request back from the receive queue -- it might already be processed
by remote_recv_thread, which lead to crashes like below:
(primary) Unable to receive reply header: Connection reset by peer.
(primary) Unable to send request (Connection reset by peer):
WRITE(954662912, 131072).
(primary) Disconnected from kopusha:7772.
(primary) Increasing localcnt to 1.
(primary) Assertion failed: (old > 0), function refcnt_release,
file refcnt.h, line 62.
Taking the request back was not necessary (it would properly be
processed by the remote_recv_thread) and only complicated things.
MFC after: 2 weeks
indication when a request can be moved to done queue, but also for
detecting the current state of memsync request.
This approach has problems, e.g. leaking a request if memsynk ack from
the secondary failed, or racy usage of write_complete, which should be
called only once per write request, but for memsync can be entered by
local_send_thread and ggate_send_thread simultaneously.
So the following approach is implemented instead:
1) Use hio_countdown only for counting components we waiting to
complete, i.e. initially it is always 2 for any replication mode.
2) To distinguish between "memsync ack" and "memsync fin" responses
from the secondary, add and use hio_memsyncacked field.
3) write_complete() in component threads is called only before
releasing hio_countdown (i.e. before the hio may be returned to the
done queue).
4) Add and use hio_writecount refcounter to detect when
write_complete() can be called in memsync case.
Reported by: Pete French petefrench ingresso.co.uk
Tested by: Pete French petefrench ingresso.co.uk
MFC after: 2 weeks
This is believed to fix hastd crashes, which might occur during
synchronization, triggered by the failed assertion:
Assertion failed: (amp->am_memtab[ext] > 0),
function activemap_write_complete, file activemap.c, line 351.
MFC after: 1 week
kept dirty to reduce the number of on-disk metadata updates. The
sequence of operations is:
1) acquire the activemap lock;
2) update in-memory map;
3) if the list of keepdirty extents is changed, update on-disk metadata;
4) release the lock.
On-disk updates are not frequent in comparison with in-memory updates,
while require much more time. So situations are possible when one
thread is updating on-disk metadata and another one is waiting for the
activemap lock just to update the in-memory map.
Improve this by introducing additional, on-disk map lock: when
in-memory map is updated and it is detected that the on-disk map needs
update too, the on-disk map lock is acquired and the on-memory lock is
released before flushing the map.
Reported by: Yamagi Burmeister yamagi.org
Tested by: Yamagi Burmeister yamagi.org
Reviewed by: pjd
Approved by: re (marius)
MFC after: 2 weeks
waiting on an empty queue as the queue may have several consumers.
Before the fix the following scenario was possible: 2 threads are
waiting on empty queue, 2 threads are inserting simultaneously. The
first inserting thread detects that the queue is empty and is going to
send the signal, but before it sends the second thread inserts
too. When the first sends the signal only one of the waiting threads
receive it while the other one may wait forever.
The scenario above is is believed to be the cause of the observed
cases, when ggate_recv_thread() was getting stuck on taking free
request, while the free queue was not empty.
Reviewed by: pjd
Tested by: Yamagi Burmeister yamagi.org
Approved by: re (marius)
MFC after: 2 weeks
why it will now be the default.
- Bump protocol version to 2 and add backward compatibility for version 1.
- Allow to specify hosts by kern.hostid as well (in addition to hostname and
kern.hostuuid) in configuration file.
Sponsored by: Panzura
Tested by: trociny
reads with native speed of the underlying provider.
There are three situations when direct reads are not used:
1. Data is being synchronized and synchronization source is the secondary
node, which means secondary node has more recent data and we should read
from it.
2. Local read failed and we have to try to read from the secondary node.
3. Local component is unavailable and all I/O requests are served from the
secondary node.
Sponsored by: Panzura, http://www.panzura.com
MFC after: 1 month
is synchronizing data that is out of date on the local component, we
should not send G_GATE_CMD_DONE acknowledge to the kernel.
This fixes the issue, observed in async mode, when on synchronization
from the remote component the worker terminated with "G_GATE_CMD_DONE
failed" error.
Reported by: Artem Kajalainen <artem kayalaynen ru>
Reviewed by: pjd
MFC after: 1 week
write. This way on first connection we will synchronize only the extents that
were modified during the lifetime of primary node, not entire GEOM provider.
