[PATCH V8 0/7] fixes for virtual address update

[PATCH V8 0/7] fixes for virtual address update

[Steve Sistare]

Fix bugs in the interfaces that allow the underlying memory object of an
iova range to be mapped in a new address space.  They allow userland to
indefinitely block vfio mediated device kernel threads, and do not
propagate the locked_vm count to a new mm.  Also fix a pre-existing bug
that allows locked_vm underflow.

The fixes impose restrictions that eliminate waiting conditions, so
revert the dead code:
  commit 898b9eaeb3fe ("vfio/type1: block on invalid vaddr")
  commit 487ace134053 ("vfio/type1: implement notify callback")
  commit ec5e32940cc9 ("vfio: iommu driver notify callback")

Changes in V2 (thanks Alex):
  * do not allow group attach while vaddrs are invalid
  * add patches to delete dead code
  * add WARN_ON for never-should-happen conditions
  * check for changed mm in unmap.
  * check for vfio_lock_acct failure in remap

Changes in V3 (ditto!):
  * return errno at WARN_ON sites, and make it unique
  * correctly check for dma task mm change
  * change dma owner to current when vaddr is updated
  * add Fixes to commit messages
  * refactored new code in vfio_dma_do_map

Changes in V4:
  * misc cosmetic changes

Changes in V5 (thanks Jason and Kevin):
  * grab mm and use it for locked_vm accounting
  * separate patches for underflow and restoring locked_vm
  * account for reserved pages
  * improve error messages

Changes in V6:
  * drop "count reserved pages" patch
  * add "track locked_vm" patch
  * grab current->mm not group_leader->mm
  * simplify vfio_change_dma_owner
  * fix commit messages

Changes in v7:
  * compare current->mm not group_leader->mm (missed one)
  * misc cosmetic changes

Changes in v8:
  * updated group_leader comment in vfio_dma_do_map
  * delete async arg from mm_lock_acct
  * pass async=false to vfio_lock_acct in vfio_pin_page_external
  * locked_vm becomes size_t
  * improved commit message in "restore locked_vm"
  * simplified flow in vfio_change_dma_owner
  * rebase to v6.2-rc6

Steve Sistare (7):
  vfio/type1: exclude mdevs from VFIO_UPDATE_VADDR
  vfio/type1: prevent underflow of locked_vm via exec()
  vfio/type1: track locked_vm per dma
  vfio/type1: restore locked_vm
  vfio/type1: revert "block on invalid vaddr"
  vfio/type1: revert "implement notify callback"
  vfio: revert "iommu driver notify callback"

 drivers/vfio/container.c        |   5 -
 drivers/vfio/vfio.h             |   7 --
 drivers/vfio/vfio_iommu_type1.c | 248 ++++++++++++++++++----------------------
 include/uapi/linux/vfio.h       |  15 ++-
 4 files changed, 120 insertions(+), 155 deletions(-)

-- 
1.8.3.1

[PATCH V8 1/7] vfio/type1: exclude mdevs from VFIO_UPDATE_VADDR

[Steve Sistare]

Disable the VFIO_UPDATE_VADDR capability if mediated devices are present.
Their kernel threads could be blocked indefinitely by a misbehaving
userland while trying to pin/unpin pages while vaddrs are being updated.

Do not allow groups to be added to the container while vaddr's are invalid,
so we never need to block user threads from pinning, and can delete the
vaddr-waiting code in a subsequent patch.

Fixes: c3cbab24db38 ("vfio/type1: implement interfaces to update vaddr")
Cc: stable@vger.kernel.org
Signed-off-by: Steve Sistare <steven.sistare@oracle.com>
Reviewed-by: Kevin Tian <kevin.tian@intel.com>
Reviewed-by: Jason Gunthorpe <jgg@nvidia.com>
---
 drivers/vfio/vfio_iommu_type1.c | 44 +++++++++++++++++++++++++++++++++++++++--
 include/uapi/linux/vfio.h       | 15 ++++++++------
 2 files changed, 51 insertions(+), 8 deletions(-)

diff --git a/drivers/vfio/vfio_iommu_type1.c b/drivers/vfio/vfio_iommu_type1.c
index 2209372..e8902e7 100644
--- a/drivers/vfio/vfio_iommu_type1.c
+++ b/drivers/vfio/vfio_iommu_type1.c
@@ -861,6 +861,12 @@ static int vfio_iommu_type1_pin_pages(void *iommu_data,
 
 	mutex_lock(&iommu->lock);
 
+	if (WARN_ONCE(iommu->vaddr_invalid_count,
+		      "vfio_pin_pages not allowed with VFIO_UPDATE_VADDR\n")) {
+		ret = -EBUSY;
+		goto pin_done;
+	}
+
 	/*
 	 * Wait for all necessary vaddr's to be valid so they can be used in
 	 * the main loop without dropping the lock, to avoid racing vs unmap.
@@ -1343,6 +1349,12 @@ static int vfio_dma_do_unmap(struct vfio_iommu *iommu,
 
 	mutex_lock(&iommu->lock);
 
