| Age | Commit message (Collapse) | Author |
|---|
|
When VIRTIO_F_ACCESS_PLATFORM changes, cached vring pointers and IOTLB
metadata are interpreted in a different address space. Keeping them
across the transition can leave stale ring mappings in use.
Clearing d->iotlb before taking the VQ locks also lets a worker observe
a transient NULL d->iotlb and fall back to d->umem while translating a
descriptor.
Add a common vhost_clear_device_iotlb() helper for vhost-net and
vhost-vsock. Take all VQ mutexes in index order before dropping the
device-wide IOTLB, invalidate each VQ's cached ring access and metadata,
clear pending IOTLB messages, and free the old table after the handoff.
This serializes the transition with workers and prevents mixed address
space mappings.
On the first direct-to-IOTLB transition, invalidate the cached vring
addresses. When an existing device IOTLB is replaced, preserve the
GIOVA ring addresses and reset only the metadata cache. After clearing
ACCESS_PLATFORM, userspace must configure the vring addresses for the
new address mode.
vhost_vq_invalidate_access() clears desc, avail, and used together.
Treat the VQ as invalidated only when all three are NULL, since a single
GIOVA address may legitimately be zero.
Fixes: 6b1e6cc7855b ("vhost: new device IOTLB API")
Fixes: e13a6915a03f ("vhost/vsock: add IOTLB API support")
Suggested-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Jia Jia <physicalmtea@gmail.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Message-ID: <20260828085721.57816-1-physicalmtea@gmail.com>
|
|
vhost allocates a message node whenever address translation misses. If
userspace reads these messages without resolving them, repeated virtqueue
kicks can grow the pending message list until the host runs out of memory.
Virtqueue processing stops at the first translation miss and cannot make
progress until userspace installs a mapping. Keep a pointer to that
outstanding message in the virtqueue and suppress additional misses until
the node is resolved or discarded.
The pointer remains set while the message is queued for reading, copied to
userspace, or waiting on the pending list. Clear it under the IOTLB lock
when the owning node is freed. This bounds outstanding miss messages by the
fixed number of virtqueues without introducing an arbitrary queue limit.
Signed-off-by: Linfeng Sun <linfeng.sun.dev@gmail.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Message-ID: <20260903-fix-kernel-panic-in-vhost_iotlb_miss_pending_list-v1-1-39b8cd427978@gmail.com>
|
|
Reject VHOST_IOTLB_INVALIDATE messages with size == 0 to prevent
iova + size - 1 from underflowing to U64_MAX, which would
incorrectly delete the entire IOTLB.
Signed-off-by: xiongweimin <xiongweimin@kylinos.cn>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Message-ID: <20260716030236.124322-1-xiongwm2026@163.com>
|
|
vq->meta_iotlb[] caches the vhost_iotlb_map that backs each vring
metadata region, and iotlb_access_ok() returns early on a cache hit,
taking the hit as proof that the region has already been validated:
if (vhost_vq_meta_fetch(vq, addr, len, type))
return true;
The cache is reset on VHOST_IOTLB_UPDATE and VHOST_IOTLB_INVALIDATE, on
device IOTLB (re)initialisation and on vq reset, but not when
VHOST_SET_VRING_ADDR replaces vq->desc, vq->avail and vq->used, nor when
VHOST_SET_VRING_NUM changes the region sizes.
With a device IOTLB attached both ioctls are accepted while the vq is
live, and neither validates the addresses at ioctl time: vq_access_ok()
and vq_log_used_access_ok() return true early because the addresses are
GIOVAs, deferring validation to prefetch time. Once the cache has been
populated that deferred validation no longer runs -- vq_meta_prefetch()
hits the stale entry and returns true -- and vhost_vq_meta_fetch() keeps
translating through the old mapping as
map->addr + addr - map->start
for an address the mapping no longer covers. vhost_copy_to_user() and
vhost_copy_from_user() consume the result with __copy_to_user() and
__copy_from_user(), which do not check it either, so a subsequent used
ring update or descriptor fetch accesses memory outside the region the
IOTLB actually maps.
Reset the metadata cache whenever the vring is reconfigured, so the new
addresses are pushed back through iotlb_access_ok()'s slow path.
Fixes: f88949138058 ("vhost: introduce O(1) vq metadata cache")
Cc: stable@vger.kernel.org
Assisted-by: tencentos-corvus-ai:kimi-k3
Signed-off-by: Jun Yang <junvyyang@tencent.com>
Message-ID: <20260803014823.68623-1-juny24602@gmail.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
|
|
vhost_iotlb_add_range_ctx() only retires an old entry when the table
has a non-zero limit, has exactly reached that limit and has
VHOST_IOTLB_FLAG_RETIRE set. Non-retiring tables can keep allocating
entries after reaching their configured limit.
Existing vhost devices allocate their IOTLB with max_iotlb_entries from
vhost.c, which defaults to 2048 and is tunable by module parameter. Use
the caller-provided limit at the allocation point instead of adding a
separate default in the common IOTLB helper, and reject non-positive
values in vhost paths that can report an error.
