On my machine the arm_replay test takes over 2 minutes to run
in a config with Rust enabled and debug enabled:
$ time (cd build/rust ; PYTHONPATH=../../python:../../tests/functional
QEMU_TEST_QEMU_BINARY=./qemu-system-arm ./pyvenv/bin/python3
../../tests/functional/test_arm_replay.py)
TAP version 13
ok 1 test_arm_replay.ArmReplay.test_cubieboard
ok 2 test_arm_replay.ArmReplay.test_vexpressa9
ok 3 test_arm_replay.ArmReplay.test_virt
1..3
real 2m16.564s
user 2m13.461s
sys 0m3.523s
Bump up the timeout to 4 minutes.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Thomas Huth <thuth@redhat.com>
Message-id: 20250310102830.3752440-1-peter.maydell@linaro.org
I'm down as the only listed maintainer for quite a lot of Arm SoC and
board types. In some cases this is only as the "maintainer of last
resort" and I'm not in practice doing anything beyond patch review
and the odd bit of tidyup.
Move these entries in MAINTAINERS from "Maintained" to "Odd Fixes",
to better represent reality. Entries for other boards and SoCs where
I do more actively care (or where there is a listed co-maintainer)
remain as they are.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Philippe Mathieu-Daudé <philmd@linaro.org>
Message-id: 20250307152838.3226398-1-peter.maydell@linaro.org
In the Arm ARM, rule R_TYTWB states that returning to AArch32
is an illegal exception return if:
* AArch32 is not supported at any exception level
* the target EL is configured for AArch64 via SCR_EL3.RW
or HCR_EL2.RW or via CPU state at reset
We check the second of these, but not the first (which can only be
relevant for the case of a return to EL0, because if AArch32 is not
supported at one of the higher ELs then the RW bits will have an
effective value of 1 and the the "configured for AArch64" condition
will hold also).
Add the missing condition. Although this is technically a bug
(because we have one AArch64-only CPU: a64fx) it isn't worth
backporting to stable because no sensible guest code will
deliberately try to return to a nonexistent execution state
to check that it gets an illegal exception return.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
We already call env_archcpu() multiple times within the
exception_return helper function, and we're about to want to
add another use of the ARMCPU pointer. Add a local variable
cpu so we can call env_archcpu() just once.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
When EL1 doesn't support AArch32, the HCR_EL2.RW bit is supposed to
be RAO/WI. Enforce the RAO/WI behaviour.
Note that we handle "reset value should honour RES1 bits" in the same
way that SCR_EL3 does, via a reset function.
We do already have some CPU types which don't implement AArch32
above EL0, so this is technically a bug; it doesn't seem worth
backporting to stable because no sensible guest code will be
deliberately attempting to set the RW bit to a value corresponding
to an unimplemented execution state and then checking that we
did the right thing.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
During normal migration, new QEMU creates and initializes memory regions,
then loads the preserved contents of the region from vmstate.
During CPR, memory regions are preserved in place, then the realize
method initializes the regions contents, losing the old contents. To
fix, skip writes to the qxl memory regions during CPR load.
Reported-by: andrey.drobyshev@virtuozzo.com
Tested-by: andrey.drobyshev@virtuozzo.com
Signed-off-by: Steve Sistare <steven.sistare@oracle.com>
Reviewed-by: Fabiano Rosas <farosas@suse.de>
Message-ID: <1741380954-341079-5-git-send-email-steven.sistare@oracle.com>
Signed-off-by: Fabiano Rosas <farosas@suse.de>
During normal migration, new QEMU creates and initializes memory regions,
then loads the preserved contents of the region from vmstate.
During CPR, memory regions are preserved in place, then the realize
method initializes the regions contents, losing the old contents. To
fix, skip the re-init during CPR.
Signed-off-by: Steve Sistare <steven.sistare@oracle.com>
Reviewed-by: Fabiano Rosas <farosas@suse.de>
Message-ID: <1741380954-341079-4-git-send-email-steven.sistare@oracle.com>
Signed-off-by: Fabiano Rosas <farosas@suse.de>
During normal migration, new QEMU creates and initializes memory regions,
then loads the preserved contents of the region from vmstate.
During CPR, memory regions are preserved in place, then the realize
method initializes the regions contents, losing the old contents. To
fix, skip the re-init during CPR.
