In times of ARC 700 performance counters didn't have support of
interrupt an so for ARC we only had support of non-sampling events.
Put simply only "perf stat" was functional.
Now with ARC HS we have support of interrupts in performance counters
which this change introduces support of.
ARC performance counters act in the following way in regard of
interrupts generation.
[1] A counter counts starting from value set in PCT_COUNT register pair
[2] Once counter reaches value set in PCT_INT_CNT interrupt is raised
Basic setup look like this:
[1] PCT_COUNT = 0;
[2] PCT_INT_CNT = __limit_value__;
[3] Enable interrupts for that counter and let it run
[4] Let counter reach its limit
[5] Handle interrupt when it happens
Note that PCT HW block is build in CPU core and so ints interrupt
line (which is basically OR of all counters IRQs) is wired directly to
top-level IRQC. That means do de-assert PCT interrupt it's required to
reset IRQs from all counters that have reached their limit values.
Acked-by: Peter Zijlstra <peterz@infradead.org>
Cc: Arnaldo Carvalho de Melo <acme@kernel.org>
Signed-off-by: Alexey Brodkin <abrodkin@synopsys.com>
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
The number of counters in PCT can never be more than 32 (while
countable conditions could be 100+) for both ARCompact and ARCv2
And while at it update copyright dates.
Acked-by: Peter Zijlstra <peterz@infradead.org>
Cc: Arnaldo Carvalho de Melo <acme@kernel.org>
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
When kernel's binary becomes large enough (32M and more) errors
may occur during the final linkage stage. It happens because
the build system uses short relocations for ARC by default.
This problem may be easily resolved by passing -mlong-calls
option to GCC to use long absolute jumps (j) instead of short
relative branchs (b).
But there are fragments of pure assembler code exist which use
branchs in inappropriate places and cause a linkage error because
of relocations overflow.
First of these fragments is .fixup insertion in futex.h and
unaligned.c. It inserts a code in the separate section (.fixup)
with branch instruction. It leads to the linkage error when
kernel becomes large.
Second of these fragments is calling scheduler's functions
(common kernel code) from entry.S of ARC's code. When kernel's
binary becomes large it may lead to the linkage error because
scheduler may occur far enough from ARC's code in the final
binary.
Signed-off-by: Yuriy Kolerov <yuriy.kolerov@synopsys.com>
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
W/o hardware assisted atomic r-m-w the best we can do is to disable
preemption.
Cc: David Hildenbrand <dahi@linux.vnet.ibm.com>
Cc: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Michel Lespinasse <walken@google.com>
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
Callers of cmpxchg_futex_value_locked() in futex code expect bimodal
return value:
!0 (essentially -EFAULT as failure)
0 (success)
Before this patch, the success return value was old value of futex,
which could very well be non zero, causing caller to possibly take the
failure path erroneously.
Fix that by returning 0 for success
(This fix was done back in 2011 for all upstream arches, which ARC
obviously missed)
Cc: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Michel Lespinasse <walken@google.com>
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
The atomic ops on futex need to provide the full barrier just like
regular atomics in kernel.
Also remove pagefault_enable/disable in futex_atomic_cmpxchg_inatomic()
as core code already does that
Cc: David Hildenbrand <dahi@linux.vnet.ibm.com>
Cc: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Michel Lespinasse <walken@google.com>
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
In case of ARCv2 CPU there're could be following configurations
that affect cache handling for data exchanged with peripherals
via DMA:
[1] Only L1 cache exists
[2] Both L1 and L2 exist, but no IO coherency unit
[3] L1, L2 caches and IO coherency unit exist
Current implementation takes care of [1] and [2].
Moreover support of [2] is implemented with run-time check
for SLC existence which is not super optimal.
This patch introduces support of [3] and rework of DMA ops
usage. Instead of doing run-time check every time a particular
DMA op is executed we'll have 3 different implementations of
DMA ops and select appropriate one during init.
