the SMP load-balancer uses the boot-time migration-cost estimation
code to attempt to improve the quality of balancing. The reason for
this code is that the discrete priority queues do not preserve
the order of scheduling accurately, so the load-balancer skips
tasks that were running on a CPU 'recently'.
this code is fundamental fragile: the boot-time migration cost detector
doesnt really work on systems that had large L3 caches, it caused boot
delays on large systems and the whole cache-hot concept made the
balancing code pretty undeterministic as well.
(and hey, i wrote most of it, so i can say it out loud that it sucks ;-)
under CFS the same purpose of cache affinity can be achieved without
any special cache-hot special-case: tasks are sorted in the 'timeline'
tree and the SMP balancer picks tasks from the left side of the
tree, thus the most cache-cold task is balanced automatically.
Signed-off-by: Ingo Molnar <mingo@elte.hu>
It's not just sun4v hypervisor platforms that should return true
for this, sun4u with UltraSPARC-IV should return true too.
Signed-off-by: David S. Miller <davem@davemloft.net>
The scheduling domain hierarchy is:
all cpus -->
cpus that share an instruction cache -->
cpus that share an integer execution unit
Signed-off-by: David S. Miller <davem@davemloft.net>
Cheetah systems can have cpuids as large as 1023, although physical
systems don't have that many cpus.
Only three limitations existed in the kernel preventing arbitrary
NR_CPUS values:
1) dcache dirty cpu state stored in page->flags on
D-cache aliasing platforms. With some build time
calculations and some build-time BUG checks on
page->flags layout, this one was easily solved.
2) The cheetah XCALL delivery code could only handle
a cpumask with up to 32 cpus set. Some simple looping
logic clears that up too.
3) thread_info->cpu was a u8, easily changed to a u16.
There are a few spots in the kernel that still put NR_CPUS
sized arrays on the kernel stack, but that's not a sparc64
specific problem.
Signed-off-by: David S. Miller <davem@davemloft.net>
Remove includes of <linux/smp_lock.h> where it is not used/needed.
Suggested by Al Viro.
Builds cleanly on x86_64, i386, alpha, ia64, powerpc, sparc,
sparc64, and arm (all 59 defconfigs).
Signed-off-by: Randy Dunlap <randy.dunlap@oracle.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Let's allow page-alignment in general for per-cpu data (wanted by Xen, and
Ingo suggested KVM as well).
Because larger alignments can use more room, we increase the max per-cpu
memory to 64k rather than 32k: it's getting a little tight.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Signed-off-by: Jeremy Fitzhardinge <jeremy@xensource.com>
Signed-off-by: Andi Kleen <ak@suse.de>
Acked-by: Ingo Molnar <mingo@elte.hu>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
I'd like to thank John Stul and others for helping
me along the way.
A lot of cleanups fell out of this. For example, the get_compare()
tick_op was totally unused, so was deleted. And the most often used
tick_op members were grouped together for cache-friendlyness.
The sparc64 TSC is given to the kernel as a one-shot timer.
tick_ops->init_timer() simply turns off the privileged bit in
the tick register (when possible), and disables the interrupt
by setting bit 63 in the compare register. The ->disable_irq()
op also sets this bit.
tick_ops->add_compare() is changed to:
1) Add the given delta to "tick" not to "compare"
2) Return a boolean which, if true, means that the tick
value read after writing the compare value was found
to have incremented past the initial tick value. This
mirrors logic used in the HPET driver's ->next_event()
method.
Each tick_ops implementation also now provides a name string.
And we feed this into the clocksource and clockevents layers.
Signed-off-by: David S. Miller <davem@davemloft.net>
Things were scattered all over the place, split between
SMP and non-SMP.
Unify it all so that dyntick support is easier to add.