Signed-off-by: Steve Sistare <steven.sistare@oracle.com>
Reviewed-by: Fabiano Rosas <farosas@suse.de>
Message-ID: <1741380954-341079-3-git-send-email-steven.sistare@oracle.com>
Signed-off-by: Fabiano Rosas <farosas@suse.de>
Define the cpr_is_incoming helper, to be used in several cpr fix patches.
Signed-off-by: Steve Sistare <steven.sistare@oracle.com>
Reviewed-by: Peter Xu <peterx@redhat.com>
Reviewed-by: Fabiano Rosas <farosas@suse.de>
Message-ID: <1741380954-341079-2-git-send-email-steven.sistare@oracle.com>
Signed-off-by: Fabiano Rosas <farosas@suse.de>
The definition of SCR_EL3.RW says that its effective value is 1 if:
- EL2 is implemented and does not support AArch32, and SCR_EL3.NS is 1
- the effective value of SCR_EL3.{EEL2,NS} is {1,0} (i.e. we are
Secure and Secure EL2 is disabled)
We implement the second of these in arm_el_is_aa64(), but forgot the
first.
Provide a new function arm_scr_rw_eff() to return the effective
value of SCR_EL3.RW, and use it in arm_el_is_aa64() and the other
places that currently look directly at the bit value.
(scr_write() enforces that the RW bit is RAO/WI if neither EL1 nor
EL2 have AArch32 support, but if EL1 does but EL2 does not then the
bit must still be writeable.)
This will mean that if code at EL3 attempts to perform an exception
return to AArch32 EL2 when EL2 is AArch64-only we will correctly
handle this as an illegal exception return: it will be caught by the
"return to an EL which is configured for a different register width"
check in HELPER(exception_return).
We do already have some CPU types which don't implement AArch32
above EL0, so this is technically a bug; it doesn't seem worth
backporting to stable because no sensible guest code will be
deliberately attempting to set the RW bit to a value corresponding
to an unimplemented execution state and then checking that we
did the right thing.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
The functions arm_current_el() and arm_el_is_aa64() are used only in
target/arm and in hw/intc/arm_gicv3_cpuif.c. They're functions that
query internal state of the CPU. Move them out of cpu.h and into
internals.h.
This means we need to include internals.h in arm_gicv3_cpuif.c, but
this is justifiable because that file is implementing the GICv3 CPU
interface, which really is part of the CPU proper; we just ended up
implementing it in code in hw/intc/ for historical reasons.
The motivation for this move is that we'd like to change
arm_el_is_aa64() to add a condition that uses cpu_isar_feature();
but we don't want to include cpu-features.h in cpu.h.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
The arm_cpu_data_is_big_endian() and related functions are now used
only in target/arm; they can be moved to internals.h.
The motivation here is that we would like to move arm_current_el()
to internals.h.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
In linux-user/arm/cpu_loop.c we define a full set of get/put
macros for both code and data (since the endianness handling
is different between the two). However the only one we actually
use is get_user_code_u32(). Remove the rest.
We leave a comment noting how data-side accesses should be handled
for big-endian, because that's a subtle point and we just removed the
macros that were effectively documenting it.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
At the top of linux-user/aarch64/cpu_loop.c we define a set of
macros for reading and writing data and code words, but we never
use these macros. Delete them.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
We would like to move arm_el_is_aa64() to internals.h; however, it is
used by access_secure_reg(). Make that function not be inline, so
that it can stay in cpu.h.
access_secure_reg() is used only in two places:
* in hflags.c
* in the user-mode arm emulators, to decide whether to store
the TLS value in the secure or non-secure banked field
The second of these is not on a super-hot path that would care about
the inlining (and incidentally will always use the NS banked field
because our user-mode CPUs never set ARM_FEATURE_EL3); put the
definition of access_secure_reg() in hflags.c, near its only use
inside target/arm.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
The A32_BANKED_REG_{GET,SET} macros are only used inside target/arm;
move their definitions to cpregs.h. There's no need to have them
defined in all the code that includes cpu.h.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
This also creates the `qapi-qsd-index` and `qapi-qga-index` QMP indices.
Signed-off-by: John Snow <jsnow@redhat.com>
Message-ID: <20250313044312.189276-12-jsnow@redhat.com>
Acked-by: Markus Armbruster <armbru@redhat.com>
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Before we enable the QGA and QSD namespaces, we need to disambiguate
some of the references that would become ambiguous as a result!
