Skip to contentSkip to conversation

media: v4l2-mem2mem: fix poll() bug

PATCH Superseded Alexandre Courbot 3 participants 12 messages Aug 25, 2020 – Aug 26, 2020 Media (V4L/DVB)

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore
#0 lore raw

cover media: v4l2-mem2mem: fix poll() bug

This addresses a very corner case that probably nobody ever encounters, but I have hit it when playing with vicoded so here is a tentative fix.

Patch 1/2 addresses the issue that when the last buffer of a m2m device has been dequeued, any attempt to poll with EPOLLOUT will result in only EPOLLIN being returned, even if OUTPUT buffers are still pending. The issue stems from the fact that the last buffer check if done first, and returns immediately if true.

Patch 2/2 builds on the first one to (hopefully) clean up the code a bit and make the function flow easier to follow. Functionally speaking it is supposed to be a no-op and it can safely be dropped if the former code is preferred - the actual fix is in 1/2.

Alexandre Courbot (2):
media: v4l2-mem2mem: consider OUTPUT queue first when polling
media: v4l2-mem2mem: simplify poll logic a bit

 drivers/media/v4l2-core/v4l2-mem2mem.c | 42 +++++++++++---------------
 1 file changed, 18 insertions(+), 24 deletions(-)
Signature · 1 line
2.28.0
#1 lore raw

PATCH 1/2 media: v4l2-mem2mem: consider OUTPUT queue first when polling

If poll() is called on a m2m device with the EPOLLOUT event after the last buffer of the CAPTURE queue is dequeued, any buffer available on OUTPUT queue will never be signaled because v4l2_m2m_poll_for_data() starts by checking whether dst_q->last_buffer_dequeued is set and returns EPOLLIN in this case, without looking at the state of the OUTPUT queue.

Fix this by checking the state of the OUTPUT queue before considering that early-return case.

This also has the side-effect of bringing the two blocks of code dealing with the CAPTURE queue next to one another, and saves us one spin lock/unlock cycle, for what it's worth.

---
 drivers/media/v4l2-core/v4l2-mem2mem.c | 23 +++++++++++------------
 1 file changed, 11 insertions(+), 12 deletions(-)
drivers/media/v4l2-core/v4l2-mem2mem.c +11 -12
index 95a8f2dc5341d..0d0192119af20 100644
@@ -862,6 +862,15 @@ static __poll_t v4l2_m2m_poll_for_data(struct file *file,
 	     list_empty(&dst_q->queued_list)))
 		return EPOLLERR;
 
+	spin_lock_irqsave(&src_q->done_lock, flags);
+	if (!list_empty(&src_q->done_list))
+		src_vb = list_first_entry(&src_q->done_list, struct vb2_buffer,
+						done_entry);
+	if (src_vb && (src_vb->state == VB2_BUF_STATE_DONE
+			|| src_vb->state == VB2_BUF_STATE_ERROR))
+		rc |= EPOLLOUT | EPOLLWRNORM;
+	spin_unlock_irqrestore(&src_q->done_lock, flags);
+
 	spin_lock_irqsave(&dst_q->done_lock, flags);
 	if (list_empty(&dst_q->done_list)) {
 		/*
@@ -870,21 +879,11 @@ static __poll_t v4l2_m2m_poll_for_data(struct file *file,
 		 */
 		if (dst_q->last_buffer_dequeued) {
 			spin_unlock_irqrestore(&dst_q->done_lock, flags);
-			return EPOLLIN | EPOLLRDNORM;
+			rc |= EPOLLIN | EPOLLRDNORM;
+			return rc;
 		}
 	}
-	spin_unlock_irqrestore(&dst_q->done_lock, flags);
 
-	spin_lock_irqsave(&src_q->done_lock, flags);
-	if (!list_empty(&src_q->done_list))
-		src_vb = list_first_entry(&src_q->done_list, struct vb2_buffer,
-						done_entry);
-	if (src_vb && (src_vb->state == VB2_BUF_STATE_DONE
-			|| src_vb->state == VB2_BUF_STATE_ERROR))
-		rc |= EPOLLOUT | EPOLLWRNORM;
-	spin_unlock_irqrestore(&src_q->done_lock, flags);
-
-	spin_lock_irqsave(&dst_q->done_lock, flags);
 	if (!list_empty(&dst_q->done_list))
 		dst_vb = list_first_entry(&dst_q->done_list, struct vb2_buffer,
 						done_entry);
Signature · 1 line
2.28.0
Reviewed-by on v1 1/2
#2 lore raw

Re: [PATCH 1/2] media: v4l2-mem2mem: consider OUTPUT queue first when polling

Hi Alexandre,

Quoted 8 lines · Alexandre Courbot

If poll() is called on a m2m device with the EPOLLOUT event after the last buffer of the CAPTURE queue is dequeued, any buffer available on OUTPUT queue will never be signaled because v4l2_m2m_poll_for_data() starts by checking whether dst_q->last_buffer_dequeued is set and returns EPOLLIN in this case, without looking at the state of the OUTPUT queue.

Quoted 8 lines

Fix this by checking the state of the OUTPUT queue before considering that early-return case.

