Reduce number of iterations used for calibrating ICR read loop. The
new number of iteration still gives the same ICR latency as before,
tested on Intel SandyBridge and Haswell machines, and on AMD. But it
significantly reduces the unneeded pause on boot in some VMs, from ~10
secs to less then 1 sec. It was reported to occur in bhyve on AMD
host.
Reported and tested by: avg
Sponsored by: The FreeBSD Foundation
MFC after: 1 week
The existing version depends on register_t and uintptr_t, which are only
available when including headers such as <sys/types.h>. As this macro is
used by <sys/socket.h>, for example, it should be written in such a way
that it doesn't depend on those types.
In r227474, this header file was changed to define SIG_ATOMIC_{MIN,MAX}
in terms of LONG_{MIN,MAX}. Unlike all of the definitions in this header
file, LONG_{MIN,MAX} is provided by <limits.h>. Remove the dependency on
<limits.h> by using __LONG_{MIN,MAX} instead and including
<machine/_limits.h>.
This change is needed to make SIG_ATOMIC_{MIN,MAX} work without
including any other header files.
This header uses __INT_MIN and __INT_MAX, which is provided by
<machine/_limits.h>. This is needed to make <stdint.h>'s WCHAR_MIN and
WCHAR_MAX work without including other headers as well.
switching between LAPIC modes is not supported, and there is no need
to wait for IPI ack in x2APIC mode. So the calibrated delay is only
needed for !x2APIC.
This saves around a second of boot time on the real hardware for
x2APIC.
Sponsored by: The FreeBSD Foundation
Add implementations of bus_map/unmap_resource to the x86 nexus driver.
Change bus_activate/deactivate_resource to honor RF_UNMAPPED and to
use bus_map/unmap_resource to create/destroy the implicit mapping when
RF_UNMAPPED is not set.
Reviewed by: cem
Differential Revision: https://reviews.freebsd.org/D5237
Currently, Application Processors (non-boot CPUs) are started by
MD code at SI_SUB_CPU, but they are kept waiting in a "pen" until
SI_SUB_SMP at which point they are released to run kernel threads.
SI_SUB_SMP is one of the last SYSINIT levels, so APs don't enter
the scheduler and start running threads until fairly late in the
boot.
This change moves SI_SUB_SMP up to just before software interrupt
threads are created allowing the APs to start executing kernel
threads much sooner (before any devices are probed). This allows
several initialization routines that need to perform initialization
on all CPUs to now perform that initialization in one step rather
than having to defer the AP initialization to a second SYSINIT run
at SI_SUB_SMP. It also permits all CPUs to be available for
handling interrupts before any devices are probed.
This last feature fixes a problem on with interrupt vector exhaustion.
Specifically, in the old model all device interrupts were routed
onto the boot CPU during boot. Later after the APs were released at
SI_SUB_SMP, interrupts were redistributed across all CPUs.
However, several drivers for multiqueue hardware allocate N interrupts
per CPU in the system. In a system with many CPUs, just a few drivers
doing this could exhaust the available pool of interrupt vectors on
the boot CPU as each driver was allocating N * mp_ncpu vectors on the
boot CPU. Now, drivers will allocate interrupts on their desired CPUs
during boot meaning that only N interrupts are allocated from the boot
CPU instead of N * mp_ncpu.
Some other bits of code can also be simplified as smp_started is
now true much earlier and will now always be true for these bits of
code. This removes the need to treat the single-CPU boot environment
as a special case.
As a transition aid, the new behavior is available under a new kernel
option (EARLY_AP_STARTUP). This will allow the option to be turned off
if need be during initial testing. I plan to enable this on x86 by
default in a followup commit in the next few days and to have all
platforms moved over before 11.0. Once the transition is complete,
the option will be removed along with the !EARLY_AP_STARTUP code.
