Re: [PATCH] USB: EHCI: fix for leaking isochronous data

From: Soeren Moch
Date: Thu Mar 14 2013 - 14:49:49 EST

On 10.03.2013 21:59, Alan Stern wrote:
On Sun, 10 Mar 2013, Soeren Moch wrote:
On Wed, 20 Feb 2013, Soeren Moch wrote:

Ok. I use 2 em2840-based usb sticks (em28xx driver) attached to a
Marvell Kirkwood-SoC with a orion-ehci usb controller. These usb sticks
stream dvb data (digital TV) employing isochronous usb transfers (user
application is vdr).

Starting from linux-3.6 I see
ERROR: 1024 KiB atomic DMA coherent pool is too small!
in the syslog after several 10 minutes (sometimes hours) of streaming
and then streaming stops.

In linux-3.6 the memory management for the arm architecture was changed,
so that atomic coherent dma allocations are served from a special pool.
This pool gets exhausted. The only user of this pool (in my test) is
orion-ehci. Although I have only 10 URBs in flight (5 for each stick,
resubmitted in the completion handler), I have 256 atomic coherent
allocations (memory from the pool is allocated in pages) from orion-ehci
when I see this error. So I think there must be a memory leak (memory
allocated atomic somewhere below the usb_submit_urb call in em28xx-core.c).

With other dvb sticks using usb bulk transfers I never see this error.

Since you already found a memory leak in the ehci driver for isoc
transfers, I hoped you can help to solve this problem. If there are
additional questions, please ask. If there is something I can test, I
would be glad to do so.

I guess the first thing is to get a dmesg log showing the problem. You
should build a kernel with CONFIG_USB_DEBUG enabled and post the part
of the dmesg output starting from when you plug in the troublesome DVB
stick.

Sorry for my late response. Now I built a kernel 3.8.0 with usb_debug
enabled. See below for the syslog of device plug-in.

It also might help to have a record of all the isochronous-related
coherent allocations and deallocations done by the ehci-hcd driver.
Are you comfortable making your own debugging changes? The allocations
are done by a call to dma_pool_alloc() in
drivers/usb/host/ehci-sched.c:itd_urb_transaction() if the device runs
at high speed and sitd_urb_transaction() if the device runs at full
speed. The deallocations are done by calls to dma_pool_free() in
ehci-timer.c:end_free_itds().


I added a debug message to
drivers/usb/host/ehci-sched.c:itd_urb_transaction() to log the
allocation flags, see log below.

But it looks like you didn't add a message to end_free_itds(), so we
don't know when the memory gets deallocated. And you didn't print out
the values of urb, num_itds, and i, or the value of itd (so we can
match up allocations against deallocations).

OK, I will implement this more detailed logging. But with several allocations per second and runtime of several hours this will result in a very long logfile.

For me this looks like nothing is
allocated atomic here, so this function should not be the root cause of
the dma coherent pool exhaustion.

I don't understand. If non-atomic allocations can't exhaust the pool,
why do we see these allocations fail?

Good point. Unfortunately I'm not familiar with the memory management details.

Arnd, can memory allocated with dma_pool_alloc() and gfp_flags 0x20000093 or 0x80000093 come from the atomic dma coherent pool?

Are there other allocation functions
which I could track?

Yes, but they wouldn't be used for isochronous transfers. See
ehci_qtd_alloc(), ehci_qtd_free(), ehci_qh_alloc(), and qh_destroy() in
ehci-mem.c, as well as some other one-time-only coherent allocations in
that file.

Alan Stern

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