Ravi Pokala [Tue, 24 Jan 2017 01:39:40 +0000 (01:39 +0000)]
Eliminate misleading comments and dead code in lacp_port_create()
Variables "fast" and "active" are both constant in lacp_port_create(), but
comments mispleadingly suggest that "fast" can be changed via ioctl. The
constant values control the value of "lp->lp_state", so it too is constant,
and the code for assigning different value to it is essentially dead.
Remove both "fast" and "active", and set "lp->lp_state" unconditionally;
that gets rid of the dead code and misleading comments.
[am335x] Use upstream tda19988 framer node from upstream DTS
Remove custom DTS duplicate of tda19988 node and use upstream-provided
one introduced by r295436. This duplication created two tdaX devices
which confused fb driver into using only 640x480 area while setting
display to native resolution.
Adrian Chadd [Mon, 23 Jan 2017 04:47:38 +0000 (04:47 +0000)]
[ath] modify cabq and per-node packet usage limits.
* limit cabq to 64 - in practice if this stays at ath_txbuf then
all buffers can be tied up by a very busy broadcast domain (eg ARP
storm, way too much MDNS/NETBIOS). It's been like this in the
freebsd-wifi-build AP project for the longest time.
* Now that I figured out the hilarity inherent in aggregate forming
and AR9380 EDMA work, change the per-node to 64 frames by default.
I'll do some more work to shorten the queue latency introduced when
doing data so TCP isn't so terrible, but it's now no longer /always/
tens of milliseconds of extra latency when doing active iperf tests.
Notes:
The reason for the extra latency is partly tx/rx taskqueue handling and
scheduling, and partly due to a lack of airtime/QoS awareness of per-node
traffic. Ideally we'd have different limits/priorities on the QoS/TID
levels per node so say, voice/video data got a better share of buffer
allocations over best effort/bulk data, but we currently don't implement
that. It's not /hard/ to do, I just need to do it.
Tested:
* AR9380 (STA), AR9580 (hostap) - both with the relevant changes.
TCP is now at around 180mbit with rate control and RTS protection
enabled. UDP stays at 355mbit at MCS23, no HT protection.
Adrian Chadd [Mon, 23 Jan 2017 04:30:08 +0000 (04:30 +0000)]
[ath] fix thresholds for deciding to queue to the software queue and populate hardware frames
This is two fixes, which establishes what I /think/ is pretty close to the
theoretical PHY maximum speed on the AR9380 devices.
* When doing A-MPDU on a TID, don't queue to the hardware directly if
the hardware queue is busy. This gives us time to get more packets
queued up (and the hardware is busy, so there's no point in queuing
more to the hardware right now) to potentially form an A-MPDU.
This fixes up the throughput issue I was seeing where a couple hundred
single frames were being sent a second interspersed between A-MPDU
frames. It just happened that the software queue had exactly one
frame in it at that point. Queuing it until the hardware finishes
transmitting isn't exactly costly.
* When determining whether to dequeue from a software node/TID queue into
the hardware queue, fix up the checks to work right for EDMA chips
(ar9380 and later.) Before it was not dispatching anything until
the FIFO was empty. Now we allow it to dispatch another aggregate
up to the hardware aggregate limit, like I intended with the earlier
work.
This allows a 5GHz HT40, short-GI, "htprotmode off" test at MCS23
to achieve 357 Mbit/sec in a one-way UDP test. The stars have to be
aligned /just right/ so there are no retries but it can happen.
Just don't expect it to work in an OTA test if your 2yo is running
around the room - MCS23 is very very sensitive to channel conditions.
Tested:
* AR9380 STA (test) -> AR9580 hostap
TODO:
* More thorough testing on pre-AR9380 chips (AR5416, AR9160, AR9280)
* (Finally) teach ath_rate_sample about throughput/latency rather than
air time, so I can get good transmit rates with a 2yo running around.
Adrian Chadd [Mon, 23 Jan 2017 04:20:37 +0000 (04:20 +0000)]
[ath] [ar9300] ensure the software scheduler is called to form more aggregates for EDMA chips
When investigating performance on UDP TX on the AR9380 I found that the
following sequence was occuring:
* INTR
* EINPROGRESS - nothing yet
* INTR
* TXSTATUS - process a TX completion for an aggregate
* INTR, INTR
* TXSTATUS - process a TX completion for an aggregate
* TXD, TXD ... populate frames from the hardware queue and submit
What should be happening is a completed TXSTATUS fires off more packets
that are queued on active TIDs.
What /was/ happening was after that first TXSTATUS the TX queue hardware queue
was still empty, so it didn't push anything into the FIFO. Only after the
second TXSTATUS did any progress get made.
