VHOST_SET_VRING_CALL releases the previous call eventfd inside vhost_vring_ioctl() -- it swaps the new context into vq->call_ctx.ctx and then eventfd_ctx_put()s the old one, which is a synchronous kfree(). The parent vdpa device is only told about the change afterwards, when vhost_vdpa_vring_ioctl() reaches ops->set_vq_cb(). Parent drivers cache the pointer handed to them in vdpa_callback::trigger and do not take a reference on it, so throughout that window the parent holds a dangling eventfd_ctx and may signal it. The documentation added with the field describes what signalling it means but says nothing about how long it stays valid. This is the same hazard that "vhost_vdpa: assign irq bypass producer token correctly" addressed for the irq bypass producer token, by moving vhost_vdpa_unsetup_vq_irq() ahead of the vhost_vring_ioctl() call. The producer token was only one of the two consumers of that pointer; the one the parent keeps via ->set_vq_cb() was left behind the free. With VDUSE the window is directly reachable from userspace, because the device emulation daemon can inject an interrupt at any time from a different fd, and neither side shares a lock with the other: VDUSE takes vq->irq_lock, vhost takes vhost_dev.mutex + vq->mutex. BUG: KASAN: slab-use-after-free in _raw_spin_lock_irqsave+0x76/0xe0 Write of size 4 at addr ffff8881084e8788 by task vduse_uaf/2987 _raw_spin_lock_irqsave+0x76/0xe0 eventfd_signal_mask+0x69/0x120 vduse_dev_ioctl+0x337/0x1a60 <- vduse_vq_signal_irqfd(), inlined __x64_sys_ioctl+0x120/0x170 <- VDUSE_VQ_INJECT_IRQ Allocated by task 2986: do_eventfd+0x50/0x200 __x64_sys_eventfd2+0x2e/0x40 kmalloc-64, freed 64-byte region [ffff8881084e8780, ffff8881084e87c0) One thread loops VHOST_SET_VRING_CALL on /dev/vhost-vdpa-N with a fresh eventfd and then unbinds it, while another loops VDUSE_VQ_INJECT_IRQ on /dev/vduse/. This reproduces in 5 out of 5 ten-second runs on v7.1.6 and 3 out of 3 on v7.2-rc6. With the patch there are no reports in 3 out of 3 runs on either, while the same workload still gets ~30000 interrupts per run delivered into live eventfds, so the path is still being exercised. Tell the parent to drop the callback before vhost_vring_ioctl() can free the eventfd, mirroring what is already done for the bypass producer, and restore it if the ioctl fails -- on failure the swap never happened, the old context is still installed, and leaving the parent without a callback would silently drop that vq's interrupts. Clearing the callback only stops a parent from starting to use the context; a handler that already loaded it can still be running. For parents with a real interrupt, wait for it with synchronize_irq() on the vq's irq. VDUSE does not implement get_vq_irq, so that case stays covered by the teardown above. Fixes: 5e68470f4e80 ("vdpa: Add eventfd for the vdpa callback") Signed-off-by: Yu Zhang --- v2, addressing the review on v1: - synchronize_irq(): valid, fixed. Clearing the callback only stops a parent from starting to use the context; a handler that already loaded it can still be running. The teardown now ends with a synchronize_irq() on ops->get_vq_irq() where the parent has one. VDUSE does not implement get_vq_irq, so it is a no-op there and that case stays covered by the teardown alone. - torn read of the callback struct yielding a NULL cb.private: I do not think it applies here. cb.trigger has exactly one consumer -- apart from vhost/vdpa.c assigning it, the only reads are in drivers/vdpa/vdpa_user/vduse_dev.c -- and VDUSE serialises both sides: vduse_vdpa_set_vq_cb() takes vq->irq_lock around the stores, and vduse_vq_irq_inject() and vduse_vq_signal_irqfd() take the same lock before using the fields, so VDUSE cannot observe a torn update. For parents that do a bare struct copy (vp_vdpa_set_vq_cb() is vring->cb = *cb) the hazard is real, but it is not introduced here: this function already performs the identical three NULL stores through ops->set_vq_cb() in the existing else branch below, which userspace reaches with VHOST_FILE_UNBIND. Defining the store order and the matching loads has to happen on the parent side, which this caller cannot do; happy to send that separately if you want it. - retested on v7.1.6 with KASAN: no reports in 5 out of 5 runs with the patch, 3 out of 3 runs report the UAF without it, and the workload is not slowed down -- 16k-19k call fd swaps and 1.5M-1.9M injections per 30s run, against 10k/0.9M measured for v1, with 48k-89k interrupts still delivered into live eventfds. - rebased on the config_ctx fixes now in mainline. drivers/vhost/vdpa.c | 35 ++++++++++++++++++++++++++++++++++- 1 file changed, 34 insertions(+), 1 deletion(-) diff --git a/drivers/vhost/vdpa.c b/drivers/vhost/vdpa.c index a317867..e1def23 100644 --- a/drivers/vhost/vdpa.c +++ b/drivers/vhost/vdpa.c @@ -767,13 +767,46 @@ static long vhost_vdpa_vring_ioctl(struct vhost_vdpa *v, unsigned int cmd, if (ops->get_status(vdpa) & VIRTIO_CONFIG_S_DRIVER_OK) vhost_vdpa_unsetup_vq_irq(v, idx); + /* + * The parent caches call_ctx.ctx in cb.trigger without + * holding a reference, so it has to stop using it + * before vhost_vring_ioctl() drops the last one. + */ + cb.callback = NULL; + cb.private = NULL; + cb.trigger = NULL; + ops->set_vq_cb(vdpa, idx, &cb); + + /* + * A parent interrupt handler that loaded the callback + * before that store can still be running, so wait for + * it to finish with the context too. + */ + if (ops->get_vq_irq) { + int irq = ops->get_vq_irq(vdpa, idx); + + if (irq >= 0) + synchronize_irq(irq); + } } break; } r = vhost_vring_ioctl(&v->vdev, cmd, argp); - if (r) + if (r) { + /* + * A failure here means the swap never happened and the old + * context is still installed, so give the parent back the + * callback that was torn down above. + */ + if (cmd == VHOST_SET_VRING_CALL && vq->call_ctx.ctx) { + cb.callback = vhost_vdpa_virtqueue_cb; + cb.private = vq; + cb.trigger = vq->call_ctx.ctx; + ops->set_vq_cb(vdpa, idx, &cb); + } return r; + } switch (cmd) { case VHOST_SET_VRING_ADDR: -- 2.43.0