Put a bandaid to prevent ixgbe(4) from completely locking up the system
under high load. Our platform has a few CPU cores and a single active
ixgbe(4) port with 4 queues. Under high enough traffic load, at about
7.5GBs and 700,000 packets/sec (outbound), the entire system would
deadlock. What we found was that each CPU was in an endless loop on a
different ix taskqueue thread. The OACTIVE flag had gotten set on each
queue, and the ixgbe_handle_queue() function was continuously rescheduling
itself via the taskqueue_enqueue. Since all CPUs were busy with their
taskqueue threads, the ixgbe_local_timer() function couldn't run to clear
the OACTIVE flag.
Submitted by: scottl
MFC after: 1 week
it turns out that it negatively affects performance. I'm stil investigating
exactly why deferring the IO causes such negative TCP performance but
doesn't affect UDP preformance.
Leave the ath_tx_kick() change in there however; it's going to be useful
to have that there for if_transmit() work.
PR: kern/168649
called to "kick" along TX.
For now, schedule a taskqueue call.
Later on I may go back to the direct call of ath_rx_tasklet() - but for
now, this will do.
I've tested UDP and TCP TX. UDP TX still achieves 240MBit, but TCP
TX gets stuck at around 100MBit or so, instead of the 150MBit it should
be at. I'll re-test with no ACPI/power/sleep states enabled at startup
and see what effect it has.
This is in preparation for supporting an if_transmit() path, which will
turn ath_tx_kick() into a NUL operation (as there won't be an ifnet
queue to service.)
Tested:
* AR9280 STA
TODO:
* test on AR5416, AR9160, AR928x STA/AP modes
PR: kern/168649
implementing parallel TX and TX/RX completion can be done without
simply abusing long-held locks.
Right now, multiple concurrent ath_start() entries can result in
frames being dequeued out of order. Well, they're dequeued in order
fine, but if there's any preemption or race between CPUs between:
* removing the frame from the ifnet, and
* calling and runningath_tx_start(), until the frame is placed on a
software or hardware TXQ
Then although dequeueing the frame is in-order, queueing it to the hardware
may be out of order.
This is solved in a lot of other drivers by just holding a TX lock over
a rather long period of time. This lets them continue to direct dispatch
without races between dequeue and hardware queue.
Note to observers: if_transmit() doesn't necessarily solve this.
It removes the ifnet from the main path, but the same issue exists if
there's some intermediary queue (eg a bufring, which as an aside also
may pull in ifnet when you're using ALTQ.)
So, until I can sit down and code up a much better way of doing parallel
TX, I'm going to leave the TX path using a deferred taskqueue task.
What I will likely head towards is doing a direct dispatch to hardware
or software via if_transmit(), but it'll require some driver changes to
allow queues to be made without using the really large ath_buf / ath_desc
entries.
TODO:
* Look at how feasible it'll be to just do direct dispatch to
ath_tx_start() from if_transmit(), avoiding doing _any_ intermediary
serialisation into a global queue. This may break ALTQ for example,
so I have to be delicate.
* It's quite likely that I should break up ath_tx_start() so it
deposits frames onto the software queues first, and then only fill
in the 802.11 fields when it's being queued to the hardware.
That will make the if_transmit() -> software queue path very
quick and lightweight.
* This has some very bad behaviour when using ACPI and Cx states.
I'll do some subsequent analysis using KTR and schedgraph and file
a follow-up PR or two.
PR: kern/168649
update for ZFS. It seems that this does not really affect anything except
the help command. Nevertheless, rearrange things so loaddev is set only
once in all cases in order to get it right.
Pointed out by: avg
MFC after: r235364
a single device to be opened multiple times concurrently unfortunately
isn't sufficient with ZFS. This is due to the fact, that ZFS may open
different partitions of a single device simultaneously. So the best we
can do in this case is to cache the lastly used device path and close
and open devices in ofwd_strategy() as needed.
PR: 165025
Submitted by: Gavin Mu
MFC after: 1 week
EARLY_BUILD macro: the -Qunused-arguments flag isn't passed anymore when
building this particular program. However, with clang 3.1 and -Werror,
such unused argument warnings are flagged as errors, causing buildkernel
to fail at this stage, due to the -nostdinc flag passed during linking.
Since the -nostdinc flag isn't actually needed, just remove it.