This is one of two commits - it ensures that the software TX queue scheduler
is called /after/ TX completion, otherwise no frames from the software staging
queues will be processed into the hardware queues.
The second commit will fix it so it populates aggregate frames correctly
when the above occurs - right now ath_txq_sched() is called, but it doesn't
populate anything because its pre-check conditions are wrong.
Justin Hibbits [Mon, 23 Jan 2017 04:03:12 +0000 (04:03 +0000)]
Avoid using non-zero argument for __builtin_frame_address().
Building kernel with devel/powerpc64-gcc (6.2.0) yields the following error:
/usr/src/sys/powerpc/powerpc/db_trace.c:299:20: error: calling
'__builtin_frame_address' with a nonzero argument is unsafe
[-Werror=frame-address]
Work around this by dereferencing the frame address manually instead.
PR: 215600
Reported by: Mark Millard <markmi AT dsl-only DOT net>
MFC after: 2 weeks
Mark Johnston [Mon, 23 Jan 2017 02:21:06 +0000 (02:21 +0000)]
Remove the DTRACEHIOC_ADD ioctl.
This ioctl has been considered legacy by upstream since the DTrace code
was first imported, and is unused. The removal also allows some
simplification of dtrace_helper_slurp().
Also remove a bogus copyout in the DTRACEHIOC_ADDDOF handler. Due to a
bug, it would overwrite an in-memory copy of the DOF header rather than
the passed-in DOF helper. Moreover, DTRACEHIOC_ADDDOF already copies the
helper back out automatically since its argument has the IOC_OUT attribute.
Kirk McKusick [Sun, 22 Jan 2017 17:49:14 +0000 (17:49 +0000)]
By default, when doing incremental restores the restore program
overwrites an existing file rather than removing it and creating a
new file. If the old and new version of the file both have extended
attributes and the extended attributes of the two versions of the
file are different, the result is that the new file ends up with
the union of the extended attributes of the old and new files.
To get the behavior of replacing the extended attributes rather
than augmenting them requires explicitly removing the old attributes
and then adding the new ones.
To get this behavior, the old file must be unlinked (which clears
out the old extended attributes). Then the new file of the same
name must be created and the new extended attributes added to it.
This behavior can be obtained by specifying the -u flag when running
restore. Rather than defaulting the -u option to on and possibly
breaking existing scripts using restore, this change simply notes
in the restore.8 manual page that the -u flag is recommended when
using restore on filesystems that contain extended attributes.
PR: 216127
Reported by: dewayne at heuristicsystems.com.au
Differential Revision: https://reviews.freebsd.org/D9208
Adrian Chadd [Sun, 22 Jan 2017 07:05:41 +0000 (07:05 +0000)]
[athalq] fix rxtimestamp wrapping; print out per-packet timestamp deltas.
The delta here is just between the current TX/RX copmletion and the previous
TX/RX completion. The metadata needed to link TX descriptor timestamps to their
/completion/ timestamp isn't there yet.
Adrian Chadd [Sun, 22 Jan 2017 05:45:42 +0000 (05:45 +0000)]
[ath] only apply the AR9300 delimiter workaround for the first sub-frame.
This is supposed to only be applied to the first subframe and only if
RTS/CTS is being done. I'm still not yet checking RTS/CTS exchange status
so it's just happening for all subframes on AR9380 and later.
This gets MCS23 throughput up from around 250mbit to 303mbit with RTS/CTS
protection enabled, and around 330mbit with no HT protection enabled.
Now, MCS23 has a PHY rate of 450mbit and we should be seeing closer to
400mbit for a straight one-way UDP test, but this beats the previous
maximum throughput.
Tested:
* AR9380 (STA) -> AR9580 (AP) - STA with the modifications, doing UDP TX
test using iperf.
Like r310481 for i386, move the objects used to create temporary
mappings for armv6 pmap zero and copy operations to the MD PCPU region.
Change sysmap initialization to only allocate KVA pages for CPUs that
are actually present.
While here, collapse CMAP3 into CMAP2 (their use was mutually exclusive
anyway) and "recover" some space in PCPU padding that has always been
available due to 64-byte cacheline padding.
Handle the rx queue stall while reading the packets from NIC (when the
descriptor state will not change anymore). This seems to eliminate the
race where we can miss a stalled queue under high load.
While here remove the unnecessary curly brackets.
Reported by: Konstantin Kormashev <konstantin@netgate.com>
MFC after: 3 days
Sponsored by: Rubicon Communications, LLC (Netgate)
Set both IEEE80211_HTCAP_LDPC and IEEE80211_HTC_TXLDPC capability flags
if LDPC is supported + set 'do_ldpc = 1' only when it is not disabled,
not just supported.
