On 10.03.2013 21:59, Alan Stern wrote:

On Sun, 10 Mar 2013, Soeren Moch wrote:

quoted

quoted

On Wed, 20 Feb 2013, Soeren Moch wrote:

quoted

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.

quoted

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.

quoted

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?

quoted

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