On Wed, 6 Apr 2011, Joerg Roedel wrote:
> Booting 2.6.38 on my test machine produces a lockdep warning
> from the usb_amd_find_chipset_info() function:
>
> WARNING: at /data/lemmy/linux.trees.git/kernel/lockdep.c:2465 lockdep_trace_alloc+0x95/0xc2()
> Hardware name: Snook
> Modules linked in:
> Pid: 959, comm: work_for_cpu Not tainted 2.6.39-rc2+ #22
> Call Trace:
> [<ffffffff8103c0d4>] warn_slowpath_common+0x80/0x98
> [<ffffffff812387e6>] ? T.492+0x24/0x26
> [<ffffffff8103c101>] warn_slowpath_null+0x15/0x17
> [<ffffffff81068667>] lockdep_trace_alloc+0x95/0xc2
> [<ffffffff810ed9ac>] slab_pre_alloc_hook+0x18/0x3b
> [<ffffffff810ef227>] kmem_cache_alloc_trace+0x25/0xba
> [<ffffffff812387e6>] T.492+0x24/0x26
> [<ffffffff81238816>] pci_get_subsys+0x2e/0x73
> sr0: scsi3-mmc drive: 48x/48x writer dvd-ram cd/rw xa/form2 cdda tray
> [<ffffffff8123886c>] pci_get_device+0x11/0x13
> [<ffffffff814082a9>] usb_amd_find_chipset_info+0x3f/0x18a
> ...
>
> It turns out that this function calls pci_get_device under a spin_lock
> with irqs disabled, but the pci_get_device function is only allowed in
> preemptible context.
>
> This patch fixes this by using temporary variables in the quirk
> algorithm and commiting them later to the struct under the lock. This
> moves all pci_get_device() invocations out of the spin_lock and fixes
> the lockdep warning for me.
>
> Cc: stable@xxxxxxxxxx
> Signed-off-by: Joerg Roedel <joerg.roedel@xxxxxxx>
> ---
> drivers/usb/host/pci-quirks.c | 67 +++++++++++++++++++++++-----------------
> 1 files changed, 38 insertions(+), 29 deletions(-)
>
> diff --git a/drivers/usb/host/pci-quirks.c b/drivers/usb/host/pci-quirks.c
> index 1d586d4..d3e5cf3 100644
> --- a/drivers/usb/host/pci-quirks.c
> +++ b/drivers/usb/host/pci-quirks.c
> @@ -84,64 +84,73 @@ int usb_amd_find_chipset_info(void)
> {
> u8 rev = 0;
> unsigned long flags;
> + struct pci_dev *nb_dev, *smbus_dev;
> + int nb_type, sb_type;
>
> spin_lock_irqsave(&amd_lock, flags);
> -
> amd_chipset.probe_count++;
> /* probe only once */
> if (amd_chipset.probe_count > 1) {
> spin_unlock_irqrestore(&amd_lock, flags);
> return amd_chipset.probe_result;
> }
The counter really should be a bool: Has the chipset already been
probed or not? After all, nobody cares how many times this routine was
called.
> + spin_unlock_irqrestore(&amd_lock, flags);
This code now contains a bug: You incremented the probe_count _before_
doing the probe. If another thread calls this routine right now, it
will get an incorrect result.
Fixing this up should be fairly easy.
Alan Stern
--
To unsubscribe from this list: send the line "unsubscribe linux-usb" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
