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
Introduce new function devstat_end_transaction_bio_bt(), adding new argument
to specify present time. Use this function to move binuptime() out of lock,
substantially reducing lock congestion when slow timecounter is used.
originally inspired by the Solaris vmem detailed in the proceedings
of usenix 2001. The NetBSD version was heavily refactored for bugs
and simplicity.
- Use this resource allocator to allocate the buffer and transient maps.
Buffer cache defrags are reduced by 25% when used by filesystems with
mixed block sizes. Ultimately this may permit dynamic buffer cache
sizing on low KVA machines.
Discussed with: alc, kib, attilio
Tested by: pho
Sponsored by: EMC / Isilon Storage Division
do not map the b_pages pages into buffer_map KVA. The use of the
unmapped buffers eliminate the need to perform TLB shootdown for
mapping on the buffer creation and reuse, greatly reducing the amount
of IPIs for shootdown on big-SMP machines and eliminating up to 25-30%
of the system time on i/o intensive workloads.
The unmapped buffer should be explicitely requested by the GB_UNMAPPED
flag by the consumer. For unmapped buffer, no KVA reservation is
performed at all. The consumer might request unmapped buffer which
does have a KVA reserve, to manually map it without recursing into
buffer cache and blocking, with the GB_KVAALLOC flag.
When the mapped buffer is requested and unmapped buffer already
exists, the cache performs an upgrade, possibly reusing the KVA
reservation.
Unmapped buffer is translated into unmapped bio in g_vfs_strategy().
Unmapped bio carry a pointer to the vm_page_t array, offset and length
instead of the data pointer. The provider which processes the bio
should explicitely specify a readiness to accept unmapped bio,
otherwise g_down geom thread performs the transient upgrade of the bio
request by mapping the pages into the new bio_transient_map KVA
submap.
The bio_transient_map submap claims up to 10% of the buffer map, and
the total buffer_map + bio_transient_map KVA usage stays the
same. Still, it could be manually tuned by kern.bio_transient_maxcnt
tunable, in the units of the transient mappings. Eventually, the
bio_transient_map could be removed after all geom classes and drivers
can accept unmapped i/o requests.
Unmapped support can be turned off by the vfs.unmapped_buf_allowed
tunable, disabling which makes the buffer (or cluster) creation
requests to ignore GB_UNMAPPED and GB_KVAALLOC flags. Unmapped
buffers are only enabled by default on the architectures where
pmap_copy_page() was implemented and tested.
In the rework, filesystem metadata is not the subject to maxbufspace
limit anymore. Since the metadata buffers are always mapped, the
buffers still have to fit into the buffer map, which provides a
reasonable (but practically unreachable) upper bound on it. The
non-metadata buffer allocations, both mapped and unmapped, is
accounted against maxbufspace, as before. Effectively, this means that
the maxbufspace is forced on mapped and unmapped buffers separately.
The pre-patch bufspace limiting code did not worked, because
buffer_map fragmentation does not allow the limit to be reached.
By Jeff Roberson request, the getnewbuf() function was split into
smaller single-purpose functions.
Sponsored by: The FreeBSD Foundation
Discussed with: jeff (previous version)
Tested by: pho, scottl (previous version), jhb, bf
MFC after: 2 weeks
It includes three parts:
1) Modifications to CAM to detect media media changes and report them to
disk(9) layer. For modern SATA (and potentially UAS) devices it utilizes
Asynchronous Notification mechanism to receive events from hardware.
Active polling with TEST UNIT READY commands with 3 seconds period is used
for incapable hardware. After that both CD and DA drivers work the same way,
detecting two conditions: "NOT READY: Medium not present" after medium was
detected previously, and "UNIT ATTENTION: Not ready to ready change, medium
may have changed". First one reported to disk(9) as media removal, second
as media insert/change. To reliably receive second event new
AC_UNIT_ATTENTION async added to make UAs broadcasted to all periphs by
generic error handling code in cam_periph_error().
2) Modifications to GEOM core to handle media remove and change events.
Media removal handled by spoiling all consumers attached to the provider.
Media change event also schedules provider retaste after spoiling to probe
new media. New flag G_CF_ORPHAN was added to consumers to reflect that
consumer is in process of destruction. It allows retaste to create new
geom instance of the same class, while previous one is still dying.
3) Modifications to some GEOM classes: DEV -- to report media change
events to devd; VFS -- to handle spoiling same as orphan to prevent
accessing replaced media. PART class already handles spoiling alike to
orphan.
