[PATCH v5 0/1] multipathd: don't leave maps read-only after write protect is cleared

[Brian Bunker <brian@purestorage.com>]

This is v5. It follows Ben's review of v4 and the discussion with
Martin on v3: dm_addmap_reload() now clears force_readonly only after a
successful reload, needs_ro_update() clears it on every false return
except when a read-only reload still has to be retried, and
uev_update_path() takes the path's ro state from sysfs instead of from
the uevent. The changes are listed at the end.

As agreed, tracking the ro state per path and dropping force_readonly
altogether, as Martin and Uday Shankar suggested, will be a separate
series on top of this fix.

A multipath map can stay read-only after write protect is cleared on
the storage, even though all of its paths are read-write. It happens
when the map's paths are removed and re-added while write protect is
set:

1. Write protect is set. The map is reloaded read-only.
   uev_update_path() sets force_readonly for that reload and clears it
   afterwards.

2. All of the map's paths are removed, and new paths are added while
   write protect is still set. ev_add_path() sees a read-only path and
   sets force_readonly = 1. Nothing ever clears it.

3. Write protect is cleared on the new paths, one after another. Each
   path's DISK_RO=0 uevent is ignored while any path of the map is
   still read-only.

4. The last path's uevent reloads the map. The reload sees the stale
   force_readonly = 1 and loads the map read-only, logging "reloaded
   (retval 0)".

force_readonly only works in one direction: ev_add_path() can set it
but never clears it. If the new paths had been added already writable,
ev_add_path() wouldn't have set the flag and the map wouldn't have
stuck. The paths have to be added while write protected and then
cleared in place.

After step 4 nothing reloads the map again. A rescan doesn't change the
paths' ro state, so it generates no DISK_RO uevent. And since
multipathd now sees the map as read-only, needs_ro_update() also
ignores the next DISK_RO=1 uevent for it.

If two paths' uevents happen to be handled after all paths are
writable, the map does recover: uev_update_path() clears
force_readonly after its reload, so the second reload is read-write.
That makes the problem look intermittent.

Example
-------

We hit this on RHEL 9 device-mapper-multipath 0.8.7, whose
ev_add_path() and uev_update_path() handle force_readonly the same way
as current master. The host's iSCSI sessions were torn down and
re-established while the array was write protecting the volume. From
the messages log, trimmed:

  05:48:27  ...114d8: reload [0 51200 multipath 0 1 alua 1 1
            service-time 0 6 1 135:272 1 66:624 1 69:720 1 129:560 1
            132:656 1 135:752 1]
  05:48:27  sdsd: map ...114d8 reloaded (retval 0)
  05:51:04  kernel: sd 24:0:0:48: [sdaer] Write Protect is on
  05:51:05  ...114d8: remaining active paths: 0
  05:51:07  kernel: sd 29:0:0:48: [sdaoi] Write Protect is on
            (and four more new paths)
  06:09:01  kernel: sd 24:0:0:48: [sdaer] Write Protect is off
  06:09:07  kernel: sd 29:0:0:48: [sdaoi] Write Protect is off
  06:09:10  sdaoi: update path write_protect to '0' (uevent)
  06:09:10  ...114d8: reload [0 51200 multipath 0 1 alua 1 1
            service-time 0 6 1 67:880 1 70:992 1 130:848 1 133:960 1
            65:1072 1 67:1056 1]
  06:09:10  sdaoi: map ...114d8 reloaded (retval 0)

The state afterwards, as recorded by the test:

  dm-106: ro
  sdaer: rw  sdagu: rw  sdaix: rw  sdala: rw  sdand: rw  sdaoi: rw

The patch
---------

The patch sets force_readonly right before each load from the state of
the map's paths. dm_addmap_create() clears it as soon as it has read
it, and dm_addmap_reload() clears it once the reload has succeeded.
needs_ro_update() sets it for the reload done for a change uevent,
using the path's ro state in sysfs, and ev_add_path() and update_map()
compute it from all of the map's paths. With the patch, the 06:09:10
reload would have found no read-only path and loaded the map
read-write.

