When safety requirements are met, it allows to avoid passing I/O requests
to GEOM g_up/g_down thread, executing them directly in the caller context.
That allows to avoid CPU bottlenecks in g_up/g_down threads, plus avoid
several context switches per I/O.
The defined now safety requirements are:
- caller should not hold any locks and should be reenterable;
- callee should not depend on GEOM dual-threaded concurency semantics;
- on the way down, if request is unmapped while callee doesn't support it,
the context should be sleepable;
- kernel thread stack usage should be below 50%.
To keep compatibility with GEOM classes not meeting above requirements
new provider and consumer flags added:
- G_CF_DIRECT_SEND -- consumer code meets caller requirements (request);
- G_CF_DIRECT_RECEIVE -- consumer code meets callee requirements (done);
- G_PF_DIRECT_SEND -- provider code meets caller requirements (done);
- G_PF_DIRECT_RECEIVE -- provider code meets callee requirements (request).
Capable GEOM class can set them, allowing direct dispatch in cases where
it is safe. If any of requirements are not met, request is queued to
g_up or g_down thread same as before.
Such GEOM classes were reviewed and updated to support direct dispatch:
CONCAT, DEV, DISK, GATE, MD, MIRROR, MULTIPATH, NOP, PART, RAID, STRIPE,
VFS, ZERO, ZFS::VDEV, ZFS::ZVOL, all classes based on g_slice KPI (LABEL,
MAP, FLASHMAP, etc).
To declare direct completion capability disk(9) KPI got new flag equivalent
to G_PF_DIRECT_SEND -- DISKFLAG_DIRECT_COMPLETION. da(4) and ada(4) disk
drivers got it set now thanks to earlier CAM locking work.
This change more then twice increases peak block storage performance on
systems with manu CPUs, together with earlier CAM locking changes reaching
more then 1 million IOPS (512 byte raw reads from 16 SATA SSDs on 4 HBAs to
256 user-level threads).
Sponsored by: iXsystems, Inc.
MFC after: 2 months
This fixes "Negative sc_ref" panic possible when sysctl_kern_geom_confxml()
is run simultaneously with destroying GATE device.
Reviewed by: pjd
MFC after: 3 days
This will allow HAST to read directly from the local component without
even communicating userland daemon.
Sponsored by: Panzura, http://www.panzura.com
MFC after: 1 month
The SYSCTL_NODE macro defines a list that stores all child-elements of
that node. If there's no SYSCTL_DECL macro anywhere else, there's no
reason why it shouldn't be static.
g_io_deliver(). In such case it increases 'pace' counter on each ENOMEM and
reschedules the request. The 'pace' counter is decreased for each request going
down, but until 'pace' is greater than zero, GEOM will handle at most 10
requests per second. For GEOM GATE users that are proxy to local GEOM providers
(like ggatel(8) and HAST) we can end up with almost permanent slow down of GEOM
down queue. This is because once we reach GEOM GATE queue limit, we return
ENOMEM to the GEOM. This means that we have, eg. 1024 I/O requests in the GEOM
GATE queue. To make room in the queue and stop returning ENOMEM we need to
proceed the requests of course, but those requests are handled by userland
daemons that handle them by reading/writing also from/to local GEOM providers.
For example with HAST, a new requests comes to /dev/hast/data, which is GEOM
GATE provider. GEOM GATE passes the request to hastd(8) and hastd(8)
reads/writes from/to /dev/da0. Once we reach GEOM GATE queue limit, to free up
a slot in GEOM GATE queue, hastd(8) has to read/write from/to /dev/da0, but
this request will also be very slow, because GEOM now slows down all the
requests. We end up with full queue that we can unload at the speed of 10
requests per second. This simply looks like a deadlock.
Fix it by allowing userland daemons that work with both GEOM GATE and local
GEOM providers to specify unlimited queue size, so GEOM GATE will never return
ENOMEM to the GEOM.
MFC after: 1 week
registered in g_gate_units array and when its sc_provider field is
filled. If during this period g_gate_units is accessed by another
thread that is checking for provider name collision the crash is
possible.
Fix this by adding sc_name field to struct g_gate_softc. In
g_gate_create() when g_gate_softc is created but sc_provider is still
not sc_name points to provider name stored in the local array.
Approved by: pjd (mentor)
Reported by: Freddie Cash <fjwcash@gmail.com>
MFC after: 1 week
No FreeBSD version bump, the userland application to query the features will
be committed last and can serve as an indication of the availablility if
needed.
Sponsored by: Google Summer of Code 2010
Submitted by: kibab
Reviewed by: silence on geom@ during 2 weeks
X-MFC after: to be determined in last commit with code from this project
HAST allows to transparently store data on two physically separated machines
connected over the TCP/IP network. HAST works in Primary-Secondary
(Master-Backup, Master-Slave) configuration, which means that only one of the
cluster nodes can be active at any given time. Only Primary node is able to
handle I/O requests to HAST-managed devices. Currently HAST is limited to two
cluster nodes in total.
HAST operates on block level - it provides disk-like devices in /dev/hast/
directory for use by file systems and/or applications. Working on block level
makes it transparent for file systems and applications. There in no difference
between using HAST-provided device and raw disk, partition, etc. All of them
are just regular GEOM providers in FreeBSD.
For more information please consult hastd(8), hastctl(8) and hast.conf(5)
manual pages, as well as http://wiki.FreeBSD.org/HAST.
Sponsored by: FreeBSD Foundation
Sponsored by: OMCnet Internet Service GmbH
Sponsored by: TransIP BV
Fix some wrong usages.
Note: this does not affect generated binaries as this argument is not used.
PR: 137213
Submitted by: Eygene Ryabinkin (initial version)
MFC after: 1 month
- hold/release device in start/done routines, this will probably slow
down things a bit, but previous code was racy;
- only release device if g_gate_destroy() failed - if it succeeded device
is dead and there is nothing to release;
- various other changes which makes forcible destruction reliable.
MFC after: 3 days
- Prefer '_' to ' ', as it results in more easily parsed results in
memory monitoring tools such as vmstat.
- Remove punctuation that is incompatible with using memory type names
as file names, such as '/' characters.
- Disambiguate some collisions by adding subsystem prefixes to some
memory types.
- Generally prefer lower case to upper case.
- If the same type is defined in multiple architecture directories,
attempt to use the same name in additional cases.
Not all instances were caught in this change, so more work is required to
finish this conversion. Similar changes are required for UMA zone names.
This flag means "wait for all pending requests before returning to userland".
There are pending events for sure, because we just created new provider and
other classes want to taste it, but we cannot answer on I/O requests until
we're here.
4 mutex operations per I/O requests.
- Use only one mutex to protect both (incoming and outgoing) queue.
As MUTEX_PROFILING(9) shows, there is no big contention for this lock.
- Protect sc_queue_count with queue mutex, instead of doing atomic
operations on it.
- Remove DROP_GIANT()/PICKUP_GIANT() - ggate is marked as MPSAFE and no
Giant there.
for unknown events.
A number of modules return EINVAL in this instance, and I have left
those alone for now and instead taught MOD_QUIESCE to accept this
as "didn't do anything".
Now, when trying to mount file system in read-only mode it tries to
opened a device for writting to be able to update to read-write mode
latter. Ehh.
Discussed with: phk
