On Sun, 17 Mar 2013, Soeren Moch wrote:

For each device only one isochronous endpoint is used (EP IN4, 1x 940 
Bytes, Interval 1).
When the ENOMEM error occurs, a huge number of iTDs is in the free_list 
of one stream. This number is much higher than the 2*M entries, which 
should be there according to your description.

Okay, but how did they get there?  With each URB requiring 9 iTDs, and
about 5 URBs active at any time, there should be about 5*9 = 45 iTDs in
use and 2*9 = 18 iTDs on the free list.  By the time each URB
completes, it should have released all 9 iTDs back to the free list,
and each time an URB is submitted, it should be able to acquire all 9
of the iTDs that it needs from the free list -- it shouldn't have to 
allocate any from the DMA pool.

Looks like you'll have to investigate what's going on inside
itd_urb_transaction().  Print out some useful information whenever the
size of stream->free_list is above 50, such as the value of num_itds,
how many of the loop iterations could get an iTD from the free list,
and the value of itd->frame in the case where the "goto alloc_itd"
statement is followed.

It might be a good idea also to print out the size of the free list in
itd_complete(), where it calls ehci_urb_done(), and include the value
of ehci->now_frame.

Alan Stern