Changes 20 to 30 on this branch are displayed below, sorted in descending chronological order.

967 days ago: This is a pretty nasty hack to make COSS optionally use the aufs
revision info | browse files | diff
Author: adrian@creative.net.au
Changelog:
This is a pretty nasty hack to make COSS optionally use the aufs
aioRead/aioWrite routines.
* Small changes to COSS APIs, nothing affecting behaviour
* Allow COSS to either use the POSIX AIO or AUFS AIO (define in store_coss.h)
* Modify aiops to use pread/pwrite so one gets access to atomic seek-read
and seek-write operations
(This again is a hack; I haven't figured out how to properly coax their
definitions out of #include <unistd.h>)
What needs to be done!
* break out the aiops routine/type definitions into a seperate file
(eg src/fs/aufs/async_io.h) to be used by other people like COSS
* Write something to return a proper "load" on the COSS directory;
at the present the load calculation is broken
* See why it crashes after a while - probably related to 32/64 bit file
offsets floating around in the aiops/aufs code
* Tidy up the whole mess
Date: Mon, 15 May 2006 10:53:37 GMT
968 days ago: More magic typecasting.
revision info | browse files | diff
Author: adrian@creative.net.au
Changelog:
More magic typecasting.
Date: Sun, 14 May 2006 12:15:09 GMT
968 days ago: Change the debug level -this message isn't an error, its just
revision info | browse files | diff
Author: adrian@creative.net.au
Changelog:
Change the debug level -this message isn't an error, its just
informational.
Date: Sun, 14 May 2006 11:52:00 GMT
968 days ago: Fix a silly typecast - I wish gcc would complain about this.
revision info | browse files | diff
Author: adrian@creative.net.au
Changelog:
Fix a silly typecast - I wish gcc would complain about this.
Ie, the Divide needed to happen at the 64 bit offset level
before it was reduced to a 32 bit number.
Date: Sun, 14 May 2006 11:43:58 GMT
968 days ago: Adjust comments in a couple places to reflect the current reality
revision info | browse files | diff
Author: adrian@creative.net.au
Changelog:
* Adjust comments in a couple places to reflect the current reality
* Add in some logic to try and track the number of pending relocates
on a given stripe, so ...
* ... We can notice when we try to write to a part of the disk that a
pending relocate exists.
I just saw a crash whilst typing this commit message out; so I've
made things unstable again (for now!)
Date: Sun, 14 May 2006 11:42:09 GMT
968 days ago: Identify a race condition which only pops up if the async io code
revision info | browse files | diff
Author: adrian@creative.net.au
Changelog:
* Identify a race condition which only pops up if the async io code
hits load and therefore turns all operations 'sync'. Membufs could
then be freed during storeCossMaybeWriteMemBuf().. and we then tried
to free the thing again. Tsk me.
* Fix this by implementing a 'dead' list. Objects are placed onto the
dead list at any point but the object isn't freed. This means the pointer
is valid until the stack has been unwound and storeCossDirCallback()
is called.
* Add in some assertions to make sure a dead membuf isn't passed in somewhere.
Date: Sun, 14 May 2006 02:15:27 GMT
969 days ago: flesh out some more debugging
revision info | browse files | diff
Author: adrian@creative.net.au
Changelog:
* flesh out some more debugging
* fix a double-free of a MemBuf; this was due to me changing the MemBuf
concept to support hanging around after writing.
Date: Sat, 13 May 2006 03:05:24 GMT
970 days ago: Trouble with the current code:
revision info | browse files | diff
Author: adrian@creative.net.au
Changelog: Trouble with the current code:
Occasionally the membuf lockcount would "stick" and clients would not have their
requests completed.
* Add some more debugging
* Delete the pending read op from the pending reloc list of said read ops
before calling its completion handler; trying to dlinkDelete() a just-freed
read op probably explains the seen behaviour
Date: Fri, 12 May 2006 05:55:30 GMT
971 days ago: Fix up debugging which I set too high earlier.
revision info | browse files | diff
Author: adrian@creative.net.au
Changelog:
Fix up debugging which I set too high earlier.
Date: Thu, 11 May 2006 08:11:51 GMT
971 days ago: Add a missing check with the new membuf state 'written'.
revision info | browse files | diff
Author: adrian@creative.net.au
Changelog:
* Add a missing check with the new membuf state 'written'.
* Attempt to deallocate the membuf after we've written it
Date: Thu, 11 May 2006 08:06:25 GMT
earlier changes recent changes later changes