Signed-off-by: John Snow <jsnow@redhat.com>
Message-ID: <20250313044312.189276-11-jsnow@redhat.com>
Acked-by: Markus Armbruster <armbru@redhat.com>
Signed-off-by: Markus Armbruster <armbru@redhat.com>
This also creates the qapi-qmp-index.html index and cross-reference
target.
Signed-off-by: John Snow <jsnow@redhat.com>
Message-ID: <20250313044312.189276-10-jsnow@redhat.com>
Acked-by: Markus Armbruster <armbru@redhat.com>
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Generate an index-per-namespace for the QAPI domain. Due to a limitation
with Sphinx's architecture, these indices must be defined during setup
time and cannot be dynamically created on-demand when a namespace
directive is encountered.
Owing to that limitation, add a configuration value to conf.py that
specifies which QAPI namespaces we'll generate indices for.
Indices will be named after their namespace, e.g. the "QMP" namespace
will generate to "qapi-qmp-index.html" and can be referenced using
`qapi-qmp-index`.
Signed-off-by: John Snow <jsnow@redhat.com>
Message-ID: <20250313044312.189276-9-jsnow@redhat.com>
Acked-by: Markus Armbruster <armbru@redhat.com>
Signed-off-by: Markus Armbruster <armbru@redhat.com>
This patch does three things:
1. Record the current namespace context in pending_xrefs so it can be
used for link resolution later,
2. Pass that recorded namespace context to find_obj() when resolving a
reference, and
3. Wildly and completely rewrite find_obj().
cross-reference support is expanded to tolerate the presence or absence
of either namespace or module, and to cope with the presence or absence
of contextual information for either.
References now work like this:
1. If the explicit reference target is recorded in the domain's object
registry, we link to that target and stop looking. We do this lookup
regardless of how fully qualified the target is, which allows direct
references to modules (which don't have a module component to their
names) or direct references to definitions that may or may not belong
to a namespace or module.
2. If contextual information is available from qapi:namespace or
qapi:module directives, try using those components to find a direct
match to the implied target name.
3. If both prior lookups fail, generate a series of regular expressions
looking for wildcard matches in order from most to least
specific. Any explicitly provided components (namespace, module)
*must* match exactly, but both contextual and entirely omitted
components are allowed to differ from the search result. Note that if
more than one result is found, Sphinx will emit a warning (a build
error for QEMU) and list all of the candidate references.
The practical upshot is that in the large majority of cases, namespace
and module information is not required when creating simple `references`
to definitions from within the same context -- even when identical
definitions exist in other contexts.
Even when using simple `references` from elsewhere in the QEMU
documentation manual, explicit namespace info is not required if there
is only one definition by that name.
Disambiguation *will* be required from outside of the QAPI documentation
when referencing e.g. block-core definitions, which are shared between
QEMU QMP and the QEMU Storage Daemon. In that case, there are two
options:
A: References can be made partially or fully explicit,
e.g. `QMP:block-dirty-bitmap-add` will link to the QEMU version of
the definition, while `QSD:block-dirty-bitmap-add` would link to the
QSD version.
B: If all of the references in a document are intended to go to the same
place, you can insert a "qapi:namespace:: QMP" directive to influence
the fuzzy-searching for later references.
Signed-off-by: John Snow <jsnow@redhat.com>
Message-ID: <20250313044312.189276-8-jsnow@redhat.com>
Acked-by: Markus Armbruster <armbru@redhat.com>
[Commit message typo fixed]
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Add a :namespace: option to the qapi-doc directive, which inserts a
qapi:namespace directive into the start of the generated document. This,
in turn, associates all auto-generated definitions by this directive
with the specified namespace.
The source info for these generated lines are credited to the start of
the qapi-doc directive, which isn't precisely correct, but I wasn't sure
how to get it more accurate without some re-parsing shenanigans.
Signed-off-by: John Snow <jsnow@redhat.com>
Message-ID: <20250313044312.189276-7-jsnow@redhat.com>
Acked-by: Markus Armbruster <armbru@redhat.com>
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Add a new directive that marks the beginning of a QAPI "namespace", for
example; "QMP", "QGA" or "QSD". This directive will associate all
subsequent QAPI directives in a document with the specified
namespace. This does not change the visual display of any of the
definitions or index entries, but does change the "Fully Qualified Name"
inside the QAPI domain's object table. This allows for two different
"namespaces" to define entities with otherwise identical names -- which
will come in handy for documenting both QEMU QMP and the QEMU Storage
Daemon.
