vfio_unpin_pages will hold a read semaphore however it is already hold
in the same thread by vfio ioctl. It will cause below warning:
[ 5102.127454] ============================================
[ 5102.133379] WARNING: possible recursive locking detected
[ 5102.139304] 4.12.0-rc4+ #3 Not tainted
[ 5102.143483] --------------------------------------------
[ 5102.149407] qemu-system-x86/1620 is trying to acquire lock:
[ 5102.155624] (&container->group_lock){++++++}, at: [<ffffffff817768c6>] vfio_unpin_pages+0x96/0xf0
[ 5102.165626]
but task is already holding lock:
[ 5102.172134] (&container->group_lock){++++++}, at: [<ffffffff8177728f>] vfio_fops_unl_ioctl+0x5f/0x280
[ 5102.182522]
other info that might help us debug this:
[ 5102.189806] Possible unsafe locking scenario:
[ 5102.196411] CPU0
[ 5102.199136] ----
[ 5102.201861] lock(&container->group_lock);
[ 5102.206527] lock(&container->group_lock);
[ 5102.211191]
*** DEADLOCK ***
[ 5102.217796] May be due to missing lock nesting notation
[ 5102.225370] 3 locks held by qemu-system-x86/1620:
[ 5102.230618] #0: (&container->group_lock){++++++}, at: [<ffffffff8177728f>] vfio_fops_unl_ioctl+0x5f/0x280
[ 5102.241482] #1: (&(&iommu->notifier)->rwsem){++++..}, at: [<ffffffff810de775>] __blocking_notifier_call_chain+0x35/0x70
[ 5102.253713] #2: (&vgpu->vdev.cache_lock){+.+...}, at: [<ffffffff8157b007>] intel_vgpu_iommu_notifier+0x77/0x120
[ 5102.265163]
stack backtrace:
[ 5102.270022] CPU: 5 PID: 1620 Comm: qemu-system-x86 Not tainted 4.12.0-rc4+ #3
[ 5102.277991] Hardware name: Intel Corporation S1200RP/S1200RP, BIOS S1200RP.86B.03.01.APER.061220151418 06/12/2015
[ 5102.289445] Call Trace:
[ 5102.292175] dump_stack+0x85/0xc7
[ 5102.295871] validate_chain.isra.21+0x9da/0xaf0
[ 5102.300925] __lock_acquire+0x405/0x820
[ 5102.305202] lock_acquire+0xc7/0x220
[ 5102.309191] ? vfio_unpin_pages+0x96/0xf0
[ 5102.313666] down_read+0x2b/0x50
[ 5102.317259] ? vfio_unpin_pages+0x96/0xf0
[ 5102.321732] vfio_unpin_pages+0x96/0xf0
[ 5102.326024] intel_vgpu_iommu_notifier+0xe5/0x120
[ 5102.331283] notifier_call_chain+0x4a/0x70
[ 5102.335851] __blocking_notifier_call_chain+0x4d/0x70
[ 5102.341490] blocking_notifier_call_chain+0x16/0x20
[ 5102.346935] vfio_iommu_type1_ioctl+0x87b/0x920
[ 5102.351994] vfio_fops_unl_ioctl+0x81/0x280
[ 5102.356660] ? __fget+0xf0/0x210
[ 5102.360261] do_vfs_ioctl+0x93/0x6a0
[ 5102.364247] ? __fget+0x111/0x210
[ 5102.367942] SyS_ioctl+0x41/0x70
[ 5102.371542] entry_SYSCALL_64_fastpath+0x1f/0xbe
put the vfio_unpin_pages in a workqueue can fix this.