This also has the side-effect of bringing the two blocks of code dealing with the CAPTURE queue next to one another, and saves us one spin lock/unlock cycle, for what it's worth.

Change looks good to me.

Do you think it qualifies for -stable? The issue has been here since the dawn of time.

Thanks,
Ezequiel

Quoted 51 lines
---
 drivers/media/v4l2-core/v4l2-mem2mem.c | 23 +++++++++++------------
 1 file changed, 11 insertions(+), 12 deletions(-)
drivers/media/v4l2-core/v4l2-mem2mem.c +11 -13
index 95a8f2dc5341d..0d0192119af20 100644
@@ -862,6 +862,15 @@ static __poll_t v4l2_m2m_poll_for_data(struct file *file,
             list_empty(&dst_q->queued_list)))
                return EPOLLERR;
 
+       spin_lock_irqsave(&src_q->done_lock, flags);
+       if (!list_empty(&src_q->done_list))
+               src_vb = list_first_entry(&src_q->done_list, struct vb2_buffer,
+                                               done_entry);
+       if (src_vb && (src_vb->state == VB2_BUF_STATE_DONE
+                       || src_vb->state == VB2_BUF_STATE_ERROR))
+               rc |= EPOLLOUT | EPOLLWRNORM;
+       spin_unlock_irqrestore(&src_q->done_lock, flags);
+
        spin_lock_irqsave(&dst_q->done_lock, flags);
        if (list_empty(&dst_q->done_list)) {
                /*
@@ -870,21 +879,11 @@ static __poll_t v4l2_m2m_poll_for_data(struct file *file,
                 */
                if (dst_q->last_buffer_dequeued) {
                        spin_unlock_irqrestore(&dst_q->done_lock, flags);
-                       return EPOLLIN | EPOLLRDNORM;
+                       rc |= EPOLLIN | EPOLLRDNORM;
+                       return rc;
                }
        }
-       spin_unlock_irqrestore(&dst_q->done_lock, flags);
 
-       spin_lock_irqsave(&src_q->done_lock, flags);
-       if (!list_empty(&src_q->done_list))
-               src_vb = list_first_entry(&src_q->done_list, struct vb2_buffer,
-                                               done_entry);
-       if (src_vb && (src_vb->state == VB2_BUF_STATE_DONE
-                       || src_vb->state == VB2_BUF_STATE_ERROR))
-               rc |= EPOLLOUT | EPOLLWRNORM;
-       spin_unlock_irqrestore(&src_q->done_lock, flags);
-
-       spin_lock_irqsave(&dst_q->done_lock, flags);
        if (!list_empty(&dst_q->done_list))
                dst_vb = list_first_entry(&dst_q->done_list, struct vb2_buffer,
                                                done_entry);
--

2.28.0

#3 lore raw

Re: [PATCH 1/2] media: v4l2-mem2mem: consider OUTPUT queue first when polling

Quoted 19 lines · Ezequiel Garcia

Hi Alexandre,

Quoted 15 lines · Alexandre Courbot

If poll() is called on a m2m device with the EPOLLOUT event after the last buffer of the CAPTURE queue is dequeued, any buffer available on OUTPUT queue will never be signaled because v4l2_m2m_poll_for_data() starts by checking whether dst_q->last_buffer_dequeued is set and returns EPOLLIN in this case, without looking at the state of the OUTPUT queue.

Fix this by checking the state of the OUTPUT queue before considering that early-return case.

This also has the side-effect of bringing the two blocks of code dealing with the CAPTURE queue next to one another, and saves us one spin lock/unlock cycle, for what it's worth.

Quoted 8 lines
Quoted 1 line

Change looks good to me.

Do you think it qualifies for -stable? The issue has been here since the dawn of time.

Indeed, and this should be quite a rare corner case. I will leave that call to the maintainers.

Quoted 55 lines

Thanks,
Ezequiel

Quoted 51 lines
---
 drivers/media/v4l2-core/v4l2-mem2mem.c | 23 +++++++++++------------
 1 file changed, 11 insertions(+), 12 deletions(-)
drivers/media/v4l2-core/v4l2-mem2mem.c +11 -13
index 95a8f2dc5341d..0d0192119af20 100644
@@ -862,6 +862,15 @@ static __poll_t v4l2_m2m_poll_for_data(struct file *file,
             list_empty(&dst_q->queued_list)))
                return EPOLLERR;
 
+       spin_lock_irqsave(&src_q->done_lock, flags);
+       if (!list_empty(&src_q->done_list))
+               src_vb = list_first_entry(&src_q->done_list, struct vb2_buffer,
+                                               done_entry);
+       if (src_vb && (src_vb->state == VB2_BUF_STATE_DONE
+                       || src_vb->state == VB2_BUF_STATE_ERROR))
+               rc |= EPOLLOUT | EPOLLWRNORM;
+       spin_unlock_irqrestore(&src_q->done_lock, flags);
+
        spin_lock_irqsave(&dst_q->done_lock, flags);
        if (list_empty(&dst_q->done_list)) {
                /*
@@ -870,21 +879,11 @@ static __poll_t v4l2_m2m_poll_for_data(struct file *file,
                 */
                if (dst_q->last_buffer_dequeued) {
                        spin_unlock_irqrestore(&dst_q->done_lock, flags);
-                       return EPOLLIN | EPOLLRDNORM;
+                       rc |= EPOLLIN | EPOLLRDNORM;
+                       return rc;
                }
        }
-       spin_unlock_irqrestore(&dst_q->done_lock, flags);
 