Signed-off-by: John Snow <jsnow@redhat.com>
Message-ID: <20250313044312.189276-6-jsnow@redhat.com>
Acked-by: Markus Armbruster <armbru@redhat.com>
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Akin to the :module: override option, the :namespace: options allows you
to forcibly override the contextual namespace associatied with a
definition.
We don't necessarily actually need this, but I felt compelled to stick
close to how the Python domain works that offers context overrides.
As of this commit, it is possible to add e.g. ":namespace: QMP" to any
QAPI directive to forcibly associate that definition with a given
namespace.
Signed-off-by: John Snow <jsnow@redhat.com>
Message-ID: <20250313044312.189276-5-jsnow@redhat.com>
Acked-by: Markus Armbruster <armbru@redhat.com>
Signed-off-by: Markus Armbruster <armbru@redhat.com>
This patch adds a namespace component to the "Fully Qualified Name", in
the form of "domain:module.name". As there are no namespace directives
or options yet, this component will simply be empty as of this patch.
Signed-off-by: John Snow <jsnow@redhat.com>
Message-ID: <20250313044312.189276-4-jsnow@redhat.com>
Acked-by: Markus Armbruster <armbru@redhat.com>
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Currently, only the definition name is stored in the tree metadata; but
the node property is confusingly called "fullname". Rectify this by
always storing the FQN in the tree metadata.
... While we're here, re-organize the code in preparation for namespace
support to make it a bit easier to add additional components of the
FQN. With this change, there is now extremely little code left that's
taken directly from the Python domain :)
Signed-off-by: John Snow <jsnow@redhat.com>
Message-ID: <20250313044312.189276-3-jsnow@redhat.com>
Acked-by: Markus Armbruster <armbru@redhat.com>
Signed-off-by: Markus Armbruster <armbru@redhat.com>
When using the annotations feature, type hints do not need to be
imported at runtime, only at type check time. Move type-check-only
imports into a conditional to reduce the number of imports needed at
runtime.
Signed-off-by: John Snow <jsnow@redhat.com>
Message-ID: <20250313044312.189276-2-jsnow@redhat.com>
Acked-by: Markus Armbruster <armbru@redhat.com>
Signed-off-by: Markus Armbruster <armbru@redhat.com>
The description of feature @unstable is three paragraphs. The second
and third became part of the description by accident in commit
9fb49daabfb (qapi: Mark unstable QMP parts with feature 'unstable').
The second paragraph describes a defect in terms of the
implementation. Fine, but doesn't belong into user-facing
documentation. Turn it into a TODO section.
Rewrite everything else for clarity and completeness.
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Message-ID: <20250311131715.1296101-1-armbru@redhat.com>
Acked-by: Alberto Garcia <berto@igalia.com>
This tool converts a disk image to qcow2, writing the result directly
to stdout. This can be used for example to send the generated file
over the network.
This is equivalent to using qemu-img to convert a file to qcow2 and
then writing the result to stdout, with the difference that this tool
does not need to create this temporary qcow2 file and therefore does
not need any additional disk space.
Implementing this directly in qemu-img is not really an option because
it expects the output file to be seekable and it is also meant to be a
generic tool that supports all combinations of file formats and image
options. Instead, this tool can only produce qcow2 files with the
basic options, without compression, encryption or other features.
The input file is read twice. The first pass is used to determine
which clusters contain non-zero data and that information is used to
create the qcow2 header, refcount table and blocks, and L1 and L2
tables. After all that metadata is created then the second pass is
used to write the guest data.
By default qcow2-to-stdout.py expects the input to be a raw file, but
if qemu-storage-daemon is available then it can also be used to read
images in other formats. Alternatively the user can also run qemu-nbd
or qemu-storage-daemon manually instead.
Signed-off-by: Alberto Garcia <berto@igalia.com>
Signed-off-by: Madeeha Javed <javed@igalia.com>
Message-ID: <20240730141552.60404-1-berto@igalia.com>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Peter Krempa and Kevin Wolf observed that iothread-vq-mapping is
confusing to use because the control and event virtqueues have a fixed
location before the command virtqueues but need to be treated
differently.
Only expose the command virtqueues via iothread-vq-mapping so that the
command-line parameter is intuitive: it controls where SCSI requests are
processed.
The control virtqueue needs to be hardcoded to the main loop thread for
technical reasons anyway. Kevin also pointed out that it's better to
place the event virtqueue in the main loop thread since its no poll
behavior would prevent polling if assigned to an IOThread.
This change is its own commit to avoid squashing the previous commit.
