8x performance improvement in a micro benchmark on a 4 socket machine.
- Get buffer headers from a per-cpu uma cache that sits in from of the
free queue.
- Use a per-cpu quantum cache in vmem to eliminate contention for kva.
- Use multiple clean queues according to buffer cache size to eliminate
clean queue lock contention.
- Introduce a bufspace daemon that attempts to prevent getnewbuf() callers
from blocking or doing direct recycling.
- Close some bufspace allocation races that could lead to endless
recycling.
- Further the transition to a more modern style of small functions grouped
by prefix in order to improve growing complexity.
Sponsored by: EMC / Isilon
Reviewed by: kib
Tested by: pho
packets and/or state transitions from each TCP socket. That would help with
narrowing down certain problems we see in the field that are hard to reproduce
without understanding the history of how we got into a certain state. This
change provides just that.
It saves copies of the last N packets in a list in the tcpcb. When the tcpcb is
destroyed, the list is freed. I thought this was likely to be more
performance-friendly than saving copies of the tcpcb. Plus, with the packets,
you should be able to reverse-engineer what happened to the tcpcb.
To enable the feature, you will need to compile a kernel with the TCPPCAP
option. Even then, the feature defaults to being deactivated. You can activate
it by setting a positive value for the number of captured packets. You can do
that on either a global basis or on a per-socket basis (via a setsockopt call).
There is no way to get the packets out of the kernel other than using kmem or
getting a coredump. I thought that would help some of the legal/privacy concerns
regarding such a feature. However, it should be possible to add a future effort
to export them in PCAP format.
I tested this at low scale, and found that there were no mbuf leaks and the peak
mbuf usage appeared to be unchanged with and without the feature.
The main performance concern I can envision is the number of mbufs that would be
used on systems with a large number of sockets. If you save five packets per
direction per socket and have 3,000 sockets, that will consume at least 30,000
mbufs just to keep these packets. I tried to reduce the concerns associated with
this by limiting the number of clusters (not mbufs) that could be used for this
feature. Again, in my testing, that appears to work correctly.
Differential Revision: D3100
Submitted by: Jonathan Looney <jlooney at juniper dot net>
Reviewed by: gnn, hiren
Add a comment describing the necessary ordering of modifications to the
NTB Limit and Base registers.
Authored by: Jon Mason
Obtained from: Linux (Dual BSD/GPL driver)
Sponsored by: EMC / Isilon Storage Division
A WARN_ON is being hit in ntb_qp_link_work due to the NTB transport link
being down while the ntb qp link is still active. This is caused by the
transport link being brought down prior to the qp link worker thread
being terminated. To correct this, shutdown the qp's prior to bringing
the transport link down. Also, only call the qp worker thread if it is
in interrupt context, otherwise call the function directly.
Authored by: Jon Mason
Obtained from: Linux (Dual BSD/GPL driver)
Sponsored by: EMC / Isilon Storage Division
The Xeon NTB-RP setup, the transparent side does not get a link up/down
interrupt. Since the presence of a NTB device on the transparent side
means that we have a NTB link up, we can work around the lack of an
interrupt by simply calling the link up function to notify the upper
layers.
Authored by: Jon Mason
Obtained from: Linux (Dual BSD/GPL driver)
Sponsored by: EMC / Isilon Storage Division
Modifications to the 14th bit of the B2BDOORBELL register will not be
mirrored to the remote system due to a hardware issue. To get around
the issue, shrink the number of available doorbell bits by 1. The max
number of doorbells was being used as a way to referencing the Link
Doorbell bit. Since this would no longer work, the driver must now
explicitly reference that bit.
This does not affect the xeon_errata_workaround case, as it is not using
the b2bdoorbell register.
Authored by: Jon Mason
Obtained from: Linux (Dual BSD/GPL driver)
Sponsored by: EMC / Isilon Storage Division
Previously, the code prefixes the chroot path to actual file paths to
simulate the effect. This, however, will not work for tzset(3) which
expects the current system have a working set of timezone data files,
and that is not always the case.
This changeset simplifies the handling of paths and use an actual
chroot(2) call to implement the effect.
PR: bin/197313
MFC after: 2 weeks
NTB-RP is not a supported configuration on BWD hardware. Remove the
code attempting to set it up.
