are mixed. Some pure context switch microbenchmarks show up to 29%
improvement. Pipe based context switch microbenchmarks show up to 7%
improvement. Real world tests are far less impressive as they are
dominated more by actual work than switch overheads, but depending on
the machine in question, workload, kernel options, phase of moon, etc, a
few percent gain might be seen.
Summary of changes:
- don't reload MSR_[FG]SBASE registers when context switching between
non-threaded userland apps. These typically cost 120 clock cycles each
on an AMD cpu (less on Barcelona/Phenom). Intel cores are probably no
faster on this.
- The above change only helps unthreaded userland apps that tend to use
the same value for gsbase. Threaded apps will get no benefit from this.
- reorder things like accessing the pcb to be in memory order, to give
prefetching a better chance of working. Operations are now in increasing
memory address order, rather than reverse or random.
- Push some lesser used code out of the main code paths. Hopefully
allowing better code density in cache lines. This is probably futile.
- (part 2 of previous item) Reorder code so that branches have a more
realistic static branch prediction hint. Both Intel and AMD cpus
default to predicting branches to lower memory addresses as being
taken, and to higher memory addresses as not being taken. This is
overridden by the limited dynamic branch prediction subsystem. A trip
through userland might overflow this.
- Futule attempt at spreading the use of the results of previous operations
in new operations. Hopefully this will allow the cpus to execute in
parallel better.
- stop wasting 16 bytes at the top of kernel stack, below the PCB.
- Never load the userland fs/gsbase registers for kthreads, but preserve
curpcb->pcb_[fg]sbase as caches for the cpu. (Thanks Jeff!)
Microbenchmarking this code seems to be really sensitive to things like
scheduling luck, timing, cache behavior, tlb behavior, kernel options,
other random code changes, etc.
While it doesn't help heavy userland workloads much, it does help high
context switch loads a little, and should help those that involve
switching via kthreads a bit more.
A special thanks to Kris for the testing and reality checks, and Jeff for
tormenting me into doing this. :)
This is still work-in-progress.
PTE if that PTE has the PG_RW bit set. However, this assumption does
not hold on recent processors from Intel. For example, consider a PTE
that has the PG_RW bit set but the PG_M bit clear. Suppose this PTE
is cached in the TLB and later the PG_RW bit is cleared in the PTE,
but the corresponding TLB entry is not (yet) invalidated.
Historically, upon a write access using this (stale) TLB entry, the
TLB would observe that the PG_RW bit had been cleared and initiate a
page fault, aborting the setting of the PG_M bit in the PTE. Now,
however, P4- and Core2-family processors will set the PG_M bit before
observing that the PG_RW bit is clear and initiating a page fault. In
other words, the write does not occur but the PG_M bit is still set.
The real impact of this difference is not that great. Specifically,
we should no longer assert that any PTE with the PG_M bit set must
also have the PG_RW bit set, and we should ignore the state of the
PG_M bit unless the PG_RW bit is set. However, these changes enable
me to remove a work-around from pmap_promote_pde(), the superpage
promotion procedure.
(Note: The AMD processors that we have tested, including the latest,
the Phenom, still exhibit the historical behavior.)
Acknowledgments: After I observed the problem, Stephan (ups) was
instrumental in characterizing the exact behavior of Intel's recent
TLBs.
Tested by: Peter Holm
The overflow causes the wraparound with consequent corruption of the
(almost) whole address space mapping.
As Alan noted, pmap_copy() does not require the wrap-around checks
because it cannot be applied to the kernel's pmap. The checks there are
included for consistency.
Reported and tested by: kris (i386/pmap.c:pmap_remove() part)
Reviewed by: alc
MFC after: 1 week
lock in the 8259A drivers as these drivers are only used on UP systems.
This slightly reduces the penalty of an SMP kernel (such as GENERIC) on
a UP x86 machine.
resource to a CPU. The default method is to pass the request up to the
parent similar to BUS_CONFIG_INTR() so that all busses don't have to
explicitly implement bus_bind_intr. A bus_bind_intr(9) wrapper routine
similar to bus_setup/teardown_intr() is added for device drivers to use.