+	/* Cannot update vaddr if mdev is present. */
+	if (invalidate_vaddr && !list_empty(&iommu->emulated_iommu_groups)) {
+		ret = -EBUSY;
+		goto unlock;
+	}
+
 	pgshift = __ffs(iommu->pgsize_bitmap);
 	pgsize = (size_t)1 << pgshift;
 
@@ -2194,11 +2206,16 @@ static int vfio_iommu_type1_attach_group(void *iommu_data,
 	struct iommu_domain_geometry *geo;
 	LIST_HEAD(iova_copy);
 	LIST_HEAD(group_resv_regions);
-	int ret = -EINVAL;
+	int ret = -EBUSY;
 
 	mutex_lock(&iommu->lock);
 
+	/* Attach could require pinning, so disallow while vaddr is invalid. */
+	if (iommu->vaddr_invalid_count)
+		goto out_unlock;
+
 	/* Check for duplicates */
+	ret = -EINVAL;
 	if (vfio_iommu_find_iommu_group(iommu, iommu_group))
 		goto out_unlock;
 
@@ -2669,6 +2686,16 @@ static int vfio_domains_have_enforce_cache_coherency(struct vfio_iommu *iommu)
 	return ret;
 }
 
+static bool vfio_iommu_has_emulated(struct vfio_iommu *iommu)
+{
+	bool ret;
+
+	mutex_lock(&iommu->lock);
+	ret = !list_empty(&iommu->emulated_iommu_groups);
+	mutex_unlock(&iommu->lock);
+	return ret;
+}
+
 static int vfio_iommu_type1_check_extension(struct vfio_iommu *iommu,
 					    unsigned long arg)
 {
@@ -2677,8 +2704,13 @@ static int vfio_iommu_type1_check_extension(struct vfio_iommu *iommu,
 	case VFIO_TYPE1v2_IOMMU:
 	case VFIO_TYPE1_NESTING_IOMMU:
 	case VFIO_UNMAP_ALL:
-	case VFIO_UPDATE_VADDR:
 		return 1;
+	case VFIO_UPDATE_VADDR:
+		/*
+		 * Disable this feature if mdevs are present.  They cannot
+		 * safely pin/unpin/rw while vaddrs are being updated.
+		 */
+		return iommu && !vfio_iommu_has_emulated(iommu);
 	case VFIO_DMA_CC_IOMMU:
 		if (!iommu)
 			return 0;
@@ -3147,6 +3179,13 @@ static int vfio_iommu_type1_dma_rw(void *iommu_data, dma_addr_t user_iova,
 	size_t done;
 
 	mutex_lock(&iommu->lock);
+
+	if (WARN_ONCE(iommu->vaddr_invalid_count,
+		      "vfio_dma_rw not allowed with VFIO_UPDATE_VADDR\n")) {
+		ret = -EBUSY;
+		goto out;
+	}
+
 	while (count > 0) {
 		ret = vfio_iommu_type1_dma_rw_chunk(iommu, user_iova, data,
 						    count, write, &done);
@@ -3158,6 +3197,7 @@ static int vfio_iommu_type1_dma_rw(void *iommu_data, dma_addr_t user_iova,
 		user_iova += done;
 	}
 
+out:
 	mutex_unlock(&iommu->lock);
 	return ret;
 }
diff --git a/include/uapi/linux/vfio.h b/include/uapi/linux/vfio.h
index 23105eb..0552e8d 100644
--- a/include/uapi/linux/vfio.h
+++ b/include/uapi/linux/vfio.h
@@ -49,7 +49,11 @@
 /* Supports VFIO_DMA_UNMAP_FLAG_ALL */
 #define VFIO_UNMAP_ALL			9
 
-/* Supports the vaddr flag for DMA map and unmap */
+/*
+ * Supports the vaddr flag for DMA map and unmap.  Not supported for mediated
+ * devices, so this capability is subject to change as groups are added or
+ * removed.
+ */
 #define VFIO_UPDATE_VADDR		10
 
 /*
@@ -1343,8 +1347,7 @@ struct vfio_iommu_type1_info_dma_avail {
  * Map process virtual addresses to IO virtual addresses using the
  * provided struct vfio_dma_map. Caller sets argsz. READ &/ WRITE required.
  *
- * If flags & VFIO_DMA_MAP_FLAG_VADDR, update the base vaddr for iova, and
- * unblock translation of host virtual addresses in the iova range.  The vaddr
+ * If flags & VFIO_DMA_MAP_FLAG_VADDR, update the base vaddr for iova. The vaddr
  * must have previously been invalidated with VFIO_DMA_UNMAP_FLAG_VADDR.  To
  * maintain memory consistency within the user application, the updated vaddr
  * must address the same memory object as originally mapped.  Failure to do so
@@ -1395,9 +1398,9 @@ struct vfio_bitmap {
  * must be 0.  This cannot be combined with the get-dirty-bitmap flag.
  *
  * If flags & VFIO_DMA_UNMAP_FLAG_VADDR, do not unmap, but invalidate host
- * virtual addresses in the iova range.  Tasks that attempt to translate an
- * iova's vaddr will block.  DMA to already-mapped pages continues.  This
- * cannot be combined with the get-dirty-bitmap flag.
+ * virtual addresses in the iova range.  DMA to already-mapped pages continues.
+ * Groups may not be added to the container while any addresses are invalid.
+ * This cannot be combined with the get-dirty-bitmap flag.
  */
 struct vfio_iommu_type1_dma_unmap {
 	__u32	argsz;
-- 
1.8.3.1

[PATCH V8 2/7] vfio/type1: prevent underflow of locked_vm via exec()

[Steve Sistare]

When a vfio container is preserved across exec, the task does not change,
but it gets a new mm with locked_vm=0, and loses the count from existing
dma mappings.  If the user later unmaps a dma mapping, locked_vm underflows
to a large unsigned value, and a subsequent dma map request fails with
ENOMEM in __account_locked_vm.

