ngie [Tue, 28 Mar 2017 06:13:16 +0000 (06:13 +0000)]
MFC r313436,r313437,r313438,r314587,r315687:
r313436:
Clarify #includes for hexdump(3) vs sbuf_hexdump(9)
hexdump(3) only requires libutil.h, whereas sbuf_hexdump(9) requires
sys/types.h (for ssize_t) and sys/sbuf.h
r313437:
Create link from hexdump(3) to sbuf_hexdump(9) as the manpage describes
sbuf_hexdump(9)'s behavior
r313438:
Clean up trailing and leading whitespace for variables to make it
consistent with the rest of the file and style.Makefile(9) a bit
more
r314587:
Correct MLINKS for sbuf_hexdump(9)
sbuf_hexdump(9) should be linked to sbuf(9), not hexdump(3). Another
review will be posted to deduplicate the sbuf_hexdump reference in
in hexdump(3) or at the very least make the information less duplicative.
r315687:
Document sbuf_hexdump(9) in just sbuf(9)
- Remove duplicate references to sbuf_hexdump(9) from hexdump(3).
sbuf_hexdump(9) already pointed back to hexdump(3) for implementation
details.
- Refer to sbuf_hexdump(9) instead of sbuf(9) for completeness
ngie [Tue, 28 Mar 2017 06:05:26 +0000 (06:05 +0000)]
MFC r315686,r315688:
r315686:
kvm_geterr: handle `kd` == NULL in a deterministic/graceful manner
Return a NUL string instead of just working by accident with kvm_geterr(3)
when MALLOC_PRODUCTION is disabled (I didn't confirm the MALLOC_PRODUCTION
being enabled path).
Document the new explicit return behavior for kvm_geterr(3), as well
as the previous implicit behavior, i.e., the buffer attached to
returned via kvm_geterr(3) would be empty if a previous error hadn't been
stored in `kd`.
r315688:
kvm_write: fix -Wcast-qual warning in pointer arithmetic argument
Cast buf to `const char *` when doing arithmetic operation to match
`cp`'s type [1].
ngie [Mon, 27 Mar 2017 18:28:17 +0000 (18:28 +0000)]
MFC r314245:
Fill MK_LIBTHR as far as lib/libthr is concerned
There are other areas of the tree that will need to be evaluated for sanity
if they're supposed to be conditionally compiled out of the build/install,
like libzpool
Relnotes: yes (this might break someone's system if have the knob set)
kp [Sun, 26 Mar 2017 18:12:50 +0000 (18:12 +0000)]
MFC 315529
pf: Fix rule evaluation after inet6 route-to
In pf_route6() we re-run the ruleset with PF_FWD if the packet goes out
of a different interface. pf_test6() needs to know that the packet was
forwarded (in case it needs to refragment so it knows whether to call
ip6_output() or ip6_forward()).
This lead pf_test6() to try to evaluate rules against the PF_FWD
direction, which isn't supported, so it needs to treat PF_FWD as PF_OUT.
Once fwdir is set correctly the correct output/forward function will be
called.
dim [Sat, 25 Mar 2017 21:46:02 +0000 (21:46 +0000)]
Synchronize libcxxrt in stable/10 with stable/11.
MFC r284553 (by emaste):
Update libcxxrt upgrade instructions
The typeinfo file no longer exists upstream.
MFC r297299:
Compile libcxxrt as C++11, since it is only really used in combination
with libc++, which is also C++11. Also change one _Static_assert (which
is really C11) back into static_assert, like upstream.
This should help when compiling libcxxrt with newer versions of gcc,
which refuse to recognize any form of static assertions, if not
compiling for C++11 or higher.
While here, add -nostdinc++ to CFLAGS, to prevent picking up any C++
headers outside the source tree.
