
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-Id: <20190201130138.94525-6-vsementsov@virtuozzo.com>
Signed-off-by: Eric Blake <eblake@redhat.com>
Use exported report, not the variable to be reused (should not really
matter).
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-Id: <20190201130138.94525-5-vsementsov@virtuozzo.com>
Signed-off-by: Eric Blake <eblake@redhat.com>
Split connection code to reuse it for reconnect.
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-Id: <20190201130138.94525-4-vsementsov@virtuozzo.com>
Signed-off-by: Eric Blake <eblake@redhat.com>
Keep all connection code in one file, to be able to implement reconnect
in further patches.
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Message-Id: <20190201130138.94525-3-vsementsov@virtuozzo.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
[eblake: format tweak]
Signed-off-by: Eric Blake <eblake@redhat.com>
To implement nbd reconnect in further patches, we need to distinguish
error codes, returned by nbd server, from channel errors, to reconnect
only in the latter case.
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-Id: <20190201130138.94525-2-vsementsov@virtuozzo.com>
Signed-off-by: Eric Blake <eblake@redhat.com>
We generally do very similar things around nbd_read: error_prepend
specifying what we have tried to read, and be_to_cpu conversion of
integers.
So, it seems reasonable to move common things to helper functions,
which:
1. simplify code a bit
2. generalize nbd_read error descriptions, all starting with
"Failed to read"
3. make it more difficult to forget to convert things from BE
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-Id: <20190128165830.165170-1-vsementsov@virtuozzo.com>
[eblake: rename macro to DEF_NBD_READ_N and formatting tweaks;
checkpatch has false positive complaint]
Signed-off-by: Eric Blake <eblake@redhat.com>
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>
Pass 'info' instead of three separate parameters related to info,
when requesting the server to set the meta context. Update the
NBDExportInfo struct to rename the received id field to match the
fact that we are currently overloading the field to match whatever
context the user supplied through the x-dirty-bitmap hack, as well
as adding a TODO comment to remind future patches about a desire
to request two contexts at once.
Signed-off-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Richard W.M. Jones <rjones@redhat.com>
Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Message-Id: <20190117193658.16413-11-eblake@redhat.com>
Refactor the 'name' parameter of nbd_receive_negotiate() from
being a separate parameter into being part of the in-out 'info'.
This also spills over to a simplification of nbd_opt_go().
The main driver for this refactoring is that an upcoming patch
would like to add support to qemu-nbd to list information about
all exports available on a server, where the name(s) will be
provided by the server instead of the client. But another benefit
is that we can now allow the client to explicitly specify the
empty export name "" even when connecting to an oldstyle server
(even if qemu is no longer such a server after commit 7f7dfe2a).
Signed-off-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Richard W.M. Jones <rjones@redhat.com>
Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Message-Id: <20190117193658.16413-10-eblake@redhat.com>
Reduce extra noise of nbd-client, change 083 correspondingly.
In various commits (be41c100 in 2.10, f140e300 in 2.11, 78a33ab
in 2.12), we added spots where qemu as an NBD client would report
problems communicating with the server to stderr, because there
was no where else to send the error to. However, this is racy,
particularly since the most common source of these errors is when
either the client or the server abruptly hangs up, leaving one
coroutine to report the error only if it wins (or loses) the
race in attempting the read from the server before another
thread completes its cleanup of a protocol error that caused the
disconnect in the first place. The race is also apparent in the
fact that differences in the flush behavior of the server can
alter the frequency of encountering the race in the client (see
commit 6d39db96).
Rather than polluting stderr, it's better to just trace these
situations, for use by developers debugging a flaky connection,
particularly since the real error that either triggers the abrupt
disconnection in the first place, or that results from the EIO
when a request can't receive a reply, DOES make it back to the
user in the normal Error propagation channels.
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Message-Id: <20181102151152.288399-4-vsementsov@virtuozzo.com>
[eblake: drop depedence on error hint, enhance commit message]
Signed-off-by: Eric Blake <eblake@redhat.com>
If nbd_client_init() fails after we are already connected,
then the server will spam logs with:
Disconnect client, due to: Unexpected end-of-file before all bytes were read
unless we gracefully disconnect before closing the connection.
