On 6/19/23 7:27?AM, Stefan Metzmacher wrote: > Am 19.06.23 um 15:09 schrieb Stefan Metzmacher: >> Am 19.06.23 um 15:05 schrieb Jens Axboe: >>> On 6/19/23 3:57?AM, Stefan Metzmacher wrote: >>>> Hi Jens, >>>> >>>>> If the application sets ->msg_control and we have to later retry this >>>>> command, or if it got queued with IOSQE_ASYNC to begin with, then we >>>>> need to retain the original msg_control value. This is due to the net >>>>> stack overwriting this field with an in-kernel pointer, to copy it >>>>> in. Hitting that path for the second time will now fail the copy from >>>>> user, as it's attempting to copy from a non-user address. >>>> >>>> I'm not 100% sure about the impact of this change. >>>> >>>> But I think the logic we need is that only the >>>> first __sys_sendmsg_sock() that returns > 0 should >>>> see msg_control. A retry because of MSG_WAITALL should >>>> clear msg_control[len] for a follow up __sys_sendmsg_sock(). >>>> And I fear the patch below would not clear it... >>>> >>>> Otherwise the receiver/socket-layer will get the same msg_control twice, >>>> which is unexpected. >>> >>> Yes agree, if we do transfer some (but not all) data and WAITALL is set, >>> it should get cleared. I'll post a patch for that. >> >> Thanks! >> >>> Note that it was also broken before, just differently broken. The most >>> likely outcome here was a full retry and now getting -EFAULT. >> >> Yes, I can see that it was broken before... > > I haven't checked myself, but I'm wondering about the recvmsg case, > I guess we would need to advance the msg_control buffer after each > iteration, in order to avoid overwritting the already received messages > on retry. > > This all gets complicated with things like MSG_CTRUNC. > > I guess it's too late to reject MSG_WAITALL together with msg_control > for io_recvmsg() because of compat reasons, > but as MSG_WAITALL is also processed in the socket layer, we could keep it > simple for now and skip the this retry logic: > > if (flags & MSG_WAITALL) > min_ret = iov_iter_count(&kmsg->msg.msg_iter); > > This might become something similar to this, > but likely more complex, as would need to record kmsg->controllen == 0 > condition already in io_recvmsg_prep: > > if (flags & MSG_WAITALL && kmsg->controllen == 0) > min_ret = iov_iter_count(&kmsg->msg.msg_iter); Yep agree, I think this is the best way - ensure that once we transfer data with cmsg, it's a one-shot kind of deal. Do you want to cut a patch for that one? -- Jens Axboe