-       spin_lock_irqsave(&src_q->done_lock, flags);
-       if (!list_empty(&src_q->done_list))
-               src_vb = list_first_entry(&src_q->done_list, struct vb2_buffer,
-                                               done_entry);
-       if (src_vb && (src_vb->state == VB2_BUF_STATE_DONE
-                       || src_vb->state == VB2_BUF_STATE_ERROR))
-               rc |= EPOLLOUT | EPOLLWRNORM;
-       spin_unlock_irqrestore(&src_q->done_lock, flags);
-
-       spin_lock_irqsave(&dst_q->done_lock, flags);
        if (!list_empty(&dst_q->done_list))
                dst_vb = list_first_entry(&dst_q->done_list, struct vb2_buffer,
                                                done_entry);
--

2.28.0

#4 lore raw

PATCH 2/2 media: v4l2-mem2mem: simplify poll logic a bit

In reply to Alexandre Courbot

Factorize redundant checks into a single code block, remove the early return, and declare variables in their innermost block. Hopefully this makes this code a little bit easier to follow.

---
 drivers/media/v4l2-core/v4l2-mem2mem.c | 35 +++++++++++---------------
 1 file changed, 15 insertions(+), 20 deletions(-)
drivers/media/v4l2-core/v4l2-mem2mem.c +15 -20
index 0d0192119af20..aeac9707123d0 100644
@@ -841,7 +841,6 @@ static __poll_t v4l2_m2m_poll_for_data(struct file *file,
 				       struct poll_table_struct *wait)
 {
 	struct vb2_queue *src_q, *dst_q;
-	struct vb2_buffer *src_vb = NULL, *dst_vb = NULL;
 	__poll_t rc = 0;
 	unsigned long flags;
 
@@ -863,33 +862,29 @@ static __poll_t v4l2_m2m_poll_for_data(struct file *file,
 		return EPOLLERR;
 
 	spin_lock_irqsave(&src_q->done_lock, flags);
-	if (!list_empty(&src_q->done_list))
-		src_vb = list_first_entry(&src_q->done_list, struct vb2_buffer,
-						done_entry);
-	if (src_vb && (src_vb->state == VB2_BUF_STATE_DONE
-			|| src_vb->state == VB2_BUF_STATE_ERROR))
-		rc |= EPOLLOUT | EPOLLWRNORM;
+	if (!list_empty(&src_q->done_list)) {
+		struct vb2_buffer *src_vb = list_first_entry(
+			&src_q->done_list, struct vb2_buffer, done_entry);
+		if (src_vb->state == VB2_BUF_STATE_DONE ||
+		    src_vb->state == VB2_BUF_STATE_ERROR)
+			rc |= EPOLLOUT | EPOLLWRNORM;
+	}
 	spin_unlock_irqrestore(&src_q->done_lock, flags);
 
 	spin_lock_irqsave(&dst_q->done_lock, flags);
-	if (list_empty(&dst_q->done_list)) {
+	if (!list_empty(&dst_q->done_list)) {
+		struct vb2_buffer *dst_vb = list_first_entry(
+			&dst_q->done_list, struct vb2_buffer, done_entry);
+		if (dst_vb->state == VB2_BUF_STATE_DONE ||
+		    dst_vb->state == VB2_BUF_STATE_ERROR)
+			rc |= EPOLLIN | EPOLLRDNORM;
+	} else if (dst_q->last_buffer_dequeued) {
 		/*
 		 * If the last buffer was dequeued from the capture queue,
 		 * return immediately. DQBUF will return -EPIPE.
 		 */
-		if (dst_q->last_buffer_dequeued) {
-			spin_unlock_irqrestore(&dst_q->done_lock, flags);
-			rc |= EPOLLIN | EPOLLRDNORM;
-			return rc;
-		}
-	}
-
-	if (!list_empty(&dst_q->done_list))
-		dst_vb = list_first_entry(&dst_q->done_list, struct vb2_buffer,
-						done_entry);
-	if (dst_vb && (dst_vb->state == VB2_BUF_STATE_DONE
-			|| dst_vb->state == VB2_BUF_STATE_ERROR))
 		rc |= EPOLLIN | EPOLLRDNORM;
+	}
 	spin_unlock_irqrestore(&dst_q->done_lock, flags);
 
 	return rc;
Signature · 1 line
2.28.0
Reviewed-by on v1 2/2
#5 lore raw

Re: [PATCH 2/2] media: v4l2-mem2mem: simplify poll logic a bit

Hi Alexandre,

Quoted 5 lines · Alexandre Courbot

Factorize redundant checks into a single code block, remove the early return, and declare variables in their innermost block. Hopefully this makes this code a little bit easier to follow.

This _definitely_ makes the poll handling more readable.

See below a nitpick.

