On 20.12.21 16:47, Emanuele Giuseppe Esposito wrote:
On 17/12/2021 13:29, Hanna Reitz wrote:
On 24.11.21 07:44, Emanuele Giuseppe Esposito wrote:
block_crypto_amend_options_generic_luks uses the block layer
permission API, therefore it should be called with the BQL held.
However, the same function is being called ib two BlockDriver
s/ ib / by /
callbacks: bdrv_amend_options (under BQL) and bdrv_co_amend (I/O).
The latter is I/O because it is invoked by block/amend.c's
blockdev_amend_run(), a .run callback of the amend JobDriver
Therefore we want to 1) change block_crypto_amend_options_generic_luks
to use the permission API only when the BQL is held, and
2) use the .pre_run JobDriver callback to check for
permissions before switching to the job aiocontext. This has also
the benefit of applying the same permission operation to all
amend implementations, not only luks.
Signed-off-by: Emanuele Giuseppe Esposito <eespo...@redhat.com>
---
block/amend.c | 20 ++++++++++++++++++++
block/crypto.c | 18 ++++++++++++------
2 files changed, 32 insertions(+), 6 deletions(-)
diff --git a/block/amend.c b/block/amend.c
index 392df9ef83..fba6add51a 100644
--- a/block/amend.c
+++ b/block/amend.c
@@ -53,10 +53,30 @@ static int coroutine_fn blockdev_amend_run(Job
*job, Error **errp)
return ret;
}
+static int blockdev_amend_refresh_perms(Job *job, Error **errp)
+{
+ BlockdevAmendJob *s = container_of(job, BlockdevAmendJob, common);
+
+ return bdrv_child_refresh_perms(s->bs, s->bs->file, errp);
+}
I miss some documentation for this function, why we do it and how it
works together with the bdrv_co_amend implementation.
I was trying to come up with an example text, but then I wondered –
how does it actually work? bdrv_child_refresh_perms() eventually ends
up in block_crypto_child_perms(). However, that will only return
exceptional permissions if crypto->updating_keys is true. But that’s
set only in block_crypto_amend_options_generic_luks() – i.e. when the
job runs. That’s exactly why that function calls
bdrv_child_refresh_perms() only after it has modified
crypto->updating_keys.
Reproducer (amend on a LUKS image with read-only=true, so it doesn’t
have the WRITE permission continuously, but needs to take it as an
exception in block_crypto_child_perms()):
$ qemu-img create \
-f luks \
--object secret,id=sec0,data=123456 \
-o key-secret=sec0 \
test.luks \
64M
Formatting 'test.luks', fmt=luks size=67108864 key-secret=sec0
$ ./qemu-system-x86_64 \
-object secret,id=sec0,data=123456 \
-object iothread,id=iothr0 \
-blockdev file,node-name=node0,filename=test.luks \
-blockdev
luks,node-name=node1,key-secret=sec0,file=node0,read-only=true \
-device virtio-blk,drive=node1,iothread=iothr0 -qmp stdio \
<<EOF
{"execute": "qmp_capabilities"}
{
"execute": "x-blockdev-amend",
"arguments": {
"job-id": "amend0",
"node-name": "node1",
"options": {
"driver": "luks",
"state": "active",
"new-secret": "sec0"
}
}
}
EOF
{"QMP": {"version": {"qemu": {"micro": 93, "minor": 1, "major": 6},
"package": "v6.2.0-rc3-50-gdb635fc4e7"}, "capabilities": ["oob"]}}
{"return": {}}
{"timestamp": {"seconds": 1639742600, "microseconds": 574641},
"event": "JOB_STATUS_CHANGE", "data": {"status": "created", "id":
"amend0"}}
{"timestamp": {"seconds": 1639742600, "microseconds": 574919},
"event": "JOB_STATUS_CHANGE", "data": {"status": "running", "id":
"amend0"}}
{"return": {}}
qemu-system-x86_64: ../block/io.c:2041: bdrv_co_write_req_prepare:
Assertion `child->perm & BLK_PERM_WRITE' failed.
[1] 55880 IOT instruction (core dumped) ./qemu-system-x86_64
-object secret,id=sec0,data=123456 -object -blockdev
I believe this means we need some new block driver function to
prepare for an amendment operation. If so, another question comes
up, which is whether this preparatory function should then also call
bdrv_child_refresh_perms(), and then whether we should have a
clean-up function for symmetry.
Yes, unfortunately it means that (see at the end of the mail for more).
I think it does not work because of crypto->updating_keys missing in
blockdev_amend_pre_run(). That is why the permission is not correctly
set and the example fails.