As for IOC support for it we need:
[a] Implement empty DMA ops because IOC takes care of cache
coherency with DMAed data
[b] Route dma_alloc_coherent() via dma_alloc_noncoherent()
This is required to make IOC work in first place and also
serves as optimization as LD/ST to coherent buffers can be
srviced from caches w/o going all the way to memory
Signed-off-by: Alexey Brodkin <abrodkin@synopsys.com>
[vgupta:
-Added some comments about IOC gains
-Marked dma ops as static,
-Massaged changelog a bit]
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
The increment of delay counter was 2 instructions:
Arithmatic Shfit Left (ASL) + set to 1 on overflow
This can be done in 1 using ROtate Left (ROL)
Suggested-by: Nigel Topham <ntopham@synopsys.com>
Cc: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
KGDB fails to build after f51e2f1911 ("ARC: make sure instruction_pointer()
returns unsigned value")
The hack to force one specific reg to unsigned backfired. There's no
reason to keep the regs signed after all.
| CC arch/arc/kernel/kgdb.o
|../arch/arc/kernel/kgdb.c: In function 'kgdb_trap':
| ../arch/arc/kernel/kgdb.c:180:29: error: lvalue required as left operand of assignment
| instruction_pointer(regs) -= BREAK_INSTR_SIZE;
Reported-by: Yuriy Kolerov <yuriy.kolerov@synopsys.com>
Fixes: f51e2f1911 ("ARC: make sure instruction_pointer() returns unsigned value")
Cc: Alexey Brodkin <abrodkin@synopsys.com>
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
The previous commit for delayed retry of SCOND needs some fine tuning
for spin locks.
The backoff from delayed retry in conjunction with spin looping of lock
itself can potentially cause the delay counter to reach high values.
So to provide fairness to any lock operation, after a lock "seems"
available (i.e. just before first SCOND try0, reset the delay counter
back to starting value of 1
Essentially reset delay to 1 for a new spin-wait-loop-acquire cycle.
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
This is to workaround the llock/scond livelock
HS38x4 could get into a LLOCK/SCOND livelock in case of multiple overlapping
coherency transactions in the SCU. The exclusive line state keeps rotating
among contenting cores leading to a never ending cycle. So break the cycle
by deferring the retry of failed exclusive access (SCOND). The actual delay
needed is function of number of contending cores as well as the unrelated
coherency traffic from other cores. To keep the code simple, start off with
small delay of 1 which would suffice most cases and in case of contention
double the delay. Eventually the delay is sufficient such that the coherency
pipeline is drained, thus a subsequent exclusive access would succeed.
Link: http://lkml.kernel.org/r/1438612568-28265-1-git-send-email-vgupta@synopsys.com
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
With LLOCK/SCOND, the rwlock counter can be atomically updated w/o need
for a guarding spin lock.
This in turn elides the EXchange instruction based spinning which causes
the cacheline transition to exclusive state and concurrent spinning
across cores would cause the line to keep bouncing around.
LLOCK/SCOND based implementation is superior as spinning on LLOCK keeps
the cacheline in shared state.
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
Current spin_lock uses EXchange instruction to implement the atomic test
and set of lock location (reads orig value and ST 1). This however forces
the cacheline into exclusive state (because of the ST) and concurrent
loops in multiple cores will bounce the line around between cores.
Instead, use LLOCK/SCOND to implement the atomic test and set which is
better as line is in shared state while lock is spinning on LLOCK
The real motivation of this change however is to make way for future
changes in atomics to implement delayed retry (with backoff).
Initial experiment with delayed retry in atomics combined with orig
EX based spinlock was a total disaster (broke even LMBench) as
struct sock has a cache line sharing an atomic_t and spinlock. The
tight spinning on lock, caused the atomic retry to keep backing off
such that it would never finish.
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
This reduces the diff in forth-coming patches and also helps understand
better the incremental changes to inline asm.
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
Extended testing of quad core configuration revealed that this fix was
insufficient. Specifically LTP open posix shm_op/23-1 would cause the
hardware livelock in llock/scond loop in update_cpu_load_active()
So remove this and make way for a proper workaround
This reverts commit a5c8b52abe.
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
With HS 2.1 release, the peripheral space register no longer contains
the uncached space specifics, causing the kernel to panic early on.
So read the newer NON VOLATILE AUX register to get that info.
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
Commit 2ae416b142 ("mm: new mm hook framework") introduced an empty
header file (mm-arch-hooks.h) for every architecture, even those which
doesn't need to define mm hooks.
As suggested by Geert Uytterhoeven, this could be cleaned through the use
of a generic header file included via each per architecture
asm/include/Kbuild file.
The PowerPC architecture is not impacted here since this architecture has
to defined the arch_remap MM hook.