X-MFC-With: r236528
The skew calculation here is exactly backwards. We were able to repro
it on a multi-package ESX server running a FreeBSD VM, where the TSCs
can be pretty evil.
MFC after: 1 week
Submitted by: Jeff Ford <jeffrey.ford2@isilon.com>
Reviewed by: avg, gnn
m_cat(), storing pointer to last mbuf in chain in local variable and
attaching new mbuf to the end of chain.
Submitter reports that CPU load dropped for > 10% on a web server
serving large files with this optimisation.
Submitted by: Sergey Budnevitch <sb nginx.com>
Constify pc_freemask[].
pmap_pv_reclaim()
Eliminate "freemask" because it was a pessimization. Add a comment about
the resident count adjustment.
free_pv_entry() [i386 only]
Merge an optimization from amd64 (r233954).
get_pv_entry()
Eliminate the move to tail of the pv_chunk on the global pv_chunks list.
(The right strategy needs more thought. Moreover, there were unintended
differences between the amd64 and i386 implementation.)
pmap_remove_pages()
Eliminate unnecessary ()'s.
structure with the first 4 registers to allow a wider range of boot
loaders to work. Future commits will make use of this to centralize
support for the different loaders.
Interface routes are refcounted as packets move through the stack,
and there's garbage collection tied to it so that route changes can
safely propagate while traffic is flowing. In our setup, we weren't
changing or deleting any routes, but the refcounting logic in
ip6_input() was wrong and caused a reference leak on every inbound
V6 packet. This eventually caused a 32bit overflow, and the resulting
0 value caused the garbage collection to run on the active route.
That then snowballed into the panic.
Reviewed by: scottl
MFC after: 3 days
- Make the device description match the driver name.
- Identify the chip variant based on the JEDEC and use that information
to use the proper values for page count, offset and size instead of
hardcoding a AT45DB642x with 2^N byte page support disabled.
- Take advantage of bioq_takefirst().
- Given that CONTINUOUS_ARRAY_READ_HF (0x0b) command isn't even mentioned
in Atmel's DataFlash Application Note, as suggested by the previous
comment may not work on all all devices and actually doesn't properly
on at least AT45DB321D (JEDEC 0x1f2701), rewrite at45d_task() to use
CONTINUOUS_ARRAY_READ (0xe8) for reading instead. This rewrite is laid
out in a way allowing to easily add support for BIO_DELETE later on.
- Add support for reads and writes not starting on a page boundary.
- Verify the flash content after writing.
- Let at45d_task() gracefully handle errors on SPI transfers and the
device not becoming ready afterwards again. [1]
- Use DEVMETHOD_END. [1]
- Use NULL instead of 0 for pointers. [1]
Additional testing by: Ian Lepore
Submitted by: Ian Lepore [1]
MFC after: 1 week
- Move DMA tag and map creature to at91_spi_activate() where the other
resource allocation also lives. [1]
- Flesh out at91_spi_deactivate(). [1]
- Work around the "Software Reset must be Written Twice" erratum.
- For now, run the bus at the slowest speed possible in order to work
around data corruption on transit even seen with 9 MHz on ETHERNUT5
(15 MHz maximum) and AT45DB321D (20 MHz maximum). This also serves as
a poor man's work-around for the "NPCSx rises if no data data is to be
transmitted" erratum of RM9200. Being able to use the appropriate bus
speed would require:
1) Adding a proper work-around for the RM9200 bug consisting of taking
the chip select control away from the SPI peripheral and managing it
directly as a GPIO line.
2) Taking the maximum frequencies supported by the actual board and the
slave devices into account and basing the whole thing on the master
clock instead of hardcoding a divisor as previously done.
3) Fixing the above mentioned data corruption.
- KASSERT that TX/RX command and data sizes match on transfers.
- Introduce a mutex ensuring that only one child device is running a SPI
transfer at a time. [1]
- Add preliminary, #ifdef'ed out support for setting the chip select. [1]
- Use the RX instead of the TX commando size when setting up the RX side
of a transfer.
- For controllers having SPI_SR_TXEMPTY, i.e. !RM9200, also wait for the
completion of the TX part of transfers before stopping the whole thing
again.
- Use DEVMETHOD_END. [1]
- Use NULL instead of 0 for pointers. [1, partially]
Additional testing by: Ian Lepore
Submitted by: Ian Lepore [1]
MFC after: 1 week