Suggested-by: Kevin Wolf <kwolf@redhat.com>
Suggested-by: Peter Krempa <pkrempa@redhat.com>
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Message-ID: <20250311132616.1049687-14-stefanha@redhat.com>
Tested-by: Peter Krempa <pkrempa@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Previously the ctrl virtqueue was handled in the AioContext where SCSI
requests are processed. When IOThread Virtqueue Mapping was added things
become more complicated because SCSI requests could run in other
AioContexts.
Simplify by handling the ctrl virtqueue in the main loop where reset
operations can be performed. Note that BHs are still used canceling SCSI
requests in their AioContexts but at least the mean loop activity
doesn't need BHs anymore.
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Message-ID: <20250311132616.1049687-13-stefanha@redhat.com>
Tested-by: Peter Krempa <pkrempa@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Allow virtio-scsi virtqueues to be assigned to different IOThreads. This
makes it possible to take advantage of host multi-queue block layer
scalability by assigning virtqueues that have affinity with vCPUs to
different IOThreads that have affinity with host CPUs. The same feature
was introduced for virtio-blk in the past:
https://developers.redhat.com/articles/2024/09/05/scaling-virtio-blk-disk-io-iothread-virtqueue-mapping
Here are fio randread 4k iodepth=64 results from a 4 vCPU guest with an
Intel P4800X SSD:
iothreads IOPS
------------------------------
1 189576
2 312698
4 346744
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Message-ID: <20250311132616.1049687-12-stefanha@redhat.com>
Tested-by: Peter Krempa <pkrempa@redhat.com>
[kwolf: Updated 051 output, virtio-scsi can now use any iothread]
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
The code that builds an array of AioContext pointers indexed by the
virtqueue is not specific to virtio-blk. virtio-scsi will need to do the
same thing, so extract the functions.
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Message-ID: <20250311132616.1049687-11-stefanha@redhat.com>
Tested-by: Peter Krempa <pkrempa@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Use noun_verb() function naming instead of verb_noun() because the
former is the most common naming style for APIs. The next commit will
move these functions into a header file so that virtio-scsi can call
them.
Shorten iothread_vq_mapping_apply()'s iothread_vq_mapping_list argument
to just "list" like in the other functions.
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Message-ID: <20250311132616.1049687-10-stefanha@redhat.com>
Tested-by: Peter Krempa <pkrempa@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
This is the cleanup function that must be called after
apply_iothread_vq_mapping() succeeds. virtio-scsi will need this
function too, so extract it.
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Message-ID: <20250311132616.1049687-9-stefanha@redhat.com>
Tested-by: Peter Krempa <pkrempa@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
With IOThread Virtqueue Mapping there will be multiple AioContexts
processing SCSI requests. scsi_req_cancel() and other SCSI request
operations must be performed from the AioContext where the request is
running.
Introduce a virtio_scsi_defer_tmf_to_aio_context() function and the
necessary VirtIOSCSIReq->remaining refcount infrastructure to move the
TMF code into the AioContext where the request is running.
For the time being there is still just one AioContext: the main loop or
the IOThread. When the iothread-vq-mapping parameter is added in a later
patch this will be changed to per-virtqueue AioContexts.
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Message-ID: <20250311132616.1049687-8-stefanha@redhat.com>
Tested-by: Peter Krempa <pkrempa@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
The block layer can invoke the resize callback from any AioContext that
is processing requests. The virtqueue is already protected but the
events_dropped field also needs to be protected against races. Cover it
using the event virtqueue lock because it is closely associated with
accesses to the virtqueue.
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Message-ID: <20250311132616.1049687-7-stefanha@redhat.com>
Tested-by: Peter Krempa <pkrempa@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Virtqueues are not thread-safe. Until now this was not a major issue
since all virtqueue processing happened in the same thread. The ctrl
queue's Task Management Function (TMF) requests sometimes need the main
loop, so a BH was used to schedule the virtqueue completion back in the
thread that has virtqueue access.
When IOThread Virtqueue Mapping is introduced in later commits, event
and ctrl virtqueue accesses from other threads will become necessary.
Introduce an optional per-virtqueue lock so the event and ctrl
virtqueues can be protected in the commits that follow.
The addition of the ctrl virtqueue lock makes
virtio_scsi_complete_req_from_main_loop() and its BH unnecessary.
Instead, take the ctrl virtqueue lock from the main loop thread.