Reviewed by: adrian
Differential Revision: https://reviews.freebsd.org/D9277
Simplify the handling of small packets padding in cpsw:
- Pad small packets to 60 bytes and not 64 (exclude the CRC bytes);
- Pad the packet using m_append(9), if the packet has enough space for
padding, which is usually true, it will not be necessary append a newly
allocated mbuf to the chain.
Alexander Motin [Sat, 21 Jan 2017 19:38:26 +0000 (19:38 +0000)]
Add initial support for CTL module unloading.
It is only a first step and not perfect, but better then nothing.
The main blocker is CAM target frontend, that can not be unloaded,
since CAM does not have mechanism to unregister periph driver now.
Note that mandoc does not use anymore sqlite3 but a home made database format
An important improvement has been made as well in makewhatis performance:
Tests on my laptop shows makewhatis on the entire system goes from 26s to 12s
Adrian Chadd [Sat, 21 Jan 2017 06:53:30 +0000 (06:53 +0000)]
[ath] ensure both iv_ampdu_limit and iv_ampdu_rxmax is set.
A recent change enforced the VAP limit as well as the peer limit.
I now need to actually set iv_ampdu_limit or we don't transmit more
than 8K sized aggregates.
This restores the expected (suboptimal, but still much faster) behaviour.
Alan Somers [Fri, 20 Jan 2017 21:40:04 +0000 (21:40 +0000)]
Fix misc Coverity defects in camdd(8)
CID 1341620 Fix a small memory leak
CID 1341630 Though this is technically a false positive, rearrange the
code for clarity.
CID 1341635 Eliminate dead code
CID 1368663 Fix a double mutex unlock in the error path
Also:
* Use sig_atomic_t for variables accessed from signal handlers
* Don't conditionalize free(3) on its argument being non-null
Alan Somers [Fri, 20 Jan 2017 21:21:41 +0000 (21:21 +0000)]
Misc Coverity fixes in camcontrol(8)
CID 1229913 Fix output of "camcontrol persist -i report_capabilities".
The reported Persistent Reservation Types were wrong in all
cases.
CID 1356029 Annotate the code so Coverity will know that this is a false
positive.
CID 1366830 Fix a memory leak in "camcontrol timestamp -s"
CID 1366832 Fix a segfault that could be caused by bad drive firmware
Also, fix the man page entry for the "camcontrol epc state" command to match
what the code does.
Fix for race leading to endless timer interrupts related to
configtimer().
During normal operation "state->nextcallopt" will always be less than
or equal to "state->nextcall" and checking only "state->nextcallopt"
before calling "callout_process()" is sufficient. However when
"configtimer()" is called a race might happen requiring both of these
binary times to be checked.
Short description of race:
1) A configtimer() call will reset both "state->nextcall" and
"state->nextcallopt" to the same binary time.
2) If a "callout_reset()" call happens between "configtimer()" and the
next "callout_process()" call, "state->nextcallopt" will get updated
and "state->nextcall" will remain at the current time. Refer to logic
inside cpu_new_callout().
3) getnextcpuevent() only respects "state->nextcall" and returns this
value over and over again, even if it is in the past, until "now >=
state->nextcallopt" becomes true. Then these two time variables are
corrected by a "callout_process()" call and the situation goes back to
normal.
The problem manifests itself in different ways. The common factor is
the timer process(es) consume all CPU on one or more CPU cores for a
long time, blocking other kernel processes from getting execution
time. This can be seen by very high interrupt counts as displayed by
"vmstat -i | grep timer" right after boot.
When EARLY_AP_STARTUP was enabled in r310177 the likelyhood of hitting
this bug apparently increased.
Example output from "vmstat -i" before patch:
cpu0:timer 7591 69
cpu9:timer 39031773 358089
cpu4:timer 9359 85
cpu3:timer 9100 83
cpu2:timer 9620 88
Example output from "vmstat -i" after patch:
cpu0:timer 4242 34
cpu6:timer 5531 44
cpu3:timer 6450 52
cpu1:timer 4545 36
cpu9:timer 7153 58
Before the patch cpu9 in the example above, was spinning in a loop in
order to reach 39 million interrupts just a few seconds after
bootup. After the patch the timer interrupt counts are more or less
consistent.