Signed-off-by: Laurent Dufour <ldufour@linux.vnet.ibm.com>
Suggested-by: Geert Uytterhoeven <geert@linux-m68k.org>
Acked-by: Geert Uytterhoeven <geert@linux-m68k.org>
Acked-by: Vineet Gupta <vgupta@synopsys.com>
Cc: Oleg Nesterov <oleg@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Currently instruction_pointer() returns pt_regs->ret and so return value
is of type "long", which implicitly stands for "signed long".
While that's perfectly fine when dealing with 32-bit values if return
value of instruction_pointer() gets assigned to 64-bit variable sign
extension may happen.
And at least in one real use-case it happens already.
In perf_prepare_sample() return value of perf_instruction_pointer()
(which is an alias to instruction_pointer() in case of ARC) is assigned
to (struct perf_sample_data)->ip (which type is "u64").
And what we see if instuction pointer points to user-space application
that in case of ARC lays below 0x8000_0000 "ip" gets set properly with
leading 32 zeros. But if instruction pointer points to kernel address
space that starts from 0x8000_0000 then "ip" is set with 32 leadig
"f"-s. I.e. id instruction_pointer() returns 0x8100_0000, "ip" will be
assigned with 0xffff_ffff__8100_0000. Which is obviously wrong.
In particular that issuse broke output of perf, because perf was unable
to associate addresses like 0xffff_ffff__8100_0000 with anything from
/proc/kallsyms.
That's what we used to see:
----------->8----------
6.27% ls [unknown] [k] 0xffffffff8046c5cc
2.96% ls libuClibc-0.9.34-git.so [.] memcpy
2.25% ls libuClibc-0.9.34-git.so [.] memset
1.66% ls [unknown] [k] 0xffffffff80666536
1.54% ls libuClibc-0.9.34-git.so [.] 0x000224d6
1.18% ls libuClibc-0.9.34-git.so [.] 0x00022472
----------->8----------
With that change perf output looks much better now:
----------->8----------
8.21% ls [kernel.kallsyms] [k] memset
3.52% ls libuClibc-0.9.34-git.so [.] memcpy
2.11% ls libuClibc-0.9.34-git.so [.] malloc
1.88% ls libuClibc-0.9.34-git.so [.] memset
1.64% ls [kernel.kallsyms] [k] _raw_spin_unlock_irqrestore
1.41% ls [kernel.kallsyms] [k] __d_lookup_rcu
----------->8----------
Signed-off-by: Alexey Brodkin <abrodkin@synopsys.com>
Cc: arc-linux-dev@synopsys.com
Cc: stable@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
ARCompact/ARCv2 ISA provide that any instructions which deals with
bitpos/count operand ASL, LSL, BSET, BCLR, BMSK .... will only consider
lower 5 bits. i.e. auto-clamp the pos to 0-31.
ARC Linux bitops exploited this fact by NOT explicitly masking out upper
bits for @nr operand in general, saving a bunch of AND/BMSK instructions
in generated code around bitops.
While this micro-optimization has worked well over years it is NOT safe
as shifting a number with a value, greater than native size is
"undefined" per "C" spec.
So as it turns outm EZChip ran into this eventually, in their massive
muti-core SMP build with 64 cpus. There was a test_bit() inside a loop
from 63 to 0 and gcc was weirdly optimizing away the first iteration
(so it was really adhering to standard by implementing undefined behaviour
vs. removing all the iterations which were phony i.e. (1 << [63..32])
| for i = 63 to 0
| X = ( 1 << i )
| if X == 0
| continue
So fix the code to do the explicit masking at the expense of generating
additional instructions. Fortunately, this can be mitigated to a large
extent as gcc has SHIFT_COUNT_TRUNCATED which allows combiner to fold
masking into shift operation itself. It is currently not enabled in ARC
gcc backend, but could be done after a bit of testing.