These changes have only been tested on x86. Other platform maintainers
are encouraged to port their architectures over as well. The main
things to check for are any uses of smp_started in MD code that can be
simplified and SI_SUB_SMP SYSINITs in MD code that can be removed in
the EARLY_AP_STARTUP case (e.g. the interrupt shuffling).
PR: kern/199321
Reviewed by: markj, gnn, kib
Sponsored by: Netflix
if specific CPU features are not present.
Some simulation environments, e.g. gem5, have been found to require more
TLB management from the kernel in certain setups. It is currently unclear why.
Turning on the workaround_erratum383 seems to help and make problems (panics)
go away.
Given this is a fairly uncommon environment so far, allowing the workaround
to be manually enabled from loader in order to make debugging and comparing
traces easier, but also to allow gem5 run FreeBSD in X86 timing mode, seems
to be the least intrusive option for now until the issue if fully understood.
Sponsored by: DARPA/AFRL
Reviewed by: kib, alc (earlier)
MFC after: 2 weeks
Differential Revision: https://reviews.freebsd.org/D6206
bus_get_cpus() returns a specified set of CPUs for a device. It accepts
an enum for the second parameter that indicates the type of cpuset to
request. Currently two valus are supported:
- LOCAL_CPUS (on x86 this returns all the CPUs in the package closest to
the device when DEVICE_NUMA is enabled)
- INTR_CPUS (like LOCAL_CPUS but only returns 1 SMT thread for each core)
For systems that do not support NUMA (or if it is not enabled in the kernel
config), LOCAL_CPUS fails with EINVAL. INTR_CPUS is mapped to 'all_cpus'
by default. The idea is that INTR_CPUS should always return a valid set.
Device drivers which want to use per-CPU interrupts should start using
INTR_CPUS instead of simply assigning interrupts to all available CPUs.
In the future we may wish to add tunables to control the policy of
INTR_CPUS (e.g. should it be local-only or global, should it ignore
SMT threads or not).
The x86 nexus driver exposes the internal set of interrupt CPUs from the
the x86 interrupt code via INTR_CPUS.
The ACPI bus driver and PCI bridge drivers use _PXM to return a suitable
LOCAL_CPUS set when _PXM exists and DEVICE_NUMA is enabled. They also and
the global INTR_CPUS set from the nexus driver with the per-domain set from
_PXM to generate a local INTR_CPUS set for child devices.
Compared to the r298933, this version uses 'struct _cpuset' in
<sys/bus.h> instead of 'cpuset_t' to avoid requiring <sys/param.h>
(<sys/_cpuset.h> still requires <sys/param.h> for MAXCPU even though
<sys/_bitset.h> does not after recent changes).
Instead of panicking when parsing an invalid ACPI SRAT table,
just ignore it, effectively disabling NUMA.
https://lists.freebsd.org/pipermail/freebsd-current/2016-May/060984.html
Reported and tested by: Bill O'Hanlon (bill.ohanlon at gmail.com)
Reviewed by: jhb
MFC after: 1 week
Relnotes: If dmesg shows "SRAT: Duplicate local APIC ID",
try updating your BIOS to fix NUMA support.
Sponsored by: Dell Inc.
bus_get_cpus() returns a specified set of CPUs for a device. It accepts
an enum for the second parameter that indicates the type of cpuset to
request. Currently two valus are supported:
- LOCAL_CPUS (on x86 this returns all the CPUs in the package closest to
the device when DEVICE_NUMA is enabled)
- INTR_CPUS (like LOCAL_CPUS but only returns 1 SMT thread for each core)
For systems that do not support NUMA (or if it is not enabled in the kernel
config), LOCAL_CPUS fails with EINVAL. INTR_CPUS is mapped to 'all_cpus'
by default. The idea is that INTR_CPUS should always return a valid set.
Device drivers which want to use per-CPU interrupts should start using
INTR_CPUS instead of simply assigning interrupts to all available CPUs.