Other vhost IOTLB users should not create zero-limit tables when entries
can be populated from userspace or guest-controlled requests. Add
caller-side max_iotlb_entries parameters for mlx5 vDPA, VDUSE and
vhost-vDPA. Reject non-positive VDUSE and vhost-vDPA values, and require
at least two entries for vdpa_sim and mlx5 vDPA paths that install
full-range mappings, since those mappings are split into two IOTLB
entries.
Handle full-range mappings in the common helper by checking that the
IOTLB can hold both split entries before inserting the first half. This
avoids returning an error after leaving a half mapping behind.
When the table is full, keep the existing retire behavior for retiring
tables and return -ENOSPC for non-retiring tables. Reuse the retired map
node instead of freeing it and allocating a replacement, so a stream of
IOTLB updates cannot keep forcing GFP_ATOMIC allocations after the table
has reached its limit. If a zero-limit IOTLB still reaches the common
helper, treat it as a configuration error and return -EINVAL.
I found this bug myself, though the patch was written with AI assistance.
Fixes: 0bbe30668d89 ("vhost: factor out IOTLB")
Assisted-by: OpenAI-Codex:GPT-5
Signed-off-by: Linfeng Sun <linfeng.sun.dev@gamil.com>
Message-ID: <AMYAtgAiKmgYcSQT5ukl-4qq.3.1781960405943.Hmail.241270009@hdu.edu.cn>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
|
|
The vhost driver has unnecessary empty module_init and
module_exit functions. Remove them. Note that if a module_init function
exists, a module_exit function must also exist; otherwise, the module
cannot be unloaded.
Signed-off-by: Ethan Nelson-Moore <enelsonmoore@gmail.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Message-ID: <20260131020010.45647-1-enelsonmoore@gmail.com>
|
|
vhost_get_avail_idx is supposed to report whether it has updated
vq->avail_idx. Instead, it returns whether all entries have been
consumed, which is usually the same. But not always - in
drivers/vhost/net.c and when mergeable buffers have been enabled, the
driver checks whether the combined entries are big enough to store an
incoming packet. If not, the driver re-enables notifications with
available entries still in the ring. The incorrect return value from
vhost_get_avail_idx propagates through vhost_enable_notify and causes
the host to livelock if the guest is not making progress, as vhost will
immediately disable notifications and retry using the available entries.
This goes back to commit d3bb267bbdcb ("vhost: cache avail index in
vhost_enable_notify()") which changed vhost_enable_notify() to compare
the freshly read avail index against vq->last_avail_idx instead of the
previously cached vq->avail_idx. Commit 7ad472397667 ("vhost: move
smp_rmb() into vhost_get_avail_idx()") then carried over the same
comparison when refactoring vhost_enable_notify() to call the unified
vhost_get_avail_idx().
The obvious fix is to make vhost_get_avail_idx do what the comment
says it does and report whether new entries have been added.
Reported-by: ShuangYu <shuangyu@yunyoo.cc>
Fixes: d3bb267bbdcb ("vhost: cache avail index in vhost_enable_notify()")
Cc: Stefan Hajnoczi <stefanha@redhat.com>
Acked-by: Jason Wang <jasowang@redhat.com>
Reviewed-by: Stefano Garzarella <sgarzare@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Message-Id: <559b04ae6ce52973c535dc47e461638b7f4c3d63.1772441455.git.mst@redhat.com>
|
|
This converts some of the visually simpler cases that have been split
over multiple lines. I only did the ones that are easy to verify the
resulting diff by having just that final GFP_KERNEL argument on the next
line.
Somebody should probably do a proper coccinelle script for this, but for
me the trivial script actually resulted in an assertion failure in the
middle of the script. I probably had made it a bit _too_ trivial.
So after fighting that far a while I decided to just do some of the
syntactically simpler cases with variations of the previous 'sed'
scripts.
The more syntactically complex multi-line cases would mostly really want
whitespace cleanup anyway.
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
This is the exact same thing as the 'alloc_obj()' version, only much
smaller because there are a lot fewer users of the *alloc_flex()
interface.
As with alloc_obj() version, this was done entirely with mindless brute
force, using the same script, except using 'flex' in the pattern rather
than 'objs*'.
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
This was done entirely with mindless brute force, using
git grep -l '\<k[vmz]*alloc_objs*(.*, GFP_KERNEL)' |
xargs sed -i 's/\(alloc_objs*(.*\), GFP_KERNEL)/\1)/'
to convert the new alloc_obj() users that had a simple GFP_KERNEL
argument to just drop that argument.
Note that due to the extreme simplicity of the scripting, any slightly
more complex cases spread over multiple lines would not be triggered:
they definitely exist, but this covers the vast bulk of the cases, and
the resulting diff is also then easier to check automatically.
For the same reason the 'flex' versions will be done as a separate
conversion.
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
This is the result of running the Coccinelle script from
scripts/coccinelle/api/kmalloc_objs.cocci. The script is designed to
avoid scalar types (which need careful case-by-case checking), and
instead replace kmalloc-family calls that allocate struct or union
object instances:
Single allocations: kmalloc(sizeof(TYPE), ...)
are replaced with: kmalloc_obj(TYPE, ...)
Array allocations: kmalloc_array(COUNT, sizeof(TYPE), ...)
are replaced with: kmalloc_objs(TYPE, COUNT, ...)
