Keeping consumers open when device is closed is very hard. We need to
open consumers sometimes to update metadata, etc.
Many hacks was introduced in the past to made it possible. You cannot
be sure that you can open consumer for writing always, even if you think
it should be allowed. If one of the mirror components is for example da0
and you try to open it, you can get EPERM when da0s1 is opened for reading
(because BSD class opens consumers (da0) with an extra 'e' bit set).
Waiting for the events queue to be empty may do the trick, but it makes
code much uglier (as you cannot always call g_waitidle()), it doesn't
solve all edge cases and it can introduce deadlocks if there are events
in the queue that wait for gmirror.
I removed those hacks. Now all consumers are open r1w1e1 always, even if
device is closed. Maybe it is less clean from GEOM perspective, but simpify
code a lot and make it much more reliable.
The only issue was retaste event which is sent when we close consumers
opened for writing. I ignore retaste event by not detaching consumer
immediately (so retaste event is not send to my class) and sending event
right after it to detach and destroy consumer.
that the events queue is empty. In other case we're able to hit the race
where for example da0s1 is tasted by some other class, which means that
da0 is open with exclusive bit set, which means that we can't open da0
for writing if it is our component.
Reported by: Attila Nagy <bra@fsn.hu> (and somebody else sometime ago,
but I cannot find who it was)
seconds of idling, DRITY flags are removed.
- If mirror is in idle state or is not open for writing, sleep without
timeout when waiting for I/O requests.
- Don't use atomic operations, for now sysctls are protected by Giant.
- Update debugs.
- Fix for good (I hope) force-stopping mirrors and some filure cases
(e.g. the last good component dies when synchronization is in progress).
Don't use ->nstart/->nend consumer's fields, as this could be racy,
because those fields are used in g_down/g_up, use ->index consumer's
field instead for tracking number of not finished requests.
Reported by: marcel
- After 5 seconds of idle time (this should be configurable) mark all
dirty providers as clean, so when mirror is not used in 5 seconds
and there will be power failure, no synchronization on boot is needed.
Idea from: sorry, I can't find who suggested this
- When there are no ACTIVE components and no NEW components destroy whole
mirror, not only provider.
- Fix one debug to show information about I/O request, before we change
its command.
providers for tasting. Before this hack, race below is possible:
SI_SUB_RAID (no not-fully-configured geoms, so don't block)
GEOM tasting (now geoms are created)
SI_SUB_MOUNT_ROOT (if root file system is placed on a mirror, it is
possible that this mirror is not fully configured yet)
There is a lot of work to do to avoid such hacks and I need a working
solution before 5.3, sorry.
Reported by: John Hay <jhay@icomtek.csir.co.za>
but it is possible:
1. Read data from good component for synchronization.
2. Write data to the same area.
3. Write synchronization data, which are now stale.
Found by: tegge
Actually, it can even cause some problems, because GEOM requires sectorsize
to be more than 0 on first access, not on provider creation, so we can skip
valid providers by doing this check here.
Reported by: Divacky Roman <xdivac02@stud.fit.vutbr.cz>
Sven Willenberger <sven@dmv.com>
- Remove kern.geom.mirror.sync_block_size sysctl. It is quite obvious that we
want to use the biggest size possible.
- Do not use UMA zone for sync data allocations. There could be only one
synchronization request per synchronized disk at a time, so allocate memory
for one request on whole synchronization process related to one disk.
Tested by synchronizing one component (out of three) and by synchronizing
two components (out of three) in parallel.
It allows to fix problems when last provider's sector is shared between few
providers.
- Bump version number for CONCAT and STRIPE and add code for backward
compatibility.
- Do not bump version number of MIRROR, as it wasn't officially introduced yet.
Even if someone started to play with it, there is no big deal, because
wrong MD5 sum of metadata will deny those providers.
- Update manual pages.
- Add version history to g_(stripe|concat).h files.
consumer and 'bio_pflags' which can be used by provider.
- Remove BIO_FLAG1 and BIO_FLAG2 flags. From now on new fields should be
used for internal flags.
- Update g_bio(9) manual page.
- Update some comments.
- Update GEOM_MIRROR, which was the only one using BIO_FLAGs.
Idea from: phk
Reviewed by: phk
The different between the new function and g_mirror_orphan() (which was
used previously) is that syncid is bumped immediately, instead of on
first write, because when consumer was spoiled, it means, that its
provider was opened for writing, so we can't trust that its data
will be valid when it will be connected again.
features. The gmirror(8) utility should be used for control of this class.
There is no manual page yet, but I'm working on it with keramida@.
Many useful tests provided by: simon (thank you!)
Some ideas from: scottl, simon, phk