Authored by: Jon Mason
Obtained from: Linux (dual BSD/GPL driver)
Sponsored by: EMC / Isilon Storage Division
This commit does not actually add NTB-RP support. Mostly it serves to
shuffle code around to match the Linux driver. Original Linux commit
log follows:
Add support for Non-Transparent Bridge connected to a PCI-E Root Port on
the remote system (also known as NTB-RP mode). This allows for a NTB
enabled system to be connected to a non-NTB enabled system/slot.
Modifications to the registers and BARs/MWs on the Secondary side by the
remote system are reflected into registers on the Primary side for the
local system. Similarly, modifications of registers and BARs/MWs on
Primary side by the local system are reflected into registers on the
Secondary side for the Remote System. This allows communication between
the 2 sides via these registers and BARs/MWs.
Note: there is not a fix for the Xeon Errata (that was already worked
around in NTB-B2B mode) for NTB-RP mode. Due to this limitation, NTB-RP
will not work on the Secondary side with the Xeon Errata workaround
enabled. To get around this, disable the workaround via the
xeon_errata_workaround=0 modparm. However, this can cause the hang
described in the errata.
Authored by: Jon Mason
Obtained from: Linux
Sponsored by: EMC / Isilon Storage Division
Many variable names in the NTB driver refer to the primary or secondary
side. However, these variables will be used to access the reverse case
when in NTB-RP mode. Make these names more generic in anticipation of
NTB-RP support.
Authored by: Jon Mason
Obtained from: Linux
Sponsored by: EMC / Isilon Storage Division
Also improve documentation.
The SUBDIR_TARGETS variable should really be named LOCAL_SUBDIR_TARGETS, but
renaming it may be a surprise for downstream vendors who use this variable.
MFC after: 2 weeks
Sponsored by: EMC / Isilon Storage Division
Before this, the target was unknown. Now it will recurse on subdirs and run
the target in the current directory. It is required to recurse as there
may be subdirs that have objs in their directory or in the object directory,
so it is not enough to just delete the objdir of the subdir parent.
MFC after: 2 weeks
Sponsored by: EMC / Isilon Storage Division
I intended to remove this before committing r288266. It works but is clearly
wrong and working by accident due to the dependencies listed in the root
Makefile.inc1 file.
This is to make the Makefile more easily extendable for new ABIs.
This also makes several other subtle changes:
- The build now is given a list of ABIs to use based on the MACHINE_ARCH or
MACHINE_CPUARCH. These ABIs have a related path in sys/ that is used
to generate their syscalls. For each ABI to build check for a
ABI.c, MACHINE_ARCH-ABI.c, or a MACHINE_CPUARCH-ABI.c. This matches
the old behavior needed for archs such as powerpc* and mips*.
- The ABI source file selection allows for simpler assignment of common
ABIs such as "fbsd32" from sys/compat/freebsd32, or cloudabi64.
- Expand 'fbsd' to 'freebsd' everywhere for consistency.
- Split out the powerpc-fbsd.c file into a powerpc64-freebsd32.c to be more
like the amd64-freebsd32.c file and to more easily allow the auto-generation
of ABI handling to work.
- Rename 'syscalls.h' to 'fbsd_syscalls.h' to lessen the ambiguity and
avoid confusion with syscall.h (such as in r288997).
- For non-native syscall header files, they are now renamed to be
ABI_syscalls.h, where ABI is what ABI the Makefile is building.
- Remove all of the makesyscalls config files. The "native" one being
name i386.conf was a long outstanding bug. They were all the same
except for the data they generated, so now it is just auto-generated
as a build artifact.
- The syscalls array is now fixed to be static in the syscalls header to
remove the compiler warning about non-extern. This was worked around
in the aarch64-fbsd.c file but not the others.
- All syscall table names are now just 'syscallnames' since they don't
need to be different as they are all static in their own ABI files. The
alternative is to name them ABI_syscallnames which does not seem
necessary.
Reviewed by: ed, jhb
MFC after: 3 weeks
Differential Revision: https://reviews.freebsd.org/D3851
The BWD NTB device will drop the link if an error is encountered on the
point-to-point PCI bridge. The link will stay down until all errors are
cleared and the link is re-established. On link down, check to see if
the error is detected, if so do the necessary housekeeping to try and
recover from the error and reestablish the link.
There is a potential race between the 2 NTB devices recovering at the
same time. If the times are synchronized, the link will not recover and
the driver will be stuck in this loop forever. Add a random interval to
the recovery time to prevent this race.