To avoid underflow, grab and save the mm at the time a dma is mapped.
Use that mm when adjusting locked_vm, rather than re-acquiring the saved
task's mm, which may have changed.  If the saved mm is dead, do nothing.

locked_vm is incremented for existing mappings in a subsequent patch.

Fixes: 73fa0d10d077 ("vfio: Type1 IOMMU implementation")
Cc: stable@vger.kernel.org
Signed-off-by: Steve Sistare <steven.sistare@oracle.com>
Reviewed-by: Kevin Tian <kevin.tian@intel.com>
---
 drivers/vfio/vfio_iommu_type1.c | 41 ++++++++++++++---------------------------
 1 file changed, 14 insertions(+), 27 deletions(-)

diff --git a/drivers/vfio/vfio_iommu_type1.c b/drivers/vfio/vfio_iommu_type1.c
index e8902e7..ee9e988 100644
--- a/drivers/vfio/vfio_iommu_type1.c
+++ b/drivers/vfio/vfio_iommu_type1.c
@@ -100,6 +100,7 @@ struct vfio_dma {
 	struct task_struct	*task;
 	struct rb_root		pfn_list;	/* Ex-user pinned pfn list */
 	unsigned long		*bitmap;
+	struct mm_struct	*mm;
 };
 
 struct vfio_batch {
@@ -420,8 +421,8 @@ static int vfio_lock_acct(struct vfio_dma *dma, long npage, bool async)
 	if (!npage)
 		return 0;
 
-	mm = async ? get_task_mm(dma->task) : dma->task->mm;
-	if (!mm)
+	mm = dma->mm;
+	if (async && !mmget_not_zero(mm))
 		return -ESRCH; /* process exited */
 
 	ret = mmap_write_lock_killable(mm);
@@ -794,8 +795,8 @@ static int vfio_pin_page_external(struct vfio_dma *dma, unsigned long vaddr,
 	struct mm_struct *mm;
 	int ret;
 
-	mm = get_task_mm(dma->task);
-	if (!mm)
+	mm = dma->mm;
+	if (!mmget_not_zero(mm))
 		return -ENODEV;
 
 	ret = vaddr_get_pfns(mm, vaddr, 1, dma->prot, pfn_base, pages);
@@ -805,7 +806,7 @@ static int vfio_pin_page_external(struct vfio_dma *dma, unsigned long vaddr,
 	ret = 0;
 
 	if (do_accounting && !is_invalid_reserved_pfn(*pfn_base)) {
-		ret = vfio_lock_acct(dma, 1, true);
+		ret = vfio_lock_acct(dma, 1, false);
 		if (ret) {
 			put_pfn(*pfn_base, dma->prot);
 			if (ret == -ENOMEM)
@@ -1180,6 +1181,7 @@ static void vfio_remove_dma(struct vfio_iommu *iommu, struct vfio_dma *dma)
 	vfio_unmap_unpin(iommu, dma, true);
 	vfio_unlink_dma(iommu, dma);
 	put_task_struct(dma->task);
+	mmdrop(dma->mm);
 	vfio_dma_bitmap_free(dma);
 	if (dma->vaddr_invalid) {
 		iommu->vaddr_invalid_count--;
@@ -1664,29 +1666,15 @@ static int vfio_dma_do_map(struct vfio_iommu *iommu,
 	 * against the locked memory limit and we need to be able to do both
 	 * outside of this call path as pinning can be asynchronous via the
 	 * external interfaces for mdev devices.  RLIMIT_MEMLOCK requires a
-	 * task_struct and VM locked pages requires an mm_struct, however
-	 * holding an indefinite mm reference is not recommended, therefore we
-	 * only hold a reference to a task.  We could hold a reference to
-	 * current, however QEMU uses this call path through vCPU threads,
-	 * which can be killed resulting in a NULL mm and failure in the unmap
-	 * path when called via a different thread.  Avoid this problem by
-	 * using the group_leader as threads within the same group require
-	 * both CLONE_THREAD and CLONE_VM and will therefore use the same
-	 * mm_struct.
-	 *
-	 * Previously we also used the task for testing CAP_IPC_LOCK at the
-	 * time of pinning and accounting, however has_capability() makes use
-	 * of real_cred, a copy-on-write field, so we can't guarantee that it
-	 * matches group_leader, or in fact that it might not change by the
-	 * time it's evaluated.  If a process were to call MAP_DMA with
-	 * CAP_IPC_LOCK but later drop it, it doesn't make sense that they
-	 * possibly see different results for an iommu_mapped vfio_dma vs
-	 * externally mapped.  Therefore track CAP_IPC_LOCK in vfio_dma at the
-	 * time of calling MAP_DMA.
+	 * task_struct. Save the group_leader so that all DMA tracking uses
+	 * the same task, to make debugging easier.  VM locked pages requires
+	 * an mm_struct, so grab the mm in case the task dies.
 	 */
 	get_task_struct(current->group_leader);
 	dma->task = current->group_leader;
 	dma->lock_cap = capable(CAP_IPC_LOCK);
+	dma->mm = current->mm;
+	mmgrab(dma->mm);
 
 	dma->pfn_list = RB_ROOT;
 