Quoted 39 lines
---
 drivers/media/v4l2-core/v4l2-mem2mem.c | 35 +++++++++++---------------
 1 file changed, 15 insertions(+), 20 deletions(-)
drivers/media/v4l2-core/v4l2-mem2mem.c +8 -8
index 0d0192119af20..aeac9707123d0 100644
@@ -841,7 +841,6 @@ static __poll_t v4l2_m2m_poll_for_data(struct file *file,
                                       struct poll_table_struct *wait)
 {
        struct vb2_queue *src_q, *dst_q;
-       struct vb2_buffer *src_vb = NULL, *dst_vb = NULL;
        __poll_t rc = 0;
        unsigned long flags;
 
@@ -863,33 +862,29 @@ static __poll_t v4l2_m2m_poll_for_data(struct file *file,
                return EPOLLERR;
 
        spin_lock_irqsave(&src_q->done_lock, flags);
-       if (!list_empty(&src_q->done_list))
-               src_vb = list_first_entry(&src_q->done_list, struct vb2_buffer,
-                                               done_entry);
-       if (src_vb && (src_vb->state == VB2_BUF_STATE_DONE
-                       || src_vb->state == VB2_BUF_STATE_ERROR))
-               rc |= EPOLLOUT | EPOLLWRNORM;
+       if (!list_empty(&src_q->done_list)) {
+               struct vb2_buffer *src_vb = list_first_entry(
+                       &src_q->done_list, struct vb2_buffer, done_entry);
+               if (src_vb->state == VB2_BUF_STATE_DONE ||
+                   src_vb->state == VB2_BUF_STATE_ERROR)
+                       rc |= EPOLLOUT | EPOLLWRNORM;
+       }
        spin_unlock_irqrestore(&src_q->done_lock, flags);
 
        spin_lock_irqsave(&dst_q->done_lock, flags);
-       if (list_empty(&dst_q->done_list)) {
+       if (!list_empty(&dst_q->done_list)) {

drivers/media/v4l2-core/v4l2-mem2mem.c static __poll_t v4l2_m2m_poll_for_data(struct file *file, · line 885

Quoted 10 lines
drivers/media/v4l2-core/v4l2-mem2mem.c +6 -0
+               struct vb2_buffer *dst_vb = list_first_entry(
+                       &dst_q->done_list, struct vb2_buffer, done_entry);
+               if (dst_vb->state == VB2_BUF_STATE_DONE ||
+                   dst_vb->state == VB2_BUF_STATE_ERROR)
+                       rc |= EPOLLIN | EPOLLRDNORM;
+       } else if (dst_q->last_buffer_dequeued) {
                /*
                 * If the last buffer was dequeued from the capture queue,
                 * return immediately. DQBUF will return -EPIPE.
                 */

The part about "returning immediately" doesn't make much sense now. Could we rephrase this, keeping the -EPIPE comment?

Thanks,
Ezequiel

Quoted 20 lines
drivers/media/v4l2-core/v4l2-mem2mem.c +1 -13
-               if (dst_q->last_buffer_dequeued) {
-                       spin_unlock_irqrestore(&dst_q->done_lock, flags);
-                       rc |= EPOLLIN | EPOLLRDNORM;
-                       return rc;
-               }
-       }
-
-       if (!list_empty(&dst_q->done_list))
-               dst_vb = list_first_entry(&dst_q->done_list, struct vb2_buffer,
-                                               done_entry);
-       if (dst_vb && (dst_vb->state == VB2_BUF_STATE_DONE
-                       || dst_vb->state == VB2_BUF_STATE_ERROR))
                rc |= EPOLLIN | EPOLLRDNORM;
+       }
        spin_unlock_irqrestore(&dst_q->done_lock, flags);
 
        return rc;
--

2.28.0

#6 lore raw

Re: [PATCH 2/2] media: v4l2-mem2mem: simplify poll logic a bit

Hi Ezequiel, thanks for the review!

Quoted 61 lines · Ezequiel Garcia

Hi Alexandre,

Quoted 5 lines · Alexandre Courbot

Factorize redundant checks into a single code block, remove the early return, and declare variables in their innermost block. Hopefully this makes this code a little bit easier to follow.

This _definitely_ makes the poll handling more readable.

See below a nitpick.

Quoted 45 lines
---
 drivers/media/v4l2-core/v4l2-mem2mem.c | 35 +++++++++++---------------
 1 file changed, 15 insertions(+), 20 deletions(-)
drivers/media/v4l2-core/v4l2-mem2mem.c +14 -8
index 0d0192119af20..aeac9707123d0 100644
@@ -841,7 +841,6 @@ static __poll_t v4l2_m2m_poll_for_data(struct file *file,
                                       struct poll_table_struct *wait)
 {
        struct vb2_queue *src_q, *dst_q;
-       struct vb2_buffer *src_vb = NULL, *dst_vb = NULL;
        __poll_t rc = 0;
        unsigned long flags;
 