Signed-off-by: David S. Miller <davem@davemloft.net>
Compiling the kernel with CONFIG_HOTPLUG = y and CONFIG_HOTPLUG_CPU = n
with CONFIG_RELOCATABLE = y generates the following modpost warnings
WARNING: vmlinux - Section mismatch: reference to .init.data: from
.text between '_cpu_up' (at offset 0xc0141b7d) and 'cpu_up'
WARNING: vmlinux - Section mismatch: reference to .init.data: from
.text between '_cpu_up' (at offset 0xc0141b9c) and 'cpu_up'
WARNING: vmlinux - Section mismatch: reference to .init.text:__cpu_up
from .text between '_cpu_up' (at offset 0xc0141bd8) and 'cpu_up'
WARNING: vmlinux - Section mismatch: reference to .init.data: from
.text between '_cpu_up' (at offset 0xc0141c05) and 'cpu_up'
WARNING: vmlinux - Section mismatch: reference to .init.data: from
.text between '_cpu_up' (at offset 0xc0141c26) and 'cpu_up'
WARNING: vmlinux - Section mismatch: reference to .init.data: from
.text between '_cpu_up' (at offset 0xc0141c37) and 'cpu_up'
This is because cpu_up, _cpu_up and __cpu_up (in some architectures) are
defined as __devinit
AND
__cpu_up calls some __cpuinit functions.
Since __cpuinit would map to __init with this kind of a configuration,
we get a .text refering .init.data warning.
This patch solves the problem by converting all of __cpu_up, _cpu_up
and cpu_up from __devinit to __cpuinit. The approach is justified since
the callers of cpu_up are either dependent on CONFIG_HOTPLUG_CPU or
are of __init type.
Thus when CONFIG_HOTPLUG_CPU=y, all these cpu up functions would land up
in .text section, and when CONFIG_HOTPLUG_CPU=n, all these functions would
land up in .init section.
Tested on a i386 SMP machine running linux-2.6.20-rc3-mm1.
Signed-off-by: Gautham R Shenoy <ego@in.ibm.com>
Cc: Vivek Goyal <vgoyal@in.ibm.com>
Cc: Mikael Starvik <starvik@axis.com>
Cc: Ralf Baechle <ralf@linux-mips.org>
Cc: Kyle McMartin <kyle@mcmartin.ca>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: "David S. Miller" <davem@davemloft.net>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Uses of smp_processor_id() get pushed earlier and earlier in
the start_kernel() sequence. So just get it working before
we call start_kernel() to avoid all possible problems.
Signed-off-by: David S. Miller <davem@davemloft.net>
for_each_cpu() actually iterates across all possible CPUs. We've had mistakes
in the past where people were using for_each_cpu() where they should have been
iterating across only online or present CPUs. This is inefficient and
possibly buggy.
We're renaming for_each_cpu() to for_each_possible_cpu() to avoid this in the
future.
This patch replaces for_each_cpu with for_each_possible_cpu.
for sparc64.
Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Acked-by: "David S. Miller" <davem@davemloft.net>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
1) Take doc-book function comment from i386 implementation.
2) cacheline align call_lock, taken from powerpc
3) Need memory barrier after setting call_data
4) Remove timeout
Signed-off-by: David S. Miller <davem@davemloft.net>
switch_mm() changes the mm state and does a tsb_context_switch()
first, then we do the cpu register state switch which changes
current_thread_info() and current().
So it's safer to check the PGD physical address stored in the
trap block (which will be updated by the tsb_context_switch() in
switch_mm()) than current->active_mm.
Technically we should never run here in between those two
updates, because interrupts are disabled during the entire
context switch operation. But some day we might like to leave
interrupts enabled during the context switch and this change
allows that to happen without any surprises.
Signed-off-by: David S. Miller <davem@davemloft.net>
When we stop allocating percpu memory for not-possible CPUs we must not touch
the percpu data for not-possible CPUs at all. The correct way of doing this
is to test cpu_possible() or to use for_each_cpu().
This patch is a kernel-wide sweep of all instances of NR_CPUS. I found very
few instances of this bug, if any. But the patch converts lots of open-coded
test to use the preferred helper macros.
