The following description relates to OpenSSH 3.5p1 as embedded in FreeBSD/i386 4.9p1. The overall behaviour (incorrectly reporting "stalled") is identical on OpenSSH 3.7.1p2 and a quick check shows that the scp blocksize and ssh hysteresis behaviour have not changed. I believe the behaviour is not platform or OS dependent. By default, scp(1) will provide a progress meter showing the transfer ETA. If the link is slow, the transfer meter will alternate between displaying "- stalled -" and unrealistically short ETAs even though the actual connection is transferring data smoothly (as shown by tcpdump). By default, the progress meter is updated every second. If there has been no apparent progress in the transfer after 5 seconds, the progress meter will report "stalled" until some progress is reported. There appear to be two issues that will result in long delays between output progress being seen by the progress meter. Firstly, output from the scp process is in filesystem blocksize blocks - the number of bytes transferred (used by the progres meter) will only be incremented when a full block of data has been transferred. Therefore if the transfer rate is less than 1.6KB/sec (old 8K filesystem) or 3.2KB/sec (newer 16KB filesystem) then the link will report as "stalled". (Identified by code inspection). Secondly, the ssh process spawned by the scp process to perform the actual encryption and transfer includes a substantial internal buffer (>64KB) and appears to implement hysteresis. ktrace output of a sample transfer shows a peak of over 96KB buffered - at which point the ssh process stops reading until the buffer drops to about 32KB. This implies that there is approximately 64KB hysteresis and a transfer rate below about 13KB/sec can result in "stalled" reports. This behaviour is a regression from from an earlier version of OpenSSH but I have not tracked down when it occurred. To reproduce: With FreeBSD, it is possible to use ipfw/dummynet to artificially reduce the outgoing ssh bandwidth to a second system. If traffic shaping is not necessary, it will be necessary to create a low speed link - eg ppp over an analogue modem or serial link. I used the following commands: # ipfw pipe 20 config queue 10 bw 80000 # ipfw add 1005 pipe 20 tcp from any to 192.168.164.18 22 # dd if=/dev/urandom of=data count=512 # scp data 192.168.164.18:/tmp The appropriate fix is unclear - the buffering in both scp and ssh as well as the hysteresis in ssh are beneficial to maximize transfer bandwidth and minimise context switching. It would not be desirable to reduce these sizes when ssh is used across a LAN. In the case of scp, changing from atomicio(write, ...) to write(...) would remove the requirement to write at least filesystem_blocksize bytes/sec to the remote system. In the case of ssh, the hysteresis needs to be adjusted based on the outgoing bandwidth - this could possibly be done by resetting the "don't read more" flag after (say) 1 second.
Created attachment 1286 [details] increment counters for short writes This patch avoids the use of atomicio on the network sockets, so it should have the chance to update the counters more frequently.
Created attachment 1302 [details] Revised scp diff There is no point incrementing the counters more quickly when the file descriptors are blocking. This diff makes them non-blocking, unifies a little code and increases the buffer size for improved performance.
patch applied - will be in openssh-4.8.
Fix shipped in 4.9/4.9p1 release.