Flex array allocations: kmalloc(struct_size(PTR, FAM, COUNT), ...)
are replaced with: kmalloc_flex(*PTR, FAM, COUNT, ...)
(where TYPE may also be *VAR)
The resulting allocations no longer return "void *", instead returning
"TYPE *".
Signed-off-by: Kees Cook <kees@kernel.org>
|
|
vhost_get_user and vhost_put_user leverage __get_user and __put_user,
respectively, which were both added in 2016 by commit 6b1e6cc7855b
("vhost: new device IOTLB API"). In a heavy UDP transmit workload on a
vhost-net backed tap device, these functions showed up as ~11.6% of
samples in a flamegraph of the underlying vhost worker thread.
Quoting Linus from [1]:
Anyway, every single __get_user() call I looked at looked like
historical garbage. [...] End result: I get the feeling that we
should just do a global search-and-replace of the __get_user/
__put_user users, replace them with plain get_user/put_user instead,
and then fix up any fallout (eg the coco code).
Switch to plain get_user/put_user in vhost, which results in a slight
throughput speedup. get_user now about ~8.4% of samples in flamegraph.
Basic iperf3 test on a Intel 5416S CPU with Ubuntu 25.10 guest:
TX: taskset -c 2 iperf3 -c <rx_ip> -t 60 -p 5200 -b 0 -u -i 5
RX: taskset -c 2 iperf3 -s -p 5200 -D
Before: 6.08 Gbits/sec
After: 6.32 Gbits/sec
As to what drives the speedup, Sean's patch [2] explains:
Use the normal, checked versions for get_user() and put_user() instead of
the double-underscore versions that omit range checks, as the checked
versions are actually measurably faster on modern CPUs (12%+ on Intel,
25%+ on AMD).
The performance hit on the unchecked versions is almost entirely due to
the added LFENCE on CPUs where LFENCE is serializing (which is effectively
all modern CPUs), which was added by commit 304ec1b05031 ("x86/uaccess:
Use __uaccess_begin_nospec() and uaccess_try_nospec"). The small
optimizations done by commit b19b74bc99b1 ("x86/mm: Rework address range
check in get_user() and put_user()") likely shave a few cycles off, but
the bulk of the extra latency comes from the LFENCE.
[1] https://lore.kernel.org/all/CAHk-=wiJiDSPZJTV7z3Q-u4DfLgQTNWqUqqrwSBHp0+Dh016FA@mail.gmail.com/
[2] https://lore.kernel.org/all/20251106210206.221558-1-seanjc@google.com/
Suggested-by: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: Sean Christopherson <seanjc@google.com>
Signed-off-by: Jon Kohler <jon@nutanix.com>
Message-Id: <20251113005529.2494066-1-jon@nutanix.com>
Acked-by: Jason Wang <jasowang@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
|
|
Pull virtio updates from Michael Tsirkin:
"Just a bunch of fixes and cleanups, mostly very simple. Several
features were merged through net-next this time around"
* tag 'for_linus' of git://git.kernel.org/pub/scm/linux/kernel/git/mst/vhost:
virtio_pci: drop kernel.h
vhost: switch to arrays of feature bits
vhost/test: add test specific macro for features
virtio: clean up features qword/dword terms
vduse: add WQ_PERCPU to alloc_workqueue users
virtio_balloon: add WQ_PERCPU to alloc_workqueue users
vdpa/pds: use %pe for ERR_PTR() in event handler registration
vhost: Fix kthread worker cgroup failure handling
virtio: vdpa: Fix reference count leak in octep_sriov_enable()
vdpa/mlx5: Fix incorrect error code reporting in query_virtqueues
virtio: fix map ops comment
virtio: fix virtqueue_set_affinity() docs
virtio: standardize Returns documentation style
virtio: fix grammar in virtio_map_ops docs
virtio: fix grammar in virtio_queue_info docs
virtio: fix whitespace in virtio_config_ops
virtio: fix typo in virtio_device_ready() comment
virtio: fix kernel-doc for mapping/free_coherent functions
virtio_vdpa: fix misleading return in void function
|
|
If we fail to attach to a cgroup we are leaking the id. This adds
a new goto to free the id.
Fixes: 7d9896e9f6d0 ("vhost: Reintroduce kthread API and add mode selection")
Signed-off-by: Mike Christie <michael.christie@oracle.com>
Acked-by: Jason Wang <jasowang@redhat.com>
Reviewed-by: Chaitanya Kulkarni <kch@nvidia.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Message-Id: <20251101194358.13605-1-michael.christie@oracle.com>
|
|
When discarding descriptors with IN_ORDER, we should rewind
next_avail_head otherwise it would run out of sync with
last_avail_idx. This would cause driver to report
"id X is not a head".
Fixing this by returning the number of descriptors that is used for
each buffer via vhost_get_vq_desc_n() so caller can use the value
while discarding descriptors.