@@ -863,33 +862,29 @@ static __poll_t v4l2_m2m_poll_for_data(struct file *file,
                return EPOLLERR;
 
        spin_lock_irqsave(&src_q->done_lock, flags);
-       if (!list_empty(&src_q->done_list))
-               src_vb = list_first_entry(&src_q->done_list, struct vb2_buffer,
-                                               done_entry);
-       if (src_vb && (src_vb->state == VB2_BUF_STATE_DONE
-                       || src_vb->state == VB2_BUF_STATE_ERROR))
-               rc |= EPOLLOUT | EPOLLWRNORM;
+       if (!list_empty(&src_q->done_list)) {
+               struct vb2_buffer *src_vb = list_first_entry(
+                       &src_q->done_list, struct vb2_buffer, done_entry);
+               if (src_vb->state == VB2_BUF_STATE_DONE ||
+                   src_vb->state == VB2_BUF_STATE_ERROR)
+                       rc |= EPOLLOUT | EPOLLWRNORM;
+       }
        spin_unlock_irqrestore(&src_q->done_lock, flags);
 
        spin_lock_irqsave(&dst_q->done_lock, flags);
-       if (list_empty(&dst_q->done_list)) {
+       if (!list_empty(&dst_q->done_list)) {
+               struct vb2_buffer *dst_vb = list_first_entry(
+                       &dst_q->done_list, struct vb2_buffer, done_entry);
+               if (dst_vb->state == VB2_BUF_STATE_DONE ||
+                   dst_vb->state == VB2_BUF_STATE_ERROR)
+                       rc |= EPOLLIN | EPOLLRDNORM;
+       } else if (dst_q->last_buffer_dequeued) {
Quoted 8 lines

drivers/media/v4l2-core/v4l2-mem2mem.c static __poll_t v4l2_m2m_poll_for_data(struct file *file, · line 885

Quoted 4 lines
drivers/media/v4l2-core/v4l2-mem2mem.c +0 -0
                /*
                 * If the last buffer was dequeued from the capture queue,
                 * return immediately. DQBUF will return -EPIPE.
                 */

The part about "returning immediately" doesn't make much sense now. Could we rephrase this, keeping the -EPIPE comment?

I understood this sentence as referring to the system call and not just this function, but maybe we can rephrase this as "... make user-space wake up immediately"?

Quoted 24 lines

Thanks,
Ezequiel

Quoted 20 lines
drivers/media/v4l2-core/v4l2-mem2mem.c +1 -13
-               if (dst_q->last_buffer_dequeued) {
-                       spin_unlock_irqrestore(&dst_q->done_lock, flags);
-                       rc |= EPOLLIN | EPOLLRDNORM;
-                       return rc;
-               }
-       }
-
-       if (!list_empty(&dst_q->done_list))
-               dst_vb = list_first_entry(&dst_q->done_list, struct vb2_buffer,
-                                               done_entry);
-       if (dst_vb && (dst_vb->state == VB2_BUF_STATE_DONE
-                       || dst_vb->state == VB2_BUF_STATE_ERROR))
                rc |= EPOLLIN | EPOLLRDNORM;
+       }
        spin_unlock_irqrestore(&dst_q->done_lock, flags);
 
        return rc;
--

2.28.0

#7 lore raw

Re: [PATCH 2/2] media: v4l2-mem2mem: simplify poll logic a bit

Quoted 75 lines · Alexandre Courbot

Hi Ezequiel, thanks for the review!

Quoted 61 lines · Ezequiel Garcia

Hi Alexandre,

Quoted 5 lines · Alexandre Courbot

Factorize redundant checks into a single code block, remove the early return, and declare variables in their innermost block. Hopefully this makes this code a little bit easier to follow.

This _definitely_ makes the poll handling more readable.

See below a nitpick.

Quoted 45 lines
---
 drivers/media/v4l2-core/v4l2-mem2mem.c | 35 +++++++++++---------------
 1 file changed, 15 insertions(+), 20 deletions(-)
drivers/media/v4l2-core/v4l2-mem2mem.c +14 -8
index 0d0192119af20..aeac9707123d0 100644
@@ -841,7 +841,6 @@ static __poll_t v4l2_m2m_poll_for_data(struct file *file,
                                       struct poll_table_struct *wait)
 {
        struct vb2_queue *src_q, *dst_q;
-       struct vb2_buffer *src_vb = NULL, *dst_vb = NULL;
        __poll_t rc = 0;
        unsigned long flags;
 
@@ -863,33 +862,29 @@ static __poll_t v4l2_m2m_poll_for_data(struct file *file,
                return EPOLLERR;
 
        spin_lock_irqsave(&src_q->done_lock, flags);
-       if (!list_empty(&src_q->done_list))
-               src_vb = list_first_entry(&src_q->done_list, struct vb2_buffer,
-                                               done_entry);
-       if (src_vb && (src_vb->state == VB2_BUF_STATE_DONE
-                       || src_vb->state == VB2_BUF_STATE_ERROR))
-               rc |= EPOLLOUT | EPOLLWRNORM;
+       if (!list_empty(&src_q->done_list)) {
+               struct vb2_buffer *src_vb = list_first_entry(
+                       &src_q->done_list, struct vb2_buffer, done_entry);
+               if (src_vb->state == VB2_BUF_STATE_DONE ||
+                   src_vb->state == VB2_BUF_STATE_ERROR)
+                       rc |= EPOLLOUT | EPOLLWRNORM;
+       }
        spin_unlock_irqrestore(&src_q->done_lock, flags);
 