Ways to trigger this:
$ opts=driver=nbd,export=foo,server.type=inet,server.host=localhost,server.port=10809
$ qemu-img map --output=json --image-opts $opts,read-only=off
$ qemu-img map --output=json --image-opts $opts,x-dirty-bitmap=nosuch:
Signed-off-by: Eric Blake <eblake@redhat.com>
Message-Id: <20181130023232.3079982-4-eblake@redhat.com>
Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
The implementation of x-dirty-bitmap in qemu 3.0 (commit 216ee365)
silently falls back to treating the server as not supporting
NBD_CMD_BLOCK_STATUS if a requested meta_context name was not
negotiated, which in turn means treating the _entire_ image as
data. Since our hack relied on using 'qemu-img map' to view
which portions of the image were dirty by seeing what the
redirected bdrv_block_status() treats as holes, this means
that our fallback treats the entire image as clean. Better
would have been to treat the entire image as dirty, or to fail
to connect because the user's request for a specific context
could not be honored. This patch goes with the latter.
Signed-off-by: Eric Blake <eblake@redhat.com>
Message-Id: <20181130023232.3079982-3-eblake@redhat.com>
Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
If read-only=off, but auto-read-only=on is given, open a read-write NBD
connection if the server provides a read-write export, but instead of
erroring out for read-only exports, just degrade to read-only.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
In order to test that the NBD server is properly advertising
dirty bitmaps, we need a bare minimum client that can request
and read the context. Since feature freeze for 3.0 is imminent,
this is the smallest workable patch, which replaces the qemu
block status report with the results of the NBD server's dirty
bitmap (making it very easy to use 'qemu-img map --output=json'
to learn where the dirty portions are). Note that the NBD
protocol defines a dirty section with the same bit but opposite
sense that normal "base:allocation" uses to report an allocated
section; so in qemu-img map output, "data":true corresponds to
clean, "data":false corresponds to dirty.
A more complete solution that allows dirty bitmaps to be queried
at the same time as normal block status will be required before
this addition can lose the x- prefix. Until then, the fact that
this replaces normal status with dirty status means actions
like 'qemu-img convert' will likely misbehave due to treating
dirty regions of the file as if they are unallocated.
The next patch adds an iotest to exercise this new code.
Signed-off-by: Eric Blake <eblake@redhat.com>
Message-Id: <20180702191458.28741-2-eblake@redhat.com>
The NBD spec is proposing a relaxation of NBD_CMD_BLOCK_STATUS
where a server may have the final extent per context give a
length beyond the original request, if it can easily prove that
subsequent bytes have the same status, on the grounds that a
client can take advantage of this information for fewer block
status requests. Since qemu 2.12 as a client always sends
NBD_CMD_FLAG_REQ_ONE, and rejects a server that sends extra
length, the upstream NBD spec will probably limit this behavior
to clients that don't request REQ_ONE semantics; but it doesn't
hurt to relax qemu to always be permissive of this server
behavior, even if it continues to use REQ_ONE.
CC: qemu-stable@nongnu.org
Signed-off-by: Eric Blake <eblake@redhat.com>
Message-Id: <20180503222626.1303410-1-eblake@redhat.com>
Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
iotests 123 and 209 fail on 32-bit platforms. The culprit:
sizeof(extent) is wrong; we want sizeof(*extent). But since
the struct is 8 bytes, it happened to work on 64-bit platforms
where the pointer is also 8 bytes (nasty).
Fixes: 78a33ab58
Reported-by: Max Reitz <mreitz@redhat.com>
Signed-off-by: Eric Blake <eblake@redhat.com>
Message-Id: <20180327210517.1804242-1-eblake@redhat.com>
Reviewed-by: Paolo Bonzini <pbonzini@redhat.com>
Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Minimal realization: only one extent in server answer is supported.
Flag NBD_CMD_FLAG_REQ_ONE is used to force this behavior.