Fixes: 67a873df0c41 ("vhost: basic in order support")
Cc: stable@vger.kernel.org
Signed-off-by: Jason Wang <jasowang@redhat.com>
Acked-by: Michael S. Tsirkin <mst@redhat.com>
Link: https://patch.msgid.link/20251120022950.10117-1-jasowang@redhat.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
Commit 7918bb2d19c9 ("vhost: basic in order support") introduces
vq->nheads to store the number of batched used buffers per used elem
but it forgets to initialize the vq->nheads to NULL in
vhost_dev_init() this will cause kfree() that would try to free it
without be allocated if SET_OWNER is not called.
Reported-by: JAEHOON KIM <jhkim@linux.ibm.com>
Reported-by: Breno Leitao <leitao@debian.org>
Fixes: 45347e79b544 ("vhost: basic in order support")
Signed-off-by: Jason Wang <jasowang@redhat.com>
Message-Id: <20250729073916.80647-1-jasowang@redhat.com>
Reviewed-by: Dawid Osuchowski <dawid.osuchowski@linux.intel.com>
Tested-by: Breno Leitao <leitao@debian.org>
Reviewed-by: Stefano Garzarella <sgarzare@redhat.com>
Tested-by: Jaehoon Kim <jhkim@linux.ibm.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
|
|
Pull virtio updates from Michael Tsirkin:
- vhost can now support legacy threading if enabled in Kconfig
- vsock memory allocation strategies for large buffers have been
improved, reducing pressure on kmalloc
- vhost now supports the in-order feature. guest bits missed the merge
window.
- fixes, cleanups all over the place
* tag 'for_linus' of git://git.kernel.org/pub/scm/linux/kernel/git/mst/vhost: (30 commits)
vsock/virtio: Allocate nonlinear SKBs for handling large transmit buffers
vsock/virtio: Rename virtio_vsock_skb_rx_put()
vhost/vsock: Allocate nonlinear SKBs for handling large receive buffers
vsock/virtio: Move SKB allocation lower-bound check to callers
vsock/virtio: Rename virtio_vsock_alloc_skb()
vsock/virtio: Resize receive buffers so that each SKB fits in a 4K page
vsock/virtio: Move length check to callers of virtio_vsock_skb_rx_put()
vsock/virtio: Validate length in packet header before skb_put()
vhost/vsock: Avoid allocating arbitrarily-sized SKBs
vhost_net: basic in_order support
vhost: basic in order support
vhost: fail early when __vhost_add_used() fails
vhost: Reintroduce kthread API and add mode selection
vdpa: Fix IDR memory leak in VDUSE module exit
vdpa/mlx5: Fix release of uninitialized resources on error path
vhost-scsi: Fix check for inline_sg_cnt exceeding preallocated limit
virtio: virtio_dma_buf: fix missing parameter documentation
vhost: Fix typos
vhost: vringh: Remove unused functions
vhost: vringh: Remove unused iotlb functions
...
|
|
This patch adds basic in order support for vhost. Two optimizations
are implemented in this patch:
1) Since driver uses descriptor in order, vhost can deduce the next
avail ring head by counting the number of descriptors that has been
used in next_avail_head. This eliminate the need to access the
available ring in vhost.
2) vhost_add_used_and_singal_n() is extended to accept the number of
batched buffers per used elem. While this increases the times of
userspace memory access but it helps to reduce the chance of
used ring access of both the driver and vhost.
Vhost-net will be the first user for this.
Acked-by: Jonah Palmer <jonah.palmer@oracle.com>
Signed-off-by: Jason Wang <jasowang@redhat.com>
Message-Id: <20250714084755.11921-3-jasowang@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Tested-by: Lei Yang <leiyang@redhat.com>
|
|
This patch fails vhost_add_used_n() early when __vhost_add_used()
fails to make sure used idx is not updated with stale used ring
information.
Reported-by: Eugenio Pérez <eperezma@redhat.com>
Signed-off-by: Jason Wang <jasowang@redhat.com>
Message-Id: <20250714084755.11921-2-jasowang@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Tested-by: Lei Yang <leiyang@redhat.com>
|
|
Since commit 6e890c5d5021 ("vhost: use vhost_tasks for worker threads"),
the vhost uses vhost_task and operates as a child of the
owner thread. This is required for correct CPU usage accounting,
especially when using containers.
However, this change has caused confusion for some legacy
userspace applications, and we didn't notice until it's too late.
Unfortunately, it's too late to revert - we now have userspace
depending both on old and new behaviour :(
To address the issue, reintroduce kthread mode for vhost workers and
provide a configuration to select between kthread and task worker.
- Add 'fork_owner' parameter to vhost_dev to let users select kthread
or task mode. Default mode is task mode(VHOST_FORK_OWNER_TASK).
- Reintroduce kthread mode support:
* Bring back the original vhost_worker() implementation,
and renamed to vhost_run_work_kthread_list().
* Add cgroup support for the kthread
* Introduce struct vhost_worker_ops:
- Encapsulates create / stop / wake‑up callbacks.
- vhost_worker_create() selects the proper ops according to
inherit_owner.