Interesting fixes: 3adaa2e Fix _Unwind_Exception cleanup functions 286776c Check exception cleanup function ptr before calling edda626 Correct exception specifications on new and delete operators
MFC r303157 (by emaste):
libcxxrt: add padding in __cxa_allocate_* to fix alignment
The addition of the referenceCount to __cxa_allocate_exception put the
unwindHeader at offset 0x58 in __cxa_exception, but it requires 16-byte
alignment. In order to avoid changing the current __cxa_exception ABI
(and thus breaking its consumers), add explicit padding in the
allocation routines (and account for it when freeing).
This is intended as a lower-risk change for FreeBSD 11. A "more correct"
fix should be prepared for upstream and -CURRENT.
Reviewed by: dim
Sponsored by: The FreeBSD Foundation
Differential Revision: https://reviews.freebsd.org/D7271
MFC r305396:
Add _US_ACTION_MASK to libcxxrt's arm-specific unwind header. This
value is used in newer versions of compiler-rt.
badger [Sat, 25 Mar 2017 20:14:08 +0000 (20:14 +0000)]
MFC r315412, r314852:
r315412:
Don't clear p_ptevents on normal SIGKILL delivery
The ptrace() user has the option of discarding the signal. In such a
case, p_ptevents should not be modified. If the ptrace() user decides to
send a SIGKILL, ptevents will be cleared in ptracestop(). procfs events
do not have the capability to discard the signal, so continue to clear
the mask in that case.
r314852:
don't stop in issignal() if P_SINGLE_EXIT is set
Suppose a traced process is stopped in ptracestop() due to receipt of a
SIGSTOP signal, and is awaiting orders from the tracing process on how
to handle the signal. Before sending any such orders, the tracing
process exits. This should kill the traced process. But suppose a second
thread handles the SIGKILL and proceeds to exit1(), calling
thread_single(). The first thread will now awaken and will have a chance
to check once more if it should go to sleep due to the SIGSTOP. It must
not sleep after P_SINGLE_EXIT has been set; this would prevent the
SIGKILL from taking effect, leaving a stopped orphan behind after the
tracing process dies.
badger [Sat, 25 Mar 2017 13:33:23 +0000 (13:33 +0000)]
MFC r313992, r314075, r314118, r315484:
r315484:
ptrace_test: eliminate assumption about thread scheduling
A couple of the ptrace tests make assumptions about which thread in a
multithreaded process will run after a halt. This makes the tests less
portable across branches, and susceptible to future breakage. Instead,
twiddle thread scheduling and priorities to match the tests'
expectation.
r314118:
Actually fix buildworlds other than i386/amd64/sparc64 after r313992
Disable offending test for platforms without a userspace visible
breakpoint().
r314075:
Fix world build for archs where __builtin_debugtrap() does not work.
The offending code was introduced in r313992.
r313992:
Defer ptracestop() signals that cannot be delivered immediately
When a thread is stopped in ptracestop(), the ptrace(2) user may request
a signal be delivered upon resumption of the thread. Heretofore, those signals
were discarded unless ptracestop()'s caller was issignal(). Fix this by
modifying ptracestop() to queue up signals requested by the ptrace user that
will be delivered when possible. Take special care when the signal is SIGKILL
(usually generated from a PT_KILL request); no new stop events should be
triggered after a PT_KILL.
Add a number of tests for the new functionality. Several tests were authored
by jhb.
mav [Sat, 25 Mar 2017 11:45:19 +0000 (11:45 +0000)]
MFC r315084: Increase device openings to tagged maximum.
Some SIMs report much less untagged device openings then tagged ones.
Target mode devices are not handled by regular probing routines, and so
there is nothing to increase queue size for them to the SIM's maximum.
To fix that resize the queue explicitly on ctl periph registration.
This radically improves performance of mpt(4) in target mode.
Also fetch and report device queue statistics in `ctladm dumpstructs`,
since regular way of `camcontrol tags` is not usable in target mode.
grehan [Sat, 25 Mar 2017 05:09:03 +0000 (05:09 +0000)]
MFC r315361 and r315364: Hide MONITORX/MWAITX from guests.
r315361
Add the AMD MONITORX/MWAITX feature definition introduced in
Bulldozer/Ryzen CPUs.
r315364
Hide the AMD MONITORX/MWAITX capability.
