2006-10-18 21:36:52

by Stefan Richter

[permalink] [raw]
Subject: Re: Unnecessary BKL contention in video1394

(added Cc: lkml to pull in some clues)

Daniel Drake wrote at linux1394-devel:
> Hi,
>
> I noticed that video1394 calls lock_kernel in it's file_operations. I
> thought about converting these into a per-instance mutex or something,
> but in the end I couldn't find a reason why this locking is needed. (The
> more important I/O system is protected by separate spinlocks)
>
> The BKL is only contended when operations such as ioctl() or release()
> are invoked. My knowledge lacks at this point, but I'm reasonably sure
> that some upper layer must ensure that these invokations are serialized,
> as opposed to leaving it up to the driver to handle nasty potential
> situations e.g. release() being called halfway through a read(). Can
> anyone clarify?
>
> Thanks,
> Daniel

I think you are right. Same with dv1394. Although we need to
double-check whether something needs replacement protection then.

> ------------------------------------------------------------------------
>
> Index: linux/drivers/ieee1394/video1394.c
> ===================================================================
> --- linux.orig/drivers/ieee1394/video1394.c
> +++ linux/drivers/ieee1394/video1394.c
> @@ -1161,9 +1161,7 @@ static int __video1394_ioctl(struct file
> static long video1394_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
> {
> int err;
> - lock_kernel();
> err = __video1394_ioctl(file, cmd, arg);
> - unlock_kernel();
> return err;
> }
>
> @@ -1181,13 +1179,11 @@ static int video1394_mmap(struct file *f
> struct file_ctx *ctx = (struct file_ctx *)file->private_data;
> int res = -EINVAL;
>
> - lock_kernel();
> if (ctx->current_ctx == NULL) {
> PRINT(KERN_ERR, ctx->ohci->host->id,
> "Current iso context not set");
> } else
> res = dma_region_mmap(&ctx->current_ctx->dma, file, vma);
> - unlock_kernel();
>
> return res;
> }
> @@ -1200,7 +1196,6 @@ static unsigned int video1394_poll(struc
> struct dma_iso_ctx *d;
> int i;
>
> - lock_kernel();
> ctx = file->private_data;
> d = ctx->current_ctx;
> if (d == NULL) {
> @@ -1221,7 +1216,6 @@ static unsigned int video1394_poll(struc
> }
> spin_unlock_irqrestore(&d->lock, flags);
> done:
> - unlock_kernel();
>
> return mask;
> }
> @@ -1257,7 +1251,6 @@ static int video1394_release(struct inod
> struct list_head *lh, *next;
> u64 mask;
>
> - lock_kernel();
> list_for_each_safe(lh, next, &ctx->context_list) {
> struct dma_iso_ctx *d;
> d = list_entry(lh, struct dma_iso_ctx, link);
> @@ -1278,7 +1271,6 @@ static int video1394_release(struct inod
> kfree(ctx);
> file->private_data = NULL;
>
> - unlock_kernel();
> return 0;
> }
>
>

--
Stefan Richter
-=====-=-==- =-=- =--=-
http://arcgraph.de/sr/


2006-10-19 13:19:40

by Daniel Drake

[permalink] [raw]
Subject: Re: Unnecessary BKL contention in video1394

On Wed, 2006-10-18 at 23:36 +0200, Stefan Richter wrote:
> (added Cc: lkml to pull in some clues)
>
> Daniel Drake wrote at linux1394-devel:
> > Hi,
> >
> > I noticed that video1394 calls lock_kernel in it's file_operations. I
> > thought about converting these into a per-instance mutex or something,
> > but in the end I couldn't find a reason why this locking is needed. (The
> > more important I/O system is protected by separate spinlocks)
> >
> > The BKL is only contended when operations such as ioctl() or release()
> > are invoked. My knowledge lacks at this point, but I'm reasonably sure
> > that some upper layer must ensure that these invokations are serialized,
> > as opposed to leaving it up to the driver to handle nasty potential
> > situations e.g. release() being called halfway through a read(). Can
> > anyone clarify?

> I think you are right. Same with dv1394. Although we need to
> double-check whether something needs replacement protection then.

Adding Andi Kleen to CC, who added the BKL around __video1394_ioctl a
long while back (when converting video1394 to compat_ioctl).

I don't feel that any replacement protection is needed, since the
critical sections (where structures are used both in interrupts and in
file_operations) are already protected by spinlocks.