- Userspace configuration interface:
* New IOCTLs:
- VHOST_SET_FORK_FROM_OWNER lets userspace select task mode
(VHOST_FORK_OWNER_TASK) or kthread mode (VHOST_FORK_OWNER_KTHREAD)
- VHOST_GET_FORK_FROM_OWNER reads the current worker mode
* Expose module parameter 'fork_from_owner_default' to allow system
administrators to configure the default mode for vhost workers
* Kconfig option CONFIG_VHOST_ENABLE_FORK_OWNER_CONTROL controls whether
these IOCTLs and the parameter are available
- The VHOST_NEW_WORKER functionality requires fork_owner to be set
to true, with validation added to ensure proper configuration
This partially reverts or improves upon:
commit 6e890c5d5021 ("vhost: use vhost_tasks for worker threads")
commit 1cdaafa1b8b4 ("vhost: replace single worker pointer with xarray")
Fixes: 6e890c5d5021 ("vhost: use vhost_tasks for worker threads"),
Signed-off-by: Cindy Lu <lulu@redhat.com>
Message-Id: <20250714071333.59794-2-lulu@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Acked-by: Jason Wang <jasowang@redhat.com>
Tested-by: Lei Yang <leiyang@redhat.com>
|
|
Fix multiple typos and improve comment clarity across vhost.c.
Spelling errors: "thead" -> "thread", "RUNNUNG" -> "RUNNING"
and "available".
Signed-off-by: Alok Tiwari <alok.a.tiwari@oracle.com>
Message-Id: <20250615173933.1610324-1-alok.a.tiwari@oracle.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Reviewed-by: Simon Horman <horms@kernel.org>
|
|
Use the extended feature type for 'acked_features' and implement
two new ioctls operation allowing the user-space to set/query an
unbounded amount of features.
The actual number of processed features is limited by VIRTIO_FEATURES_MAX
and attempts to set features above such limit fail with
EOPNOTSUPP.
Note that: the legacy ioctls implicitly truncate the negotiated
features to the lower 64 bits range and the 'acked_backend_features'
field don't need conversion, as the only negotiated feature there
is in the low 64 bit range.
Acked-by: Jason Wang <jasowang@redhat.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
|
|
Currently, the only user of vhost_log_write() is vhost-net. The 'len'
argument prevents logging of pages that are not tainted by the RX path.
Adjustments are needed since more drivers (i.e. vhost-scsi) begin using
vhost_log_write(). So far vhost-net RX path may only partially use pages
shared via the last vring descriptor. Unlike vhost-net, vhost-scsi always
logs all pages shared via vring descriptors. To accommodate this,
use (len == U64_MAX) to indicate whether the driver would log all pages of
vring descriptors, or only pages that are tainted by the driver.
In addition, removes BUG().
Suggested-by: Joao Martins <joao.m.martins@oracle.com>
Signed-off-by: Dongli Zhang <dongli.zhang@oracle.com>
Message-Id: <20250403063028.16045-5-dongli.zhang@oracle.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
|
|
Lets callers distinguish why the vhost task creation failed. No one
currently cares why it failed, so no real runtime change from this
patch, but that will not be the case for long.
Signed-off-by: Keith Busch <kbusch@kernel.org>
Message-ID: <20250227230631.303431-2-kbusch@meta.com>
Reviewed-by: Mike Christie <michael.christie@oracle.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
|
|
All callers of vhost_get_avail_idx() use smp_rmb() to
order the available ring entry read and avail_idx read.
Make vhost_get_avail_idx() call smp_rmb() itself whenever the avail_idx
is accessed. This way, the callers don't need to worry about the memory
barrier. As a side benefit, we also validate the index on all paths now,
which will hopefully help prevent/catch earlier future bugs.
Note that current code is inconsistent in how the errors are handled.
They are treated as an empty ring in some places, but as non-empty
ring in other places. This patch doesn't attempt to change the existing
behaviour.
No functional change intended.
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Reviewed-by: Gavin Shan <gshan@redhat.com>
Acked-by: Will Deacon <will@kernel.org>
Message-Id: <20240429232748.642356-1-gshan@redhat.com>
|
|
Instead of lingering until the device is closed, this has us handle
SIGKILL by:
1. marking the worker as killed so we no longer try to use it with
new virtqueues and new flush operations.
2. setting the virtqueue to worker mapping so no new works are queued.
3. running all the exiting works.
Suggested-by: Edward Adam Davis <eadavis@qq.com>
Reported-and-tested-by: syzbot+98edc2df894917b3431f@syzkaller.appspotmail.com
Message-Id: <tencent_546DA49414E876EEBECF2C78D26D242EE50A@qq.com>
Signed-off-by: Mike Christie <michael.christie@oracle.com>
Message-Id: <20240316004707.45557-9-michael.christie@oracle.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
|
|
In the next patches where the worker can be killed while in use, we
need to be able to take the worker mutex and kill queued works for
new IO and flushes, and set some new flags to prevent new
__vhost_vq_attach_worker calls from swapping in/out killed workers.
If we are holding the worker mutex during a flush and the flush's work
is still in the queue, the worker code that will handle the SIGKILL
cleanup won't be able to take the mutex and perform it's cleanup. So
this patch has us drop the worker mutex while waiting for the flush
to complete.
Signed-off-by: Mike Christie <michael.christie@oracle.com>
Message-Id: <20240316004707.45557-8-michael.christie@oracle.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
|
|
__vhost_vq_attach_worker uses the vhost_dev mutex to serialize the
swapping of a virtqueue's worker. This was done for simplicity because
we are already holding that mutex.
