At this point we have not rendezvous'ed with the mirror worker thread, and
I/O may still be in flight. Various I/O completion paths expect to be able
to obtain a reference to the mirror softc from the GEOM, so setting it to
NULL may result in various NULL pointer dereferences if the mirror is
stopped with -f or the kernel is shut down while a mirror is
synchronizing. The worker thread will clear the softc pointer before
exiting.
Tested by: pho
MFC after: 2 weeks
Sponsored by: Dell EMC Isilon
We are otherwise susceptible to a race with a concurrent teardown of the
mirror provider, causing the I/O to be left uncompleted after the mirror
started withering.
Tested by: pho
MFC after: 2 weeks
Sponsored by: Dell EMC Isilon
Regular I/O requests may be blocked by concurrent synchronization requests
targeted to the same LBAs, in which case they are moved to a holding queue
until the conflicting I/O completes. We therefore want to stop
synchronization before completing pending I/O in g_mirror_destroy_provider()
since this ensures that blocked I/O requests are completed as well.
Tested by: pho
MFC after: 2 weeks
Sponsored by: Dell EMC Isilon
Entries may be removed and freed if an I/O error occurs during mirror
synchronization, so we cannot assume that all entries of ds_bios are
valid.
Also ensure that a synchronization BIO's array index is preserved after
a successful write.
Reported and tested by: pho
MFC after: 2 weeks
Sponsored by: Dell EMC Isilon
A request may be queued while the queue lock is dropped when the mirror is
being destroyed. The corresponding wakeup would be lost, possibly resulting
in an apparent hang of the mirror worker thread.
Tested by: pho (part of a larger patch)
MFC after: 1 week
Sponsored by: Dell EMC Isilon
The worker thread will destroy the mirror provider as part of its teardown
sequence. The call made sense in the initial revision of gmirror, but
became unnecessary in r137248.
Tested by: pho (part of a larger diff)
MFC afteR: 2 weeks
Sponsored by: Dell EMC Isilon
- Don't execute any of g_mirror_shutdown_post_sync() when panicking. We
cannot safely idle the mirror or stop synchronization in that state, and
the current attempts to do so complicate debugging of gmirror itself.
- Check for a non-NULL panicstr instead of using SCHEDULER_STOPPED(). The
latter was added for use in the locking primitives.
Reviewed by: mav, pjd
MFC after: 2 weeks
Sponsored by: Dell EMC Isilon
It is quite specific mode of operation without storing on-disk metadata.
It can be useful in some cases in combination with some external control
tools handling mirror creation and disks hot-plug.
MFC after: 2 weeks
Sponsored by: iXsystems, Inc.
Normally gmirror allows colliding requests to proceed whenever a
synchronization request completes and advances to the next offset. However
if an I/O request collides with one of the final g_mirror_syncreqs, nothing
releases it once synchronization completes, resulting in an apparent I/O
hang. The same problem can occur if synchronization is aborted by an
I/O error. Therefore, be sure to requeue pending requests when
mirror synchronization is stopped for any reason.
While here, remove some dead code from g_mirror_regular_release().
MFC after: 2 weeks
Sponsored by: Dell EMC Isilon
Introduce internal counter to track opens. Using provider's counters is
not very successfull after calling g_wither_provider().
MFC after: 2 weeks
Sponsored by: iXsystems, Inc.
When a syncid bump is pending, any write to the mirror results in the
updated syncid being written to each component's metadata block. However,
the update was only being performed after the writes to the mirror
componenents were queued. Instead, synchronously update the metadata block
first.
MFC after: 3 weeks
Sponsored by: Dell EMC Isilon
Consider a mirror with two components, m1 and m2. Suppose a hardware error
results in the removal of m2, with m1's genid bumped. Suppose further that
a replacement mirror component m3 is created and synchronized, after which
the system is shut down uncleanly. During a subsequent bootup, if gmirror
tastes m1 and m2 first, m2 will be removed from the mirror because it is
broken, but the mirror will be started without bumping the syncid on m1
because all elements of the mirror are accounted for. Then m3 will be
added to the already-running mirror with the same syncid as m1, so the
components will not be synchronized despite the unclean shutdown.
Handle this scenario by bumping the syncid of healthy components if any
broken mirrors are discovered during mirror startup.
MFC after: 3 weeks
Sponsored by: Dell EMC Isilon
Such errors can occur as the result of a write error or because the disk
backing the mirror element was removed. They result in a generation ID bump
on all active elements of the mirror, so we can safely disconnect the mirror
component rather than destroy it.
MFC after: 2 weeks
Sponsored by: EMC / Isilon Storage Division
Differential Revision: https://reviews.freebsd.org/D7750
These are useful for testing changes to I/O error handling, and for
reproducing existing bugs in a controlled manner. The fail points are
g_mirror_regular_request_read
g_mirror_regular_request_write
g_mirror_sync_request_read
g_mirror_sync_request_write
g_mirror_metadata_write
They all effectively allow one to inject an error value into the bio_error
field of a corresponding BIO request as it is being completed.