Unbinding an interrupt is done by binding it to NOCPU. The IRQ resource
must be allocated, but it can happen in any order with respect to
bus_setup_intr(). Currently it is only supported on amd64 and i386 via
nexus(4) methods that simply call the intr_bind() routine.
Tested by: gallatin
- Always include the ie_disable and ie_eoi methods in 'struct intr_event'
and collapse down to one intr_event_create() routine. The disable and
eoi hooks simply aren't used currently in the !INTR_FILTER case.
- Expand 'disab' to 'disable' in a few places.
- Use function casts for arm and i386:intr_eoi_src() instead of wrapper
routines since to trim one extra indirection.
Compiled on: {arm,amd64,i386,ia64,ppc,sparc64} x {FILTER, !FILTER}
Tested on: {amd64,i386} x {FILTER, !FILTER}
after each SYSINIT() macro invocation. This makes a number of
lightweight C parsers much happier with the FreeBSD kernel
source, including cflow's prcc and lxr.
MFC after: 1 month
Discussed with: imp, rink
- Add a new intr_event method ie_assign_cpu() that is invoked when the MI
code wishes to bind an interrupt source to an individual CPU. The MD
code may reject the binding with an error. If an assign_cpu function
is not provided, then the kernel assumes the platform does not support
binding interrupts to CPUs and fails all requests to do so.
- Bind ithreads to CPUs on their next execution loop once an interrupt
event is bound to a CPU. Only shared ithreads are bound. We currently
leave private ithreads for drivers using filters + ithreads in the
INTR_FILTER case unbound.
- A new intr_event_bind() routine is used to bind an interrupt event to
a CPU.
- Implement binding on amd64 and i386 by way of the existing pic_assign_cpu
PIC method.
- For x86, provide a 'intr_bind(IRQ, cpu)' wrapper routine that looks up
an interrupt source and binds its interrupt event to the specified CPU.
MI code can currently (ab)use this by doing:
intr_bind(rman_get_start(irq_res), cpu);
however, I plan to add a truly MI interface (probably a bus_bind_intr(9))
where the implementation in the x86 nexus(4) driver would end up calling
intr_bind() internally.
Requested by: kmacy, gallatin, jeff
Tested on: {amd64, i386} x {regular, INTR_FILTER}
receivers from being given interrupts if any CPUs in the system were not
tagged as interrupt receivers that I introduced when switching the x86
interrupt code to track CPUs via FreeBSD CPU IDs rather than local APIC
IDs. In practice this only affects systems with Hyperthreading (though
disabling HTT in the BIOS would workaround the issue) as that is the only
case currently where one can have CPUs that aren't tagged as interrupt
receivers. On a Dell SC1425 test box with 2 x Xeon w/ HTT (so 4 logical
CPUs of which 2 were interrupt receivers) the result was that all
device interrupts were sent to CPU 0.
MFC after: 1 week
Pointy hat to: jhb
different "platforms" on x86 machines. The existing code already handles
having two platforms: ACPI and legacy. However, the existing approach was
rather hardcoded and difficult to extend. These changes take the approach
that each x86 hardware platform should provide its own nexus(4) driver (it
can inherit most of its behavior from the default legacy nexus(4) driver)
which is responsible for probing for the platform and performing
appropriate platform-specific setup during attach (such as adding a
platform-specific bus device). This does mean changing the x86 platform
busses to no longer use an identify routine for probing, but to move that
logic into their matching nexus(4) driver instead.
- Make the default nexus(4) driver in nexus.c on i386 and amd64 handle the
legacy platform. It's probe routine now returns BUS_PROBE_GENERIC so it
can be overriden.
- Expose a nexus_init_resources() routine which initializes the various
resource managers so that subclassed nexus(4) drivers can invoke it from
their attach routine.
- The legacy nexus(4) driver explicitly adds a legacy0 device in its
attach routine.
- The ACPI driver no longer contains an new-bus identify method. Instead
it exposes a public function (acpi_identify()) which is a probe routine
that the MD nexus(4) drivers can use to probe for ACPI. All of the
probe logic in acpi_probe() is now moved into acpi_identify() and
acpi_probe() is just a stub.
