When processing a taskqueue and a task has associated epoch, then
enter for duration of the task. If consecutive tasks belong to the
same epoch, batch them. Now we are talking about the network epoch
only.
Shrink the ta_priority size to 8-bits. No current consumers use
a priority that won't fit into 8 bits. Also complexity of
taskqueue_enqueue() is a square of maximum value of priority, so
we unlikely ever want to go over UCHAR_MAX here.
Reviewed by: hselasky
Differential Revision: https://reviews.freebsd.org/D23518
Eliminate recursion from most thread_lock consumers. Return from
sched_add() without the thread_lock held. This eliminates unnecessary
atomics and lock word loads as well as reducing the hold time for
scheduler locks. This will eventually allow for lockless remote adds.
Discussed with: kib
Reviewed by: jhb
Tested by: pho
Differential Revision: https://reviews.freebsd.org/D22626
- Optimize enqueue for two task priority values by adding new tq_hint
field, pointing to the last task inserted into the middle of the list.
In case of more then two priority values it should halve average search.
- Move tq_active insert/remove out of the taskqueue_run_locked loop.
Instead of dirtying few shared cache lines per task introduce different
mechanism to drain active tasks, based on task sequence number counter,
that uses only cache lines already present in cache. Since the new
mechanism does not need ordering, switch tq_active from TAILQ to LIST.
- Move static and dynamic struct taskqueue fields into different cache
lines. Move lock into its own cache line, so that heavy lock spinning
by multiple waiting threads would not affect the running thread.
- While there, correct some TQ_SLEEP() wait messages.
This change fixes certain ZFS write workloads, causing huge congestion
on taskqueue lock. Those workloads combine some large block writes to
saturate the pool and trigger allocation throttling, which uses higher
priority tasks to requeue the delayed I/Os, with many small blocks to
generate deep queue of small tasks for taskqueue to sort.
MFC after: 1 week
Sponsored by: iXsystems, Inc.
So far, intr_{g,s}etaffinity(9) take a single int for identifying
a device interrupt. This approach doesn't work on all architectures
supported, as a single int isn't sufficient to globally specify a
device interrupt. In particular, with multiple interrupt controllers
in one system as found on e. g. arm and arm64 machines, an interrupt
number as returned by rman_get_start(9) may be only unique relative
to the bus and, thus, interrupt controller, a certain device hangs
off from.
In turn, this makes taskqgroup_attach{,_cpu}(9) and - internal to
the gtaskqueue implementation - taskqgroup_attach_deferred{,_cpu}()
not work across architectures. Yet in turn, iflib(4) as gtaskqueue
consumer so far doesn't fit architectures where interrupt numbers
aren't globally unique.
However, at least for intr_setaffinity(..., CPU_WHICH_IRQ, ...) as
employed by the gtaskqueue implementation to bind an interrupt to a
particular CPU, using bus_bind_intr(9) instead is equivalent from
a functional point of view, with bus_bind_intr(9) taking the device
and interrupt resource arguments required for uniquely specifying a
device interrupt.
Thus, change the gtaskqueue implementation to employ bus_bind_intr(9)
instead and intr_{g,s}etaffinity(9) to take the device and interrupt
resource arguments required respectively. This change also moves
struct grouptask from <sys/_task.h> to <sys/gtaskqueue.h> and wraps
struct gtask along with the gtask_fn_t typedef into #ifdef _KERNEL
as userland likes to include <sys/_task.h> or indirectly drags it
in - for better or worse also with _KERNEL defined -, which with
device_t and struct resource dependencies otherwise is no longer
as easily possible now.
The userland inclusion problem probably can be improved a bit by
introducing a _WANT_TASK (as well as a _WANT_MOUNT) akin to the
existing _WANT_PRISON etc., which is orthogonal to this change,
though, and likely needs an exp-run.
While at it:
- Change the gt_cpu member in the grouptask structure to be of type
int as used elswhere for specifying CPUs (an int16_t may be too
narrow sooner or later),
- move the gtaskqueue_enqueue_fn typedef from <sys/gtaskqueue.h> to
the gtaskqueue implementation as it's only used and needed there,
- change the GTASK_INIT macro to use "gtask" rather than "task" as
argument given that it actually operates on a struct gtask rather
than a struct task, and
- let subr_gtaskqueue.c consistently use __func__ to print functions
names.