MFC after: 2 weeks
Sponsored by: EMC / Isilon Storage Division
This will result in lock recursion and is more generally incorrect since
the completion handlers will just reinsert the BIOs into the queue we're
trying to drain.
Reviewed by: imp, ngie
Approved by: re (gjb)
MFC after: 3 weeks
Sponsored by: EMC / Isilon Storage Division
Differential Revision: https://reviews.freebsd.org/D6908
otherwise the system will hang.
This is a temporarily least intrusive crutch to get certain panicing systems
dumping. The proper fix should question is g_mirror_destroy() should be called
on a panicing system at all.
Discussed with: mav
In struct:gctl_req, nargs is unsigned.
In mirror:
g_mirror_syncreqs is unsigned.
In raid:
in struct:g_raid_volume, v_disks_count is unsigned.
In virstor:
in struct:g_virstor_softc, n_components is unsigned.
MFC after: 2 weeks
for all struct bio you get back from g_{new,alloc}_bio. Temporary
bios that you create on the stack or elsewhere should use this before
first use of the bio, and between uses of the bio. At the moment, it
is nothing more than a wrapper around bzero, but that may change in
the future. The wrapper also removes one place where we encode the
size of struct bio in the KBI.
years for head. However, it is continuously misused as the mpsafe argument
for callout_init(9). Deprecate the flag and clean up callout_init() calls
to make them more consistent.
Differential Revision: https://reviews.freebsd.org/D2613
Reviewed by: jhb
MFC after: 2 weeks
When CPU is not busy, those queues are typically empty. When CPU is busy,
then one more extra sorting is the last thing it needs. If specific device
(HDD) really needs sorting, then it will be done later by CAM.
This supposed to fix livelock reported for mirror of two SSDs, when UFS
fires zillion of BIO_DELETE requests, that totally blocks I/O subsystem by
pointless sorting of requests and responses under single mutex lock.
MFC after: 2 weeks
Do not report GEOM::candelete if none of providers support BIO_DELETE.
If consumer still requests BIO_DELETE, report error instead of hanging.
MFC after: 2 weeks
These changes prevent sysctl(8) from returning proper output,
such as:
1) no output from sysctl(8)
2) erroneously returning ENOMEM with tools like truss(1)
or uname(1)
truss: can not get etype: Cannot allocate memory
there is an environment variable which shall initialize the SYSCTL
during early boot. This works for all SYSCTL types both statically and
dynamically created ones, except for the SYSCTL NODE type and SYSCTLs
which belong to VNETs. A new flag, CTLFLAG_NOFETCH, has been added to
be used in the case a tunable sysctl has a custom initialisation
function allowing the sysctl to still be marked as a tunable. The
kernel SYSCTL API is mostly the same, with a few exceptions for some
special operations like iterating childrens of a static/extern SYSCTL
node. This operation should probably be made into a factored out
common macro, hence some device drivers use this. The reason for
changing the SYSCTL API was the need for a SYSCTL parent OID pointer
and not only the SYSCTL parent OID list pointer in order to quickly
generate the sysctl path. The motivation behind this patch is to avoid
parameter loading cludges inside the OFED driver subsystem. Instead of
adding special code to the OFED driver subsystem to post-load tunables
into dynamically created sysctls, we generalize this in the kernel.
Other changes:
- Corrected a possibly incorrect sysctl name from "hw.cbb.intr_mask"
to "hw.pcic.intr_mask".
- Removed redundant TUNABLE statements throughout the kernel.
- Some minor code rewrites in connection to removing not needed
TUNABLE statements.
- Added a missing SYSCTL_DECL().
- Wrapped two very long lines.
- Avoid malloc()/free() inside sysctl string handling, in case it is
called to initialize a sysctl from a tunable, hence malloc()/free() is
not ready when sysctls from the sysctl dataset are registered.
- Bumped FreeBSD version to indicate SYSCTL API change.
MFC after: 2 weeks
Sponsored by: Mellanox Technologies
This fixes the problem, when gmirror starts again just after stop.
The problem occurs when gmirror's component has geom label with equal size.
E.g. gpt and gptid have the same size as partition, diskid has the same
size as entire disk. When gmirror's geom has been destroyed, glabel
creates its providers and this initiate retaste.
Now "gmirror destroy" command is available. It destroys geom and also
erases gmirror's metadata.
MFC after: 2 weeks
Now it is easy to expand the size of the mirror when all its components
are replaced. Also add g_resize method to geom_mirror class. It will write
updated metadata to new last sector, when parent provider is resized.
Silence from: geom@
MFC after: 1 month
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
as clean on shutdown and move that action from shutdown_pre_sync stage to
shutdown_post_sync to avoid extra flapping.
ZFS tends to not close devices on shutdown, that doesn't allow GEOM RAID
to shutdown gracefully. To handle that, mark volume as clean just when
shutdown time comes and there are no active writes.
PR: kern/113957
MFC after: 2 weeks
bytes syncronized.