In the future we may wish to add tunables to control the policy of
INTR_CPUS (e.g. should it be local-only or global, should it ignore
SMT threads or not).
The x86 nexus driver exposes the internal set of interrupt CPUs from the
the x86 interrupt code via INTR_CPUS.
The ACPI bus driver and PCI bridge drivers use _PXM to return a suitable
LOCAL_CPUS set when _PXM exists and DEVICE_NUMA is enabled. They also and
the global INTR_CPUS set from the nexus driver with the per-domain set from
_PXM to generate a local INTR_CPUS set for child devices.
Reviewed by: wblock (manpage)
Differential Revision: https://reviews.freebsd.org/D5519
This is going to be used by the Xen clock on Dom0 in order to set the RTC of
the host. The current logic in atrtc_settime is moved to atrtc_set and the
unused device_t parameter is removed from the atrtc_set function call so it
can be safely used by other callers.
Sponsored by: Citrix Systems R&D
Reviewed by: kib, jhb
Differential revision: https://reviews.freebsd.org/D6067
rounddown2 tends to produce longer lines than the original code
and when the code has a high indentation level it was not really
advantageous to do the replacement.
This tries to strike a balance between readability using the macros
and flexibility of having the expressions, so not everything is
converted.
If we reached MAXMEMDOM, we would previously try to insert an additional
element and only detect overflow after causing (probably trivial) memory
overflow. Instead, detect the ndomain > MAXMEMDOM case before we write past
the end.
Reported by: Coverity
CID: 1354783
Sponsored by: EMC / Isilon Storage Division
which queued invalidation completion interrupt is requested with
regard to the queued invalidation requests. In other words, setting
the value of the knob to N requests completion interrupt after N items
are processed. Existing behaviour is restored by setting
hw.dmar.batch_coalesce=1.
The knob significantly decreases the DMAR qi interrupt rate at the
cost of slightly longer DMAR map entries recycling.
Sponsored by: The FreeBSD Foundation
initially configured in the TSC deadline mode, eventtimer subsystem
can be switched to periodic, and then DCR register is loaded with
unitialized value.
Reset the LAPIC eventtimer frequency and min/max periods when changing
between deadline and counted periodic modes.
Reported and tested by: Vladimir Zakharov <zakharov.vv@gmail.com>
Sponsored by: The FreeBSD Foundation
Revert r292255 because it can create bounced regions without contiguous
page offsets, which is needed for USB devices.
Another solution would be to force bouncing the full buffer always (even
when only one page requires bouncing), but this seems overly complicated and
unnecessary, and it will probably involve using more bounce pages than the
current code.
Reported by: phk
system. This uses the hints mechnanism. This mostly works today
because when there's no static hints (the default), this value can be
fetched from the hint. When there is a static hints file, the hint
passed from the boot loader to the kernel is ignored, but for the BIOS
case we're able to find it anyway. However, with UEFI, the fallback
doesn't work, so we get a panic instead.
Switch to acpi.rsdp and use TUNABLE_ULONG_FETCH instead. Continue to
generate the old values to allow for transitions. In addition, fall
back to the old method if the new method isn't present.
Add comments about all this.
Differential Revision: https://reviews.freebsd.org/D5866
Some BIOSes disable AMD Topology extension on AMD Family 15h notebook
processors. We re-enable the extension, so that we can properly discover
core and cache topology. Linux seems to do the same.
Reported by: Johannes Dieterich <dieterich.joh@gmail.com>
Reviewed by: jhb, kib
Tested by: Johannes Dieterich <dieterich.joh@gmail.com>
(earlier version)
MFC after: 3 weeks
Differential Revision: https://reviews.freebsd.org/D5883
VM_NUMA_ALLOC is used to enable use of domain-aware memory allocation in
the virtual memory system. DEVICE_NUMA is used to enable affinity
reporting for devices such as bus_get_domain().