Otherwise, recent Linux guests will use these instructions, resulting
in #UD exceptions since bhyve doesn't implement MONITOR/MWAIT exits.
This fixes boot-time hangs in recent Linux guests on Ryzen CPUs
(and probably Bulldozer aka AMD FX as well).
dim [Fri, 24 Mar 2017 18:28:13 +0000 (18:28 +0000)]
Pull in r283944 from upstream libc++ trunk (by Eric Fiselier):
Fix std::pair on FreeBSD
Summary:
FreeBSD ships an old ABI for std::pair which requires that it have
non-trivial copy/move constructors. Currently the non-trivial
copy/move is achieved by providing explicit definitions of the
constructors. This is problematic because it means the constructors
don't SFINAE properly. In order to SFINAE copy/move constructors they
have to be explicitly defaulted and hense non-trivial.
This patch attempts to provide SFINAE'ing copy/move constructors for
std::pair while still making them non-trivial. It does this by adding
a base class with a non-trivial copy constructor and then allowing
pair's constructors to be generated by the compiler. This also allows
the constructors to be constexpr.
This should fix building www/chromium 57.0.2987.110 on stable/10 and
stable/9 without having to use -D_LIBCPP_TRIVIAL_PAIR_COPY_CTOR=1 (which
changes the ABI). Direct commit to stable/10 and stable/9, since head
already has libc++ 4.0, which includes this fix.
des [Fri, 24 Mar 2017 14:26:01 +0000 (14:26 +0000)]
MFH (r313974,r314596): open .netrc early in case we want to drop privs
MFH (r314396,r315143): fix a crash caused by an incorrect format string
MFH (r314701): fix handling of 416 errors when requesting a range
MFH (r315455): fix parsing of IP literals (square brackets)
mav [Fri, 24 Mar 2017 07:02:03 +0000 (07:02 +0000)]
MFC r315030: Abort all ATIOs and INOTs queued to SIM on LUN disable.
Some SIMs may not abort them implicitly, that either fail the LUN disable
request or just make us wait for those CCBs forever. With this change
I can successfully disable LUNs on mpt(4). For isp(4), which aborts them
implicitly, this change should be irrelevant.
davidcs [Fri, 24 Mar 2017 03:11:11 +0000 (03:11 +0000)]
MFC r314365
1. state checks in bxe_tx_mq_start_locked() and bxe_tx_mq_start() to
sync threads during interface down or detach.
2. add sysctl to set pause frame parameters
3. increase max segs for TSO packets to BXE_TSO_MAX_SEGMENTS (32)
4. add debug messages for PHY
5. HW LRO support restricted to FreeBSD versions 8.x and above.
np [Thu, 23 Mar 2017 20:23:00 +0000 (20:23 +0000)]
MFC r314814 and r315325.
r314814:
cxgbe/iw_cxgbe: Abort connection if there is an error during c4iw_modify_qp.
r315325:
cxgbe/iw_cxgbe: Use the socket and not the toepcb to reach for the
inpcb. t4_tom detaches the inpcb from the toepcb as soon as the
hardware is done with the connection (in final_cpl_received) but the
socket is around as long as the cm_id and the rest of iWARP state is.
This fixes an intermittent NULL dereference during abort.
mav [Thu, 23 Mar 2017 06:55:32 +0000 (06:55 +0000)]
MFC r315067: Partially fix target task management requests handling.
- XPT_NOTIFY_ACKNOWLEDGE was not handled, causing stuck requests.
- XPT_ABORT was not even trying to abort active ATIOs/INOTs.
- Initiator's tag was not stored and not used where needed.
- List of TM request types needed update.
- mpt_scsi_tgt_status() missed some useful debugging.
After this change global TM requests, such as reset, should work properly.
ABORT TASK (ABTS) requests are still not passes to CTL, that is not good
and should be fixed.
mav [Thu, 23 Mar 2017 06:49:01 +0000 (06:49 +0000)]
MFC r314998: Fix FC target mode in mpt(4), broken in multiple ways.