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Message-Id: <20180312152126.286890-6-vsementsov@virtuozzo.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
[eblake: grammar tweaks, fix min_block check and 32-bit cap, use -1
instead of errno on failure in nbd_negotiate_simple_meta_context,
ensure that block status makes progress on success]
Signed-off-by: Eric Blake <eblake@redhat.com>
It is ok, that fatal error hides previous not fatal, but hiding
first fatal error is a bad feature.
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-Id: <20180312152126.286890-5-vsementsov@virtuozzo.com>
Signed-off-by: Eric Blake <eblake@redhat.com>
Commit 79ba8c98 (v2.7) changed the setting of request_alignment
to occur only during bdrv_refresh_limits(), rather than at at
bdrv_open() time; but at the time, NBD was unaffected, because
it still used sector-based callbacks, so the block layer
defaulted NBD to use 512 request_alignment.
Later, commit 70c4fb26 (also v2.7) changed NBD to use byte-based
callbacks, without setting request_alignment. This resulted in
NBD using request_alignment of 1, which works great when the
server supports it (as is the case for qemu-nbd), but falls apart
miserably if the server requires alignment (but only if qemu
actually sends a sub-sector request; qemu-io can do it, but
most qemu operations still perform on sectors or larger).
Even later, the NBD protocol was updated to document that clients
should learn the server's minimum alignment during NBD_OPT_GO;
and recommended that clients should assume a minimum size of 512
unless the server understands NBD_OPT_GO and replied with a smaller
size. Commit 081dd1fe (v2.10) attempted to do that, by assigning
request_alignment to whatever was learned from the server; but
it has two flaws: the assignment is done during bdrv_open() so
it gets unconditionally wiped out back to 1 during any later
bdrv_refresh_limits(); and the code is not using a default of 512
when the server did not report a minimum size.
Fix these issues by moving the assignment to request_alignment
to the right function, and by using a sane default when the
server does not advertise a minimum size.
CC: qemu-stable@nongnu.org
Signed-off-by: Eric Blake <eblake@redhat.com>
Message-Id: <20180215032905.27146-1-eblake@redhat.com>
Reviewed-by: Vladimir Sementsov-Ogievskiy<vsementsov@virtuozzo.com>
If a server fails a read, for example with EIO, but the connection
is still live, then we would crash trying to print a non-existent
error message in nbd_client_co_preadv(). For consistency, also
change the error printout in nbd_read_reply_entry(), although that
instance does not crash. Bug introduced in commit f140e300.
Signed-off-by: Eric Blake <eblake@redhat.com>
Message-Id: <20171112013936.5942-1-eblake@redhat.com>
Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Ensure that the server is not sending unexpected chunk lengths
for either the NONE or the OFFSET_DATA chunk, nor unexpected
hole length for OFFSET_HOLE. This will flag any server as
broken that responds to a zero-length read with an OFFSET_DATA
(what our server currently does, but that's about to be fixed)
or with OFFSET_HOLE, even though we previously fixed our client
to never be able to send such a request over the wire.
Signed-off-by: Eric Blake <eblake@redhat.com>
Message-Id: <20171108215703.9295-7-eblake@redhat.com>
Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
The NBD spec was recently clarified to state that clients should
not send 0-length requests to the server, as the server behavior
is undefined [1]. We know that qemu-nbd's behavior is a successful
no-op (once it has filtered for read-only exports), but other NBD
implementations might return an error. To avoid any questionable
server implementations, it is better to just short-circuit such
requests on the client side (we are relying on the block layer to
already filter out requests such as invalid offset, write to a
read-only volume, and so forth); do the short-circuit as late as
possible to still benefit from protections from assertions that
the block layer is not violating our assumptions.
[1] https://github.com/NetworkBlockDevice/nbd/commit/ee926037
Signed-off-by: Eric Blake <eblake@redhat.com>
Message-Id: <20171108215703.9295-6-eblake@redhat.com>
Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
The NBD spec says that clients should not try to write/trim to
an export advertised as read-only by the server. But we failed
to check that, and would allow the block layer to use NBD with
BDRV_O_RDWR even when the server is read-only, which meant we
were depending on the server sending a proper EPERM failure for
various commands, and also exposes a leaky abstraction: using
qemu-io in read-write mode would succeed on 'w -z 0 0' because
of local short-circuiting logic, but 'w 0 0' would send a
request over the wire (where it then depends on the server, and
fails at least for qemu-nbd but might pass for other NBD
implementations).