MAXMEMDOM must still be set to a value greater than for any NUMA support
to be effective. Note that 'cpuset -gd' always works if MAXMEMDOM is
enabled and the system supports NUMA.
Reviewed by: kib
Differential Revision: https://reviews.freebsd.org/D5782
It is needed by the hypervisor FreeBSD guest to allocate/free private
interrupt vectors.
Reviewed by: kib, jhb, Dexuan Cui <decui microsoft com>
Sponsored by: Microsoft OSTC
Differential Revision: https://reviews.freebsd.org/D5849
Previously, the code determined a topology of processing units
(hardware threads, cores, packages) and then deduced a cache topology
using certain assumptions. The new code builds a topology that
includes both processing units and caches using the information
provided by the hardware.
At the moment, the discovered full topology is used only to creeate
a scheduling topology for SCHED_ULE.
There is no KPI for other kernel uses.
Summary:
- based on APIC ID derivation rules for Intel and AMD CPUs
- can handle non-uniform topologies
- requires homogeneous APIC ID assignment (same bit widths for ID
components)
- topology for dual-node AMD CPUs may not be optimal
- topology for latest AMD CPU models may not be optimal as the code is
several years old
- supports only thread/package/core/cache nodes
Todo:
- AMD dual-node processors
- latest AMD processors
- NUMA nodes
- checking for homogeneity of the APIC ID assignment across packages
- more flexible cache placement within topology
- expose topology to userland, e.g., via sysctl nodes
Long term todo:
- KPI for CPU sharing and affinity with respect to various resources
(e.g., two logical processors may share the same FPU, etc)
Reviewed by: mav
Tested by: mav
MFC after: 1 month
Differential Revision: https://reviews.freebsd.org/D2728
Simplify and unify placeholder type definitions.
Reviewed by: jhb
Sponsored by: The FreeBSD Foundation
Differential revision: https://reviews.freebsd.org/D5771
- fix UP build [1]
- do not obliterate initial reading of rdtsc by the loop counter [2]
- restore the meaning of the argument -1 to native_lapic_ipi_wait()
as wait until LAPIC acknowledge without timeout
- correct formula for calculating loop iteration count for 1us, it was
inverted, and ensure that even on unlikely slow CPUs at least one
check for ack is performed.
Reported by: Michael Butler <imb@protected-networks.net> [1], rpokala[2],
jhb[3]
Tested by: Michael Butler
Pointy hat to: kib
Sponsored by: The FreeBSD Foundation
and avoid a delay while waiting for IPI delivery acknowledgement in
xAPIC mode. This makes the loop exit immediately after the delivery
bit in APIC_ICR register is set, instead of waiting for some
microseconds.
We only need to ensure that some amount of time is allowed for the
LAPIC to react to the command, and we need that the wait time is
finite and reasonable. For that reasons, it is irrelevant if the CPU
frequency or throttling decrease the speed and make the loop,
calibrated for full CPU speed at boot time, execute somewhat slower.
Discussed with: bde, jhb
Tested by: pho
Sponsored by: The FreeBSD Foundation
deadline mode the divide configuration is not used and
lapic_timer_divisor is not set.
Reported by: dhw, mav
Tested by: mav
Sponsored by: The FreeBSD Foundation
LAPIC timer iinterrupt when TSC reaches the value written to the
IA32_TSC_DEADLINE MSR. To arm or reset the timer in deadline mode, a
single non-serializing MSR write is enough. This is an advance from
the one-shot mode of LAPIC, where timer operated with the FSB
frequency and required two (serialized in case of xAPIC) writes to the
APIC registers.
The LVT_TIMER register value is cached to avoid unneeded writes in the
deadline mode. Unused arguments to specify period (which is passed in
struct lapic as la_timer_period) and interrupt enable (which is always
enabled) are removed from lapic_timer_{oneshot,periodic,deadline}
functions. Instead, special lapic_timer_oneshot_nointr() function for
interrupt-less one-shot calibration is added.