- Not set BufferLength caused receive of empty ATIOs.
- CDB length guessing was broken at least for RC16.
- mpt_req_untimeout() was called with wrong req parameter.
- Sense data reporting was broken in several ways.
With this change my LSI7204EP-LC can pass at least basic tests as target.
The code is still far from perfect, but finally I found second hw/driver
after isp(4) that really can work in CAM target mode.
The sim_vid, hba_vid, and dev_name fields of struct ccb_pathinq are
fixed-length strings. AFAICT the only place they're read is in
sbin/camcontrol/camcontrol.c, which assumes they'll be null-terminated.
However, the kernel doesn't null-terminate them. A bunch of copy-pasted code
uses strncpy to write them, and doesn't guarantee null-termination. For at
least 4 drivers (mpr, mps, ciss, and hyperv), the hba_vid field actually
overflows. You can see the result by doing "camcontrol negotiate da0 -v".
This change null-terminates those fields everywhere they're set in the
kernel. It also shortens a few strings to ensure they'll fit within the
16-character field.
mav [Thu, 23 Mar 2017 06:37:23 +0000 (06:37 +0000)]
MFC r308423 (by scottl):
Fix the fallout from r308268 (mpt driver causes endless witness warnings in
VMWare and elsewhere) with the precision of a dull, rusty butter knife.
ngie [Thu, 23 Mar 2017 04:54:31 +0000 (04:54 +0000)]
MFC r315360:
Return NULL instead of 0 on failure in _kvm_open, kvm_open{,2,files}
This is being done for the following reasons:
- kvm_open(3), etc says they will return NULL.
- NULL by definition is (void*)0 per POSIX, but can be redefined,
depending on the compiler, etc.
ngie [Thu, 23 Mar 2017 04:50:50 +0000 (04:50 +0000)]
MFC r315363,r315365:
r315363:
Fix manlint errors
- Add missing comma after kvm_dpcpu_setcpu .Nm macro use (multiple
.Nm entries should be separated by commas)
- Add missing section for kvm_dpcpu_setcpu Xr.
r315365:
Tweak r315363 slightly
I noticed after commit that kvm_dpcpu_setcpu was defined in the manpage.
Thus, the correct macro for the function reference is .Fn, not .Xr.
ngie [Thu, 23 Mar 2017 04:47:43 +0000 (04:47 +0000)]
MFC r315654:
Fix linking /rescue/rescue to multiple programs in usr.bin after r315113
I meant for the line that conditionally added in /usr/bin/nc support to
be `+=', not `=`. This restores hardlinks for all programs in usr.bin
specified before nc(1), e.g., bunzip2 and tar.
bdrewery [Wed, 22 Mar 2017 18:35:48 +0000 (18:35 +0000)]
MFC r314709,r314790,r314794:
r314709:
Fix bootstrapping mtree after r313404 for older systems.
r314790:
Added comments for why nmtree/libmd are bootstrapped.
r314794:
Fix bootstrapping libmd on older systems after r314709.
ngie [Wed, 22 Mar 2017 01:11:13 +0000 (01:11 +0000)]
MFC r315320:
Start adding basic tests for cam(3)
This change contains several negative and positive tests for:
- cam_open_device
- cam_close_device
- cam_getccb
- cam_freeccb
This also contains a test for the failure case noted in bug 217649,
i.e., O_RDWR must be specified because pass(4) requires it.