With this patch, a client MUST request read-only mode to access
a server that is doing a read-only export, or else it will get
a message like:
can't open device nbd://localhost:10809/foo: request for write access conflicts with read-only export
It is no longer possible to even attempt writes over the wire
(including the corner case of 0-length writes), because the block
layer enforces the explicit read-only request; this matches the
behavior of qcow2 when backed by a read-only POSIX file.
Fix several iotests to comply with the new behavior (since
qemu-nbd of an internal snapshot, as well as nbd-server-add over QMP,
default to a read-only export, we must tell blockdev-add/qemu-io to
set up a read-only client).
CC: qemu-stable@nongnu.org
Signed-off-by: Eric Blake <eblake@redhat.com>
Message-Id: <20171108215703.9295-3-eblake@redhat.com>
Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Provide missing spaces that are required when using string
concatenation to break error messages across source lines.
Introduced in commit f140e300.
Signed-off-by: Eric Blake <eblake@redhat.com>
Message-Id: <20171108215703.9295-2-eblake@redhat.com>
Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Minimal implementation: for structured error only error_report error
message.
Note that test 83 is now more verbose, because the implementation
prints more warnings about unexpected communication errors; perhaps
future patches should tone things down by using trace messages
instead of traces, but the common case of successful communication
is no noisier than before.
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Signed-off-by: Eric Blake <eblake@redhat.com>
Message-Id: <20171027104037.8319-13-eblake@redhat.com>
In following patch nbd_receive_reply will be used both for simple
and structured reply header receiving.
NBDReply is altered into union of simple reply header and structured
reply chunk header, simple error translation moved to block/nbd-client
to be consistent with further structured reply error translation.
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Signed-off-by: Eric Blake <eblake@redhat.com>
Message-Id: <20171027104037.8319-11-eblake@redhat.com>
Pass handle parameter directly, as the whole request isn't needed.
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-Id: <20171012095319.136610-3-vsementsov@virtuozzo.com>
Signed-off-by: Eric Blake <eblake@redhat.com>
Also improve the assertion: check that qiov is NULL for other commands
than CMD_READ and CMD_WRITE.
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-Id: <20171012095319.136610-2-vsementsov@virtuozzo.com>
Signed-off-by: Eric Blake <eblake@redhat.com>
It's incorrect to return success rc >= 0 if we skip qio_channel_writev_all()
call due to s->quit.
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-Id: <20170920124507.18841-4-vsementsov@virtuozzo.com>
Signed-off-by: Eric Blake <eblake@redhat.com>
If we are woken up from while() loop in nbd_read_reply_entry
handles must be equal. If we are woken up from
nbd_recv_coroutines_wake_all s->quit must be true, so we do
not need checking handles equality.
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-Id: <20170920124507.18841-3-vsementsov@virtuozzo.com>
Signed-off-by: Eric Blake <eblake@redhat.com>
"NBDReply *reply" parameter of nbd_co_receive_reply is used only
to pass return value for nbd_co_request (reply.error). Remove it
and use function return value instead.
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-Id: <20170920124507.18841-2-vsementsov@virtuozzo.com>
Signed-off-by: Eric Blake <eblake@redhat.com>
If 'bs' is a complex expression, we were only casting the front half
rather than the full expression. Luckily, none of the callers were
passing bad arguments, but it's better to be robust up front.
Signed-off-by: Eric Blake <eblake@redhat.com>
Message-Id: <20170918214649.17550-1-eblake@redhat.com>
Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Rather than open-coding our own read/write-all functions, we
can make use of the recently-added qio code. It slightly
changes the error message in one of the iotests.
