diff mbox

[v2,2/2] dma-buf: Add user interfaces for dmabuf sync support

Message ID 1377081215-5948-3-git-send-email-inki.dae@samsung.com (mailing list archive)
State New, archived
Headers show

Commit Message

Inki Dae Aug. 21, 2013, 10:33 a.m. UTC
This patch adds lock and poll callbacks to dma buf file operations,
and these callbacks will be called by fcntl and select system calls.

fcntl and select system calls can be used to wait for the completion
of DMA or CPU access to a shared dmabuf. The difference of them is
fcntl system call takes a lock after the completion but select system
call doesn't. So in case of fcntl system call, it's useful when a task
wants to access a shared dmabuf without any broken. On the other hand,
it's useful when a task wants to just wait for the completion.

Changelog v2:
- Add select system call support.
  . The purpose of this feature is to wait for the completion of DMA or
    CPU access to a dmabuf without that caller locks the dmabuf again
    after the completion.
    That is useful when caller wants to be aware of the completion of
    DMA access to the dmabuf, and the caller doesn't use intefaces for
    the DMA device driver.

Signed-off-by: Inki Dae <inki.dae@samsung.com>
Signed-off-by: Kyungmin Park <kyungmin.park@samsung.com>
---
 drivers/base/dma-buf.c |   81 ++++++++++++++++++++++++++++++++++++++++++++++++
 1 files changed, 81 insertions(+), 0 deletions(-)

Comments

David Herrmann Aug. 21, 2013, 1:17 p.m. UTC | #1
Hi

On Wed, Aug 21, 2013 at 12:33 PM, Inki Dae <inki.dae@samsung.com> wrote:
> This patch adds lock and poll callbacks to dma buf file operations,
> and these callbacks will be called by fcntl and select system calls.
>
> fcntl and select system calls can be used to wait for the completion
> of DMA or CPU access to a shared dmabuf. The difference of them is
> fcntl system call takes a lock after the completion but select system
> call doesn't. So in case of fcntl system call, it's useful when a task
> wants to access a shared dmabuf without any broken. On the other hand,
> it's useful when a task wants to just wait for the completion.

1)
So how is that supposed to work in user-space? I don't want to block
on a buffer, but get notified once I can lock it. So I do:
  select(..dmabuf..)
Once it is finished, I want to use it:
  flock(..dmabuf..)
However, how can I guarantee the flock will not block? Some other
process might have locked it in between. So I do a non-blocking
flock() and if it fails I wait again? Looks ugly and un-predictable.

2)
What do I do if some user-space program holds a lock and dead-locks?

3)
How do we do modesetting in atomic-context in the kernel? There is no
way to lock the object. But this is required for panic-handlers and
more importantly the kdb debugging hooks.
Ok, I can live with that being racy, but would still be nice to be considered.

4)
Why do we need locks? Aren't fences enough? That is, in which
situation is a lock really needed?
If we assume we have two writers A and B (DMA, CPU, GPU, whatever) and
they have no synchronization on their own. What do we win by
synchronizing their writes? Ok, yeah, we end up with either A or B and
not a mixture of both. But if we cannot predict whether we get A or B,
I don't know why we care at all? It's random, so a mixture would be
fine, too, wouldn't it?

So if user-space doesn't have any synchronization on its own, I don't
see why we need an implicit sync on a dma-buf. Could you describe a
more elaborate use-case?

I think the problems we need to fix are read/write syncs. So we have a
write that issues the DMA+write plus a fence and passes the buf plus
fence to the reader. The reader waits for the fence and then issues
the read plus fence. It passes the fence back to the writer. The
writer waits for the fence again and then issues the next write if
required.

This has the following advantages:
 - fences are _guaranteed_ to finish in a given time period. Locks, on
the other hand, might never be freed (of the holder dead-locks, for
instance)
 - you avoid any stalls. That is, if a writer releases a buffer and
immediately locks it again, the reader side might stall if it didn't
lock it in exactly the given window. You have no control to guarantee
the reader ever gets access. You would need a synchronization in
user-space between the writer and reader to guarantee that. This makes
the whole lock useles, doesn't it?