Fixes STAR 9000866918 ("unsafe "undefined behavior" code in kernel")
Reported-by: Noam Camus <noamc@ezchip.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
Merge third patchbomb from Andrew Morton:
- the rest of MM
- scripts/gdb updates
- ipc/ updates
- lib/ updates
- MAINTAINERS updates
- various other misc things
* emailed patches from Andrew Morton <akpm@linux-foundation.org>: (67 commits)
genalloc: rename of_get_named_gen_pool() to of_gen_pool_get()
genalloc: rename dev_get_gen_pool() to gen_pool_get()
x86: opt into HAVE_COPY_THREAD_TLS, for both 32-bit and 64-bit
MAINTAINERS: add zpool
MAINTAINERS: BCACHE: Kent Overstreet has changed email address
MAINTAINERS: move Jens Osterkamp to CREDITS
MAINTAINERS: remove unused nbd.h pattern
MAINTAINERS: update brcm gpio filename pattern
MAINTAINERS: update brcm dts pattern
MAINTAINERS: update sound soc intel patterns
MAINTAINERS: remove website for paride
MAINTAINERS: update Emulex ocrdma email addresses
bcache: use kvfree() in various places
libcxgbi: use kvfree() in cxgbi_free_big_mem()
target: use kvfree() in session alloc and free
IB/ehca: use kvfree() in ipz_queue_{cd}tor()
drm/nouveau/gem: use kvfree() in u_free()
drm: use kvfree() in drm_free_large()
cxgb4: use kvfree() in t4_free_mem()
cxgb3: use kvfree() in cxgb_free_mem()
...
ARCv2 is the next generation ISA from Synopsys and basis for the
HS3{4,6,8} families of processors which retain the traditional ARC mantra of
low power and configurability and are now more performant and feature rich.
HS38x is a 10 stage pipeline core which supports MMU (with huge pages) and
SMP (upto 4 cores) among other features.
+ www.synopsys.com/dw/ipdir.php?ds=arc-hs38-processor
+ http://news.synopsys.com/2014-10-14-New-DesignWare-ARC-HS38-Processor-Doubles-Performance-for-Embedded-Linux-Applications
+ http://www.embedded.com/electronics-news/4435975/Synopsys-ARC-HS38-core-gives-2X-boost-to-Linux-based-apps
- Support for ARC SDP (Software Development platform): Main Board + CPU Cards
= AXS101: CPU Card with ARC700 in silicon @ 700 MHz
= AXS103: CPU Card with HS38x in FPGA
- Refactoring of ARCompact port to accomodate new ARCv2 ISA
- Miscll updates/cleanups
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1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=KwNo
-----END PGP SIGNATURE-----
Merge tag 'arc-4.2-rc1' of git://git.kernel.org/pub/scm/linux/kernel/git/vgupta/arc
Pull ARC architecture updates from Vineet Gupta:
- support for HS38 cores based on ARCv2 ISA
ARCv2 is the next generation ISA from Synopsys and basis for the
HS3{4,6,8} families of processors which retain the traditional ARC mantra of
low power and configurability and are now more performant and feature rich.
HS38x is a 10 stage pipeline core which supports MMU (with huge pages) and
SMP (upto 4 cores) among other features.
+ www.synopsys.com/dw/ipdir.php?ds=arc-hs38-processor
+ http://news.synopsys.com/2014-10-14-New-DesignWare-ARC-HS38-Processor-Doubles-Performance-for-Embedded-Linux-Applications
+ http://www.embedded.com/electronics-news/4435975/Synopsys-ARC-HS38-core-gives-2X-boost-to-Linux-based-apps
- support for ARC SDP (Software Development platform): Main Board + CPU Cards
= AXS101: CPU Card with ARC700 in silicon @ 700 MHz
= AXS103: CPU Card with HS38x in FPGA
- refactoring of ARCompact port to accomodate new ARCv2 ISA
- misc updates/cleanups
* tag 'arc-4.2-rc1' of git://git.kernel.org/pub/scm/linux/kernel/git/vgupta/arc: (72 commits)
ARC: Fix build failures for ARCompact in linux-next after ARCv2 support
ARCv2: Allow older gcc to cope with new regime of ARCv2/ARCompact support
ARCv2: [vdk] dts files and defconfig for HS38 VDK
ARCv2: [axs103] Support ARC SDP FPGA platform for HS38x cores
ARC: [axs101] Prepare for AXS103
ARCv2: [nsim*hs*] Support simulation platforms for HS38x cores
ARCv2: All bits in place, allow ARCv2 builds
ARCv2: SLC: Handle explcit flush for DMA ops (w/o IO-coherency)
ARCv2: STAR 9000837815 workaround hardware exclusive transactions livelock
ARC: Reduce bitops lines of code using macros
ARCv2: barriers
arch: conditionally define smp_{mb,rmb,wmb}
ARC: add smp barriers around atomics per Documentation/atomic_ops.txt
ARC: add compiler barrier to LLSC based cmpxchg
ARCv2: SMP: intc: IDU 2nd level intc for dynamic IRQ distribution
ARCv2: SMP: clocksource: Enable Global Real Time counter
ARCv2: SMP: ARConnect debug/robustness
ARCv2: SMP: Support ARConnect (MCIP) for Inter-Core-Interrupts et al
ARC: make plat_smp_ops weak to allow over-rides
ARCv2: clocksource: Introduce 64bit local RTC counter
...