Cc: Mikael Starvik <starvik@axis.com>
Cc: David Howells <dhowells@redhat.com>
Acked-by: Kyle McMartin <kyle@parisc-linux.org>
Cc: Anton Blanchard <anton@samba.org>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Martin Schwidefsky <schwidefsky@de.ibm.com>
Cc: Heiko Carstens <heiko.carstens@de.ibm.com>
Cc: Paul Mundt <lethal@linux-sh.org>
Cc: "David S. Miller" <davem@davemloft.net>
Cc: William Lee Irwin III <wli@holomorphy.com>
Cc: Andi Kleen <ak@muc.de>
Cc: Christian Zankel <chris@zankel.net>
Cc: Philippe Elie <phil.el@wanadoo.fr>
Cc: Nathan Scott <nathans@sgi.com>
Cc: Jens Axboe <axboe@suse.de>
Cc: Eric Dumazet <dada1@cosmosbay.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
The mapping is a simple "(cpuid >> 2) == core" for now.
Later we'll add more sophisticated code that will walk
the sun4v machine description and figure this out from
there.
We should also add core mappings for jaguar and panther
processors.
Signed-off-by: David S. Miller <davem@davemloft.net>
Don't piggy back the SMP receive signal code to do the
context version change handling.
Instead allocate another fixed PIL number for this
asynchronous cross-call. We can't use smp_call_function()
because this thing is invoked with interrupts disabled
and a few spinlocks held.
Also, fix smp_call_function_mask() to count "cpus" correctly.
There is no guarentee that the local cpu is in the mask
yet that is exactly what this code was assuming.
Signed-off-by: David S. Miller <davem@davemloft.net>
This cpu mondo sending interface isn't all that easy to
use correctly...
We were clearing out the wrong bits from the "mask" after getting
something other than EOK from the hypervisor.
It turns out the hypervisor can just be resent the same cpu_list[]
array, with the 0xffff "done" entries still in there, and it will do
the right thing.
So don't update or try to rebuild the cpu_list[] array to condense it.
This requires the "forward_progress" check to be done slightly
differently, but this new scheme is less bug prone than what we were
doing before.
Signed-off-by: David S. Miller <davem@davemloft.net>
There were several bugs in the SUN4V cpu mondo dispatch code.
In fact, if we ever got a EWOULDBLOCK or other error from
the hypervisor call, we'd potentially send a cpu mondo multiple
times to the same cpu and even worse we could loop until the
timeout resending the same mondo over and over to such cpus.
So let's bulletproof this thing as follows:
1) Implement cpu_mondo_send() and cpu_state() hypervisor calls
in arch/sparc64/kernel/entry.S, add prototypes to asm/hypervisor.h
2) Don't build and update the cpulist using inline functions, this
was causing the cpu mask to not get updated in the caller.
3) Disable interrupts during the entire mondo send, otherwise our
cpu list and/or mondo block could get overwritten if we take
an interrupt and do a cpu mondo send on the current cpu.
4) Check for all possible error return types from the cpu_mondo_send()
hypervisor call. In particular:
HV_EOK) Our work is done, all cpus have received the mondo.
HV_CPUERROR) One or more of the cpus in the cpu list we passed
to the hypervisor are in error state. Use cpu_state()
calls over the entries in the cpu list to see which
ones. Record them in "error_mask" and report this
after we are done sending the mondo to cpus which are
not in error state.
HV_EWOULDBLOCK) We need to keep trying.
Any other error we consider fatal, we report the event and exit
immediately.
5) We only timeout if forward progress is not made. Forward progress
is defined as having at least one cpu get the mondo successfully
in a given cpu_mondo_send() call. Otherwise we bump a counter
and delay a little. If the counter hits a limit, we signal an
error and report the event.
Also, smp_call_function_mask() error handling reports the number
of cpus incorrectly.
Signed-off-by: David S. Miller <davem@davemloft.net>
1) We must flush the TLB, duh.