> > ------------------------------------------------------------------------
> >
> > Index: linux/drivers/ieee1394/video1394.c
> > ===================================================================
> > --- linux.orig/drivers/ieee1394/video1394.c
> > +++ linux/drivers/ieee1394/video1394.c
> > @@ -1161,9 +1161,7 @@ static int __video1394_ioctl(struct file
> > static long video1394_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
> > {
> > int err;
> > - lock_kernel();
> > err = __video1394_ioctl(file, cmd, arg);
> > - unlock_kernel();
> > return err;
> > }
> >
> > @@ -1181,13 +1179,11 @@ static int video1394_mmap(struct file *f
> > struct file_ctx *ctx = (struct file_ctx *)file->private_data;
> > int res = -EINVAL;
> >
> > - lock_kernel();
> > if (ctx->current_ctx == NULL) {
> > PRINT(KERN_ERR, ctx->ohci->host->id,
> > "Current iso context not set");
> > } else
> > res = dma_region_mmap(&ctx->current_ctx->dma, file, vma);
> > - unlock_kernel();
> >
> > return res;
> > }
> > @@ -1200,7 +1196,6 @@ static unsigned int video1394_poll(struc
> > struct dma_iso_ctx *d;
> > int i;
> >
> > - lock_kernel();
> > ctx = file->private_data;
> > d = ctx->current_ctx;
> > if (d == NULL) {
> > @@ -1221,7 +1216,6 @@ static unsigned int video1394_poll(struc
> > }
> > spin_unlock_irqrestore(&d->lock, flags);
> > done:
> > - unlock_kernel();
> >
> > return mask;
> > }
> > @@ -1257,7 +1251,6 @@ static int video1394_release(struct inod
> > struct list_head *lh, *next;
> > u64 mask;
> >
> > - lock_kernel();
> > list_for_each_safe(lh, next, &ctx->context_list) {
> > struct dma_iso_ctx *d;
> > d = list_entry(lh, struct dma_iso_ctx, link);
> > @@ -1278,7 +1271,6 @@ static int video1394_release(struct inod
> > kfree(ctx);
> > file->private_data = NULL;
> >
> > - unlock_kernel();
> > return 0;
> > }
> >
> >
>

2006-10-19 13:27:50

by Andi Kleen

[permalink] [raw]
Subject: Re: Unnecessary BKL contention in video1394


> Adding Andi Kleen to CC, who added the BKL around __video1394_ioctl a
> long while back (when converting video1394 to compat_ioctl).
>
> I don't feel that any replacement protection is needed, since the
> critical sections (where structures are used both in interrupts and in
> file_operations) are already protected by spinlocks.

Fine by me. I just did it to preserve old semantics because I didn't want
to audit the 1394 locking. But if you think it's not needed feel free to remove
them.

-andi

2006-10-19 14:36:13

by Stefan Richter

[permalink] [raw]
Subject: Re: Unnecessary BKL contention in video1394

Andi Kleen wrote:
[Daniel Drake wrote:]
>> Adding Andi Kleen to CC, who added the BKL around __video1394_ioctl a
>> long while back (when converting video1394 to compat_ioctl).
>>
>> I don't feel that any replacement protection is needed, since the
>> critical sections (where structures are used both in interrupts and in
>> file_operations) are already protected by spinlocks.
>
> Fine by me. I just did it to preserve old semantics because I didn't want
> to audit the 1394 locking. But if you think it's not needed feel free to remove
> them.

Thanks for the info. Daniel, do you want to resend a signed-off patch?
And __video1394_ioctl and its wrapper video1394_ioctl can certainly be
merged then.
--
Stefan Richter
-=====-=-==- =-=- =--==
http://arcgraph.de/sr/

2006-10-19 15:32:41

by Daniel Drake

[permalink] [raw]
Subject: Re: Unnecessary BKL contention in video1394

On Thu, 2006-10-19 at 16:36 +0200, Stefan Richter wrote:
> Thanks for the info. Daniel, do you want to resend a signed-off patch?
> And __video1394_ioctl and its wrapper video1394_ioctl can certainly be
> merged then.

Yep, I had already made that change locally. I will run it overnight on
several cameras just to be sure, and will send a patch tomorrow.

I also did some more investigation and straightened out my knowledge of
file_operations: release() can never be called while inside a read() or
ioctl(): This would imply that separate threads are in use, and the
driver's release() function is not called until *all* threads have
closed the fd. In other words I'm now much more confident that this
patch is not removing any necessary locking.

Daniel