        spin_lock_irqsave(&dst_q->done_lock, flags);
-       if (list_empty(&dst_q->done_list)) {
+       if (!list_empty(&dst_q->done_list)) {
+               struct vb2_buffer *dst_vb = list_first_entry(
+                       &dst_q->done_list, struct vb2_buffer, done_entry);
+               if (dst_vb->state == VB2_BUF_STATE_DONE ||
+                   dst_vb->state == VB2_BUF_STATE_ERROR)
+                       rc |= EPOLLIN | EPOLLRDNORM;
+       } else if (dst_q->last_buffer_dequeued) {
Quoted 8 lines

drivers/media/v4l2-core/v4l2-mem2mem.c static __poll_t v4l2_m2m_poll_for_data(struct file *file, · line 885

Quoted 4 lines
drivers/media/v4l2-core/v4l2-mem2mem.c +0 -0
                /*
                 * If the last buffer was dequeued from the capture queue,
                 * return immediately. DQBUF will return -EPIPE.
                 */

The part about "returning immediately" doesn't make much sense now. Could we rephrase this, keeping the -EPIPE comment?

Quoted 4 lines

I understood this sentence as referring to the system call and not just this function, but maybe we can rephrase this as "... make user-space wake up immediately"?

But is this really about user-space wakeup? I am under the impression that past poll_wait on both queues, we are already about to return (and wakeup).

The way I see it, the original commit intention was to skip any done_list handling, returning immediately on the last buffer condition.

How about just

"""
If the last buffer was dequeued from the capture queue, signal userspace. DQBUF will return -EPIPE. """

?

Quoted 24 lines
Quoted 24 lines

Thanks,
Ezequiel

Quoted 20 lines
drivers/media/v4l2-core/v4l2-mem2mem.c +1 -13
-               if (dst_q->last_buffer_dequeued) {
-                       spin_unlock_irqrestore(&dst_q->done_lock, flags);
-                       rc |= EPOLLIN | EPOLLRDNORM;
-                       return rc;
-               }
-       }
-
-       if (!list_empty(&dst_q->done_list))
-               dst_vb = list_first_entry(&dst_q->done_list, struct vb2_buffer,
-                                               done_entry);
-       if (dst_vb && (dst_vb->state == VB2_BUF_STATE_DONE
-                       || dst_vb->state == VB2_BUF_STATE_ERROR))
                rc |= EPOLLIN | EPOLLRDNORM;
+       }
        spin_unlock_irqrestore(&dst_q->done_lock, flags);
 
        return rc;
--

2.28.0

#8 lore raw

Re: [PATCH 2/2] media: v4l2-mem2mem: simplify poll logic a bit

Quoted 85 lines · Ezequiel Garcia
Quoted 75 lines · Alexandre Courbot

Hi Ezequiel, thanks for the review!

Quoted 61 lines · Ezequiel Garcia

Hi Alexandre,

Quoted 5 lines · Alexandre Courbot

Factorize redundant checks into a single code block, remove the early return, and declare variables in their innermost block. Hopefully this makes this code a little bit easier to follow.

This _definitely_ makes the poll handling more readable.

See below a nitpick.

Quoted 45 lines
---
 drivers/media/v4l2-core/v4l2-mem2mem.c | 35 +++++++++++---------------
 1 file changed, 15 insertions(+), 20 deletions(-)
drivers/media/v4l2-core/v4l2-mem2mem.c +14 -8
index 0d0192119af20..aeac9707123d0 100644
@@ -841,7 +841,6 @@ static __poll_t v4l2_m2m_poll_for_data(struct file *file,
                                       struct poll_table_struct *wait)
 {
        struct vb2_queue *src_q, *dst_q;
-       struct vb2_buffer *src_vb = NULL, *dst_vb = NULL;
        __poll_t rc = 0;
        unsigned long flags;
 
@@ -863,33 +862,29 @@ static __poll_t v4l2_m2m_poll_for_data(struct file *file,
                return EPOLLERR;
 
        spin_lock_irqsave(&src_q->done_lock, flags);
-       if (!list_empty(&src_q->done_list))
-               src_vb = list_first_entry(&src_q->done_list, struct vb2_buffer,
-                                               done_entry);
-       if (src_vb && (src_vb->state == VB2_BUF_STATE_DONE
-                       || src_vb->state == VB2_BUF_STATE_ERROR))
-               rc |= EPOLLOUT | EPOLLWRNORM;
+       if (!list_empty(&src_q->done_list)) {
+               struct vb2_buffer *src_vb = list_first_entry(
+                       &src_q->done_list, struct vb2_buffer, done_entry);
+               if (src_vb->state == VB2_BUF_STATE_DONE ||
+                   src_vb->state == VB2_BUF_STATE_ERROR)
+                       rc |= EPOLLOUT | EPOLLWRNORM;
+       }
        spin_unlock_irqrestore(&src_q->done_lock, flags);
 
        spin_lock_irqsave(&dst_q->done_lock, flags);
-       if (list_empty(&dst_q->done_list)) {
+       if (!list_empty(&dst_q->done_list)) {
+               struct vb2_buffer *dst_vb = list_first_entry(
+                       &dst_q->done_list, struct vb2_buffer, done_entry);
+               if (dst_vb->state == VB2_BUF_STATE_DONE ||
+                   dst_vb->state == VB2_BUF_STATE_ERROR)
+                       rc |= EPOLLIN | EPOLLRDNORM;
+       } else if (dst_q->last_buffer_dequeued) {
Quoted 8 lines

drivers/media/v4l2-core/v4l2-mem2mem.c static __poll_t v4l2_m2m_poll_for_data(struct file *file, · line 885