Authored by: Jon Mason
Obtained from: Linux
Sponsored by: EMC / Isilon Storage Division
checked with dumpfs(8). This may lead to other problems, f.e. geom_label kernel
module sanity checks do not like zero fs_old_size value and skips such UFS1
file system while tasting (fs_old_size derives from sblock.fs_providersize).
PR: 203704
Submitted by: eugen@grosbein.net
Reviewed by: marcel
[x86] Fix wrong lowering of vsetcc nodes (PR25080).
Function LowerVSETCC (in X86ISelLowering.cpp) worked under the wrong
assumption that for non-AVX512 targets, the source type and destination type
of a type-legalized setcc node were always the same type.
This assumption was unfortunately incorrect; the type legalizer is not always
able to promote the return type of a setcc to the same type as the first
operand of a setcc.
In the case of a vsetcc node, the legalizer firstly checks if the first input
operand has a legal type. If so, then it promotes the return type of the vsetcc
to that same type. Otherwise, the return type is promoted to the 'next legal
type', which, for vectors of MVT::i1 is always a 128-bit integer vector type.
Example (-mattr=+avx):
%0 = trunc <8 x i32> %a to <8 x i23>
%1 = icmp eq <8 x i23> %0, zeroinitializer
The initial selection dag for the code above is:
v8i1 = setcc t5, t7, seteq:ch
t5: v8i23 = truncate t2
t2: v8i32,ch = CopyFromReg t0, Register:v8i32 %vreg1
t7: v8i32 = build_vector of all zeroes.
The type legalizer would firstly check if 't5' has a legal type. If so, then it
would reuse that same type to promote the return type of the setcc node.
Unfortunately 't5' is of illegal type v8i23, and therefore it cannot be used to
promote the return type of the setcc node. Consequently, the setcc return type
is promoted to v8i16. Later on, 't5' is promoted to v8i32 thus leading to the
following dag node:
v8i16 = setcc t32, t25, seteq:ch
where t32 and t25 are now values of type v8i32.
Before this patch, function LowerVSETCC would have wrongly expanded the setcc
to a single X86ISD::PCMPEQ. Surprisingly, ISel was still able to match an
instruction. In our case, ISel would have matched a VPCMPEQWrr:
t37: v8i16 = X86ISD::VPCMPEQWrr t36, t25
However, t36 and t25 are both VR256, while the result type is instead of class
VR128. This inconsistency ended up causing the insertion of COPY instructions
like this:
%vreg7<def> = COPY %vreg3; VR128:%vreg7 VR256:%vreg3
Which is an invalid full copy (not a sub register copy).
Eventually, the backend would have hit an UNREACHABLE "Cannot emit physreg copy
instruction" in the attempt to expand the malformed pseudo COPY instructions.
This patch fixes the problem adding the missing logic in LowerVSETCC to handle
the corner case of a setcc with 128-bit return type and 256-bit operand type.
This problem was originally reported by Dimitry as PR25080. It has been latent
for a very long time. I have added the minimal reproducible from that bugzilla
as test setcc-lowering.ll.
Differential Revision: http://reviews.llvm.org/D13660
This should fix the "Cannot emit physreg copy instruction" errors when
compiling contrib/wpa/src/common/ieee802_11_common.c, and CPUTYPE is set
to a CPU supporting AVX (e.g. sandybridge, ivybridge).
There is a Xeon hardware errata related to writes to SDOORBELL or B2BDOORBELL
in conjunction with inbound access to NTB MMIO Space, which may hang the
system. To workaround this issue, use one of the memory windows to access the
interrupt and scratch pad registers on the remote system. This bypasses the
issue, but removes one of the memory windows from use by the transport. This
reduction of MWs necessitates adding some logic to determine the number of
available MWs.
Since some NTB usage methodologies may have unidirectional traffic, the ability
to disable the workaround via modparm has been added.
See BF113 in
http://www.intel.com/content/dam/www/public/us/en/documents/specification-updates/xeon-c5500-c3500-spec-update.pdf
See BT119 in
http://www.intel.com/content/dam/www/public/us/en/documents/specification-updates/xeon-e5-family-spec-update.pdf
Authored by: Jon Mason
Obtained from: Linux
Sponsored by: EMC / Isilon Storage Division
Due to ambiguous documentation, the USD/DSD identification is backward
when compared to the setting in BIOS. Correct the bits to match the
BIOS setting.
Authored by: Jon Mason
Obtained from: Linux
Sponsored by: EMC / Isilon Storage Division
The NTB Xeon hardware has 16 scratch pad registers and 16 back-to-back
scratch pad registers. Correct the #define to represent this and update
the variable names to reflect their usage.
