Jason and Yonatan, On Mon, Feb 27, 2017 at 11:29 AM, Jason Gunthorpe <jgunthorpe@xxxxxxxxxxxxxxxxxxxx> wrote: > On Sun, Feb 26, 2017 at 06:09:34PM +0200, Yonatan Cohen wrote: > >> I bisected the rdma-core library and figured out that the following commit >> introduced this regression: >> 6b26a9e24739 Use C11 atomics instead of wmb/rmb macros for CPU-only atomics >> >> I haven't debugged this yet and would appreciate Jason's input. > > Oops, I think I typo'd it here: > Ie deleted pad_3[31] by mistake! I just confirmed that reverting the C11 atomics commit (6b26a9e24739) fixes ibv_rc_pingpong on my two at91 ARM boards. For some reason the first few packets seem to send slowly, but once it gets going the rest send quickly. Youngjae, I suspect this may correct the issue you reported in http://www.spinics.net/lists/linux-rdma/msg46451.html. IMPORTANT: I had previously found the pad_3[31] issue and corrected it. That resulted in wr_id showing up in the kernel with the correct value, but ibv_rc_pingpong would still sometimes (30% or so?) fail with "Couldn't post send" "parse WC failed 1" on one side. Weirdly, it seems to fail more often just after a reboot, and only occasionally once I run several tests. Jason, was it intentional that rmb() was removed with no replacement in rxe_post_one_recv()? See https://github.com/linux-rdma/rdma-core/commit/6b26a9e24739576ac3f4ae308485389a5b285497?diff=split#diff-f6b2d2321c2b3273e3453d055a62fa98 for details. Unfortunately, even after reverting the C11 atomics commit, I still seem to observe "Couldn't post send" failures which kill the ping occasionally. Is this a known issue? Thanks, -G -- To unsubscribe from this list: send the line "unsubscribe linux-rdma" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