2) Even if the sw context was seen to be valid, the local cpu's
hw context can be out of date, so reload it unconditionally.
Signed-off-by: David S. Miller <davem@davemloft.net>
The context allocation scheme we use depends upon there being a 1<-->1
mapping from cpu to physical TLB for correctness. Chips like Niagara
break this assumption.
So what we do is notify all cpus with a cross call when the context
version number changes, and if necessary this makes them allocate
a valid context for the address space they are running at the time.
Stress tested with make -j1024, make -j2048, and make -j4096 kernel
builds on a 32-strand, 8 core, T2000 with 16GB of ram.
Signed-off-by: David S. Miller <davem@davemloft.net>
Set, but never used.
We used to use this for dynamic IRQ retargetting, but that
code died a long time ago.
Signed-off-by: David S. Miller <davem@davemloft.net>
The sibling cpu bringup is extremely fragile. We can only
perform the most basic calls until we take over the trap
table from the firmware/hypervisor on the new cpu.
This means no accesses to %g4, %g5, %g6 since those can't be
TLB translated without our trap handlers.
In order to achieve this:
1) Change sun4v_init_mondo_queues() so that it can operate in
several modes.
It can allocate the queues, or install them in the current
processor, or both.
The boot cpu does both in it's call early on.
Later, the boot cpu allocates the sibling cpu queue, starts
the sibling cpu, then the sibling cpu loads them in.
2) init_cur_cpu_trap() is changed to take the current_thread_info()
as an argument instead of reading %g6 directly on the current
cpu.
3) Create a trampoline stack for the sibling cpus. We do our basic
kernel calls using this stack, which is locked into the kernel
image, then go to our proper thread stack after taking over the
trap table.
4) While we are in this delicate startup state, we put 0xdeadbeef
into %g4/%g5/%g6 in order to catch accidental accesses.
5) On the final prom_set_trap_table*() call, we put &init_thread_union
into %g6. This is a hack to make prom_world(0) work. All that
wants to do is restore the %asi register using
get_thread_current_ds().
Longer term we should just do the OBP calls to set the trap table by
hand just like we do for everything else. This would avoid that silly
prom_world(0) issue, then we can remove the init_thread_union hack.
Signed-off-by: David S. Miller <davem@davemloft.net>
Use prom_startcpu_cpuid() on SUN4V instead of prom_startcpu().
We should really test for "SUNW,start-cpu-by-cpuid" presence
and use it if present even on SUN4U.
Signed-off-by: David S. Miller <davem@davemloft.net>
We have to use bootmem during init_IRQ and page alloc
for sibling cpu calls.
Also, fix incorrect hypervisor call return value
checks in the hypervisor SMP cpu mondo send code.
Signed-off-by: David S. Miller <davem@davemloft.net>
Technically the hypervisor call supports sending in a list
of all cpus to get the cross-call, but I only pass in one
cpu at a time for now.
The multi-cpu support is there, just ifdef'd out so it's easy to
enable or delete it later.
Signed-off-by: David S. Miller <davem@davemloft.net>
And more consistently check cheetah{,_plus} instead
of assuming anything not spitfire is cheetah{,_plus}.
Signed-off-by: David S. Miller <davem@davemloft.net>
On uniprocessor, it's always zero for optimize that.
On SMP, the jmpl to the stub kills the return address stack in the cpu
branch prediction logic, so expand the code sequence inline and use a
code patching section to fix things up. This also always better and
explicit register selection, which will be taken advantage of in a
future changeset.
The hard_smp_processor_id() function is big, so do not inline it.
Fix up tests for Jalapeno to also test for Serrano chips too. These
tests want "jbus Ultra-IIIi" cases to match, so that is what we should
test for.
Signed-off-by: David S. Miller <davem@davemloft.net>
It is totally unnecessary complexity. After we take over
the trap table, we handle all PROM tlb misses fully.