Reported by: mmel
Reviewed by: mmel
Differential Revision: https://reviews.freebsd.org/D19139
taskqgroup_detach() would remove the task even if it was running or
enqueued, which could lead to panics (see D17404). With this change,
taskqgroup_detach() drains the task and sets a new flag which prevents the
task from being scheduled again.
I've added grouptask_block() and grouptask_unblock() to allow control
over the flag from other locations as well.
Reviewed by: Jeffrey Pieper <jeffrey.e.pieper@intel.com>
MFC after: 1 week
Sponsored by: Limelight Networks
Differential Revision: https://reviews.freebsd.org/D17674
That commit is causing kernel panics in em(4), so this will be reverted
until those are fixed.
Reported by: ae@, pho@, et al
Sponsored by: Intel Corporation
The taskqgroup_detach function does not check if task is already enqueued when
detaching it. This may lead to kernel panic if enqueued task starts after
context state lock is destroyed. Ensure that the already enqueued admin tasks
are executed before detaching them.
The issue was discovered during validation of D16429. Unloading of if_ixlv
followed by immediate removal of VFs with iovctl -D may lead to panic on
NODEBUG kernel.
As well, check if iflib is in detach before enqueueing new admin or iov
tasks, to prevent new tasks from executing while the taskqgroup tasks
are being drained.
Submitted by: Krzysztof Galazka <krzysztof.galazka@intel.com>
Reviewed by: shurd@, erj@
Sponsored by: Intel Corporation
Differential Revision: https://reviews.freebsd.org/D17404
On PowerPC (and possibly other architectures), that doesn't use
EARLY_AP_STARTUP, the config task queue may be used initialized.
This was observed while trying to mount the root fs from NFS, as
reported here: https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=230168.
This patch has 2 main changes:
1- Perform a basic initialization of qgroup_config, similar to
what is done in taskqgroup_adjust, but simpler.
This makes qgroup_config ready to be used during NFS root mount.
2- When EARLY_AP_STARTUP is not used, call inm_init() and
in6m_init() right before SI_SUB_ROOT_CONF, because bootp needs
to send multicast packages to request an IP.
PR: Bug 230168
Reported by: sbruno
Reviewed by: jhibbits, mmacy, sbruno
Approved by: jhibbits
Differential Revision: D16633
The *name parameter passed to iflib_irq_alloc_generic and
iflib_softirq_alloc_generic is never modified. Many places in code pass
string literals and thus should not be modified.
Mark the *name parameter as a const char * instead, so that we enforce
that the name is not modified before passing to bus_describe_intr()
Submitted by: Jacob Keller <jacob.e.keller@intel.com>
Reviewed by: kmacy
Sponsored by: Intel Corporation
Differential Revision: https://reviews.freebsd.org/D15343
Multicast incorrectly calls in to drivers with a mutex held causing drivers
to have to go through all manner of contortions to use a non sleepable lock.
Serialize multicast updates instead.
Submitted by: mmacy <mmacy@mattmacy.io>
Reviewed by: shurd, sbruno
Sponsored by: Limelight Networks
Differential Revision: https://reviews.freebsd.org/D14969
Improved logging added in r323879 exposed an error during
attach. We need the irq, not the rid to work correctly. em uses
shared irqs, so it will use the same irq for TX as RX. bnxt does
not use shared irqs, or TX irqs at all, so there's no need to set
the TX irq affinity.
Reviewed by: sbruno
Approved by: sbruno (mentor)
Sponsored by: Limelight Networks
Differential Revision: https://reviews.freebsd.org/D12496
Previously, it was just a pointer which was copied, but
some callers pass in a stack variable which will go out of scope.
Add GROUPTASK_NAMELEN macro (32) and snprintf() the name into it,
using "grouptask" if name is NULL. We can now safely include
gtask->gt_name in console messages.
Reviewed by: sbruno
Approved by: sbruno (mentor)
Sponsored by: Limelight Networks
Differential Revision: https://reviews.freebsd.org/D12449
It doesn't appear to be safe to use gtask->gt_name.