Signed-off-by: Eric Blake <eblake@redhat.com>
Message-Id: <20170905191114.5959-4-eblake@redhat.com>
Reviewed-by: Daniel P. Berrange <berrange@redhat.com>
Add nbd_co_request, to remove code duplications in
nbd_client_co_{pwrite,pread,...} functions. Also this is
needed for further refactoring.
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Message-Id: <20170804151440.320927-8-vsementsov@virtuozzo.com>
[eblake: make nbd_co_request a wrapper, rather than merging two
existing functions]
Signed-off-by: Eric Blake <eblake@redhat.com>
Rename nbd_recv_coroutines_enter_all to nbd_recv_coroutines_wake_all,
as it most probably just adds all recv coroutines into co_queue_wakeup,
rather than directly enter them.
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Message-Id: <20170804151440.320927-9-vsementsov@virtuozzo.com>
[eblake: tweak commit message]
Signed-off-by: Eric Blake <eblake@redhat.com>
Use int variable for nbd_co_send_request return value (as
nbd_co_send_request returns int).
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Message-Id: <20170804151440.320927-6-vsementsov@virtuozzo.com>
Signed-off-by: Eric Blake <eblake@redhat.com>
The following segfault is encountered if the NBD server closes the UNIX
domain socket immediately after negotiation:
Program terminated with signal SIGSEGV, Segmentation fault.
#0 aio_co_schedule (ctx=0x0, co=0xd3c0ff2ef0) at util/async.c:441
441 QSLIST_INSERT_HEAD_ATOMIC(&ctx->scheduled_coroutines,
(gdb) bt
#0 0x000000d3c01a50f8 in aio_co_schedule (ctx=0x0, co=0xd3c0ff2ef0) at util/async.c:441
#1 0x000000d3c012fa90 in nbd_coroutine_end (bs=bs@entry=0xd3c0fec650, request=<optimized out>) at block/nbd-client.c:207
#2 0x000000d3c012fb58 in nbd_client_co_preadv (bs=0xd3c0fec650, offset=0, bytes=<optimized out>, qiov=0x7ffc10a91b20, flags=0) at block/nbd-client.c:237
#3 0x000000d3c0128e63 in bdrv_driver_preadv (bs=bs@entry=0xd3c0fec650, offset=offset@entry=0, bytes=bytes@entry=512, qiov=qiov@entry=0x7ffc10a91b20, flags=0) at block/io.c:836
#4 0x000000d3c012c3e0 in bdrv_aligned_preadv (child=child@entry=0xd3c0ff51d0, req=req@entry=0x7f31885d6e90, offset=offset@entry=0, bytes=bytes@entry=512, align=align@entry=1, qiov=qiov@entry=0x7ffc10a91b20, f
+lags=0) at block/io.c:1086
#5 0x000000d3c012c6b8 in bdrv_co_preadv (child=0xd3c0ff51d0, offset=offset@entry=0, bytes=bytes@entry=512, qiov=qiov@entry=0x7ffc10a91b20, flags=flags@entry=0) at block/io.c:1182
#6 0x000000d3c011cc17 in blk_co_preadv (blk=0xd3c0ff4f80, offset=0, bytes=512, qiov=0x7ffc10a91b20, flags=0) at block/block-backend.c:1032
#7 0x000000d3c011ccec in blk_read_entry (opaque=0x7ffc10a91b40) at block/block-backend.c:1079
#8 0x000000d3c01bbb96 in coroutine_trampoline (i0=<optimized out>, i1=<optimized out>) at util/coroutine-ucontext.c:79
#9 0x00007f3196cb8600 in __start_context () at /lib64/libc.so.6
The problem is that nbd_client_init() uses
nbd_client_attach_aio_context() -> aio_co_schedule(new_context,
client->read_reply_co). Execution of read_reply_co is deferred to a BH
which doesn't run until later.
In the mean time blk_co_preadv() can be called and nbd_coroutine_end()
calls aio_wake() on read_reply_co. At this point in time
read_reply_co's ctx isn't set because it has never been entered yet.