This test unfortunately cannot assume that cam-capable devices are
present, so the user must explicitly provide a device via
`test_suites.FreeBSD.cam_test_device`. In the future, a test kernel
module might be shipped, or ctl(4) might be used, as a test device
when testing out libcam, which will allow the tests to do away with
having to specify an explicit test device.
vangyzen [Tue, 21 Mar 2017 01:24:56 +0000 (01:24 +0000)]
MFC r315510
nanosleep: plug a kernel memory disclosure
nanosleep() updates rmtp on EINVAL. In that case, kern_nanosleep()
has not updated rmt, so sys_nanosleep() updates the user-space rmtp
by copying garbage from its stack frame. This is not only a kernel
memory disclosure, it's also not POSIX-compliant. Fix it to update
rmtp only on EINTR.
ngie [Mon, 20 Mar 2017 03:13:03 +0000 (03:13 +0000)]
MFC r315206:
bsnmpd: fix segfault when trans_insert_port(..) is called with multiple
out of order addresses
Move `port->transport` initialization before the TAILQ_FOREACH(..) loop
to ensure that the value is properly initialized before it's inserted
into the TAILQ.
ngie [Mon, 20 Mar 2017 03:01:23 +0000 (03:01 +0000)]
MFC r315202:
lib/libcam/cam_cdbparse.3: fix manpage warnings
- Add comma before and after 'e.g.'; remove surrounding parentheses
that were unnecessary after this change [1].
- Add .Mt when referencing ken and Peter Dufault's email addresses
[2].
- Sprinkle around .An use where proper [2].
ngie [Mon, 20 Mar 2017 03:00:22 +0000 (03:00 +0000)]
MFC r315132,r315133,r315186:
r315132:
Use .Dv when referencing NULL
This is the correct markup macro, as opposed to .Va (variable names)
While here, annotate several bare references to `NULL` with .Dv.
r315133:
lib/libcam/cam.3: fix manpage warnings
- spelling: "mis-named" should be "misnamed".
- delete spaces interspersed in literal representation of
`struct cam_device` as hard-tabs separate the types and fields.
- Add commas after `e.g.`.
r315186:
lib/libcam/cam.3: note that cam_freeccb(3) with ccb == NULL is a no-op
This allows me to accurately test this scenario, and for others to rely
on the behavior, instead of relying on knowledge obtained via code
inspection.
pfg [Mon, 20 Mar 2017 00:55:24 +0000 (00:55 +0000)]
MFC r315212, r315213, r315214, r315215:
mkimg(1): let calloc(3) do the multiplication.
nscd(8): let calloc(3) do the multiplying.
mpsutil(8): let calloc(3) do the multiplying.
ypbind(8): let calloc(3) do the multiplying.
erj [Thu, 16 Mar 2017 23:18:57 +0000 (23:18 +0000)]
ixgbe(4): Fix VF build failure on i386 caused by r315333
SYSCTL_ADD_UQUAD on i386 expects a uint64_t pointer, and not an
unsigned long pointer. So change the data type of the original variable to
u64 (uint64_t) from unsigned long, since the statistics are intended to be
64 bits, anyway.
mm [Thu, 16 Mar 2017 23:08:18 +0000 (23:08 +0000)]
MFC r314571:
Update libarchive to version 3.3.1 (and sync with latest vendor dist)
Notable vendor changes:
PR #501: improvements in ACL path handling
PR #724: fix hang when reading malformed cpio files
PR #864: fix out of bounds read with malformed GNU tar archives
Documentation, style, test suite improvements and typo fixes.
New options to bsdtar that enable or disable reading and/or writing of:
Access Control Lists (--acls, --no-acls)
Extended file flags (--fflags, --no-fflags)
Extended attributes (--xattrs, --no-xattrs)
Mac OS X metadata (Mac OS X only) (--mac-metadata, --no-mac-metadata)
mav [Thu, 16 Mar 2017 07:11:35 +0000 (07:11 +0000)]
MFC r314549: Execute last ZIO of log commit synchronously.
For short transactions overhead of context switch can be too large.
Skipping it gives significant latency reduction. For large ones,
including multiple ZIOs, latency is less critical, while throughput
there may become limited by checksumming speed of single CPU core.
To get best of both cases, execute last ZIO directly from calling
thread context to save latency, while all others (if there are any)
enqueue to taskqueues in traditional way.
ngie [Thu, 16 Mar 2017 01:59:43 +0000 (01:59 +0000)]
MFC r314924:
sbin/devfs: clarify usage
- Note existence of -m option.