Cheers
David

> Changelog v2:
> - Add select system call support.
>   . The purpose of this feature is to wait for the completion of DMA or
>     CPU access to a dmabuf without that caller locks the dmabuf again
>     after the completion.
>     That is useful when caller wants to be aware of the completion of
>     DMA access to the dmabuf, and the caller doesn't use intefaces for
>     the DMA device driver.
>
> Signed-off-by: Inki Dae <inki.dae@samsung.com>
> Signed-off-by: Kyungmin Park <kyungmin.park@samsung.com>
> ---
>  drivers/base/dma-buf.c |   81 ++++++++++++++++++++++++++++++++++++++++++++++++
>  1 files changed, 81 insertions(+), 0 deletions(-)
>
> diff --git a/drivers/base/dma-buf.c b/drivers/base/dma-buf.c
> index 4aca57a..f16a396 100644
> --- a/drivers/base/dma-buf.c
> +++ b/drivers/base/dma-buf.c
> @@ -29,6 +29,7 @@
>  #include <linux/export.h>
>  #include <linux/debugfs.h>
>  #include <linux/seq_file.h>
> +#include <linux/poll.h>
>  #include <linux/dmabuf-sync.h>
>
>  static inline int is_dma_buf_file(struct file *);
> @@ -80,9 +81,89 @@ static int dma_buf_mmap_internal(struct file *file, struct vm_area_struct *vma)
>         return dmabuf->ops->mmap(dmabuf, vma);
>  }
>
> +static unsigned int dma_buf_poll(struct file *filp,
> +                                       struct poll_table_struct *poll)
> +{
> +       struct dma_buf *dmabuf;
> +       struct dmabuf_sync_reservation *robj;
> +       int ret = 0;
> +
> +       if (!is_dma_buf_file(filp))
> +               return POLLERR;
> +
> +       dmabuf = filp->private_data;
> +       if (!dmabuf || !dmabuf->sync)
> +               return POLLERR;
> +
> +       robj = dmabuf->sync;
> +
> +       mutex_lock(&robj->lock);
> +
> +       robj->polled = true;
> +
> +       /*
> +        * CPU or DMA access to this buffer has been completed, and
> +        * the blocked task has been waked up. Return poll event
> +        * so that the task can get out of select().
> +        */
> +       if (robj->poll_event) {
> +               robj->poll_event = false;
> +               mutex_unlock(&robj->lock);
> +               return POLLIN | POLLOUT;
> +       }
> +
> +       /*
> +        * There is no anyone accessing this buffer so just return.
> +        */
> +       if (!robj->locked) {
> +               mutex_unlock(&robj->lock);
> +               return POLLIN | POLLOUT;
> +       }
> +
> +       poll_wait(filp, &robj->poll_wait, poll);
> +
> +       mutex_unlock(&robj->lock);
> +
> +       return ret;
> +}
> +
> +static int dma_buf_lock(struct file *file, int cmd, struct file_lock *fl)
> +{
> +       struct dma_buf *dmabuf;
> +       unsigned int type;
> +       bool wait = false;
> +
> +       if (!is_dma_buf_file(file))
> +               return -EINVAL;
> +
> +       dmabuf = file->private_data;
> +
> +       if ((fl->fl_type & F_UNLCK) == F_UNLCK) {
> +               dmabuf_sync_single_unlock(dmabuf);
> +               return 0;
> +       }
> +
> +       /* convert flock type to dmabuf sync type. */
> +       if ((fl->fl_type & F_WRLCK) == F_WRLCK)
> +               type = DMA_BUF_ACCESS_W;
> +       else if ((fl->fl_type & F_RDLCK) == F_RDLCK)
> +               type = DMA_BUF_ACCESS_R;
> +       else
> +               return -EINVAL;
> +
> +       if (fl->fl_flags & FL_SLEEP)
> +               wait = true;
> +
> +       /* TODO. the locking to certain region should also be considered. */
> +
> +       return dmabuf_sync_single_lock(dmabuf, type, wait);
> +}
> +
>  static const struct file_operations dma_buf_fops = {
>         .release        = dma_buf_release,
>         .mmap           = dma_buf_mmap_internal,
> +       .poll           = dma_buf_poll,
> +       .lock           = dma_buf_lock,
>  };
>
>  /*
> --
> 1.7.5.4
>
> --
> To unsubscribe from this list: send the line "unsubscribe linux-fbdev" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
Inki Dae Aug. 22, 2013, 8:46 a.m. UTC | #2
Thanks for your comments,
Inki Dae