In the next patches where the worker can be killed while in use, we need
finer grained locking because some drivers will hold the vhost_dev mutex
while flushing. However in the SIGKILL handler in the next patches, we
will need to be able to swap workers (set current one to NULL), kill
queued works and stop new flushes while flushes are in progress.
To prepare us, this has us use the virtqueue mutex for swapping workers
instead of the vhost_dev one.
Signed-off-by: Mike Christie <michael.christie@oracle.com>
Message-Id: <20240316004707.45557-7-michael.christie@oracle.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
|
|
vhost_vq_flush is no longer used so remove it.
Signed-off-by: Mike Christie <michael.christie@oracle.com>
Message-Id: <20240316004707.45557-5-michael.christie@oracle.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
|
|
Guest moved avail idx not used idx when we need to print log if
'(vq->avail_idx - last_avail_idx) > vq->num', so fix it.
Signed-off-by: Xianting Tian <xianting.tian@linux.alibaba.com>
Message-Id: <20240311082109.46773-1-xianting.tian@linux.alibaba.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
|
|
A smp_rmb() has been missed in vhost_enable_notify(), inspired by
Will. Otherwise, it's not ensured the available ring entries pushed
by guest can be observed by vhost in time, leading to stale available
ring entries fetched by vhost in vhost_get_vq_desc(), as reported by
Yihuang Yu on NVidia's grace-hopper (ARM64) platform.
/home/gavin/sandbox/qemu.main/build/qemu-system-aarch64 \
-accel kvm -machine virt,gic-version=host -cpu host \
-smp maxcpus=1,cpus=1,sockets=1,clusters=1,cores=1,threads=1 \
-m 4096M,slots=16,maxmem=64G \
-object memory-backend-ram,id=mem0,size=4096M \
: \
-netdev tap,id=vnet0,vhost=true \
-device virtio-net-pci,bus=pcie.8,netdev=vnet0,mac=52:54:00:f1:26:b0
:
guest# netperf -H 10.26.1.81 -l 60 -C -c -t UDP_STREAM
virtio_net virtio0: output.0:id 100 is not a head!
Add the missed smp_rmb() in vhost_enable_notify(). When it returns true,
it means there's still pending tx buffers. Since it might read indices,
so it still can bypass the smp_rmb() in vhost_get_vq_desc(). Note that
it should be safe until vq->avail_idx is changed by commit d3bb267bbdcb
("vhost: cache avail index in vhost_enable_notify()").
Fixes: d3bb267bbdcb ("vhost: cache avail index in vhost_enable_notify()")
Cc: <stable@kernel.org> # v5.18+
Reported-by: Yihuang Yu <yihyu@redhat.com>
Suggested-by: Will Deacon <will@kernel.org>
Signed-off-by: Gavin Shan <gshan@redhat.com>
Acked-by: Jason Wang <jasowang@redhat.com>
Message-Id: <20240328002149.1141302-3-gshan@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Reviewed-by: Stefano Garzarella <sgarzare@redhat.com>
|
|
A smp_rmb() has been missed in vhost_vq_avail_empty(), spotted by
Will. Otherwise, it's not ensured the available ring entries pushed
by guest can be observed by vhost in time, leading to stale available
ring entries fetched by vhost in vhost_get_vq_desc(), as reported by
Yihuang Yu on NVidia's grace-hopper (ARM64) platform.
/home/gavin/sandbox/qemu.main/build/qemu-system-aarch64 \
-accel kvm -machine virt,gic-version=host -cpu host \
-smp maxcpus=1,cpus=1,sockets=1,clusters=1,cores=1,threads=1 \
-m 4096M,slots=16,maxmem=64G \
-object memory-backend-ram,id=mem0,size=4096M \
: \
-netdev tap,id=vnet0,vhost=true \
-device virtio-net-pci,bus=pcie.8,netdev=vnet0,mac=52:54:00:f1:26:b0
:
guest# netperf -H 10.26.1.81 -l 60 -C -c -t UDP_STREAM
virtio_net virtio0: output.0:id 100 is not a head!
Add the missed smp_rmb() in vhost_vq_avail_empty(). When tx_can_batch()
returns true, it means there's still pending tx buffers. Since it might
read indices, so it still can bypass the smp_rmb() in vhost_get_vq_desc().
Note that it should be safe until vq->avail_idx is changed by commit
275bf960ac697 ("vhost: better detection of available buffers").
Fixes: 275bf960ac69 ("vhost: better detection of available buffers")
Cc: <stable@kernel.org> # v4.11+
Reported-by: Yihuang Yu <yihyu@redhat.com>
Suggested-by: Will Deacon <will@kernel.org>
Signed-off-by: Gavin Shan <gshan@redhat.com>
Acked-by: Jason Wang <jasowang@redhat.com>
Message-Id: <20240328002149.1141302-2-gshan@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Reviewed-by: Stefano Garzarella <sgarzare@redhat.com>
|
|
Ever since the eventfd type was introduced back in 2007 in commit
e1ad7468c77d ("signal/timer/event: eventfd core") the eventfd_signal()
function only ever passed 1 as a value for @n. There's no point in
keeping that additional argument.