Discussed with: mav @
Reported by: several people
MFC after: 1 week
Sponsored by: Mellanox Technologies
Ryan Stone [Fri, 20 Jan 2017 17:16:48 +0000 (17:16 +0000)]
Fix reference to free memory in ixgbe/if_media.c
When ixgbe receives an interrupt indicating that a new optical module
may have been inserted, it discards all of its current media types
by calling ifmedia_removeall() and then creates a new set of media
types for the supported media on the new module. However,
ifmedia_removeall() was maintaining a pointer to whatever the
current media type was before the call to ifmedia_removealL().
The result of this was that any attempt to read the current media
type of the interface (e.g. via ifconfig) would return potentially
garbage data from free memory (or if one were particularly unlucky
on an architecture that does not malloc() from a direct map, page
fault the kernel).
Fix this by NULL'ing out the current media field in if_media.c,
and have ixgbe update the current media type after recreating
them.
Allow transmit packet bufring in software to be disabled.
- Add new sysctl node to control the transmit packet bufring.
- Add optimised version of the transmit routine which output packets
directly to the DMA ring instead of using bufring in case the transmit
lock is congested. This can reduce the number of taskswitches which in
turn influence the overall system CPU usage, depending on the
workload.
- Add " TX" suffix to debug name for transmit mutexes to silence some
witness warnings about aquiring duplicate locks having same name.
https://www.illumos.org/issues/6569
The core issue I've found is that there is no throttle for how many
deletes get assigned to one TXG. As a results when deleting large files
we end up filling consecutive TXGs with deletes/frees, then write
throttling other (more important) ops.
There is an easy test case for this problem. Try deleting several
large files (at least 1/2 TB) while you do write ops on the same
pool. What we've seen is performance of these write ops (let's
call it sideload I/O) would drop to zero.
More specifically the problem is that dmu_free_long_range_impl()
can/will fill up all of the dirty data in the pool "instantly",
before many of the sideload ops can get in. So sideload
performance will be impacted until all the files are freed.
The solution we have tested at Nexenta (with positive results)
creates a relatively simple throttle for how many "free" ops we let
into one TXG.
However this solution exposes other problems that should also be
addressed. If we are to slow down freeing of data that means one
has to wait even longer (assuming vnode ref count of 1) to get shell
back after an rm or for NFS thread to finish the free-ing op.
To avoid this the proposed solution is to call zfs_inactive() async
for "large" files. Async freeing then begs for the reclaimed space
to be accounted for in the zpool's "freeing" prop.
The other issue with having a longer delete is the inability to
export/unmount for a longer period of time. The proposed solution
is to interrupt freeing of blocks when a fs is unmounted.
Author: Alek Pinchuk <alek@nexenta.com>
Reviewed by: Matt Ahrens <mahrens@delphix.com>
Reviewed by: Sanjay Nadkarni <sanjay.nadkarni@nexenta.com>
Reviewed by: Pavel Zakharov <pavel.zakharov@delphix.com>
Approved by: Dan McDonald <danmcd@omniti.com>
Andriy Gapon [Fri, 20 Jan 2017 13:39:07 +0000 (13:39 +0000)]
don't abort writing of a core dump after EFAULT
It's possible to get EFAULT when writing a segment backed by a file
if the segment extends beyond the file.
The core dump could still be useful if we skip the rest of the segment
and proceed to other segements.
The skipped segment (or a portion of it) will be zero-filled.
While there, use 'const' to signify that core_write() only reads the
buffer and use __DECONST before calling vn_rdwr_inchunks() because it
can be used for both reading and writing.
Before the change:
kernel: Failed to write core file for process mmap_trunc_core (error 14)
kernel: pid 77718 (mmap_trunc_core), uid 1001: exited on signal 6
After the change:
kernel: Failed to fully fault in a core file segment at VA 0x800645000 with size 0x4000 to be written at offset 0x29000 for process mmap_trunc_core
kernel: pid 4901 (mmap_trunc_core), uid 1001: exited on signal 6 (core dumped)
Reviewed by: julian, kib
Obtained from: Panzura (older version of the change)
MFC after: 5 days
Sponsored by: Panzura
Differential Revision: https://reviews.freebsd.org/D9233
Andriy Gapon [Fri, 20 Jan 2017 13:21:27 +0000 (13:21 +0000)]
vmm_dev: work around a bogus error with gcc 6.3.0
The error is:
vmm_dev.c: In function 'alloc_memseg':
vmm_dev.c:261:11: error: null argument where non-null required (argument 1) [-Werror=nonnull]
Apparently, the gcc is unable to figure out that if a ternary operator
produced a non-NULL value once, then the operator with exactly the same
operands would produce the same value again.
Add runtime support for modifying the SQ and RQ completion event
moderation mode. The presence of this feature is indicated through the
firmware capabilities.