The cmd virtqueue does not have a lock because the entirety of SCSI
command processing happens in one thread. Only one thread accesses the
cmd virtqueue and a lock is unnecessary.
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Message-ID: <20250311132616.1049687-6-stefanha@redhat.com>
Tested-by: Peter Krempa <pkrempa@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
SCSIDevice keeps track of in-flight requests for device reset and Task
Management Functions (TMFs). The request list requires protection so
that multi-threaded SCSI emulation can be implemented in commits that
follow.
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Message-ID: <20250311132616.1049687-5-stefanha@redhat.com>
Tested-by: Peter Krempa <pkrempa@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Until now, a SCSIDevice's I/O requests have run in a single AioContext.
In order to support multiple IOThreads it will be necessary to move to
the concept of a per-SCSIRequest AioContext.
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Message-ID: <20250311132616.1049687-4-stefanha@redhat.com>
Tested-by: Peter Krempa <pkrempa@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
In the past a single AioContext was used for block I/O and it was
fetched using blk_get_aio_context(). Nowadays the block layer supports
running I/O from any AioContext and multiple AioContexts at the same
time. Remove the dma_blk_io() AioContext argument and use the current
AioContext instead.
This makes calling the function easier and enables multiple IOThreads to
use dma_blk_io() concurrently for the same block device.
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Message-ID: <20250311132616.1049687-3-stefanha@redhat.com>
Tested-by: Peter Krempa <pkrempa@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Commit 71544d30a6f8 ("scsi: push request restart to SCSIDevice") removed
the only user of SCSIDiskState->bh.
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Philippe Mathieu-Daudé <philmd@linaro.org>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Message-ID: <20250311132616.1049687-2-stefanha@redhat.com>
Tested-by: Peter Krempa <pkrempa@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
qsd-migrate is currently only working for raw, qcow2 and qed.
Other formats are failing, e.g. because they don't support migration.
Thus let's limit this test to the three usable formats now.
Suggested-by: Kevin Wolf <kwolf@redhat.com>
Signed-off-by: Thomas Huth <thuth@redhat.com>
Message-ID: <20250224214058.205889-1-thuth@redhat.com>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
aio_dispatch_handler() adds handlers to ctx->poll_aio_handlers if
polling should be enabled. If we call adjust_polling_time() for all
polling handlers before this, new polling handlers are still left at
poll->ns = 0 and polling is only actually enabled after the next event.
Move the adjust_polling_time() call after aio_dispatch_handler().
This fixes test-nested-aio-poll, which expects that polling becomes
effective the first time around.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Message-ID: <20250311141912.135657-1-kwolf@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Adaptive polling has a big problem: It doesn't consider that an event
loop can wait for many different events that may have very different
typical latencies.
For example, think of a guest that tends to send a new I/O request soon
after the previous I/O request completes, but the storage on the host is
rather slow. In this case, getting the new request from guest quickly
means that polling is enabled, but the next thing is performing the I/O
request on the backend, which is slow and disables polling again for the
next guest request. This means that in such a scenario, polling could
help for every other event, but is only ever enabled when it can't
succeed.
In order to fix this, keep a separate AioPolledEvent for each
AioHandler. We will then know that the backend file descriptor always
has a high latency and isn't worth polling for, but we also know that
the guest is always fast and we should poll for it. This solves at least
half of the problem, we can now keep polling for those cases where it
makes sense and get the improved performance from it.
Since the event loop doesn't know which event will be next, we still do
some unnecessary polling while we're waiting for the slow disk. I made
some attempts to be more clever than just randomly growing and shrinking
the polling time, and even to let callers be explicit about when they
expect a new event, but so far this hasn't resulted in improved
performance or even caused performance regressions. For now, let's just
fix the part that is easy enough to fix, we can revisit the rest later.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Message-ID: <20250307221634.71951-6-kwolf@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Message-ID: <20250307221634.71951-5-kwolf@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
As a preparation for having multiple adaptive polling states per
AioContext, move the 'ns' field into a separate struct.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Message-ID: <20250307221634.71951-4-kwolf@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
For block drivers that don't advertise FUA support, we already call
bdrv_co_flush(), which considers BDRV_O_NO_FLUSH. However, drivers that
do support FUA still see the FUA flag with BDRV_O_NO_FLUSH and get the
associated performance penalty that cache.no-flush=on was supposed to
avoid.
Clear FUA for write requests if BDRV_O_NO_FLUSH is set.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Message-ID: <20250307221634.71951-3-kwolf@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
