lkml.org 
[lkml]   [2009]   [Jul]   [16]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
SubjectRe: [PATCH 1/2] vga: implements VGA arbitration on Linux
From
On Thu, Jul 16, 2009 at 6:48 PM, Alan Cox<alan@lxorguk.ukuu.org.uk> wrote:
>> >> +     pci_read_config_word(pdev, PCI_COMMAND, &cmd);
>> >> +     if (rsrc & (VGA_RSRC_LEGACY_IO | VGA_RSRC_NORMAL_IO))
>> >> +             cmd |= PCI_COMMAND_IO;
>> >> +     if (rsrc & (VGA_RSRC_LEGACY_MEM | VGA_RSRC_NORMAL_MEM))
>> >> +             cmd |= PCI_COMMAND_MEMORY;
>> >> +     pci_write_config_word(pdev, PCI_COMMAND, cmd);
>> >
>> > Locking question - what locks this lot against hotplug also touching
>> > bridge settings ?
>>
>> well here we just bang on device config space registers which means we
>> can probably
>> race against lots of other things that rmw the PCI_COMMAND not just hotplug.
>>
>> Perhaps we need some sort per device PCI command space lock,
>> granted this still means we race against anyone directly hacking it
>> behind our backs.
>
> I suspect the right thing to do is to move that function into the
> drivers/pci code and lock it properly there. That would keep all the
> locking detail internal and private (and someone else's problem ;))

I'll have a look, Jesse any ideas? the hotplug bridge bashing looks unfun.

I noticed a fair few drivers seem to bash these things, and really
pci_enable_device
could possible race with hotplug.

>
>> >> +             pdev = vga_default_device();
>> >
>> > What if the BIOS provided device was hot unplugged ?
>>
>> we just use the pdev as a cookie, if it was hot unplugged we'll
>> have gotten a callback to remove it from the VGA device list
>> and the lookup which happens 5 lines later inside the spinlock
>> will fail.
>
> What if I inserted a new device - the allocator might give me a new
> device with the same pointer if its reusing the same slab entry for that
> size. Unlikely but given a zillion boxes this starts to occur 8(

A zillion boxes with hotplug VGA devices, I wish :-), but I'll fix it up
to do the pci_dev_get/puts.

>
>> >> +             /* We have a conflict, we wait until somebody kicks the
>> >> +              * work queue. Currently we have one work queue that we
>> >
>> > If two drivers own half the resources and both are waiting for the rest
>> > what handles the deadlock
>> >
>
> Not commented on - but a serious question would be "do we actually care
> enough or are there really devices with just I/O and just vga memory
> access used ?"
>

Oops yes I want to avoid doing anything except VGA resources on/off
I don't think the granularity matters, will talk to BenH tomorrow see what his
original idea was.

>> > PCI device refcounting ?
>>
>> Again its just used a cookie for a later lookup in our vgadev array,
>> its gone away it'll have been removed from the array,
>>
>> We could use pci_dev_get/pci_dev_put I suppose but its only used as
>> a cookie so far.
>
> Your cookie validity is suspect I fear. Also holding the device reference
> means you stop that exact set of resources being reissued too early and
> you (or clients) scribbling on them through unfortunate timing. So I
> think you actually do need to grab references properly, Doesn't make the
> code much more complex that I can see.

I'll fix it seems like a good plan just in case.

Dave.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

\
 
 \ /
  Last update: 2009-07-16 12:41    [from the cache]
©2003-2011 Jasper Spaans