Signed-off-by: David S. Miller <davem@davemloft.net>
As the RSS grows, grow the TSB in order to reduce the likelyhood
of hash collisions and thus poor hit rates in the TSB.
This definitely needs some serious tuning.
Signed-off-by: David S. Miller <davem@davemloft.net>
UltraSPARC has special sets of global registers which are switched to
for certain trap types. There is one set for MMU related traps, one
set of Interrupt Vector processing, and another set (called the
Alternate globals) for all other trap types.
For what seems like forever we've hard coded the values in some of
these trap registers. Some examples include:
1) Interrupt Vector global %g6 holds current processors interrupt
work struct where received interrupts are managed for IRQ handler
dispatch.
2) MMU global %g7 holds the base of the page tables of the currently
active address space.
3) Alternate global %g6 held the current_thread_info() value.
Such hardcoding has resulted in some serious issues in many areas.
There are some code sequences where having another register available
would help clean up the implementation. Taking traps such as
cross-calls from the OBP firmware requires some trick code sequences
wherein we have to save away and restore all of the special sets of
global registers when we enter/exit OBP.
We were also using the IMMU TSB register on SMP to hold the per-cpu
area base address, which doesn't work any longer now that we actually
use the TSB facility of the cpu.
The implementation is pretty straight forward. One tricky bit is
getting the current processor ID as that is different on different cpu
variants. We use a stub with a fancy calling convention which we
patch at boot time. The calling convention is that the stub is
branched to and the (PC - 4) to return to is in register %g1. The cpu
number is left in %g6. This stub can be invoked by using the
__GET_CPUID macro.
We use an array of per-cpu trap state to store the current thread and
physical address of the current address space's page tables. The
TRAP_LOAD_THREAD_REG loads %g6 with the current thread from this
table, it uses __GET_CPUID and also clobbers %g1.
TRAP_LOAD_IRQ_WORK is used by the interrupt vector processing to load
the current processor's IRQ software state into %g6. It also uses
__GET_CPUID and clobbers %g1.
Finally, TRAP_LOAD_PGD_PHYS loads the physical address base of the
current address space's page tables into %g7, it clobbers %g1 and uses
__GET_CPUID.
Many refinements are possible, as well as some tuning, with this stuff
in place.
Signed-off-by: David S. Miller <davem@davemloft.net>
We now use the TSB hardware assist features of the UltraSPARC
MMUs.
SMP is currently knowingly broken, we need to find another place
to store the per-cpu base pointers. We hid them away in the TSB
base register, and that obviously will not work any more :-)
Another known broken case is non-8KB base page size.
Also noticed that flush_tlb_all() is not referenced anywhere, only
the internal __flush_tlb_all() (local cpu only) is used by the
sparc64 port, so we can get rid of flush_tlb_all().
The kernel gets it's own 8KB TSB (swapper_tsb) and each address space
gets it's own private 8K TSB. Later we can add code to dynamically
increase the size of per-process TSB as the RSS grows. An 8KB TSB is
good enough for up to about a 4MB RSS, after which the TSB starts to
incur many capacity and conflict misses.
We even accumulate OBP translations into the kernel TSB.
Another area for refinement is large page size support. We could use
a secondary address space TSB to handle those.
Signed-off-by: David S. Miller <davem@davemloft.net>
The change to kernel/sched.c's init code to use for_each_cpu()
requires that the cpu_possible_map be setup much earlier.
Set it up via setup_arch(), constrained to NR_CPUS, and later
constrain it to max_cpus in smp_prepare_cpus().
This fixes SMP booting on sparc64.
Signed-off-by: David S. Miller <davem@davemloft.net>
Noticed by Tom 'spot' Callaway.
Even on uniprocessor we always reported the number of physical
cpus in the system via /proc/cpuinfo. But when this got changed
to use num_possible_cpus() it always reads as "1" on uniprocessor.
This change was unintentional.
So scan the firmware device tree and count the number of cpu
nodes, and report that, as we always did.
Signed-off-by: David S. Miller <davem@davemloft.net>
