Prev: [tip:perf/core] perf probe: Fix memory leaks in add_perf_probe_events
Next: [PATCH 4/6] lib/list_sort: test: improve errors handling
From: Dan Carpenter on 7 Aug 2010 03:30
On Fri, Aug 06, 2010 at 11:49:10PM +0400, Kulikov Vasiliy wrote:
> Driver should call disable_pci_device() if it returns from pci_probe()
> with error. Also it must not be called if pci_request_region() fails as
> it means that somebody uses device resources and rules the device.
>

I think we should disable it actually. The comments on
pci_enable_device() and pci_disable_device() say that only the first and
last callers actually enable and disable it. The others just increment
or decrement a counter.

regards,
dan carpenter

> Move pci_enable_device() after checking moxa_boards[] to remove senseless
> pci_enable_device()/pci_disable_device().
>
> Signed-off-by: Kulikov Vasiliy <segooon(a)gmail.com>
> ---
> drivers/char/moxa.c | 12 +++++++-----
> 1 files changed, 7 insertions(+), 5 deletions(-)
>
> diff --git a/drivers/char/moxa.c b/drivers/char/moxa.c
> index 107b0bd..0ea5aeb 100644
> --- a/drivers/char/moxa.c
> +++ b/drivers/char/moxa.c
> @@ -939,11 +939,6 @@ static int __devinit moxa_pci_probe(struct pci_dev *pdev,
> int board_type = ent->driver_data;
> int retval;
>
> - retval = pci_enable_device(pdev);
> - if (retval) {
> - dev_err(&pdev->dev, "can't enable pci device\n");
> - goto err;
> - }
>
> for (i = 0; i < MAX_BOARDS; i++)
> if (moxa_boards[i].basemem == NULL)
> @@ -956,6 +951,12 @@ static int __devinit moxa_pci_probe(struct pci_dev *pdev,
> goto err;
> }
>
> + retval = pci_enable_device(pdev);
> + if (retval) {
> + dev_err(&pdev->dev, "can't enable pci device\n");
> + goto err;
> + }
> +
> board = &moxa_boards[i];
>
> retval = pci_request_region(pdev, 2, "moxa-base");
> @@ -1001,6 +1002,7 @@ err_base:
> board->basemem = NULL;
> err_reg:
> pci_release_region(pdev, 2);
> + pci_disable_device(pdev);
> err:
> return retval;
> }
> --
> 1.7.0.4
>
> --
> To unsubscribe from this list: send the line "unsubscribe kernel-janitors" in
> the body of a message to majordomo(a)vger.kernel.org
> More majordomo info at http://vger.kernel.org/majordomo-info.html
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo(a)vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
From: Dan Carpenter on 7 Aug 2010 06:00
On Sat, Aug 07, 2010 at 12:55:12PM +0400, Vasiliy Kulikov wrote:
> On Sat, Aug 07, 2010 at 09:22 +0200, Dan Carpenter wrote:
> > On Fri, Aug 06, 2010 at 11:49:10PM +0400, Kulikov Vasiliy wrote:
> > > Driver should call disable_pci_device() if it returns from pci_probe()
> > > with error. Also it must not be called if pci_request_region() fails as
> > > it means that somebody uses device resources and rules the device.
> > >
> >
> > I think we should disable it actually. The comments on
> > pci_enable_device() and pci_disable_device() say that only the first and
> > last callers actually enable and disable it. The others just increment
> > or decrement a counter.
>
> See this thread: http://lkml.org/lkml/2005/2/13/82
>
> Specifically this mail:
>
> Date Mon, 14 Feb 2005 14:51:26 -0500
> From Jeff Garzik <>
>
> ...
> You also need to consider situations such as out-of-tree drivers
> for the same hardware (might not use PCI API), and situations where you
> have peer devices discovered and used (PCI API doesn't have "hey, <this>
> device is associated with <current driver>, too" capability).
> ...
>
> Searching for 'pci_disable_device() inurl:lkml' doesn't give me newer info
> aboud this problem, so I think it's better to play safe.
>

That's ancient. That's a couple months before the start of git.

git show v2.6.12:drivers/pci/pci.c

In those days pci_enable/disable_device() were not nestable. These days
we can just unwind normally so it's a big improvement.

regards,
dan carpenter
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo(a)vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
From: Dan Carpenter on 7 Aug 2010 15:10
On Sat, Aug 07, 2010 at 10:02:52PM +0400, Vasiliy Kulikov wrote:
>
> The thing is not in that pci_enable_device() is nestable, but in that
> some buggy non-mainline drivers may not use PCI api, but use
> request_*(). In this case PCI bus doesn't see this driver and tries to
> call mainline driver probe(). It sees that resources are already
> grabbed. The only reason can be another driver that rules this device
> (or our kernel is crazy). As we are robust driver we don't want to break
> this buggy third party driver :-)
>

What Jeff Garzik was saying in those days if two people called
pci_enable_device() it wasn't a problem, but if one person then called
pci_disable_device() the other person who still wanted it enabled was
screwed. That model was "*first* person out of the building lock the
door."

That was half a decade ago though. The fix that Jeff proposed wasn't
really a good one because there is no way people would get that right.
None of that applies any more. These days we just call enable when we
want it enabled and disable when we are done. All the dependencies are
handled for us.

In fact, if we try to use the old hacks and work arounds, that will cause
bugs in the new code...

regards,
dan carpenter

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo(a)vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
 | 
Pages: 1
Prev: [tip:perf/core] perf probe: Fix memory leaks in add_perf_probe_events
Next: [PATCH 4/6] lib/list_sort: test: improve errors handling