This patch simplifies the nbd_co_send_request() ->
nbd_co_receive_reply() -> nbd_coroutine_end() lifecycle to just
nbd_co_send_request() -> nbd_co_receive_reply(). The request is "ended"
if an error occurs at any point. Callers no longer have to invoke
nbd_coroutine_end().
This cleanup also eliminates the segfault because we don't call
aio_co_schedule() to wake up s->read_reply_co if sending the request
failed. It is only necessary to wake up s->read_reply_co if a reply was
received.
Note this only happens with UNIX domain sockets on Linux. It doesn't
seem possible to reproduce this with TCP sockets.
Suggested-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Message-Id: <20170829122745.14309-2-stefanha@redhat.com>
Signed-off-by: Eric Blake <eblake@redhat.com>
The following scenario leads to an assertion failure in
qio_channel_yield():
1. Request coroutine calls qio_channel_yield() successfully when sending
would block on the socket. It is now yielded.
2. nbd_read_reply_entry() calls nbd_recv_coroutines_enter_all() because
nbd_receive_reply() failed.
3. Request coroutine is entered and returns from qio_channel_yield().
Note that the socket fd handler has not fired yet so
ioc->write_coroutine is still set.
4. Request coroutine attempts to send the request body with nbd_rwv()
but the socket would still block. qio_channel_yield() is called
again and assert(!ioc->write_coroutine) is hit.
The problem is that nbd_read_reply_entry() does not distinguish between
request coroutines that are waiting to receive a reply and those that
are not.
This patch adds a per-request bool receiving flag so
nbd_read_reply_entry() can avoid spurious aio_wake() calls.
Reported-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Message-Id: <20170822125113.5025-1-stefanha@redhat.com>
Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Tested-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Eric Blake <eblake@redhat.com>
travis builds fail at HEAD at rc3 master with
block/nbd-client.c: In function ‘nbd_read_reply_entry’:
block/nbd-client.c:110:8: error: ‘ret’ may be used uninitialized in this function [-Werror=uninitialized]
fix it by initializing 'ret' to 0
Signed-off-by: Igor Mammedov <imammedo@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
When we switched NBD to use coroutines for qemu 2.9 (in particular,
commit a12a712a), we introduced a regression: if a server sends us
garbage (such as a corrupted magic number), we quit the read loop
but do not stop sending further queued commands, resulting in the
client hanging when it never reads the response to those additional
commands. In qemu 2.8, we properly detected that the server is no
longer reliable, and cancelled all existing pending commands with
EIO, then tore down the socket so that all further command attempts
get EPIPE.
Restore the proper behavior of quitting (almost) all communication
with a broken server: Once we know we are out of sync or otherwise
can't trust the server, we must assume that any further incoming
data is unreliable and therefore end all pending commands with EIO,
and quit trying to send any further commands. As an exception, we
still (try to) send NBD_CMD_DISC to let the server know we are going
away (in part, because it is easier to do that than to further
refactor nbd_teardown_connection, and in part because it is the
only command where we do not have to wait for a reply).
Based on a patch by Vladimir Sementsov-Ogievskiy.
A malicious server can be created with the following hack,
followed by setting NBD_SERVER_DEBUG to a non-zero value in the
environment when running qemu-nbd:
| --- a/nbd/server.c
| +++ b/nbd/server.c
| @@ -919,6 +919,17 @@ static int nbd_send_reply(QIOChannel *ioc, NBDReply *reply, Error **errp)
| stl_be_p(buf + 4, reply->error);
| stq_be_p(buf + 8, reply->handle);
|
| + static int debug;
| + static int count;
| + if (!count++) {
| + const char *str = getenv("NBD_SERVER_DEBUG");
| + if (str) {
| + debug = atoi(str);
| + }
| + }
| + if (debug && !(count % debug)) {
| + buf[0] = 0;
| + }
| return nbd_write(ioc, buf, sizeof(buf), errp);
| }
Reported-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Signed-off-by: Eric Blake <eblake@redhat.com>
Message-Id: <20170814213426.24681-1-eblake@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Eric Blake
Kevin Wolf
Vladimir Sementsov-Ogievskiy
Stefan Hajnoczi
Igor Mammedov