Reported by: Mark Johnston, Jenkins
Reviewed by: sbruno
Approved by: sbruno (mentor)
Sponsored by: Limelight Networks
Differential Revision: https://reviews.freebsd.org/D12448
Check the return code of intr_setaffinity() and log any errors
it returns. When a qid is not located, log an error before returning
failure. Also, use __func__ rather than hardcoding the function name
Reviewed by: sbruno
Approved by: sbruno (mentor)
Sponsored by: Limelight Networks
Differential Revision: https://reviews.freebsd.org/D12436
Previously had the same short and long description as taskqueues.
This could cause problems with memguard(9) and vmstat -m which use
the short description as a unique identifier.
Reviewed by: sbruno
Approved by: sbruno (mentor)
MFC after: 3 days
Sponsored by: Limelight Networks
Differential Revision: https://reviews.freebsd.org/D12438
This was really too big of a commit even if everything worked, but there
are multiple new issues introduced in the one huge commit, so it's not
worth keeping this until it's fixed.
I'll work on splitting this up into logical chunks and introduce them one
at a time over the next week or two.
Approved by: sbruno (mentor)
Sponsored by: Limelight Networks
by Matt Macy as well as other changes which he has accepted via pull
request to his github repo at https://github.com/mattmacy/networking/
This should bring -CURRENT and the github repo into close enough sync to
allow small feature branches rather than a large chain of interdependant
patches being developed out of tree. The reset of the synchronization
should be able to be completed on github by splitting the remaining
changes that are not yet ready into short feature branches for later
review as smaller commits.
Here is a summary of changes included in this patch:
1) More checks when INVARIANTS are enabled for eariler problem
detection
2) Group Task Queue cleanups
- Fix use of duplicate shortdesc for gtaskqueue malloc type.
Some interfaces such as memguard(9) use the short description to
identify malloc types, so duplicates should be avoided.
3) Allow gtaskqueues to use ithreads in addition to taskqueues
- In some cases, this can improve performance
4) Better logging when taskqgroup_attach*() fails to set interrupt
affinity.
5) Do not start gtaskqueues until they're needed
6) Have mp_ring enqueue function enter the ABDICATED rather than BUSY
state. This moves the TX to the gtaskq and allows processing to
continue faster as well as make TX batching more likely.
7) Add an ift_txd_errata function to struct if_txrx. This allows
drivers to inspect/modify mbufs before transmission.
8) Add a new IFLIB_NEED_ZERO_CSUM for drivers to indicate they need
checksums zeroed for checksum offload to work. This avoids modifying
packet data in the TX path when possible.
9) Use ithreads for iflib I/O instead of taskqueues
10) Clean up ioctl and support async ioctl functions
11) Prefetch two cachlines from each mbuf instead of one up to 128B. We
often need to parse packet header info beyond 64B.
12) Fix potential memory corruption due to fence post error in
bit_nclear() usage.
13) Improved hang detection and handling
14) If the packet is smaller than MTU, disable the TSO flags.
This avoids extra packet parsing when not needed.
15) Move TCP header parsing inside the IS_TSO?() test.
This avoids extra packet parsing when not needed.
16) Pass chains of mbufs that are not consumed by lro to if_input()
rather call if_input() for each mbuf.
17) Re-arrange packet header loads to get as much work as possible done
before a cache stall.
18) Lock the context when calling IFDI_ATTACH_PRE()/IFDI_ATTACH_POST()/
IFDI_DETACH();
19) Attempt to distribute RX/TX tasks across cores more sensibly,
especially when RX and TX share an interrupt. RX will attempt to
take the first threads on a core, and TX will attempt to take
successive threads.
20) Allow iflib_softirq_alloc_generic() to request affinity to the same
cpus an interrupt has affinity with. This allows TX queues to
ensure they are serviced by the socket the device is on.
21) Add new iflib sysctls to net.iflib:
- timer_int - interval at which to run per-queue timers in ticks
- force_busdma
22) Add new per-device iflib sysctls to dev.X.Y.iflib
- rx_budget allows tuning the batch size on the RX path
- watchdog_events Count of watchdog events seen since load
23) Fix error where netmap_rxq_init() could get called before
IFDI_INIT()
24) e1000: Fixed version of r323008: post-cold sleep instead of DELAY
when waiting for firmware
- After interrupts are enabled, convert all waits to sleeps
- Eliminates e1000 software/firmware synchronization busy waits after
startup
25) e1000: Remove special case for budget=1 in em_txrx.c
- Premature optimization which may actually be incorrect with
multi-segment packets
26) e1000: Split out TX interrupt rather than share an interrupt for
RX and TX.