> -----Original Message-----
> From: David Herrmann [mailto:dh.herrmann@gmail.com]
> Sent: Wednesday, August 21, 2013 10:17 PM
> To: Inki Dae
> Cc: dri-devel@lists.freedesktop.org; linux-fbdev@vger.kernel.org; linux-
> arm-kernel@lists.infradead.org; linux-media@vger.kernel.org; linaro-
> kernel@lists.linaro.org; Maarten Lankhorst; Sumit Semwal;
> kyungmin.park@samsung.com; myungjoo.ham@samsung.com
> Subject: Re: [PATCH v2 2/2] dma-buf: Add user interfaces for dmabuf sync
> support
> 
> Hi
> 
> On Wed, Aug 21, 2013 at 12:33 PM, Inki Dae <inki.dae@samsung.com> wrote:
> > This patch adds lock and poll callbacks to dma buf file operations,
> > and these callbacks will be called by fcntl and select system calls.
> >
> > fcntl and select system calls can be used to wait for the completion
> > of DMA or CPU access to a shared dmabuf. The difference of them is
> > fcntl system call takes a lock after the completion but select system
> > call doesn't. So in case of fcntl system call, it's useful when a task
> > wants to access a shared dmabuf without any broken. On the other hand,
> > it's useful when a task wants to just wait for the completion.
> 
> 1)
> So how is that supposed to work in user-space? I don't want to block
> on a buffer, but get notified once I can lock it. So I do:
>   select(..dmabuf..)
> Once it is finished, I want to use it:
>   flock(..dmabuf..)
> However, how can I guarantee the flock will not block? Some other
> process might have locked it in between. So I do a non-blocking
> flock() and if it fails I wait again?

s/flock/fcntl

Yes, it does if you wanted to do a non-blocking fcntl. The fcntl() call will
return -EAGAIN if some other process have locked first. So user process
could retry to lock or do other work. This user process called fcntl() with
non-blocking mode so in this case, I think the user should consider two
things. One is that the fcntl() call couldn't be failed, and other is that
the call could take a lock successfully. Isn't fcntl() with a other fd also,
not dmabuf, take a same action?

>Looks ugly and un-predictable.
> 

So I think this is reasonable. However, for select system call, I'm not sure
that this system call is needed yet. So I can remove it if unnecessary.

> 2)
> What do I do if some user-space program holds a lock and dead-locks?
> 

I think fcntl call with a other fd also could lead same situation, and the
lock will be unlocked once the user-space program is killed because when the
process is killed, all file descriptors of the process are closed.

> 3)
> How do we do modesetting in atomic-context in the kernel? There is no
> way to lock the object. But this is required for panic-handlers and
> more importantly the kdb debugging hooks.
> Ok, I can live with that being racy, but would still be nice to be
> considered.

Yes,  The lock shouldn't be called in the atomic-context. For this, will add
comments enough.