This replaces the plain loop over the sglist array with for_each_sg()
macro which consists of sg_next() function calls. Since arc doesn't
select ARCH_HAS_SG_CHAIN, it is not necessary to use for_each_sg() in
order to loop over each sg element. But this can help find problems with
drivers that do not properly initialize their sg tables when
CONFIG_DEBUG_SG is enabled.
Signed-off-by: Akinobu Mita <akinobu.mita@gmail.com>
Acked-by: Vineet Gupta <vgupta@synopsys.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Pull asm/scatterlist.h removal from Jens Axboe:
"We don't have any specific arch scatterlist anymore, since parisc
finally switched over. Kill the include"
* 'for-4.2/sg' of git://git.kernel.dk/linux-block:
remove scatterlist.h generation from arch Kbuild files
remove <asm/scatterlist.h>
CRIU is recreating the process memory layout by remapping the checkpointee
memory area on top of the current process (criu). This includes remapping
the vDSO to the place it has at checkpoint time.
However some architectures like powerpc are keeping a reference to the
vDSO base address to build the signal return stack frame by calling the
vDSO sigreturn service. So once the vDSO has been moved, this reference
is no more valid and the signal frame built later are not usable.
This patch serie is introducing a new mm hook framework, and a new
arch_remap hook which is called when mremap is done and the mm lock still
hold. The next patch is adding the vDSO remap and unmap tracking to the
powerpc architecture.
This patch (of 3):
This patch introduces a new set of header file to manage mm hooks:
- per architecture empty header file (arch/x/include/asm/mm-arch-hooks.h)
- a generic header (include/linux/mm-arch-hooks.h)
The architecture which need to overwrite a hook as to redefine it in its
header file, while architecture which doesn't need have nothing to do.
The default hooks are defined in the generic header and are used in the
case the architecture is not defining it.
In a next step, mm hooks defined in include/asm-generic/mm_hooks.h should
be moved here.
Signed-off-by: Laurent Dufour <ldufour@linux.vnet.ibm.com>
Suggested-by: Andrew Morton <akpm@linux-foundation.org>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Pavel Emelyanov <xemul@parallels.com>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
L2 cache on ARCHS processors is called SLC (System Level Cache)
For working DMA (in absence of hardware assisted IO Coherency) we need
to manage SLC explicitly when buffers transition between cpu and
controllers.
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
A quad core SMP build could get into hardware livelock with concurrent
LLOCK/SCOND. Workaround that by adding a PREFETCHW which is serialized by
SCU (System Coherency Unit). It brings the cache line in Exclusive state
and makes others invalidate their lines. This gives enough time for
winner to complete the LLOCK/SCOND, before others can get the line back.
The prefetchw in the ll/sc loop is not nice but this is the only
software workaround for current version of RTL.
Cc: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Will Deacon <will.deacon@arm.com>
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
ARCv2 based HS38 cores are weakly ordered and thus explicit barriers for
kernel proper.
SMP barrier is provided by DMB instruction which also guarantees local
barrier hence used as backend of smp_*mb() as well as *mb() APIs
Also hookup barriers into MMIO accessors to avoid ordering issues in IO
Cc: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Will Deacon <will.deacon@arm.com>
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
- arch_spin_lock/unlock were lacking the ACQUIRE/RELEASE barriers
Since ARCv2 only provides load/load, store/store and all/all, we need
the full barrier
- LLOCK/SCOND based atomics, bitops, cmpxchg, which return modified
values were lacking the explicit smp barriers.