- Allows better performance by keeping RX and TX paths separate
27) e1000: Separate igb from em code where suitable
Much easier to understand separate functions and "if (is_igb)" than
previous tests like "if (reg_icr & (E1000_ICR_RXSEQ | E1000_ICR_LSC))"
#blamebruno
Reviewed by: sbruno
Approved by: sbruno (mentor)
Sponsored by: Limelight Networks
Differential Revision: https://reviews.freebsd.org/D12235
Add IRQ placement-only and ithread-only API variants. intr_event_bind
has been extended with sibling methods, as it has many more callsites in
existing code.
Reviewed by: kib@, adrian@ (earlier version)
Sponsored by: Dell EMC Isilon
Differential Revision: https://reviews.freebsd.org/D10586
Add internal tracking of smp startup status to reliably figure out
what methods are to be used to get gtaskqueue up and running.
e1000:
Calculating this pointer gives undefined behaviour when (last == -1)
(it is before the buffer). The pointer is always followed. Panics
occurred when it points to an unmapped page. Otherwise, the pointed-to
garbage tends to not have the E1000_TXD_STAT_DD bit set in it, so in the
broken case the loop was usually null and the function just returned, and
this was acidentally correct.
Submitted by: bde
Reported by: Matt Macy <mmacy@nextbsd.org>
Add internal tracking of smp startup status to reliably figure out
what methods are to be used to get gtaskqueue up and running.
e1000:
Calculating this pointer gives undefined behaviour when (last == -1)
(it is before the buffer). The pointer is always followed. Panics
occurred when it points to an unmapped page. Otherwise, the pointed-to
garbage tends to not have the E1000_TXD_STAT_DD bit set in it, so in the
broken case the loop was usually null and the function just returned, and
this was acidentally correct.
Submitted by: bde
Reviewed by: Matt Macy <mmacy@nextbsd.org>
gtaskqueue bits at SI_SUB_INIT_IF instead of waiting until SI_SUB_SMP
which is far too late.
Add an assertion in taskqgroup_attach() to catch startup initialization
failures in the future.
Reported by: kib bde
sys/net/iflib.c:
Add ctx to filter_info and don't skpi interrupt early on unless we're on an
SMP system
sys/kern/subr_gtaskqueue.c:
Skip smp check if we're running UP
Submitted by: Matt Macy <mmacy@nextbsd.org>
Reported by: emaste bde
- iflib - add checksum in place support (mmacy)
- iflib - initialize IP for TSO (going to be needed for e1000) (mmacy)
- iflib - move isc_txrx from shared context to softc context (mmacy)
- iflib - Normalize checks in TXQ drainage. (shurd)
- iflib - Fix queue capping checks (mmacy)
- iflib - Fix invalid assert, em can need 2 sentinels (mmacy)
- iflib - let the driver determine what capabilities are set and what
tx csum flags are used (mmacy)
- add INVARIANTS debugging hooks to gtaskqueue enqueue (mmacy)
- update bnxt(4) to support the changes to iflib (shurd)
Some other various, sundry updates. Slightly more verbose changelog:
Submitted by: mmacy@nextbsd.org
Reviewed by: shurd
mFC after:
Sponsored by: LimeLight Networks and Dell EMC Isilon
ref: 535865d02c
Fix cpu assignment by assuring stride is non-zero, assert that all tasks
have a valid taskqueue.
ref: db39817623
Start cpu assignment from zero.
ref: d99d39b6b6
Submitted by: mmacy@nextbsd.org
- Move group task queue into kern/subr_gtaskqueue.c
- Change intr_enable to return an int so it can be detected if it's not
implemented
- Allow different TX/RX queues per set to be different sizes
- Don't split up TX mbufs before transmit
- Allow a completion queue for TX as well as RX
- Pass the RX budget to isc_rxd_available() to allow an earlier return
and avoid multiple calls
Submitted by: shurd
Reviewed by: gallatin
Approved by: scottl
Differential Revision: https://reviews.freebsd.org/D7393
