On 03/21/13 22:12, Alan Stern wrote:

On Thu, 21 Mar 2013, Alan Stern wrote:

quoted

On Thu, 21 Mar 2013, Soeren Moch wrote:

quoted

Now I found out what is going on here:

In itd_urb_transaction() we allocate 9 iTDs for each URB with
number_of_packets == 64 in my case. The iTDs are added to
sched->td_list. For a frame-aligned scheduling we need 8 iTDs, the 9th
one is released back to the front of the streams free_list in
iso_sched_free(). This iTD was cleared after allocation and has a frame
number of 0 now. So for each allocation when now_frame == 0 we allocate
from the dma_pool, not from the free_list.

Okay, that is a problem.  But it shouldn't be such a big problem,
because now_frame should not be equal to 0 very often.

Oh, wait, now I get it.  We never reach a steady state, because the
free list never shrinks, but occasionally it does increase when
now_frame is equal to 0.  Even though that doesn't happen very often,
the effects add up.

Correct. Approximately twice a second we allocate 9 iTDs from the 
dma_pool in my case (the frame numbers run from 0 to 511 - not up to 
2047, which I thought they should). But since we allocate iTDs every 8 
frames in my case (for each active stream), the pool allocations occur 
only when we hit the allocation sequence frame 0,8,16,... (not one of 
the other 7 opportunities frame 1,9,17,... - frame 7,15,23,...)

Very good; tomorrow I will send your patch in.

Alan Stern

Thank you. And also thanks to all who helped to track down this issue.

  Soeren