@@ -3131,9 +3119,8 @@ static int vfio_iommu_type1_dma_rw_chunk(struct vfio_iommu *iommu,
 			!(dma->prot & IOMMU_READ))
 		return -EPERM;
 
-	mm = get_task_mm(dma->task);
-
-	if (!mm)
+	mm = dma->mm;
+	if (!mmget_not_zero(mm))
 		return -EPERM;
 
 	if (kthread)
-- 
1.8.3.1

[PATCH V8 3/7] vfio/type1: track locked_vm per dma

[Steve Sistare]

Track locked_vm per dma struct, and create a new subroutine, both for use
in a subsequent patch.  No functional change.

Fixes: c3cbab24db38 ("vfio/type1: implement interfaces to update vaddr")
Cc: stable@vger.kernel.org
Signed-off-by: Steve Sistare <steven.sistare@oracle.com>
Reviewed-by: Kevin Tian <kevin.tian@intel.com>
---
 drivers/vfio/vfio_iommu_type1.c | 23 +++++++++++++++++------
 1 file changed, 17 insertions(+), 6 deletions(-)

diff --git a/drivers/vfio/vfio_iommu_type1.c b/drivers/vfio/vfio_iommu_type1.c
index ee9e988..5a08346 100644
--- a/drivers/vfio/vfio_iommu_type1.c
+++ b/drivers/vfio/vfio_iommu_type1.c
@@ -101,6 +101,7 @@ struct vfio_dma {
 	struct rb_root		pfn_list;	/* Ex-user pinned pfn list */
 	unsigned long		*bitmap;
 	struct mm_struct	*mm;
+	size_t			locked_vm;
 };
 
 struct vfio_batch {
@@ -413,6 +414,19 @@ static int vfio_iova_put_vfio_pfn(struct vfio_dma *dma, struct vfio_pfn *vpfn)
 	return ret;
 }
 
+static int mm_lock_acct(struct task_struct *task, struct mm_struct *mm,
+			bool lock_cap, long npage)
+{
+	int ret = mmap_write_lock_killable(mm);
+
+	if (ret)
+		return ret;
+
+	ret = __account_locked_vm(mm, abs(npage), npage > 0, task, lock_cap);
+	mmap_write_unlock(mm);
+	return ret;
+}
+
 static int vfio_lock_acct(struct vfio_dma *dma, long npage, bool async)
 {
 	struct mm_struct *mm;
@@ -425,12 +439,9 @@ static int vfio_lock_acct(struct vfio_dma *dma, long npage, bool async)
 	if (async && !mmget_not_zero(mm))
 		return -ESRCH; /* process exited */
 
-	ret = mmap_write_lock_killable(mm);
-	if (!ret) {
-		ret = __account_locked_vm(mm, abs(npage), npage > 0, dma->task,
-					  dma->lock_cap);
-		mmap_write_unlock(mm);
-	}
+	ret = mm_lock_acct(dma->task, mm, dma->lock_cap, npage);
+	if (!ret)
+		dma->locked_vm += npage;
 
 	if (async)
 		mmput(mm);
-- 
1.8.3.1

[PATCH V8 4/7] vfio/type1: restore locked_vm

[Steve Sistare]

When a vfio container is preserved across exec or fork-exec, the new
task's mm has a locked_vm count of 0.  After a dma vaddr is updated using
VFIO_DMA_MAP_FLAG_VADDR, locked_vm remains 0, and the pinned memory does
not count against the task's RLIMIT_MEMLOCK.

To restore the correct locked_vm count, when VFIO_DMA_MAP_FLAG_VADDR is
used and the dma's mm has changed, add the dma's locked_vm count to
the new mm->locked_vm, subject to the rlimit, and subtract it from the
old mm->locked_vm.

Fixes: c3cbab24db38 ("vfio/type1: implement interfaces to update vaddr")
Cc: stable@vger.kernel.org
Signed-off-by: Steve Sistare <steven.sistare@oracle.com>
Reviewed-by: Kevin Tian <kevin.tian@intel.com>
---
 drivers/vfio/vfio_iommu_type1.c | 35 +++++++++++++++++++++++++++++++++++
 1 file changed, 35 insertions(+)

diff --git a/drivers/vfio/vfio_iommu_type1.c b/drivers/vfio/vfio_iommu_type1.c
index 5a08346..7fa68dc 100644
--- a/drivers/vfio/vfio_iommu_type1.c
+++ b/drivers/vfio/vfio_iommu_type1.c
@@ -1591,6 +1591,38 @@ static bool vfio_iommu_iova_dma_valid(struct vfio_iommu *iommu,
 	return list_empty(iova);
 }
 