Quoted 4 lines
drivers/media/v4l2-core/v4l2-mem2mem.c +0 -0
                /*
                 * If the last buffer was dequeued from the capture queue,
                 * return immediately. DQBUF will return -EPIPE.
                 */

The part about "returning immediately" doesn't make much sense now. Could we rephrase this, keeping the -EPIPE comment?

Quoted 4 lines

I understood this sentence as referring to the system call and not just this function, but maybe we can rephrase this as "... make user-space wake up immediately"?

But is this really about user-space wakeup? I am under the impression that past poll_wait on both queues, we are already about to return (and wakeup).

Quoted 8 lines

The way I see it, the original commit intention was to skip any done_list handling, returning immediately on the last buffer condition.

How about just

"""
If the last buffer was dequeued from the capture queue, signal userspace. DQBUF will return -EPIPE.

I'd write 'DQBUF(CAPTURE)' here to emphasize that only the capture queue will return -EPIPE when you try to dequeue from it.

Also note that the original text was a copy-and-paste from vb2_core_poll(). The phrase 'return immediately' makes sense in that context since that poll code deals with a single queue. In this case you have two queues, and 'return immediately' no longer applies (in fact, that effectively is the bug that being fixed here!).

Regards,

Hans

Quoted 28 lines

"""

?

Quoted 24 lines
Quoted 24 lines

Thanks,
Ezequiel

Quoted 20 lines
drivers/media/v4l2-core/v4l2-mem2mem.c +1 -13
-               if (dst_q->last_buffer_dequeued) {
-                       spin_unlock_irqrestore(&dst_q->done_lock, flags);
-                       rc |= EPOLLIN | EPOLLRDNORM;
-                       return rc;
-               }
-       }
-
-       if (!list_empty(&dst_q->done_list))
-               dst_vb = list_first_entry(&dst_q->done_list, struct vb2_buffer,
-                                               done_entry);
-       if (dst_vb && (dst_vb->state == VB2_BUF_STATE_DONE
-                       || dst_vb->state == VB2_BUF_STATE_ERROR))
                rc |= EPOLLIN | EPOLLRDNORM;
+       }
        spin_unlock_irqrestore(&dst_q->done_lock, flags);
 
        return rc;
--

2.28.0

#9 lore raw

Re: [PATCH 2/2] media: v4l2-mem2mem: simplify poll logic a bit

In reply to Alexandre Courbot
Quoted 26 lines · Alexandre Courbot

Factorize redundant checks into a single code block, remove the early return, and declare variables in their innermost block. Hopefully this makes this code a little bit easier to follow.

---
 drivers/media/v4l2-core/v4l2-mem2mem.c | 35 +++++++++++---------------
 1 file changed, 15 insertions(+), 20 deletions(-)
drivers/media/v4l2-core/v4l2-mem2mem.c +0 -2
index 0d0192119af20..aeac9707123d0 100644
@@ -841,7 +841,6 @@ static __poll_t v4l2_m2m_poll_for_data(struct file *file,
 				       struct poll_table_struct *wait)
 {
 	struct vb2_queue *src_q, *dst_q;
-	struct vb2_buffer *src_vb = NULL, *dst_vb = NULL;
 	__poll_t rc = 0;
 	unsigned long flags;
 
@@ -863,33 +862,29 @@ static __poll_t v4l2_m2m_poll_for_data(struct file *file,
 		return EPOLLERR;
 
 	spin_lock_irqsave(&src_q->done_lock, flags);
-	if (!list_empty(&src_q->done_list))

drivers/media/v4l2-core/v4l2-mem2mem.c static __poll_t v4l2_m2m_poll_for_data(struct file *file, · line 869

Quoted 10 lines
drivers/media/v4l2-core/v4l2-mem2mem.c +5 -5
-		src_vb = list_first_entry(&src_q->done_list, struct vb2_buffer,
-						done_entry);
-	if (src_vb && (src_vb->state == VB2_BUF_STATE_DONE
-			|| src_vb->state == VB2_BUF_STATE_ERROR))
-		rc |= EPOLLOUT | EPOLLWRNORM;
+	if (!list_empty(&src_q->done_list)) {
+		struct vb2_buffer *src_vb = list_first_entry(
+			&src_q->done_list, struct vb2_buffer, done_entry);
+		if (src_vb->state == VB2_BUF_STATE_DONE ||
+		    src_vb->state == VB2_BUF_STATE_ERROR)

This test is unnecessary: only buffers in state DONE or ERROR can be on the done_list.