Link: https://lore.kernel.org/r/20231122-vfs-eventfd-signal-v2-2-bd549b14ce0c@kernel.org
Acked-by: Xu Yilun <yilun.xu@intel.com>
Acked-by: Andrew Donnellan <ajd@linux.ibm.com> # ocxl
Acked-by: Eric Farman <farman@linux.ibm.com> # s390
Reviewed-by: Jan Kara <jack@suse.cz>
Reviewed-by: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Christian Brauner <brauner@kernel.org>
|
|
Commit e2ae38cf3d91 ("vhost: fix hung thread due to erroneous iotlb
entries") Forbade vhost iotlb msg with null size to prevent entries
with size = start = 0 and last = ULONG_MAX to end up in the iotlb.
Then commit 95932ab2ea07 ("vhost: allow batching hint without size")
only applied the check for VHOST_IOTLB_UPDATE and VHOST_IOTLB_INVALIDATE
message types to fix a regression observed with batching hit.
Still, the introduction of that check introduced a regression for
some users attempting to invalidate the whole ULONG_MAX range by
setting the size to 0. This is the case with qemu/smmuv3/vhost
integration which does not work anymore. It Looks safe to partially
revert the original commit and allow VHOST_IOTLB_INVALIDATE messages
with null size. vhost_iotlb_del_range() will compute a correct end
iova. Same for vhost_vdpa_iotlb_unmap().
Signed-off-by: Eric Auger <eric.auger@redhat.com>
Fixes: e2ae38cf3d91 ("vhost: fix hung thread due to erroneous iotlb entries")
Cc: stable@vger.kernel.org # v5.17+
Acked-by: Jason Wang <jasowang@redhat.com>
Message-Id: <20230927140544.205088-1-eric.auger@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
|
|
This patch drops the requirement that we can only switch workers if work
has not been queued by using RCU for the vq based queueing paths and a
mutex for the device wide flush.
We can also use this to support SIGKILL properly in the future where we
should exit almost immediately after getting that signal. With this
patch, when get_signal returns true, we can set the vq->worker to NULL
and do a synchronize_rcu to prevent new work from being queued to the
vhost_task that has been killed.
Signed-off-by: Mike Christie <michael.christie@oracle.com>
Message-Id: <20230626232307.97930-18-michael.christie@oracle.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
|
|
For vhost-scsi with 3 vqs or more and a workload that tries to use
them in parallel like:
fio --filename=/dev/sdb --direct=1 --rw=randrw --bs=4k \
--ioengine=libaio --iodepth=128 --numjobs=3
the single vhost worker thread will become a bottlneck and we are stuck
at around 500K IOPs no matter how many jobs, virtqueues, and CPUs are
used.
To better utilize virtqueues and available CPUs, this patch allows
userspace to create workers and bind them to vqs. You can have N workers
per dev and also share N workers with M vqs on that dev.
This patch adds the interface related code and the next patch will hook
vhost-scsi into it. The patches do not try to hook net and vsock into
the interface because:
1. multiple workers don't seem to help vsock. The problem is that with
only 2 virtqueues we never fully use the existing worker when doing
bidirectional tests. This seems to match vhost-scsi where we don't see
the worker as a bottleneck until 3 virtqueues are used.
2. net already has a way to use multiple workers.
Signed-off-by: Mike Christie <michael.christie@oracle.com>
Message-Id: <20230626232307.97930-16-michael.christie@oracle.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
|
|
The next patch allows userspace to create multiple workers per device,
so this patch replaces the vhost_worker pointer with an xarray so we
can store mupltiple workers and look them up.
Signed-off-by: Mike Christie <michael.christie@oracle.com>
Message-Id: <20230626232307.97930-15-michael.christie@oracle.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
|
|
The next patches add new vhost worker ioctls which will need to get a
vhost_virtqueue from a userspace struct which specifies the vq's index.
This moves the vhost_vring_ioctl code to do this to a helper so it can
be shared.
Signed-off-by: Mike Christie <michael.christie@oracle.com>
Message-Id: <20230626232307.97930-14-michael.christie@oracle.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
|
|
vhost_work_queue is no longer used. Each driver is using the poll or vq
based queueing, so remove vhost_work_queue.
Signed-off-by: Mike Christie <michael.christie@oracle.com>
Message-Id: <20230626232307.97930-13-michael.christie@oracle.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
|
|
This has the drivers pass in their poll to vq mapping and then converts
the core poll code to use the vq based helpers. In the next patches we
will allow vqs to be handled by different workers, so to allow drivers
to execute operations like queue, stop, flush, etc on specific polls/vqs
we need to know the mappings.
Signed-off-by: Mike Christie <michael.christie@oracle.com>
Message-Id: <20230626232307.97930-8-michael.christie@oracle.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
|
|
This patch has the core work flush function take a worker. When we
support multiple workers we can then flush each worker during device
removal, stoppage, etc. It also adds a helper to flush specific
virtqueues, so vhost-scsi can flush IO vqs from it's ctl vq.
Signed-off-by: Mike Christie <michael.christie@oracle.com>
Message-Id: <20230626232307.97930-7-michael.christie@oracle.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
|
|
This patch has the core work queueing function take a worker for when we
support multiple workers. It also adds a helper that takes a vq during
queueing so modules can control which vq/worker to queue work on.