+static int vfio_change_dma_owner(struct vfio_dma *dma)
+{
+	struct task_struct *task = current->group_leader;
+	struct mm_struct *mm = current->mm;
+	long npage = dma->locked_vm;
+	bool lock_cap;
+	int ret;
+
+	if (mm == dma->mm)
+		return 0;
+
+	lock_cap = capable(CAP_IPC_LOCK);
+	ret = mm_lock_acct(task, mm, lock_cap, npage);
+	if (ret)
+		return ret;
+
+	if (mmget_not_zero(dma->mm)) {
+		mm_lock_acct(dma->task, dma->mm, dma->lock_cap, -npage);
+		mmput(dma->mm);
+	}
+
+	if (dma->task != task) {
+		put_task_struct(dma->task);
+		dma->task = get_task_struct(task);
+	}
+	mmdrop(dma->mm);
+	dma->mm = mm;
+	mmgrab(dma->mm);
+	dma->lock_cap = lock_cap;
+	return 0;
+}
+
 static int vfio_dma_do_map(struct vfio_iommu *iommu,
 			   struct vfio_iommu_type1_dma_map *map)
 {
@@ -1640,6 +1672,9 @@ static int vfio_dma_do_map(struct vfio_iommu *iommu,
 			   dma->size != size) {
 			ret = -EINVAL;
 		} else {
+			ret = vfio_change_dma_owner(dma);
+			if (ret)
+				goto out_unlock;
 			dma->vaddr = vaddr;
 			dma->vaddr_invalid = false;
 			iommu->vaddr_invalid_count--;
-- 
1.8.3.1

[PATCH V8 5/7] vfio/type1: revert "block on invalid vaddr"

[Steve Sistare]

Revert this dead code:
  commit 898b9eaeb3fe ("vfio/type1: block on invalid vaddr")

Signed-off-by: Steve Sistare <steven.sistare@oracle.com>
Reviewed-by: Kevin Tian <kevin.tian@intel.com>
---
 drivers/vfio/vfio_iommu_type1.c | 94 +++--------------------------------------
 1 file changed, 5 insertions(+), 89 deletions(-)

diff --git a/drivers/vfio/vfio_iommu_type1.c b/drivers/vfio/vfio_iommu_type1.c
index 7fa68dc..be1dcad 100644
--- a/drivers/vfio/vfio_iommu_type1.c
+++ b/drivers/vfio/vfio_iommu_type1.c
@@ -72,7 +72,6 @@ struct vfio_iommu {
 	unsigned int		vaddr_invalid_count;
 	uint64_t		pgsize_bitmap;
 	uint64_t		num_non_pinned_groups;
-	wait_queue_head_t	vaddr_wait;
 	bool			v2;
 	bool			nesting;
 	bool			dirty_page_tracking;
@@ -154,8 +153,6 @@ struct vfio_regions {
 #define DIRTY_BITMAP_PAGES_MAX	 ((u64)INT_MAX)
 #define DIRTY_BITMAP_SIZE_MAX	 DIRTY_BITMAP_BYTES(DIRTY_BITMAP_PAGES_MAX)
 
-#define WAITED 1
-
 static int put_pfn(unsigned long pfn, int prot);
 
 static struct vfio_iommu_group*
@@ -607,61 +604,6 @@ static int vaddr_get_pfns(struct mm_struct *mm, unsigned long vaddr,
 	return ret;
 }
 
-static int vfio_wait(struct vfio_iommu *iommu)
-{
-	DEFINE_WAIT(wait);
-
-	prepare_to_wait(&iommu->vaddr_wait, &wait, TASK_KILLABLE);
-	mutex_unlock(&iommu->lock);
-	schedule();
-	mutex_lock(&iommu->lock);
-	finish_wait(&iommu->vaddr_wait, &wait);
-	if (kthread_should_stop() || !iommu->container_open ||
-	    fatal_signal_pending(current)) {
-		return -EFAULT;
-	}
-	return WAITED;
-}
-
-/*
- * Find dma struct and wait for its vaddr to be valid.  iommu lock is dropped
- * if the task waits, but is re-locked on return.  Return result in *dma_p.
- * Return 0 on success with no waiting, WAITED on success if waited, and -errno
- * on error.
- */
-static int vfio_find_dma_valid(struct vfio_iommu *iommu, dma_addr_t start,
-			       size_t size, struct vfio_dma **dma_p)
-{
-	int ret = 0;
-
-	do {
-		*dma_p = vfio_find_dma(iommu, start, size);
-		if (!*dma_p)
-			return -EINVAL;
-		else if (!(*dma_p)->vaddr_invalid)
-			return ret;
-		else
-			ret = vfio_wait(iommu);
-	} while (ret == WAITED);
-
-	return ret;
-}
-
-/*
- * Wait for all vaddr in the dma_list to become valid.  iommu lock is dropped
- * if the task waits, but is re-locked on return.  Return 0 on success with no
- * waiting, WAITED on success if waited, and -errno on error.
- */
-static int vfio_wait_all_valid(struct vfio_iommu *iommu)
-{
-	int ret = 0;
-
-	while (iommu->vaddr_invalid_count && ret >= 0)
-		ret = vfio_wait(iommu);
-
-	return ret;
-}
-
 /*
  * Attempt to pin pages.  We really don't want to track all the pfns and
  * the iommu can only map chunks of consecutive pfns anyway, so get the
@@ -862,7 +804,6 @@ static int vfio_iommu_type1_pin_pages(void *iommu_data,
 	unsigned long remote_vaddr;
 	struct vfio_dma *dma;
 	bool do_accounting;
-	dma_addr_t iova;
 
 	if (!iommu || !pages)
 		return -EINVAL;
@@ -879,22 +820,6 @@ static int vfio_iommu_type1_pin_pages(void *iommu_data,
 		goto pin_done;
 	}
 