- On i386 and amd64, an ACPI-specific nexus(4) driver checks for ACPI via
acpi_identify() and claims the nexus0 device if the probe succeeds. It
then explicitly adds an acpi0 device in its attach routine.
- The legacy(4) driver no longer knows anything about the acpi0 device.
- On ia64 if acpi_identify() fails you basically end up with no devices.
This matches the previous behavior where the old acpi_identify() would
fail to add an acpi0 device again leaving you with no devices.
Discussed with: imp
Silence on: arch@
ABI and the direction flag, that is it now assumes that the direction
flag is cleared at the entry of a function and it doesn't clear once
more if needed. This new behaviour conforms to the i386/amd64 ABI.
Modify the signal handler frame setup code to clear the DF {e,r}flags
bit on the amd64/i386 for the signal handlers.
jhb@ noted that it might break old apps if they assumed DF == 1 would be
preserved in the signal handlers, but that such apps should be rare and
that older versions of gcc would not generate such apps.
Submitted by: Aurelien Jarno <aurelien aurel32 net>
PR: 121422
Reviewed by: jhb
MFC after: 2 weeks
PhysMask fields based on the number of physical address bits supported
by the current CPU. The old code assumed 36 bits on i386 and 40 bits on
amd64. In truth, all Intel CPUs up until recently used 36 bits (a newer
Intel CPU uses 38 bits) and all the Opteron CPUs used 40 bits.
In at least one case (the new Intel CPU) having the size of the mask field
wrong resulted in writing questionable values into the MTRR registers on
the application processors (BSP as well if you modify the MTRRs via
memcontrol or running X, etc.). The result of the questionable physmask
was that all of memory was apparently treated as uncached rather than
write-back resulting in a very significant performance hit.
Fix this by constructing a run-time mask for the PhysBase and PhysMask
fields based on the number of physical address bits supported by the CPU.
All 64-bit capable CPUs provide a count of PA bits supported via the
0x80000008 extended CPUID feature, so use that if it is available. If that
feature is not available, then assume 36 PA bits.
While I'm here, expand the (now-unused) macros for the PhysBase and
PhysMask fields to the current largest possible value (52 PA bits).
MFC after: 1 week
PR: i386/120516
Reported by: Nokia
While the KSE project was quite successful in bringing threading to
FreeBSD, the M:N approach taken by the kse library was never developed
to its full potential. Backwards compatibility will be provided via
libmap.conf for dynamically linked binaries and static binaries will
be broken.
the cpufreq drivers to reliably use properties of PCI devices for quirks,
etc.
- For the legacy drivers, add CPU devices via an identify routine in the
CPU driver itself rather than in the legacy driver's attach routine.
- Add CPU devices after Host-PCI bridges in the acpi bus driver.
- Change the ichss(4) driver to use pci_find_bsf() to locate the ICH and
check its device ID rather than having a bogus PCI attachment that only
checked for the ID in probe and always failed. As a side effect, you
can now kldload ichss after boot.
- Fix the ichss(4) driver to use the correct device_t for the ICH (and not
for ichss0) when doing PCI config space operations to enable SpeedStep.
MFC after: 2 weeks
Reviewed by: njl, Andriy Gapon avg of icyb.net.ua
mappings. Automatic promotion can be enabled by setting the tunable
"vm.pmap.pg_ps_enabled" to a non-zero value. By default, automatic
promotion is disabled. (Expect this to change.)
Reviewed by: ups
Tested by: kris, Peter Holm
tree structure that encodes the level of cache sharing and other
properties.
- Provide several convenience functions for creating one and two level
cpu trees as well as a default flat topology. The system now always
has some topology.
- On i386 and amd64 create a seperate level in the hierarchy for HTT
and multi-core cpus. This will allow the scheduler to intelligently
load balance non-uniform cores. Presently we don't detect what level
of the cache hierarchy is shared at each level in the topology.
- Add a mechanism for testing common topologies that have more information
than the MD code is able to provide via the kern.smp.topology tunable.
This should be considered a debugging tool only and not a stable api.
Sponsored by: Nokia
PAGE_SIZE or less, the bounce page counting logic was flawed and wouldn't
reserve any pages. Adjust to be correct. Review of other architectures is
forthcoming.