drivers/media/v4l2-core/v4l2-mem2mem.c static __poll_t v4l2_m2m_poll_for_data(struct file *file, · line 879

Quoted 11 lines
drivers/media/v4l2-core/v4l2-mem2mem.c +7 -1
+			rc |= EPOLLOUT | EPOLLWRNORM;
+	}
 	spin_unlock_irqrestore(&src_q->done_lock, flags);
 
 	spin_lock_irqsave(&dst_q->done_lock, flags);
-	if (list_empty(&dst_q->done_list)) {
+	if (!list_empty(&dst_q->done_list)) {
+		struct vb2_buffer *dst_vb = list_first_entry(
+			&dst_q->done_list, struct vb2_buffer, done_entry);
+		if (dst_vb->state == VB2_BUF_STATE_DONE ||
+		    dst_vb->state == VB2_BUF_STATE_ERROR)

Ditto.

Regards,

Hans

Quoted 24 lines
drivers/media/v4l2-core/v4l2-mem2mem.c +3 -12
+			rc |= EPOLLIN | EPOLLRDNORM;
+	} else if (dst_q->last_buffer_dequeued) {
 		/*
 		 * If the last buffer was dequeued from the capture queue,
 		 * return immediately. DQBUF will return -EPIPE.
 		 */
-		if (dst_q->last_buffer_dequeued) {
-			spin_unlock_irqrestore(&dst_q->done_lock, flags);
-			rc |= EPOLLIN | EPOLLRDNORM;
-			return rc;
-		}
-	}
-
-	if (!list_empty(&dst_q->done_list))
-		dst_vb = list_first_entry(&dst_q->done_list, struct vb2_buffer,
-						done_entry);
-	if (dst_vb && (dst_vb->state == VB2_BUF_STATE_DONE
-			|| dst_vb->state == VB2_BUF_STATE_ERROR))
 		rc |= EPOLLIN | EPOLLRDNORM;
+	}
 	spin_unlock_irqrestore(&dst_q->done_lock, flags);
 
 	return rc;
#10 lore raw
In reply to Alexandre Courbot

Hello Alex,

Thanks for the patch.

Quoted 4 lines · Alexandre Courbot

This addresses a very corner case that probably nobody ever encounters, but I have hit it when playing with vicoded so here is a tentative fix.

I'll try to make a more complete review soon, but meanwhile
I was thinking if it was possible to include a little kselftest program for this issue, something CIs can pick-up and test corner cases like this, making sure we don't regress on the issue.

(Or alternatively, v4l2-compliance?)

Thanks!
Ezequiel

Quoted 21 lines

Patch 1/2 addresses the issue that when the last buffer of a m2m device has been dequeued, any attempt to poll with EPOLLOUT will result in only EPOLLIN being returned, even if OUTPUT buffers are still pending. The issue stems from the fact that the last buffer check if done first, and returns immediately if true.

Patch 2/2 builds on the first one to (hopefully) clean up the code a bit and make the function flow easier to follow. Functionally speaking it is supposed to be a no-op and it can safely be dropped if the former code is preferred - the actual fix is in 1/2.

Alexandre Courbot (2):
media: v4l2-mem2mem: consider OUTPUT queue first when polling
media: v4l2-mem2mem: simplify poll logic a bit

 drivers/media/v4l2-core/v4l2-mem2mem.c | 42 +++++++++++---------------
 1 file changed, 18 insertions(+), 24 deletions(-)

--
2.28.0

#11 lore raw
Quoted 11 lines · Ezequiel Garcia

Hello Alex,

Thanks for the patch.

Quoted 4 lines · Alexandre Courbot

This addresses a very corner case that probably nobody ever encounters, but I have hit it when playing with vicoded so here is a tentative fix.

Quoted 6 lines

I'll try to make a more complete review soon, but meanwhile
I was thinking if it was possible to include a little kselftest program for this issue, something CIs can pick-up and test corner cases like this, making sure we don't regress on the issue.

(Or alternatively, v4l2-compliance?)

I am not very familiar with kselftest, but IIUC the conditions that lead to this issue are easier to reproduce using v4l2-compliance, where we can have a user-space driving the queues. It would also have the benefit to exercise all drivers. I'll think about adding such a test, thanks for the suggestion!

Quoted 25 lines

Thanks!
Ezequiel

Quoted 21 lines

Patch 1/2 addresses the issue that when the last buffer of a m2m device has been dequeued, any attempt to poll with EPOLLOUT will result in only EPOLLIN being returned, even if OUTPUT buffers are still pending. The issue stems from the fact that the last buffer check if done first, and returns immediately if true.

Patch 2/2 builds on the first one to (hopefully) clean up the code a bit and make the function flow easier to follow. Functionally speaking it is supposed to be a no-op and it can safely be dropped if the former code is preferred - the actual fix is in 1/2.

Alexandre Courbot (2):
media: v4l2-mem2mem: consider OUTPUT queue first when polling
media: v4l2-mem2mem: simplify poll logic a bit

 drivers/media/v4l2-core/v4l2-mem2mem.c | 42 +++++++++++---------------
 1 file changed, 18 insertions(+), 24 deletions(-)

--
2.28.0