-	/*
-	 * Wait for all necessary vaddr's to be valid so they can be used in
-	 * the main loop without dropping the lock, to avoid racing vs unmap.
-	 */
-again:
-	if (iommu->vaddr_invalid_count) {
-		for (i = 0; i < npage; i++) {
-			iova = user_iova + PAGE_SIZE * i;
-			ret = vfio_find_dma_valid(iommu, iova, PAGE_SIZE, &dma);
-			if (ret < 0)
-				goto pin_done;
-			if (ret == WAITED)
-				goto again;
-		}
-	}
-
 	/* Fail if no dma_umap notifier is registered */
 	if (list_empty(&iommu->device_list)) {
 		ret = -EINVAL;
@@ -910,6 +835,7 @@ static int vfio_iommu_type1_pin_pages(void *iommu_data,
 
 	for (i = 0; i < npage; i++) {
 		unsigned long phys_pfn;
+		dma_addr_t iova;
 		struct vfio_pfn *vpfn;
 
 		iova = user_iova + PAGE_SIZE * i;
@@ -1194,10 +1120,8 @@ static void vfio_remove_dma(struct vfio_iommu *iommu, struct vfio_dma *dma)
 	put_task_struct(dma->task);
 	mmdrop(dma->mm);
 	vfio_dma_bitmap_free(dma);
-	if (dma->vaddr_invalid) {
+	if (dma->vaddr_invalid)
 		iommu->vaddr_invalid_count--;
-		wake_up_all(&iommu->vaddr_wait);
-	}
 	kfree(dma);
 	iommu->dma_avail++;
 }
@@ -1678,7 +1602,6 @@ static int vfio_dma_do_map(struct vfio_iommu *iommu,
 			dma->vaddr = vaddr;
 			dma->vaddr_invalid = false;
 			iommu->vaddr_invalid_count--;
-			wake_up_all(&iommu->vaddr_wait);
 		}
 		goto out_unlock;
 	} else if (dma) {
@@ -1753,10 +1676,6 @@ static int vfio_iommu_replay(struct vfio_iommu *iommu,
 	unsigned long limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
 	int ret;
 
-	ret = vfio_wait_all_valid(iommu);
-	if (ret < 0)
-		return ret;
-
 	/* Arbitrarily pick the first domain in the list for lookups */
 	if (!list_empty(&iommu->domain_list))
 		d = list_first_entry(&iommu->domain_list,
@@ -2656,7 +2575,6 @@ static void *vfio_iommu_type1_open(unsigned long arg)
 	mutex_init(&iommu->lock);
 	mutex_init(&iommu->device_list_lock);
 	INIT_LIST_HEAD(&iommu->device_list);
-	init_waitqueue_head(&iommu->vaddr_wait);
 	iommu->pgsize_bitmap = PAGE_MASK;
 	INIT_LIST_HEAD(&iommu->emulated_iommu_groups);
 
@@ -3153,13 +3071,12 @@ static int vfio_iommu_type1_dma_rw_chunk(struct vfio_iommu *iommu,
 	struct vfio_dma *dma;
 	bool kthread = current->mm == NULL;
 	size_t offset;
-	int ret;
 
 	*copied = 0;
 
-	ret = vfio_find_dma_valid(iommu, user_iova, 1, &dma);
-	if (ret < 0)
-		return ret;
+	dma = vfio_find_dma(iommu, user_iova, 1);
+	if (!dma)
+		return -EINVAL;
 
 	if ((write && !(dma->prot & IOMMU_WRITE)) ||
 			!(dma->prot & IOMMU_READ))
@@ -3268,7 +3185,6 @@ static void vfio_iommu_type1_notify(void *iommu_data,
 	mutex_lock(&iommu->lock);
 	iommu->container_open = false;
 	mutex_unlock(&iommu->lock);
-	wake_up_all(&iommu->vaddr_wait);
 }
 
 static const struct vfio_iommu_driver_ops vfio_iommu_driver_ops_type1 = {
-- 
1.8.3.1

[PATCH V8 6/7] vfio/type1: revert "implement notify callback"

[Steve Sistare]

This is dead code.  Revert it.
  commit 487ace134053 ("vfio/type1: implement notify callback")