Reviewed by: silence on geom@ and scsi@
Tested by: avg
Sponsored by: iXsystems, Inc. / PC-BSD
MFC after: 2 months
Add the BIO_ORDERED flag for struct bio and update bio clients to use it.
The barrier semantics of bioq_insert_tail() were broken in two ways:
o In bioq_disksort(), an added bio could be inserted at the head of
the queue, even when a barrier was present, if the sort key for
the new entry was less than that of the last queued barrier bio.
o The last_offset used to generate the sort key for newly queued bios
did not stay at the position of the barrier until either the
barrier was de-queued, or a new barrier (which updates last_offset)
was queued. When a barrier is in effect, we know that the disk
will pass through the barrier position just before the
"blocked bios" are released, so using the barrier's offset for
last_offset is the optimal choice.
sys/geom/sched/subr_disk.c:
sys/kern/subr_disk.c:
o Update last_offset in bioq_insert_tail().
o Only update last_offset in bioq_remove() if the removed bio is
at the head of the queue (typically due to a call via
bioq_takefirst()) and no barrier is active.
o In bioq_disksort(), if we have a barrier (insert_point is non-NULL),
set prev to the barrier and cur to it's next element. Now that
last_offset is kept at the barrier position, this change isn't
strictly necessary, but since we have to take a decision branch
anyway, it does avoid one, no-op, loop iteration in the while
loop that immediately follows.
o In bioq_disksort(), bypass the normal sort for bios with the
BIO_ORDERED attribute and instead insert them into the queue
with bioq_insert_tail(). bioq_insert_tail() not only gives
the desired command order during insertion, but also provides
barrier semantics so that commands disksorted in the future
cannot pass the just enqueued transaction.
sys/sys/bio.h:
Add BIO_ORDERED as bit 4 of the bio_flags field in struct bio.
sys/cam/ata/ata_da.c:
sys/cam/scsi/scsi_da.c
Use an ordered command for SCSI/ATA-NCQ commands issued in
response to bios with the BIO_ORDERED flag set.
sys/cam/scsi/scsi_da.c
Use an ordered tag when issuing a synchronize cache command.
Wrap some lines to 80 columns.
sys/cddl/contrib/opensolaris/uts/common/fs/zfs/vdev_geom.c
sys/geom/geom_io.c
Mark bios with the BIO_FLUSH command as BIO_ORDERED.
Sponsored by: Spectra Logic Corporation
MFC after: 1 month
Previsouly this condition was reported with EIO by bio_offset > mediasize
check.
Perhaps that check should be extended to bio_offset+bio_length > mediasize.
MFC after: 1 week
Remove msleep() timeout from g_io_schedule_up/down(). It works fine
without it, saving few percents of CPU on high request rates without
need to rearm callout twice per request.
struct bio to store classification information, and a hook
for classifier functions that can be called by g_io_request().
This code is from Fabio Checconi as part of his GSOC work.
state where sleeping on a sleep queue is not allowed. The facility
doesn't support recursion but uses a simple private per-thread flag
(TDP_NOSLEEPING). The sleepq_add() function will panic if the flag is
set and INVARIANTS is enabled.
- Use this new facility to replace the g_xup and g_xdown mutexes that were
(ab)used to achieve similar behavior.
- Disallow sleeping in interrupt threads when invoking interrupt handlers.
MFC after: 1 week
Reviewed by: phk
the assumption that performance was more important that beancounter
quality statistics.
As it transpires the microoptimization is not measurable in the
real world and the inconsistent statistics confuse users, so revert
the decision.
MT6 candidate: possibly
MT5 candidate: possibly
in the g_up and g_down threads. Each time a bio is propelled up and
down the stack, an event is generating showing the provider, offset,
and length, as well as thread wakeup and work status information.
and bio_inbed fields to 0. Without this change we can end up with
I/O leakage in some rare situations.
I tested this change by putting failure probability mechanism simlar
to this used in NOP class into g_clone_bio(9) function, so it was
able to return NULL with the given probability.
Discussed with: phk
to warn about attempts to sleep in the I/O path. This change pushes the
definition and use of 'mymutex' behind #ifdef WITNESS to avoid the cost
in non-debugging cases. This results in a clear .22% performance win for
512 byte and 1k I/O tests on my SMP test box. Not much, but every bit
counts.