Submitted by: Joseph Golio
aligned (or at least not cross a page boundary). However, it turns out
that on at least one machine one table header does cross a page boundary.
This caused problems with the MADT early probe as it uses the crash dump
map to load ACPI tables by loading the RSDT/XSDT into pages 1 ... N and
loading the header of each ACPI table header into page 0 looking for the
MADT. However, if a table header crossed a page boundary, then page 1
would get trashed resulting in a panic. Fix this by reserving the first
2 pages for ACPI table headers (headers are less than a page in size,
so 2 pages will be sufficient) and use pages 2 .. N for the RSDT and XSDT.
Note: amd64 should probably be simplified to just use pmap_mapbios()
for all these tables which will use the direct map and not need the
crash dump hack.
MFC after: 5 days
Tested on: i386
Reported by: Pete French petefrench of ticketswitch.com
compiled under PMAP_DIAGNOSTIC are now KASSERT()s. (Note: The kernel
option DIAGNOSTIC still disables inlining of certain pmap functions.)
Eliminate dead code from pmap_enter(). This code implemented an assertion.
On i386, an equivalent check is already implemented. However, on amd64,
a small change is required to implement an equivalent check.
Eliminate \n from a nearby panic string.
Use KASSERT() to reimplement pmap_copy()'s two assertions.
in the range and precision of their type(s) on amd64, but FLT_EVAL_METHOD
said that they were evalated in the "interesting" (buggy) i387 methods.
float_t was broken compatibly with FLT_EVAL_METHOD.
These definitions seem to be broken on powerpc and possibly on arm.
float_t is float on powerpc with gcc [-notraditional] according to
glibc, and FLT_EVAL_METHOD is marked with XXX on arm.
the kernel's direct map instead of the pmap's recursive mapping to access
the lowest level in the page table. The direct map is preferable for two
reasons: (1) The TLB is more likely to hold the required direct mapping
because pmap_enter() has already used the direct map to access a nearby
PTE and (2) loading a direct mapping into the TLB involves walking only 2
or 3 levels of the page table instead of 4.
Unmasked exceptions (which can be fixed up using fpset*() before they
trap) are very rare, especially on amd64 since SSE exceptions trap
synchronously, but I want to merge the faster amd64 implementations of
fpset*() back to i386 without introducing the bug on i386.
The i386 implementation has always avoided the trap automatically by
changing things using load/store of the FP environment, but this is
very slow. Most changes only affect the control word, so they can
usually be done much more efficiently, and amd64 has always done this,
but loading the control word can trap.
This version use the fast method only in the usual case where it will
not trap. This only costs a couple of integer instructions (including
one branch which I haven't optimized carefully yet) in the usual case,
but bloats the inlines a lot. The inlines were already a bit too large
to handle both the FPU and SSE.
- fix a previous style fix: shifts should be in the correct direction even
if they are null.
- restore a comment about namespace pollution from floatingpoint.h 1.12 and
update it.
- remove unused namespace pollution FP_*REG.
- improve some comments.
- sort macro definitions for entry points.
- don't use underscores for macro args.
- fix this to compile with C++ by casting ints to enums in a few places
and by using the correct parameter type for _fpsetprec(). Remove
__cplusplus ifdefs which disabled the buggy code.
- remove __CC_SUPPORTS___INLINE ifdefs. `__inline' vs `inline', and either
of these #defined away, are supposed to be handled by very old ifdefs
in <sys/cdefs.h>. Thus the __CC_SUPPORTS___INLINE macro is not needed
here (or anywhere else that it used). It is less needed here than in
most places, since this file is userland-only and userland is far from
supporting INTEL_COMPILER. The __CC_SUPPORTS___INLINE__ macro which
was used here is even less needed. It is to support spelling `inline'
as `__inline__' instead of the usual spelling `__inline'.
Fix some style bugs that I missed in the previous commit (remove unused
asms and sort more variables).
pv_list_count from struct md_page. Ever since Peter rewrote the pv
entry allocator for amd64 and i386 pv_list_count has been correctly
maintained but otherwise unused.