v2:
- use for style instead of do{}while(1). (Zhenyu)
v3:
- rename gvt_cache_mark to gvt_cache_mark_remove. (Zhenyu)
Fixes: 659643f7d814 ("drm/i915/gvt/kvmgt: add vfio/mdev support to KVMGT")
Signed-off-by: Chuanxiao Dong <chuanxiao.dong@xxxxxxxxx>
Cc: Zhenyu Wang <zhenyuw@xxxxxxxxxxxxxxx>
Cc: stable@xxxxxxxxxxxxxxx # v4.10+
---
drivers/gpu/drm/i915/gvt/gvt.h | 3 +++
drivers/gpu/drm/i915/gvt/kvmgt.c | 55 ++++++++++++++++++++++++++++++++--------
2 files changed, 48 insertions(+), 10 deletions(-)
diff --git a/drivers/gpu/drm/i915/gvt/gvt.h b/drivers/gpu/drm/i915/gvt/gvt.h
index 3a74e79..28d817e 100644
--- a/drivers/gpu/drm/i915/gvt/gvt.h
+++ b/drivers/gpu/drm/i915/gvt/gvt.h
@@ -182,6 +182,9 @@ struct intel_vgpu {
struct kvm *kvm;
struct work_struct release_work;
atomic_t released;
+ struct work_struct unpin_work;
+ spinlock_t unpin_lock; /* To protect unpin_list */
+ struct list_head unpin_list;
} vdev;
#endif
};
diff --git a/drivers/gpu/drm/i915/gvt/kvmgt.c b/drivers/gpu/drm/i915/gvt/kvmgt.c
index 1ae0b40..56a8d6a 100644
--- a/drivers/gpu/drm/i915/gvt/kvmgt.c
+++ b/drivers/gpu/drm/i915/gvt/kvmgt.c
@@ -78,6 +78,7 @@ struct gvt_dma {
struct rb_node node;
gfn_t gfn;
unsigned long iova;
+ struct list_head list;
};
static inline bool handle_valid(unsigned long handle)
@@ -166,6 +167,7 @@ static void gvt_cache_add(struct intel_vgpu *vgpu, gfn_t gfn,
new->gfn = gfn;
new->iova = iova;
+ INIT_LIST_HEAD(&new->list);
mutex_lock(&vgpu->vdev.cache_lock);
while (*link) {
@@ -197,26 +199,52 @@ static void __gvt_cache_remove_entry(struct intel_vgpu *vgpu,
kfree(entry);
}
-static void gvt_cache_remove(struct intel_vgpu *vgpu, gfn_t gfn)
+static void intel_vgpu_unpin_work(struct work_struct *work)
{
+ struct intel_vgpu *vgpu = container_of(work, struct intel_vgpu,
+ vdev.unpin_work);
struct device *dev = mdev_dev(vgpu->vdev.mdev);
struct gvt_dma *this;
- unsigned long g1;
- int rc;
+ unsigned long gfn;
+
+ for (;;) {
+ spin_lock(&vgpu->vdev.unpin_lock);
+ if (list_empty(&vgpu->vdev.unpin_list)) {
+ spin_unlock(&vgpu->vdev.unpin_lock);
+ break;
+ }
+ this = list_first_entry(&vgpu->vdev.unpin_list,
+ struct gvt_dma, list);
+ list_del(&this->list);
+ spin_unlock(&vgpu->vdev.unpin_lock);
+
+ gfn = this->gfn;
+ vfio_unpin_pages(dev, &gfn, 1);
+ kfree(this);
+ }
+}
+
+static bool gvt_cache_mark_remove(struct intel_vgpu *vgpu, gfn_t gfn)
+{
+ struct gvt_dma *this;
mutex_lock(&vgpu->vdev.cache_lock);
this = __gvt_cache_find(vgpu, gfn);
if (!this) {
mutex_unlock(&vgpu->vdev.cache_lock);
- return;
+ return false;
}
-
- g1 = gfn;
gvt_dma_unmap_iova(vgpu, this->iova);
- rc = vfio_unpin_pages(dev, &g1, 1);
- WARN_ON(rc != 1);
- __gvt_cache_remove_entry(vgpu, this);
+ /* remove this from rb tree */
+ rb_erase(&this->node, &vgpu->vdev.cache);
mutex_unlock(&vgpu->vdev.cache_lock);
+
+ /* put this to the unpin_list */
+ spin_lock(&vgpu->vdev.unpin_lock);
+ list_move_tail(&this->list, &vgpu->vdev.unpin_list);
+ spin_unlock(&vgpu->vdev.unpin_lock);
+
+ return true;
}
static void gvt_cache_init(struct intel_vgpu *vgpu)
@@ -453,6 +481,9 @@ static int intel_vgpu_create(struct kobject *kobj, struct mdev_device *mdev)
}
INIT_WORK(&vgpu->vdev.release_work, intel_vgpu_release_work);
+ INIT_WORK(&vgpu->vdev.unpin_work, intel_vgpu_unpin_work);
+ spin_lock_init(&vgpu->vdev.unpin_lock);
+ INIT_LIST_HEAD(&vgpu->vdev.unpin_list);
vgpu->vdev.mdev = mdev;
mdev_set_drvdata(mdev, vgpu);
@@ -482,6 +513,7 @@ static int intel_vgpu_iommu_notifier(struct notifier_block *nb,
struct intel_vgpu *vgpu = container_of(nb,
struct intel_vgpu,
vdev.iommu_notifier);
+ bool sched_unmap = false;
if (action == VFIO_IOMMU_NOTIFY_DMA_UNMAP) {
struct vfio_iommu_type1_dma_unmap *unmap = data;
@@ -491,7 +523,10 @@ static int intel_vgpu_iommu_notifier(struct notifier_block *nb,
end_gfn = gfn + unmap->size / PAGE_SIZE;
while (gfn < end_gfn)
- gvt_cache_remove(vgpu, gfn++);
+ sched_unmap |= gvt_cache_mark_remove(vgpu, gfn++);
+
+ if (sched_unmap)
+ schedule_work(&vgpu->vdev.unpin_work);
}
return NOTIFY_OK;
--
2.7.4