The rationale behind this is the following: for large disks the
percent synchronisation counter ticks too seldom, and monitoring
software (as well as human operator) can't tell whether
synchronisation goes on or one of disks got stuck. On an idle
server one can look into gstat and see whether synchronisation goes
on or not, but on a busy server that won't work. Also, new value
monitored can be differentiated obtaining the synchronisation speed
quite precisely.
Submitted by: Konstantin Kukushkin <dark ramtel.ru>
Reviewed by: pjd
we need to pass BIO_DELETE requests down to providers that support
it. Also, we need to announce our support for BIO_DELETE to upper
consumer. This requires:
- In g_mirror_start() return true for "GEOM::candelete" request.
- In g_mirror_init_disk() probe below provider for "GEOM::candelete"
attribute, and mark disk with a flag if it does support BIO_DELETE.
- In g_mirror_register_request() distribute BIO_DELETE requests only
to those disks, that do support it.
Note that we announce "GEOM::candelete" as true unconditionally of
whether we have TRIM-capable media down below or not. This is made
intentionally, because upper consumer (usually UFS) requests the
attribite only once at mount time. And if user ever migrates his
mirror from HDDs to SSDs, then he/she would get TRIM working without
remounting filesystem.
Reviewed by: pjd
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.
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
- Instead of measuring last request execution time for each drive and
choosing one with smallest time, use averaged number of requests, running
on each drive. This information is more accurate and timely. It allows to
distribute load between drives in more even and predictable way.
- For each drive track offset of the last submitted request. If new request
offset matches previous one or close for some drive, prefer that drive.
It allows to significantly speedup simultaneous sequential reads.
PR: kern/113885
Reviewed by: sobomax
to kproc_xxx as they actually make whole processes.
Thos makes way for us to add REAL kthread_create() and friends
that actually make theads. it turns out that most of these
calls actually end up being moved back to the thread version
when it's added. but we need to make this cosmetic change first.
I'd LOVE to do this rename in 7.0 so that we can eventually MFC the
new kthread_xxx() calls.
- Use thread_lock() rather than sched_lock for per-thread scheduling
sychronization.
- Use the per-process spinlock rather than the sched_lock for per-process
scheduling synchronization.
Tested by: kris, current@
Tested on: i386, amd64, ULE, 4BSD, libthr, libkse, PREEMPTION, etc.
Discussed with: kris, attilio, kmacy, jhb, julian, bde (small parts each)
gmirror and graid3 in a way that it is not resynchronized after a
power failure or system crash.
It is safe when gjournal is running on top of gmirror/graid3.
request can still have bio_to set to sc_provider (this is READ part of a
synchronization request) and in this case g_{mirror,raid3}_sync() wasn't
called as it should be.
MFC after: 1 week
add count of active and total components to the launched line so you can
see at a glance if your mirror/raid3 is complete...
now:
GEOM_MIRROR: Device mirror/sam launched (2/2).
Reviewed by: pjd
- Comment possible event miss, which isn't critical, but probably can be
fixed by replacing the event lock usage with the queue lock.
MFC after: 2 weeks
which means that devices will be destroyed on last close.
This fixes destruction order problems when, eg. RAID3 array is build on
top of RAID1 arrays.
Requested, reviewed and tested by: ru
MFC after: 2 weeks
requests in parallel.
+ Add kern.geom.mirror.sync_requests tunable which defines how many parallel
I/O requests should be used.
+ Retire kern.geom.mirror.reqs_per_sync and kern.geom.mirror.syncs_per_sec
sysctls.
- Fix race between regular and synchronization requests.
- Reimplement mirror's data synchronization - do not use the topology lock
for this purpose, as it may case deadlocks.
- Stop synchronization from pre-sync hook.
- Fix some other minor issues.
MFC after: 3 days
means that old problem was triggered (when two providers end at the same
offset, eg. ad0 and ad0s1 and the wrong was is picked up by gmirror/graid3).
Reported by: Michal Suszko <dry@dry.pl>
MFC after: 3 days
to preserve currect behaviour). When set to 0, components are not
disconnected - gmirror will try to still use them (only first error will
be logged). This is helpful when we have two broken components, but in
different places, so actually all data is available.
Such buggy component will be visible in 'gmirror list' output with flag
BROKEN.
- Never disconnect the last valid component. If we detect errors there we
will just pass them up. This wasn't reasonable to deny access to the
whole provider because of one broken sector.
Prodded by: ru
MFC after: 3 days
o The only indication of error condition is NULL value returned by
the function;
o value pointed to by error argument is undefined in the case when
operation completes successfully.
Discussed with: phk
- 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.
shared-last-sector problem.
After this change, even if there is more than one provider with the same
last sector, the proper one will be chosen based on its size.
It still doesn't fix the 'c' partition problem (when da0s1 can be confused
with da0s1c) and situation when 'a' partition starts at offset 0
(then da0s1a can be confused with da0s1 and da0s1c). One can use '-h'
option there, when creating device or avoid sharing last sector.
Actually, when providers share the same last sector and their size is equal,
they provide exactly the same data, so the name (da0s1, da0s1a, da0s1c)
isn't important at all.
- Provide backward compatibility.
- Update copyright's year.
MFC after: 1 week