> 
> 4)
> Why do we need locks? Aren't fences enough? That is, in which
> situation is a lock really needed?
> If we assume we have two writers A and B (DMA, CPU, GPU, whatever) and
> they have no synchronization on their own. What do we win by
> synchronizing their writes? Ok, yeah, we end up with either A or B and
> not a mixture of both. But if we cannot predict whether we get A or B,
> I don't know why we care at all? It's random, so a mixture would be
> fine, too, wouldn't it?

I think not so. There are some cases that the mixture wouldn't be fine. For
this, will describe it at below.

> 
> So if user-space doesn't have any synchronization on its own, I don't
> see why we need an implicit sync on a dma-buf. Could you describe a
> more elaborate use-case?

Ok, first, I think I described that enough though [PATCH 0/2]. For this, you
can refer to the below link,
http://lwn.net/Articles/564208/ 

Anyway, there are some cases that user-space process needs the
synchronization on its own. In case of Tizen platform[1], one is between X
Client and X Server; actually, Composite Manager. Other is between Web app
based on HTML5 and Web Browser.

Please, assume that X Client draws something in a window buffer using CPU,
and then the X Client requests SWAP to X Server. And then X Server notifies
a damage event to Composite Manager. And then Composite Manager composes the
window buffer with its own back buffer using GPU. In this case, Composite
Manager calls eglSwapBuffers; internally, flushing GL commands instead of
finishing them for more performance.

As you may know, the flushing doesn't wait for the complete event from GPU
driver. And at the same time, the X Client could do other work, and also
draw something in the same buffer again. At this time, The buffer could be
broken. Because the X Client can't aware of when GPU access to the buffer is
completed without out-of-band hand shaking; the out-of-band hand shaking is
quite big overhead. That is why we need user-space locking interface, fcntl
system call.

And also there is same issue in case of Web app: Web app draws something in
a buffer using CPU, and Web browser composes the buffer with its own buffer
using GPU. At the above, you mentioned that a mixture would be fine but it's
not fine with such reasons.

[1] https://www.tizen.org/

> 
> I think the problems we need to fix are read/write syncs. So we have a
> write that issues the DMA+write plus a fence and passes the buf plus
> fence to the reader. The reader waits for the fence and then issues
> the read plus fence. It passes the fence back to the writer. The
> writer waits for the fence again and then issues the next write if
> required.
> 
> This has the following advantages:
>  - fences are _guaranteed_ to finish in a given time period. Locks, on
> the other hand, might never be freed (of the holder dead-locks, for
> instance)

Not so. If never unlock since locked then the buffer will be unlocked by
timeout worker queue handler in a given time period.

>  - you avoid any stalls. That is, if a writer releases a buffer and
> immediately locks it again, the reader side might stall if it didn't
> lock it in exactly the given window.
> You have no control to guarantee
> the reader ever gets access. You would need a synchronization in
> user-space between the writer and reader to guarantee that. This makes
> the whole lock useles, doesn't it?
> 

Ah..... right. The lock mechanism cannot guarantee the write-and-then-read
order. When the writer tries to take a lock again after the reader tried to
take a lock for read but blocked, the writer don't know the fact that the
reader tried to take a lock for read so in turn, the writer would take a
lock, and the reader side would stall as you mentioned above.

I think we wouldn't need the synchronization in user-space between the write
and reader to guarantee that if we use a mechanism such as DMA fence
additionally; wound/wait mutex to avoid dead lock issue, and DMA fence to
guarantee the write-and-then-read order. For this, I will consider using the
DMA fence. Maybe it takes much time.