+
+static int blockdev_amend_pre_run(Job *job, Error **errp)
+{
+ return blockdev_amend_refresh_perms(job, errp);
+}
+
+static void blockdev_amend_clean(Job *job)
+{
+ Error *errp;
+ blockdev_amend_refresh_perms(job, &errp);
Do we really want to ignore this error? If so, we shouldn’t pass a
pointer to an unused local variable, but NULL.
If we don’t want to ignore it, we have the option of doing what you
do here and then at least reporting a potential error with
error_report_err(), and then freeing it, and we also must initialize
errp to NULL in this case.
Going with this one above, thanks.
If we expect no error to happen (e.g. because we require the amend
implementation to only release/share permissions and not
acquire/unshare them), then I’d expect passing &error_abort here.
+}
+
static const JobDriver blockdev_amend_job_driver = {
.instance_size = sizeof(BlockdevAmendJob),
.job_type = JOB_TYPE_AMEND,
.run = blockdev_amend_run,
+ .pre_run = blockdev_amend_pre_run,
+ .clean = blockdev_amend_clean,
};
void qmp_x_blockdev_amend(const char *job_id,
diff --git a/block/crypto.c b/block/crypto.c
index c8ba4681e2..82f154516c 100644
--- a/block/crypto.c
+++ b/block/crypto.c
@@ -780,6 +780,7 @@
block_crypto_get_specific_info_luks(BlockDriverState *bs, Error **errp)
static int
block_crypto_amend_options_generic_luks(BlockDriverState *bs,
QCryptoBlockAmendOptions *amend_options,
+ bool under_bql,
This name makes sense in the context of this series, but not so much
outside of it.
I’d rename it to e.g. “in_amend_job” (and invert its value), and then
explain that we don’t need to refresh the child permissions when
running in an amend job, because that job has already taken care of
that.
OTOH, given that I believe we need some separate preparatory function
anyway, perhaps we should just pull out the
bdrv_child_refresh_perms() from this function altogether, so that we
have:
block_crypto_amend_options_luks():
/* sets updating_keys to true, and invokes bdrv_child_refresh_perms() */
block_crypto_amend_options_prepare();
block_crypto_amend_options_generic_luks();
/* sets updating_keys to false, and invokes
bdrv_child_refresh_perms() */
block_crypto_amend_options_clean();
block_crypto_co_amend_luks():
/* No need to prepare or clean up, that is taken care of by the amend
job */
block_crypto_amend_options_generic_luks();
(If we decide not to put bdrv_child_refresh_perms() into
prepare()/clean(), then it would need to be called by
block_crypto_amend_options_luks(); and if we decide not to have a
block_crypto_amend_options_clean(), then we’d need to inline it fully.)
So a couple of things I will change (according with your feedbacks):
- Remove the assertion job->aio_context == qemu_in_main_thread() done
in job_co_entry, as it is wrong. I don't know why I added that, but we
cannot assume that job->run() always run in the main context, because
the job aiocontext can be different. I don't think there is a test
doing that now, but it is possible. If run() was in the main context,
then bdrv_co_amend (called only in blockdev_amend_run) would be GS
too, but it isn't, also according with your comment in v4:
"[...] .bdrv_co_amend very much strikes me like a GS function, but
it isn’t. I’m afraid it must work on nodes that are not in the main
context, and it launches a job, so AFAIU we absolutely cannot run it
under the BQL."
- Introduce block_crypto_amend_options_prepare and
block_crypto_amend_options_clean, as you suggested above. These fix
the GS call stack of block_crypto_amend_options_generic_luks()
- Introduce .bdrv_pre_run() and .bdrv_cleanup(), respectively called
by .job_pre_run() and .job_cleanup(). The reason is that we need to
set crypto->updating_keys, otherwise the job amend won't temporary
give the write permission so the example above would fail.
So for the I/O callstack of block_crypto_amend_options_generic_luks()
we will have:
job->pre_run():
.bdrv_pre_run();
crypto->update_keys = true;
blockdev_amend_refresh_perms()
job->run():
block_crypto_amend_options_generic_luks()
job->cleanup():
.bdrv_cleanup();
crypto->update_keys = false;
blockdev_amend_refresh_perms()
Sounds good! The only adjustment I’d make is to add “amend” somewhere
in the .bdrv functions (e.g. “.bdrv_amend_pre_run” and
“.bdrv_amend_cleanup”), because AFAIU they’ll still be amend-specific,
right?
(Happy holidays :))
Hanna