- Non LLOCK/SCOND varaints don't need the explicit barriers since that
is implicity provided by the spin locks used to implement the
critical section (the spin lock barriers in turn are also fixed in
this commit as explained above
Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Cc: stable@vger.kernel.org
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
When auditing cmpxchg call sites, Chuck noted that gcc was optimizing
away some of the desired LDs.
| do {
| new = old = *ipi_data_ptr;
| new |= 1U << msg;
| } while (cmpxchg(ipi_data_ptr, old, new) != old);
was generating to below
| 8015cef8: ld r2,[r4,0] <-- First LD
| 8015cefc: bset r1,r2,r1
|
| 8015cf00: llock r3,[r4] <-- atomic op
| 8015cf04: brne r3,r2,8015cf10
| 8015cf08: scond r1,[r4]
| 8015cf0c: bnz 8015cf00
|
| 8015cf10: brne r3,r2,8015cf00 <-- Branch doesn't go to orig LD
Although this was fixed by adding a ACCESS_ONCE in this call site, it
seems safer (for now at least) to add compiler barrier to LLSC based
cmpxchg
Reported-by: Chuck Jordan <cjordan@synopsys,com>
Cc: <stable@vger.kernel.org>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
Pull x86 core updates from Ingo Molnar:
"There were so many changes in the x86/asm, x86/apic and x86/mm topics
in this cycle that the topical separation of -tip broke down somewhat -
so the result is a more traditional architecture pull request,
collected into the 'x86/core' topic.
The topics were still maintained separately as far as possible, so
bisectability and conceptual separation should still be pretty good -
but there were a handful of merge points to avoid excessive
dependencies (and conflicts) that would have been poorly tested in the
end.
The next cycle will hopefully be much more quiet (or at least will
have fewer dependencies).
The main changes in this cycle were:
* x86/apic changes, with related IRQ core changes: (Jiang Liu, Thomas
Gleixner)
- This is the second and most intrusive part of changes to the x86
interrupt handling - full conversion to hierarchical interrupt
domains:
[IOAPIC domain] -----
|
[MSI domain] --------[Remapping domain] ----- [ Vector domain ]
| (optional) |
[HPET MSI domain] ----- |
|
[DMAR domain] -----------------------------
|
[Legacy domain] -----------------------------
This now reflects the actual hardware and allowed us to distangle
the domain specific code from the underlying parent domain, which
can be optional in the case of interrupt remapping. It's a clear
separation of functionality and removes quite some duct tape
constructs which plugged the remap code between ioapic/msi/hpet
and the vector management.
- Intel IOMMU IRQ remapping enhancements, to allow direct interrupt
injection into guests (Feng Wu)
* x86/asm changes:
- Tons of cleanups and small speedups, micro-optimizations. This
is in preparation to move a good chunk of the low level entry
code from assembly to C code (Denys Vlasenko, Andy Lutomirski,
Brian Gerst)
- Moved all system entry related code to a new home under
arch/x86/entry/ (Ingo Molnar)
- Removal of the fragile and ugly CFI dwarf debuginfo annotations.
Conversion to C will reintroduce many of them - but meanwhile
they are only getting in the way, and the upstream kernel does
not rely on them (Ingo Molnar)
- NOP handling refinements. (Borislav Petkov)
* x86/mm changes:
- Big PAT and MTRR rework: making the code more robust and
preparing to phase out exposing direct MTRR interfaces to drivers -
in favor of using PAT driven interfaces (Toshi Kani, Luis R
Rodriguez, Borislav Petkov)
- New ioremap_wt()/set_memory_wt() interfaces to support
Write-Through cached memory mappings. This is especially
important for good performance on NVDIMM hardware (Toshi Kani)
* x86/ras changes:
- Add support for deferred errors on AMD (Aravind Gopalakrishnan)
This is an important RAS feature which adds hardware support for
poisoned data. That means roughly that the hardware marks data
which it has detected as corrupted but wasn't able to correct, as
poisoned data and raises an APIC interrupt to signal that in the
form of a deferred error. It is the OS's responsibility then to
take proper recovery action and thus prolonge system lifetime as
far as possible.
