This came up when I tested my latest changes to D.R.Packs: I wondered why
this is so extremely slow when I cloned a remote repo over ssh from my home
machine. Then I added --debug and saw that re-using the connection failed
each time.

1 patch for repository http://darcs.net/screened:

patch f4d8302036b82a26296f5201cc6959c836909595
Author: Ben Franksen <benjamin.franksen@helmholtz-berlin.de>
Date:   Sun Jun 28 03:53:30 CEST 2015
  * made connection caching in Darcs.Repository.Ssh thread safe
  
  The module in question had a global variable implemented as an IORef which
  does not play well with concurrent threads. The revised implementation uses
  MVars and also takes care that a connection is only used by one thread at a
  time. This fixes severe performance problems and intermittent failures when
  cloning repos over ssh due to concurrent access from D.R.Packs.
  
  Another problem was that the connections were identified by the user@host
  part only, but not by the directory of the repo. This causes problems when
  e.g. pulling from multiple ssh repos on the same user@host. We now use a
  (user@host, repodir) pair as the key for cached connections.

Attachments

• made-connection-caching-in-darcs_repository_ssh-thread-safe.dpatch
• patch-preview.txt
• unnamed

Following up with a small re-factoring that depends on the fix.

2 patches for repository http://darcs.net/screened:

patch f4d8302036b82a26296f5201cc6959c836909595
Author: Ben Franksen <benjamin.franksen@helmholtz-berlin.de>
Date:   Sun Jun 28 03:53:30 CEST 2015
  * made connection caching in Darcs.Repository.Ssh thread safe
  
  The module in question had a global variable implemented as an IORef which
  does not play well with concurrent threads. The revised implementation uses
  MVars and also takes care that a connection is only used by one thread at a
  time. This fixes severe performance problems and intermittent failures when
  cloning repos over ssh due to concurrent access from D.R.Packs.
  
  Another problem was that the connections were identified by the user@host
  part only, but not by the directory of the repo. This causes problems when
  e.g. pulling from multiple ssh repos on the same user@host. We now use a
  (user@host, repodir) pair as the key for cached connections.

patch 09283ac15a43a4f182726a583a8901fc474f48b9
Author: Ben Franksen <benjamin.franksen@helmholtz-berlin.de>
Date:   Sun Jun 28 03:44:43 CEST 2015
  * use transferModeHeader from D.R.Ssh in D.U.C.TransferMode

Attachments

• made-connection-caching-in-darcs_repository_ssh-thread-safe.dpatch
• patch-preview.txt
• unnamed

Bugfix, so should go into 2.10.1.

I agree it would be nice to have (and it applies cleanly to 2.10).  I
have quickly looked at it and it looks sane. It would be nice if
someone could review it.

Guillaume

A small re-factoring on top of the ssh connection caching fix.

1 patch for repository http://darcs.net/screened:

patch 050d575189bd701f7a214b7e1d01aebe17ea2ad9
Author: Ben Franksen <benjamin.franksen@helmholtz-berlin.de>
Date:   Thu Jul  2 15:06:28 CEST 2015
  * simplified code in D.R.Ssh by using withMVar

Attachments

• patch-preview.txt
• simplified-code-in-d_r_ssh-by-using-withmvar.dpatch
• unnamed

Looks good to me, sorry for not reviewing it sooner.