Thanks,
Inki Dae

> Cheers
> David
> 
> > Changelog v2:
> > - Add select system call support.
> >   . The purpose of this feature is to wait for the completion of DMA or
> >     CPU access to a dmabuf without that caller locks the dmabuf again
> >     after the completion.
> >     That is useful when caller wants to be aware of the completion of
> >     DMA access to the dmabuf, and the caller doesn't use intefaces for
> >     the DMA device driver.
> >
> > Signed-off-by: Inki Dae <inki.dae@samsung.com>
> > Signed-off-by: Kyungmin Park <kyungmin.park@samsung.com>
> > ---
> >  drivers/base/dma-buf.c |   81
> ++++++++++++++++++++++++++++++++++++++++++++++++
> >  1 files changed, 81 insertions(+), 0 deletions(-)
> >
> > diff --git a/drivers/base/dma-buf.c b/drivers/base/dma-buf.c
> > index 4aca57a..f16a396 100644
> > --- a/drivers/base/dma-buf.c
> > +++ b/drivers/base/dma-buf.c
> > @@ -29,6 +29,7 @@
> >  #include <linux/export.h>
> >  #include <linux/debugfs.h>
> >  #include <linux/seq_file.h>
> > +#include <linux/poll.h>
> >  #include <linux/dmabuf-sync.h>
> >
> >  static inline int is_dma_buf_file(struct file *);
> > @@ -80,9 +81,89 @@ static int dma_buf_mmap_internal(struct file *file,
> struct vm_area_struct *vma)
> >         return dmabuf->ops->mmap(dmabuf, vma);
> >  }
> >
> > +static unsigned int dma_buf_poll(struct file *filp,
> > +                                       struct poll_table_struct *poll)
> > +{
> > +       struct dma_buf *dmabuf;
> > +       struct dmabuf_sync_reservation *robj;
> > +       int ret = 0;
> > +
> > +       if (!is_dma_buf_file(filp))
> > +               return POLLERR;
> > +
> > +       dmabuf = filp->private_data;
> > +       if (!dmabuf || !dmabuf->sync)
> > +               return POLLERR;
> > +
> > +       robj = dmabuf->sync;
> > +
> > +       mutex_lock(&robj->lock);
> > +
> > +       robj->polled = true;
> > +
> > +       /*
> > +        * CPU or DMA access to this buffer has been completed, and
> > +        * the blocked task has been waked up. Return poll event
> > +        * so that the task can get out of select().
> > +        */
> > +       if (robj->poll_event) {
> > +               robj->poll_event = false;
> > +               mutex_unlock(&robj->lock);
> > +               return POLLIN | POLLOUT;
> > +       }
> > +
> > +       /*
> > +        * There is no anyone accessing this buffer so just return.
> > +        */
> > +       if (!robj->locked) {
> > +               mutex_unlock(&robj->lock);
> > +               return POLLIN | POLLOUT;
> > +       }
> > +
> > +       poll_wait(filp, &robj->poll_wait, poll);
> > +
> > +       mutex_unlock(&robj->lock);
> > +
> > +       return ret;
> > +}
> > +
> > +static int dma_buf_lock(struct file *file, int cmd, struct file_lock
> *fl)
> > +{
> > +       struct dma_buf *dmabuf;
> > +       unsigned int type;
> > +       bool wait = false;
> > +
> > +       if (!is_dma_buf_file(file))
> > +               return -EINVAL;
> > +
> > +       dmabuf = file->private_data;
> > +
> > +       if ((fl->fl_type & F_UNLCK) == F_UNLCK) {
> > +               dmabuf_sync_single_unlock(dmabuf);
> > +               return 0;
> > +       }
> > +
> > +       /* convert flock type to dmabuf sync type. */
> > +       if ((fl->fl_type & F_WRLCK) == F_WRLCK)
> > +               type = DMA_BUF_ACCESS_W;
> > +       else if ((fl->fl_type & F_RDLCK) == F_RDLCK)
> > +               type = DMA_BUF_ACCESS_R;
> > +       else
> > +               return -EINVAL;
> > +
> > +       if (fl->fl_flags & FL_SLEEP)
> > +               wait = true;
> > +
> > +       /* TODO. the locking to certain region should also be
considered.
> */
> > +
> > +       return dmabuf_sync_single_lock(dmabuf, type, wait);
> > +}
> > +
> >  static const struct file_operations dma_buf_fops = {
> >         .release        = dma_buf_release,
> >         .mmap           = dma_buf_mmap_internal,
> > +       .poll           = dma_buf_poll,
> > +       .lock           = dma_buf_lock,
> >  };
> >
> >  /*
> > --
> > 1.7.5.4
> >
> > --
> > To unsubscribe from this list: send the line "unsubscribe linux-fbdev"
> in
> > the body of a message to majordomo@vger.kernel.org
> > More majordomo info at  http://vger.kernel.org/majordomo-info.html
diff mbox

