On 28.2.2019 9.09, Greg Kroah-Hartman wrote:
On Wed, Feb 27, 2019 at 03:19:17PM -0700, Daniel Kurtz wrote:
In cases such as xhci_abort_cmd_ring(), xhci_handshake() is called with
a spin lock held (and local interrupts disabled) with a huge 5 second
timeout. This can translates to 5 million calls to udelay(1). By its
very nature, udelay() is not meant to be precise, it only guarantees to
delay a minimum of 1 microsecond. Therefore the actual delay of
xhci_handshake() can be significantly longer. If the average udelay(1)
is greater than 2.2 us, the total time in xhci_handshake() - with
interrupts disabled can be > 11 seconds triggering the kernel's soft lockup
detector.
To avoid this, let's replace the open coded io polling loop with one from
iopoll.h that uses a loop timed with the more presumably reliable ktime
infrastructure.
Signed-off-by: Daniel Kurtz <djkurtz@xxxxxxxxxxxx>
Looks sane to me, nice fixup.
Reviewed-by: Greg Kroah-Hartman <gregkh@xxxxxxxxxxxxxxxxxxx>
Is this causing problems on older kernels/devices today such that we
should backport this?
A very similar patch was submitted some weeks ago by Andrey Smirnov.
https://lore.kernel.org/lkml/20190208014816.21869-1-andrew.smirnov@xxxxxxxxx/
His commit message only mentions that readl_poll_timeout_atomic() does the same job,
not about any issues with the loop, so I was going to send it forward to usb-next
after 5.1-rc (to 5.2).
-Mathias
