[Drbd-dev] DRBD-8: FOR REVIEW; proposed phase-I fixes to remove drbd_panic() calls

[Graham, Simon]

 <<drbd-panic.patch>> I'm not done testing yet (because it's been hard
to keep up with the changes recently ;-) but I think it's time to get
some review of the first phase of panic removal I am proposing - in the
end, the changes are actually fairly small for this phase and basically
fall into the following areas:

1. In the case of meta-data failures, I took the approach of forcibly
detaching 
   the disk even if the on-error setting is PassOn AND I made sure that
this is 
   done on ALL meta-data errors.
2. To do this, I added a new Boolean parameter to drbd_chk_io_error and
drbd_io_error 
   that indicates if a detach should be forced - all meta-data cases
pass TRUE 
   and all user data cases pass FALSE.
3. Apart from making sure that chk_io_error and io_error are called for
all meta 
   data cases, I also removed the panic()s from these failure cases.
4. In order to test this, I introduced some fault insertion code -
controlled by 
   a new config macro, DRBD_ENABLE_FAULTS, off by default. This adds a
couple of 
   module parameters;
   a. fault_rate - integer is the % of times the specified fault should
be 
      inserted - the idea is that if you run enough tests with each
fault enabled, 
      eventually all failure code paths will be tested...
   b. enable_faults - bitmap of enabled faults - I broke it down into 6
classes 
      so far - meta-data, resync and data reads and writes.
   Every time an I/O is sent to the block layer, the code tests for the
fault being 
   active and if so it completes the bio with an error instead of
sending it down.

Patch against trunk attached - all comments gratefully received...
Simon

PS: There are also a few other minor fixes:
1. when reading the bitmap, I clear the BM_MD_IO_ERROR flag before
starting - otherwise 
   if this fails once, it will fail every subsequent time.
2. some changes in tracing to help me debug - including fixing the
packet dump trace 
   code - this fix got lost somehow and received frames were printed
incorrectly.
3. At the end of drbd_nl_disk_conf, if a failure occurs AFTER the point
of no return, 
   I think it's necessary to set the local nbc value to NULL and NOT
free it - since 
   it has been put into the mdev->bc by this point, the error handling
in 
   drbd_force_state() will free the bc object and we'd end up freeing it
twice (I THINK!)
4. drbd_al_to_on_disk_bm() - if inc_local_if_state() returns 0 pay
attention!



-------------- next part --------------
A non-text attachment was scrubbed...
Name: drbd-panic.patch
Type: application/octet-stream
Size: 27704 bytes
Desc: drbd-panic.patch
Url : http://lists.linbit.com/pipermail/drbd-dev/attachments/20060914/edf01e33/drbd-panic-0001.obj