HEADS UP - recent kernel work - HAMMER fix, lvm/dm work, etc

Matthew Dillon dillon at apollo.backplane.com
Sun Aug 15 23:36:13 PDT 2010


    * A rare HAMMER bug was found and fixed, and this was also MFC'd
      to the 2.6.x branch.

      The bug is related to HAMMER going a bit overboard on its use
      of cluster_read(), creating a situation where buffers with
      overlapping address spaces in the data zone could sometimes
      be created.

      The bug would normally result in an occassional DATA CRC error
      in the reblocker or in the filesystem, required a umount/remount
      to clear up.  Typically no on-media corruption occurs.  However,
      there might have been other unknown (but rare) side effects to the
      bug so an upgrade is recommended.

      In addition, the HEAD branch now proactively asserts that no
      overlapping buffer is created.


		Remaining enhancements are only in the HEAD branch


    * The HEAD branch now does a lot better on the stress2 stress testing
      suite (/usr/src/test/stress/stress2).  Currently only the tcp test
      still creates resource exhaustion issues with default
      net.inet.tcp.{send,recv}space values.  With lowered values stress2
      survives an overnight run.

      - A per-user file descriptor limit has been added as a sysctl.
	kern.maxfilesperuser.

      - The per-user file descriptor and process limits have been reduced
	somewhat, primarily to prevent pipe(2)'s from blowing out KVA
	on i386 boxes.

    * Some serious kernel memory leaks have been fixed, particularly
      related to PTYs.

      - Memory leaks fixed

      - Numerous panics related to PTY and devfs operation have been
	fixed.

    * Alex Hornung's LVM/DM work in the HEAD branch has gone through
      several rounds of bug fixes, performance enhancements, and
      testing.  Both the crypt target and the stripe target have seen
      significant work.

      This infrastructure is looking like a nice solid addition to
      DragonFly!

      - The stripe target is no longer limited to just two drives.

      - Several resource exhaustion deadlocks have been fixed.

      - Significant testing with HAMMER on a crypt + stripe x 5
	setup has been completed without generating any errors.

      - No soft-raid features have been ported or implemented yet.

    * The cluster read-ahead code has been tuned to do a better job
      with a DM stripe target (tested with strip x 5 drives).

    * Samuel J. Greear's select/poll/kqueue work has gone through
      numerous revisions and is a lot more stable now.

    * PF has been updated to OpenBSD 4.1 (upgrading to OpenBSD-current
      is being discussed), and has also gone through a few bug-fixing
      passes since its import.

    * The wifi/wlan code has gone through a few rounds of bug fixes
      as well, including a fix for kernel memory corruption related
      to bringing a wlan interface up or down.

    That's the basics for the last month or so.  Our next release will
    be in mid-September, and hopefully a few more goodies will get in
    before then.

					-Matt
					Matthew Dillon 
					<dillon at backplane.com>





More information about the Users mailing list