Reviewed by: mav (previous version)
Tested by: pho
Sponsored by: The FreeBSD Foundation
Differential revision: https://reviews.freebsd.org/D5738
This moves the enabling of interrupts slightly earlier (the old location
was still before devices were enumerated and probed) and does it in the
interrupt code (rather than in the device configuration code). This
also avoids tripping over an assertion on the first TLB shootdown with
earlier AP startup.
Reviewed by: kib
Sponsored by: Netflix
Differential Revision: https://reviews.freebsd.org/D5710
An IPI cannot be sent via the local APIC if a previous IPI is still
being delivered. Attempts to send an IPI will wait for a pending IPI
to clear. Prior to r278325 these checks used a spin loop with a
hardcoded maximum count which broke AP startup on some systems.
However, r278325 also enforced a minimum latency of 5 microseconds if an
IPI was still pending which resulted in a measurable performance hit.
This change reduces that minimum latency to 1 microsecond.
Tested by: stas
MFC after: 3 days
On some architectures, u_long isn't large enough for resource definitions.
Particularly, powerpc and arm allow 36-bit (or larger) physical addresses, but
type `long' is only 32-bit. This extends rman's resources to uintmax_t. With
this change, any resource can feasibly be placed anywhere in physical memory
(within the constraints of the driver).
Why uintmax_t and not something machine dependent, or uint64_t? Though it's
possible for uintmax_t to grow, it's highly unlikely it will become 128-bit on
32-bit architectures. 64-bit architectures should have plenty of RAM to absorb
the increase on resource sizes if and when this occurs, and the number of
resources on memory-constrained systems should be sufficiently small as to not
pose a drastic overhead. That being said, uintmax_t was chosen for source
clarity. If it's specified as uint64_t, all printf()-like calls would either
need casts to uintmax_t, or be littered with PRI*64 macros. Casts to uintmax_t
aren't horrible, but it would also bake into the API for
resource_list_print_type() either a hidden assumption that entries get cast to
uintmax_t for printing, or these calls would need the PRI*64 macros. Since
source code is meant to be read more often than written, I chose the clearest
path of simply using uintmax_t.
Tested on a PowerPC p5020-based board, which places all device resources in
0xfxxxxxxxx, and has 8GB RAM.
Regression tested on qemu-system-i386
Regression tested on qemu-system-mips (malta profile)
Tested PAE and devinfo on virtualbox (live CD)
Special thanks to bz for his testing on ARM.
Reviewed By: bz, jhb (previous)
Relnotes: Yes
Sponsored by: Alex Perez/Inertial Computing
Differential Revision: https://reviews.freebsd.org/D4544
Summary:
The idea behind this is '~0ul' is well-defined, and casting to uintmax_t, on a
32-bit platform, will leave the upper 32 bits as 0. The maximum range of a
resource is 0xFFF.... (all bits of the full type set). By dropping the 'ul'
suffix, C type promotion rules apply, and the sign extension of ~0 on 32 bit
platforms gets it to a type-independent 'unsigned max'.
Reviewed By: cem
Sponsored by: Alex Perez/Inertial Computing
Differential Revision: https://reviews.freebsd.org/D5255
taskqueue_enqueue() was changed to support both fast and non-fast
taskqueues 10 years ago in r154167. It has been a compat shim ever
since. It's time for the compat shim to go.
Submitted by: Howard Su <howard0su@gmail.com>
Reviewed by: sephe
Differential Revision: https://reviews.freebsd.org/D5131
Summary:
As part of the migration of rman_res_t to be typed to uintmax_t, memory ranges
must be clamped appropriately for the bus, to prevent completely bogus addresses
from being used.
This is extracted from D4544.
Reviewed By: cem
Sponsored by: Alex Perez/Inertial Computing
Differential Revision: https://reviews.freebsd.org/D5134