MFC after: 3 days
reordering won't make the actual write to be committed before marking
the coresponding extent as dirty.
It can be disabled in configuration file.
If BIO_FLUSH is not supported by the underlying file system we log a warning
and never send BIO_FLUSH again to that GEOM provider.
MFC after: 3 days
disk if needed. This should fix a potential case when extents are cleared in
activemap but metadata is not updated on disk.
Suggested by: pjd
Approved by: pjd (mentor)
only receiving the data. In r220271 the unused directions were
disabled using shutdown(2).
Unfortunately, this broke automatic receive buffer sizing, which
currently works only for connections in ETASBLISHED state. It was a
root cause of the issue reported by users, when connection between
primary and secondary could get stuck.
Disable the code introduced in r220271 until the issue with automatic
buffer sizing is not resolved.
Reported by: Daniel Kalchev <daniel@digsys.bg>, danger, sobomax
Tested by: Daniel Kalchev <daniel@digsys.bg>, danger
Approved by: pjd (mentor)
MFC after: 1 week
requests as well as number of activemap updates.
Number of BIO_WRITEs and activemap updates are especially interesting, because
if those two are too close to each other, it means that your workload needs
bigger number of dirty extents. Activemap should be updated as rarely as
possible.
MFC after: 1 week
because we need to do ioctl(2)s, which are not permitted in the capability
mode. What we do now is to chroot(2) to /var/empty, which restricts access
to file system name space and we drop privileges to hast user and hast
group.
This still allows to access to other name spaces, like list of processes,
network and sysvipc.
To address that, use jail(2) instead of chroot(2). Using jail(2) will restrict
access to process table, network (we use ip-less jails) and sysvipc (if
security.jail.sysvipc_allowed is turned off). This provides much better
separation.
MFC after: 1 week
secondary role. It is possible that the remote node is primary, but only
because there was a role change and it didn't finish cleaning up (unmounting
file systems, etc.). If we detect such situation, wait for the remote node
to switch the role to secondary before accepting I/Os. If we don't wait for
it in that case, we will most likely cause split-brain.
MFC after: 1 week
- We have two nodes connected and synchronized (local counters on both sides
are 0).
- We take secondary down and recreate it.
- Primary connects to it and starts synchronization (but local counters are
still 0).
- We switch the roles.
- Synchronization restarts but data is synchronized now from new primary
(because local counters are 0) that doesn't have new data yet.
This fix this issue we bump local counter on primary when we discover that
connected secondary was recreated and has no data yet.
Reported by: trociny
Discussed with: trociny
Tested by: trociny
MFC after: 1 week
equal to secondary counters:
primary_localcnt = secondary_remotecnt
primary_remotecnt = secondary_localcnt
Previously it was done wrong and split-brain was observed after
primary had synchronized up-to-date data from secondary.
Approved by: pjd (mentor)
MFC after: 1 week
We can use capsicum for secondary worker processes and hastctl.
When working as primary we drop privileges using chroot+setgid+setuid
still as we need to send ioctl(2)s to ggate device, for which capsicum
doesn't allow (yet).
X-MFC after: capsicum is merged to stable/8
This way we know how to connect to secondary node when we are primary.
The same variable is used by the secondary node - it only accepts
connections from the address stored in 'remote' variable.
In cluster configurations it is common that each node has its individual
IP address and there is one addtional shared IP address which is assigned
to primary node. It seems it is possible that if the shared IP address is
from the same network as the individual IP address it might be choosen by
the kernel as a source address for connection with the secondary node.
Such connection will be rejected by secondary, as it doesn't come from
primary node individual IP.
Add 'source' variable that allows to specify source IP address we want to
bind to before connecting to the secondary node.
MFC after: 1 week
connection so the worker will exit if it does not receive packets from
the primary during this interval.
Reported by: Christian Vogt <Christian.Vogt@haw-hamburg.de>
Tested by: Christian Vogt <Christian.Vogt@haw-hamburg.de>
Approved by: pjd (mentor)
MFC after: 1 week
- HOLE - it simply turns all-zero blocks into few bytes header;
it is extremely fast, so it is turned on by default;
it is mostly intended to speed up initial synchronization
where we expect many zeros;
- LZF - very fast algorithm by Marc Alexander Lehmann, which shows
very decent compression ratio and has BSD license.
MFC after: 2 weeks