Authored by: Jon Mason
Obtained from: Linux
Sponsored by: EMC / Isilon Storage Division
The current default listen queue for casperd is too small (8) and
hard-coded.
This patch increases the default to SOMAXCONN, and introduces a command
line flag that can used to further increase or decrease the queue length.
PR: bin/202147
Submitted by: <lidl@pix.net>
image up to.
From ticket:
While trying to run FreeBSD/mips on some device having very small flash media,
one is forced to compress file system with mkulzma(8) utility. It is desirable
to specify small UFS block/fragment sizes like 4096/512 bytes for makefs(8)
and big compression block size like 65535 bytes to mkulzma at the same time.
Then one obtains very good comression ratios (like 75% and more) but faces
the following problem.
geom_uncompress kernel module reports GEOM provider size rounded up to its
compression block size. Generally, this changes original media size and now
it fails to match the size of embedded UFS file system that leads to other
problems, f.e. geom_label kernel module does not like this and skips the
file system while tasting the GEOM and looking for UFS label.
This makes it impossible to refer to the file system using known UFS label
instead of something like /dev/map/rootfs.uncompress.
The following patch introduces new command line option "-r roundup" for makefs
that makes it round up the image to specified block size. Hence, geom_uncompress
does not change GEOM media size for images rounded that way and geom_label
accepts such GEOMs just fine.
With the patch applied, one can use following commands:
$ makefs -t ffs -r 65536 -o bsize=4096,fsize=512,label=flash optimization=space fs.img fs
$ mkulzma -s 65536 -o fs.img.ulzma fs.img
PR: bin/203707
Submitted by: <eugen@grosbein.net>
target. This is the feeder for mergemaster / etcupdate. This change
makes installworld/mergemaster/etcupdate behave the same regardless of
whether SENDMAIL_MC or SENDMAIL_CF is used.
If you use a custom SENDMAIL_MC/CF in make.conf and excluded it from
mergemaster.rc/etcupdate.conf to work around the conflicts, you may wish
to revert that or change it from 'ignore' to 'always install'.
If you do not use a custom SENDMAIL_MC/CF, there should be no change in
behavior.
the FreeBSD test suite
functional_test.sh was ported from bin/sh/tests/functional_test.sh, as a
small wrapper around libarchive_test, bsdcpio_test, and bsdtar_test provided
by upstream.
A handful of testcases in lib/libarchive/tests have been disabled as they
were failing when run with kyua test (see BROKEN_TESTS in
lib/libarchive/tests/Makefile)
As a sidenote: this removes the check/test targets from the Makefiles as they
don't match the pattern used in the rest of the FreeBSD test suite.
MFC after: 2 weeks
Sponsored by: EMC / Isilon Storage Division
Reviewed by: Matthew Ahrens <mahrens@delphix.com>
Reviewed by: Paul Dagnelie <pcd@delphix.com>
Reviewed by: Alexander Motin <mav@freebsd.org>
Reviewed by: Dan McDonald <danmcd@omniti.com>
Reviewed by: Justin Gibbs <gibbs@scsiguy.com>
Reviewed by: Xin Li <delphij@freebsd.org>
Approved by: Gordon Ross <gordon.ross@nexenta.com>
Author: George Wilson <george.wilson@delphix.com>
illumos/illumos-gate@632802744e
Reviewed by: Matthew Ahrens <mahrens@delphix.com>
Reviewed by: Prakash Surya <prakash.surya@delphix.com>
Reviewed by: Simon Klinkert <simon.klinkert@gmail.com>
Reviewed by: Richard Elling <Richard.Elling@RichardElling.com>
Reviewed by: Albert Lee <trisk@omniti.com>
Reviewed by: Xin Li <delphij@freebsd.org>
Approved by: Garrett D'Amore <garrett@damore.org>
Author: George Wilson <george.wilson@delphix.com>
illumos/illumos-gate@139510fb6e
Reviewed by: Matthew Ahrens <mahrens@delphix.com>
Reviewed by: Prakash Surya <prakash.surya@delphix.com>
Reviewed by: Albert Lee <trisk@omniti.com>
Reviewed by: Xin Li <delphij@freebsd.org>
Approved by: Garrett D'Amore <garrett@damore.org>
Author: George Wilson <george.wilson@delphix.com>
illumos/illumos-gate@b10bba7246
