Just in case anyone encounters this issue at some point - the problem was a concurrency issue. When I call the request termination callback, target_cmd_io_done accesses an internal list in tgt, which is at the same time accessed by tgt threads. The access is not thread-safe. When I accessed it in a protected manner using a mutex, the issue was resolved On Sun, Jun 16, 2013 at 1:06 PM, osishkin osishkin <osishkin@xxxxxxxxx> wrote: > I'm pretty certain now that the problem is me not performing the > requests in a sync manner. > I wrote a mock code, in which on receipt I insert all r/w requests to > a queue. A secondary thread simply polls in small intervals (50us) on > the queue, dequeues all pneding requests and writes data to RAM. > This is the most basic async setup possible, and it still doesn't pass > my file system benchmark successfully. I should note that when simply > performing the write to RAM and returning (without the extra queuing > phase) the benchmark is successful. > > I see that there is bs_aio.c which does a very similar thing to what I > do, but using polling on a pipe. I'm not familiar with this method, > but it doesnt sound inherently different from what I do with threads > and mutexes. > So I wonder why the behavior changes so drastically between the two > implementations. > > I'd appreciate your help. > Thank you -- To unsubscribe from this list: send the line "unsubscribe stgt" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