Changes since v4:

- dm_addmap_reload() no longer clears force_readonly up front and sets
  it back on failure. It clears the flag only once the reload and
  resume have succeeded, as Ben suggested. A failed reload leaves it
  set, so the next reload tries read-only again.

- needs_ro_update() clears force_readonly on every false return,
  except while the map is read-write and one of its paths is still
  read-only, which means a read-only reload failed and must be
  retried. This covers the two cases Martin listed (map and path both
  read-only, and path read-write on a read-only map that still has a
  read-only path). In the dmi.read_only == ro case, it only walks the
  paths if the flag is actually set. The early returns for ro < 0 and
  for a failed dm_get_info() leave the flag alone, since there is no
  valid state to decide from there.

- uev_update_path() takes the path's ro state from sysfs instead of
  from the uevent's DISK_RO. The uevent only says that something
  changed. The kernel updates the ro attribute before it sends the
  uevent, so sysfs is current when the uevent is handled, while the
  uevent itself may have been queued behind later changes. If sysfs
  can't be read, the uevent's DISK_RO isn't trusted either, and
  nothing is done. This also gives needs_ro_update() a valid ro value
  for every change uevent, as Ben asked.

  Note that this changes behavior: any change uevent for a path in a
  map, not only one with DISK_RO, can now trigger a read-only reload
  if sysfs disagrees with the map. It also avoids the pointless
  reload that 6e40bf92 describes when a rescan's DISK_RO=0 and
  DISK_RO=1 uevents are handled after the rescan has finished, since
  sysfs then shows read-only for both. The cost is one sysfs read per
  change uevent, next to the pathinfo() call uev_update_path() already
  makes.

- dm_addmap_create() is unchanged and keeps the read-only shortcut.
  Ben, I read your "There's much to be gained by pulling out the
  dm_addmap_create() code" as "not much", in agreement with Martin.
  Please correct me if I got that wrong.

Changes since v3:

- If a reload fails, dm_addmap_reload() sets force_readonly back to
  the value it had on entry. The map still has its old table, so a map
  that was supposed to become read-only is tried again on the next
  reload. This covers both a failed load and a failed resume.

- In needs_ro_update(), the early return for dmi.read_only == ro now
  clears force_readonly if the map and the uevent are both read-write
  and none of the map's paths is read-only, so the restored flag can't
  make a later, unrelated reload load the map read-only.

- Move the walk over the map's path groups into map_has_ro_path(),
  since needs_ro_update() now needs it in two places.

Changes since v2:

- In needs_ro_update(), fetch the dm info with dm_get_info() if the
  map has none, and return false if that fails, instead of assuming a
  reload is needed. This also removes the separate path walk for that
  case.

- Clear force_readonly in dm_addmap_create() and dm_addmap_reload() as
  soon as they have read it, instead of in dm_addmap() on success. A
  failed load can no longer leave the flag set for a later, unrelated
  reload, such as one for a path group change. needs_ro_update() no
  longer needs to clear the flag itself.

- I didn't add per-path tracking of the ro state. force_readonly is
  only computed from sysfs right before a load, which is rare, so the
  cost of reading sysfs there is small. Per-path tracking could be a
  follow-up if you want it.

Changes since v1:

- Set force_readonly in needs_ro_update() when it returns true,
  instead of in uev_update_path(). The DISK_RO=1 case uses the uevent
  again, rather than rereading sysfs.

- Clear force_readonly in dm_addmap() on success, as for need_reload,
  instead of after the loads in ev_add_path() and uev_update_path().

- In ev_add_path(), only compute force_readonly when need_do_map is
  set.

- Walk mpp->pg rather than mpp->paths in needs_ro_update(), since
  mpp->paths may be stale there. ev_add_path() and update_map() still
  use mpp->paths: both call adopt_paths() first, and mpp->pg doesn't
  contain the new paths until setup_map().

- Move sysfs_get_ro() above has_ro_path() instead of adding a forward
  declaration. The function is unchanged.

Brian Bunker (1):
  multipathd: don't leave maps read-only after write protect is cleared

 libmultipath/devmapper.c |  18 ++++-
 multipathd/main.c        | 144 ++++++++++++++++++++++++++-------------
 2 files changed, 111 insertions(+), 51 deletions(-)

-- 
2.56.0