Signed-off-by: Steve Sistare <steven.sistare@oracle.com>
Reviewed-by: Kevin Tian <kevin.tian@intel.com>
---
 drivers/vfio/vfio_iommu_type1.c | 15 ---------------
 1 file changed, 15 deletions(-)

diff --git a/drivers/vfio/vfio_iommu_type1.c b/drivers/vfio/vfio_iommu_type1.c
index be1dcad..c845ee5 100644
--- a/drivers/vfio/vfio_iommu_type1.c
+++ b/drivers/vfio/vfio_iommu_type1.c
@@ -75,7 +75,6 @@ struct vfio_iommu {
 	bool			v2;
 	bool			nesting;
 	bool			dirty_page_tracking;
-	bool			container_open;
 	struct list_head	emulated_iommu_groups;
 };
 
@@ -2571,7 +2570,6 @@ static void *vfio_iommu_type1_open(unsigned long arg)
 	INIT_LIST_HEAD(&iommu->iova_list);
 	iommu->dma_list = RB_ROOT;
 	iommu->dma_avail = dma_entry_limit;
-	iommu->container_open = true;
 	mutex_init(&iommu->lock);
 	mutex_init(&iommu->device_list_lock);
 	INIT_LIST_HEAD(&iommu->device_list);
@@ -3175,18 +3173,6 @@ static int vfio_iommu_type1_dma_rw(void *iommu_data, dma_addr_t user_iova,
 	return domain;
 }
 
-static void vfio_iommu_type1_notify(void *iommu_data,
-				    enum vfio_iommu_notify_type event)
-{
-	struct vfio_iommu *iommu = iommu_data;
-
-	if (event != VFIO_IOMMU_CONTAINER_CLOSE)
-		return;
-	mutex_lock(&iommu->lock);
-	iommu->container_open = false;
-	mutex_unlock(&iommu->lock);
-}
-
 static const struct vfio_iommu_driver_ops vfio_iommu_driver_ops_type1 = {
 	.name			= "vfio-iommu-type1",
 	.owner			= THIS_MODULE,
@@ -3201,7 +3187,6 @@ static void vfio_iommu_type1_notify(void *iommu_data,
 	.unregister_device	= vfio_iommu_type1_unregister_device,
 	.dma_rw			= vfio_iommu_type1_dma_rw,
 	.group_iommu_domain	= vfio_iommu_type1_group_iommu_domain,
-	.notify			= vfio_iommu_type1_notify,
 };
 
 static int __init vfio_iommu_type1_init(void)
-- 
1.8.3.1

[PATCH V8 7/7] vfio: revert "iommu driver notify callback"

[Steve Sistare]

Revert this dead code:
  commit ec5e32940cc9 ("vfio: iommu driver notify callback")

Signed-off-by: Steve Sistare <steven.sistare@oracle.com>
Reviewed-by: Kevin Tian <kevin.tian@intel.com>
---
 drivers/vfio/container.c | 5 -----
 drivers/vfio/vfio.h      | 7 -------
 2 files changed, 12 deletions(-)

diff --git a/drivers/vfio/container.c b/drivers/vfio/container.c
index b7a9560..82ece00 100644
--- a/drivers/vfio/container.c
+++ b/drivers/vfio/container.c
@@ -383,11 +383,6 @@ static int vfio_fops_open(struct inode *inode, struct file *filep)
 static int vfio_fops_release(struct inode *inode, struct file *filep)
 {
 	struct vfio_container *container = filep->private_data;
-	struct vfio_iommu_driver *driver = container->iommu_driver;
-
-	if (driver && driver->ops->notify)
-		driver->ops->notify(container->iommu_data,
-				    VFIO_IOMMU_CONTAINER_CLOSE);
 
 	filep->private_data = NULL;
 
diff --git a/drivers/vfio/vfio.h b/drivers/vfio/vfio.h
index f8219a4..d5fa896 100644
--- a/drivers/vfio/vfio.h
+++ b/drivers/vfio/vfio.h
@@ -89,11 +89,6 @@ int vfio_device_set_group(struct vfio_device *device,
 void vfio_group_cleanup(void);
 
 #if IS_ENABLED(CONFIG_VFIO_CONTAINER)
-/* events for the backend driver notify callback */
-enum vfio_iommu_notify_type {
-	VFIO_IOMMU_CONTAINER_CLOSE = 0,
-};
-
 /**
  * struct vfio_iommu_driver_ops - VFIO IOMMU driver callbacks
  */
@@ -124,8 +119,6 @@ struct vfio_iommu_driver_ops {
 				  void *data, size_t count, bool write);
 	struct iommu_domain *(*group_iommu_domain)(void *iommu_data,
 						   struct iommu_group *group);
-	void		(*notify)(void *iommu_data,
-				  enum vfio_iommu_notify_type event);
 };
 
 struct vfio_iommu_driver {
-- 
1.8.3.1

Re: [PATCH V8 0/7] fixes for virtual address update

[Alex Williamson]