- Note that -s applies to rule keyword, only, by adding usage text
specifically for the `rule` and `ruleset` keywords.
Don't go into any further detail in usage(..) -- it's best that one
reads the manpage to get a better idea of how things work as there are
a number of different option-specific keywords and arguments, as well
as some rule grammar.
ngie [Thu, 16 Mar 2017 01:47:26 +0000 (01:47 +0000)]
MFC r274130:
r274130 (by bapt):
Avoid installing security.functions with executable bits, periodic(8) will
try to execute all files with an executable bit in /etc/periodic/*/ while
this file is supposed only to be sourced by others
badger [Thu, 16 Mar 2017 01:41:36 +0000 (01:41 +0000)]
MFC r313733:
sleepq_catch_signals: do thread suspension before signal check
Since locks are dropped when a thread suspends, it's possible for another
thread to deliver a signal to the suspended thread. If the thread awakens from
suspension without checking for signals, it may go to sleep despite having
a pending signal that should wake it up. Therefore the suspension check is
done first, so any signals sent while suspended will be caught in the
subsequent signal check.
ngie [Thu, 16 Mar 2017 01:38:04 +0000 (01:38 +0000)]
MFC r314830:
mergemaster: fix description of -p
-p only handles updating /etc/master.passwd and /etc/group . No more,
no less.
Also, mergemaster (and no other portions of the vanilla FreeBSD build
process) should be messing with __MAKECONF or SRCCONF as part of the
installworld or distribution process. Don't insinuate that mergemaster
does that as it's a false claim.
erj [Wed, 15 Mar 2017 21:20:17 +0000 (21:20 +0000)]
ixgbe(4): Update to 3.2.11-k
Includes:
1. Support for X553 (Denverton) Ethernet devices.
2. Initialization code refactoring.
3. ixgbe-specific netmap code moved to ixgbe folder
(but not removed from dev/netmap/).
4. VF driver full split. Some cleanup can be had due to redundant code,
but this split goes inline with the refactoring.
NOTE: This is a direct commit to stable/10. Newer branches will receive an
iflib-converted version of ixgbe, with these updates included, instead.
dim [Wed, 15 Mar 2017 19:50:58 +0000 (19:50 +0000)]
MFC r310232:
After r310171, the kernel version of sscanf() has format string checking
enabled. This results in a -Werror warning in mlx4ib:
sys/dev/mlx4/mlx4_ib/mlx4_ib_sysfs.c:90:22: error: format specifies type 'unsigned long long *' but the argument has type 'u64 *' (aka 'unsigned long *') [-Werror,-Wformat]
sscanf(buf, "%llx", &sysadmin_ag_val);
~~~~ ^~~~~~~~~~~~~~~~
Change sysadmin_ag_val to unsigned long long to avoid the warning.
hselasky [Tue, 14 Mar 2017 15:56:19 +0000 (15:56 +0000)]
MFC r314328:
Fix startup race initialising ACPI CM battery structures on MacBookPro.
During acpi_cmbat_attach() the acpi_cmbat_init_battery() notification
handler is registered. It has been observed this notification handler
can be called instantly, before the attach routine has returned. In
the notification handler there is a call to device_is_attached() which
returns false. Because the softc is set we know an attach is in
progress and the fix is simply to wait and try again in this case.
hselasky [Tue, 14 Mar 2017 15:50:36 +0000 (15:50 +0000)]
MFC r313941:
Make sure the thread constructor and destructor eventhandlers are
called for all threads belonging to a procedure. Currently the first
thread in a procedure is kept around as an optimisation step and is
never freed. Because the first thread in a procedure is never freed
nor allocated, its destructor and constructor callbacks are never
called which means per thread structures allocated by dtrace and the
Linux emulation layers for example, might be present for threads which
don't need these structures.
This patch adds a thread construction and destruction call for the
first thread in a procedure.
hselasky [Tue, 14 Mar 2017 15:37:29 +0000 (15:37 +0000)]
MFC r312551:
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
Sponsored by: Mellanox Technologies