This temp leaves vhost_work_queue. It will be removed when the drivers
are converted in the next patches.
Signed-off-by: Mike Christie <michael.christie@oracle.com>
Message-Id: <20230626232307.97930-6-michael.christie@oracle.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
|
|
In the next patches each vq might have different workers so one could
have work but others do not. For net, we only want to check specific vqs,
so this adds a helper to check if a vq has work pending and converts
vhost-net to use it.
Signed-off-by: Mike Christie <michael.christie@oracle.com>
Acked-by: Jason Wang <jasowang@redhat.com>
Message-Id: <20230626232307.97930-5-michael.christie@oracle.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
|
|
This patchset allows userspace to map vqs to different workers. This
patch adds a worker pointer to the vq so in later patches in this set
we can queue/flush specific vqs and their workers.
Signed-off-by: Mike Christie <michael.christie@oracle.com>
Message-Id: <20230626232307.97930-4-michael.christie@oracle.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
|
|
This patchset allows us to allocate multiple workers, so this has us
move from the vhost_worker that's embedded in the vhost_dev to
dynamically allocating it.
Signed-off-by: Mike Christie <michael.christie@oracle.com>
Message-Id: <20230626232307.97930-3-michael.christie@oracle.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
|
|
vsock can start queueing work after VHOST_VSOCK_SET_GUEST_CID, so
after we have called vhost_worker_create it can be calling
vhost_work_queue and trying to access the vhost worker/task. If
vhost_dev_alloc_iovecs fails, then vhost_worker_free could free
the worker/task from under vsock.
This moves vhost_worker_create to the end of vhost_dev_set_owner
where we know we can no longer fail in that path. If it fails
after the VHOST_SET_OWNER and userspace closes the device, then
the normal vsock release handling will do the right thing.
Signed-off-by: Mike Christie <michael.christie@oracle.com>
Message-Id: <20230626232307.97930-2-michael.christie@oracle.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
|
|
Use the right structs for PACKED or split vqs when setting and
getting the vring base.
Fixes: 4c8cf31885f6 ("vhost: introduce vDPA-based backend")
Signed-off-by: Shannon Nelson <shannon.nelson@amd.com>
Message-Id: <20230424225031.18947-3-shannon.nelson@amd.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Acked-by: Jason Wang <jasowang@redhat.com>
|
|
We can race where we have added work to the work_list, but
vhost_task_fn has passed that check but not yet set us into
TASK_INTERRUPTIBLE. wake_up_process will see us in TASK_RUNNING and
just return.
This bug was intoduced in commit f9010dbdce91 ("fork, vhost: Use
CLONE_THREAD to fix freezer/ps regression") when I moved the setting
of TASK_INTERRUPTIBLE to simplfy the code and avoid get_signal from
logging warnings about being in the wrong state. This moves the setting
of TASK_INTERRUPTIBLE back to before we test if we need to stop the
task to avoid a possible race there as well. We then have vhost_worker
set TASK_RUNNING if it finds work similar to before.
Fixes: f9010dbdce91 ("fork, vhost: Use CLONE_THREAD to fix freezer/ps regression")
Signed-off-by: Mike Christie <michael.christie@oracle.com>
Message-Id: <20230607192338.6041-3-michael.christie@oracle.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
|
|
If userspace does VHOST_VSOCK_SET_GUEST_CID before VHOST_SET_OWNER we
can race where:
1. thread0 calls vhost_transport_send_pkt -> vhost_work_queue
2. thread1 does VHOST_SET_OWNER which calls vhost_worker_create.
3. vhost_worker_create will set the dev->worker pointer before setting
the worker->vtsk pointer.
4. thread0's vhost_work_queue will see the dev->worker pointer is
set and try to call vhost_task_wake using not yet set worker->vtsk
pointer.
5. We then crash since vtsk is NULL.
Before commit 6e890c5d5021 ("vhost: use vhost_tasks for worker
threads"), we only had the worker pointer so we could just check it to
see if VHOST_SET_OWNER has been done. After that commit we have the
vhost_worker and vhost_task pointer, so we can now hit the bug above.
This patch embeds the vhost_worker in the vhost_dev and moves the work
list initialization back to vhost_dev_init, so we can just check the
worker.vtsk pointer to check if VHOST_SET_OWNER has been done like
before.
Fixes: 6e890c5d5021 ("vhost: use vhost_tasks for worker threads")
Signed-off-by: Mike Christie <michael.christie@oracle.com>
Message-Id: <20230607192338.6041-2-michael.christie@oracle.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Reported-by: syzbot+d0d442c22fa8db45ff0e@syzkaller.appspotmail.com
Reviewed-by: Stefano Garzarella <sgarzare@redhat.com>
|
|
Use kzalloc() to allocate new zeroed out msg node instead of
memsetting a node allocated with kmalloc().
Signed-off-by: Prathu Baronia <prathubaronia2011@gmail.com>
Message-Id: <20230522085019.42914-1-prathubaronia2011@gmail.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Reviewed-by: Stefano Garzarella <sgarzare@redhat.com>
|