On Tue, 31 Jan 2023 08:58:02 -0800
Steve Sistare <steven.sistare@oracle.com> wrote:
> Changes in v8:
>   * updated group_leader comment in vfio_dma_do_map
>   * delete async arg from mm_lock_acct
>   * pass async=false to vfio_lock_acct in vfio_pin_page_external
>   * locked_vm becomes size_t
>   * improved commit message in "restore locked_vm"
>   * simplified flow in vfio_change_dma_owner
>   * rebase to v6.2-rc6
> 
> Steve Sistare (7):
>   vfio/type1: exclude mdevs from VFIO_UPDATE_VADDR
>   vfio/type1: prevent underflow of locked_vm via exec()
>   vfio/type1: track locked_vm per dma
>   vfio/type1: restore locked_vm
>   vfio/type1: revert "block on invalid vaddr"
>   vfio/type1: revert "implement notify callback"
>   vfio: revert "iommu driver notify callback"
> 
>  drivers/vfio/container.c        |   5 -
>  drivers/vfio/vfio.h             |   7 --
>  drivers/vfio/vfio_iommu_type1.c | 248 ++++++++++++++++++----------------------
>  include/uapi/linux/vfio.h       |  15 ++-
>  4 files changed, 120 insertions(+), 155 deletions(-)

LGTM, I'll pause to give Jason a chance to review as well.  Thanks for
working on this!

Alex

Re: [PATCH V8 0/7] fixes for virtual address update

[Jason Gunthorpe]

On Tue, Jan 31, 2023 at 08:58:02AM -0800, Steve Sistare wrote:
> Fix bugs in the interfaces that allow the underlying memory object of an
> iova range to be mapped in a new address space.  They allow userland to
> indefinitely block vfio mediated device kernel threads, and do not
> propagate the locked_vm count to a new mm.  Also fix a pre-existing bug
> that allows locked_vm underflow.
> 
> The fixes impose restrictions that eliminate waiting conditions, so
> revert the dead code:
>   commit 898b9eaeb3fe ("vfio/type1: block on invalid vaddr")
>   commit 487ace134053 ("vfio/type1: implement notify callback")
>   commit ec5e32940cc9 ("vfio: iommu driver notify callback")

I would still rather we delete this API. Something that doesn't work
with iommufd, and doesn't work with mdevs doesn't seem like it should
be in the kernel.

But it is up to Alex, and the code looks fine:

Reviewed-by: Jason Gunthorpe <jgg@nvidia.com>

Jason

Re: [PATCH V8 0/7] fixes for virtual address update

[Alex Williamson]

On Tue, 31 Jan 2023 08:58:02 -0800
Steve Sistare <steven.sistare@oracle.com> wrote:

> Fix bugs in the interfaces that allow the underlying memory object of an
> iova range to be mapped in a new address space.  They allow userland to
> indefinitely block vfio mediated device kernel threads, and do not
> propagate the locked_vm count to a new mm.  Also fix a pre-existing bug
> that allows locked_vm underflow.
> 
> The fixes impose restrictions that eliminate waiting conditions, so
> revert the dead code:
>   commit 898b9eaeb3fe ("vfio/type1: block on invalid vaddr")
>   commit 487ace134053 ("vfio/type1: implement notify callback")
>   commit ec5e32940cc9 ("vfio: iommu driver notify callback")
> 
> Changes in V2 (thanks Alex):
>   * do not allow group attach while vaddrs are invalid
>   * add patches to delete dead code
>   * add WARN_ON for never-should-happen conditions
>   * check for changed mm in unmap.
>   * check for vfio_lock_acct failure in remap
> 
> Changes in V3 (ditto!):
>   * return errno at WARN_ON sites, and make it unique
>   * correctly check for dma task mm change
>   * change dma owner to current when vaddr is updated
>   * add Fixes to commit messages
>   * refactored new code in vfio_dma_do_map
> 
> Changes in V4:
>   * misc cosmetic changes
> 
> Changes in V5 (thanks Jason and Kevin):
>   * grab mm and use it for locked_vm accounting
>   * separate patches for underflow and restoring locked_vm
>   * account for reserved pages
>   * improve error messages
> 
> Changes in V6:
>   * drop "count reserved pages" patch
>   * add "track locked_vm" patch
>   * grab current->mm not group_leader->mm
>   * simplify vfio_change_dma_owner
>   * fix commit messages
> 
> Changes in v7:
>   * compare current->mm not group_leader->mm (missed one)
>   * misc cosmetic changes
> 
> Changes in v8:
>   * updated group_leader comment in vfio_dma_do_map
>   * delete async arg from mm_lock_acct
>   * pass async=false to vfio_lock_acct in vfio_pin_page_external
>   * locked_vm becomes size_t
>   * improved commit message in "restore locked_vm"
>   * simplified flow in vfio_change_dma_owner
>   * rebase to v6.2-rc6
> 
> Steve Sistare (7):
>   vfio/type1: exclude mdevs from VFIO_UPDATE_VADDR
>   vfio/type1: prevent underflow of locked_vm via exec()
>   vfio/type1: track locked_vm per dma
>   vfio/type1: restore locked_vm
>   vfio/type1: revert "block on invalid vaddr"
>   vfio/type1: revert "implement notify callback"
>   vfio: revert "iommu driver notify callback"
> 
>  drivers/vfio/container.c        |   5 -
>  drivers/vfio/vfio.h             |   7 --
>  drivers/vfio/vfio_iommu_type1.c | 248 ++++++++++++++++++----------------------
>  include/uapi/linux/vfio.h       |  15 ++-
>  4 files changed, 120 insertions(+), 155 deletions(-)
> 

Applied to vfio next branch for v6.3.  Thanks,

Alex