Patch

diff --git a/drivers/base/dma-buf.c b/drivers/base/dma-buf.c
index 4aca57a..f16a396 100644
--- a/drivers/base/dma-buf.c
+++ b/drivers/base/dma-buf.c
@@ -29,6 +29,7 @@ 
 #include <linux/export.h>
 #include <linux/debugfs.h>
 #include <linux/seq_file.h>
+#include <linux/poll.h>
 #include <linux/dmabuf-sync.h>
 
 static inline int is_dma_buf_file(struct file *);
@@ -80,9 +81,89 @@  static int dma_buf_mmap_internal(struct file *file, struct vm_area_struct *vma)
 	return dmabuf->ops->mmap(dmabuf, vma);
 }
 
+static unsigned int dma_buf_poll(struct file *filp,
+					struct poll_table_struct *poll)
+{
+	struct dma_buf *dmabuf;
+	struct dmabuf_sync_reservation *robj;
+	int ret = 0;
+
+	if (!is_dma_buf_file(filp))
+		return POLLERR;
+
+	dmabuf = filp->private_data;
+	if (!dmabuf || !dmabuf->sync)
+		return POLLERR;
+
+	robj = dmabuf->sync;
+
+	mutex_lock(&robj->lock);
+
+	robj->polled = true;
+
+	/*
+	 * CPU or DMA access to this buffer has been completed, and
+	 * the blocked task has been waked up. Return poll event
+	 * so that the task can get out of select().
+	 */
+	if (robj->poll_event) {
+		robj->poll_event = false;
+		mutex_unlock(&robj->lock);
+		return POLLIN | POLLOUT;
+	}
+
+	/*
+	 * There is no anyone accessing this buffer so just return.
+	 */
+	if (!robj->locked) {
+		mutex_unlock(&robj->lock);
+		return POLLIN | POLLOUT;
+	}
+
+	poll_wait(filp, &robj->poll_wait, poll);
+
+	mutex_unlock(&robj->lock);
+
+	return ret;
+}
+
+static int dma_buf_lock(struct file *file, int cmd, struct file_lock *fl)
+{
+	struct dma_buf *dmabuf;
+	unsigned int type;
+	bool wait = false;
+
+	if (!is_dma_buf_file(file))
+		return -EINVAL;
+
+	dmabuf = file->private_data;
+
+	if ((fl->fl_type & F_UNLCK) == F_UNLCK) {
+		dmabuf_sync_single_unlock(dmabuf);
+		return 0;
+	}
+
+	/* convert flock type to dmabuf sync type. */
+	if ((fl->fl_type & F_WRLCK) == F_WRLCK)
+		type = DMA_BUF_ACCESS_W;
+	else if ((fl->fl_type & F_RDLCK) == F_RDLCK)
+		type = DMA_BUF_ACCESS_R;
+	else
+		return -EINVAL;
+
+	if (fl->fl_flags & FL_SLEEP)
+		wait = true;
+
+	/* TODO. the locking to certain region should also be considered. */
+
+	return dmabuf_sync_single_lock(dmabuf, type, wait);
+}
+
 static const struct file_operations dma_buf_fops = {
 	.release	= dma_buf_release,
 	.mmap		= dma_buf_mmap_internal,
+	.poll		= dma_buf_poll,
+	.lock		= dma_buf_lock,
 };
 
 /*