- Add support for Intel "Local MCE"s: upcoming CPUs will support
CPU-local MCE interrupts, as opposed to the traditional system-
wide broadcasted MCE interrupts (Ashok Raj)
- Misc cleanups (Borislav Petkov)
* x86/platform changes:
- Intel Atom SoC updates
... and lots of other cleanups, fixlets and other changes - see the
shortlog and the Git log for details"
* 'x86-core-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip: (222 commits)
x86/hpet: Use proper hpet device number for MSI allocation
x86/hpet: Check for irq==0 when allocating hpet MSI interrupts
x86/mm/pat, drivers/infiniband/ipath: Use arch_phys_wc_add() and require PAT disabled
x86/mm/pat, drivers/media/ivtv: Use arch_phys_wc_add() and require PAT disabled
x86/platform/intel/baytrail: Add comments about why we disabled HPET on Baytrail
genirq: Prevent crash in irq_move_irq()
genirq: Enhance irq_data_to_desc() to support hierarchy irqdomain
iommu, x86: Properly handle posted interrupts for IOMMU hotplug
iommu, x86: Provide irq_remapping_cap() interface
iommu, x86: Setup Posted-Interrupts capability for Intel iommu
iommu, x86: Add cap_pi_support() to detect VT-d PI capability
iommu, x86: Avoid migrating VT-d posted interrupts
iommu, x86: Save the mode (posted or remapped) of an IRTE
iommu, x86: Implement irq_set_vcpu_affinity for intel_ir_chip
iommu: dmar: Provide helper to copy shared irte fields
iommu: dmar: Extend struct irte for VT-d Posted-Interrupts
iommu: Add new member capability to struct irq_remap_ops
x86/asm/entry/64: Disentangle error_entry/exit gsbase/ebx/usermode code
x86/asm/entry/32: Shorten __audit_syscall_entry() args preparation
x86/asm/entry/32: Explain reloading of registers after __audit_syscall_entry()
...
Caveats about cache flush on ARCv2 based cores
- dcache is PIPT so paddr is sufficient for cache maintenance ops (no
need to setup PTAG reg
- icache is still VIPT but only aliasing configs need PTAG setup
So basically this is departure from MMU-v3 which always need vaddr in
line ops registers (DC_IVDL, DC_FLDL, IC_IVIL) but paddr in DC_PTAG,
IC_PTAG respectively.
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
The issue was, on HS when interrupt is taken, IRQ_ACT is set and that is
NOT cleared unless we do RTIE (or manually clear it). Linux interrupt
handling has top and bottom halves. Latter lead to softirqs (which can
reschedule) AND expect interrupts to be REALLY re-enabled which was NOT
happening for us since we only SETI, dont clear IRQ_ACT
So we can have a state when both cores have taken interrupt (IRQ_ACT set),
get rescheduled, both send IPI and wait in CSD lock which will never be
cleared as cores can't take the pending IPI IRQ due to existing IRQ_ACT
set.
So local_irq_enable() now drops the IRQ_ACT.act bit to re-enable IRQs.
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
Reported by Anton as LTP:munmap01 failing with Illegal Instruction
Exception.
--------------------->8--------------------------------------
mmap2(NULL, 24576, PROT_READ|PROT_WRITE, MAP_SHARED, 3, 0) = 0x200d2000
munmap(0x200d2000, 24576) = 0
--- SIGSEGV {si_signo=SIGSEGV, si_code=SEGV_MAPERR, si_addr=0x200d2000}
---
potentially unexpected fatal signal 4.
Path: /munmap01
CPU: 0 PID: 61 Comm: munmap01 Not tainted 3.13.0-g5d5c46d9a556 #8
task: 9f1a8000 ti: 9f154000 task.ti: 9f154000
[ECR ]: 0x00020100 => Illegal Insn
[EFA ]: 0x0001354c
[BLINK ]: 0x200515d4
[ERET ]: 0x1354c
@off 0x1354c in [/munmap01]
VMA: 0x00010000 to 0x00018000
[STAT32]: 0x800802c0
...
--------------------->8--------------------------------------
The issue was
1. munmap01 accessed unmapped memory (on purpose) with signal handler
installed for SIGSEGV
2. The faulting instruction happened to be in Delay Slot
00011864 <main>:
11908: bl.d 13284 <tst_resm>
1190c: stb r16,[r2]
3. kernel sets up the reg file for signal handler and correctly clears
the DE bit in pt_regs->status32 placeholder
4. However RESTORE_CALLEE_SAVED_USER macro is not adjusted for ARCv2,
and it over-writes the above with orig/stale value of status32
5. After RTIE, userspace signal handler executes a non branch
instruction with DE bit set, triggering Illegal Instruction Exception.
Reported-by: Anton Kolesov <akolesov@synopsys.com>
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
Previously this macro was overloaded with stack switching, saving SP at right
slot in pt_regs, saving/setup of r25 and setting SP baseline to where
pt_regs->sp is saved (vs. bottom of pt_regs)
Now it only does SP switch, and leaves SP pointing to bottom of pt_regs.
r25 saving is no longer done here to allow for future reordering of
regfile in pt_regs w/o touching this macro
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
Elide the need to re-read ECR in Trap handler by ensuring that
EXCEPTION_PROLOGUE does that at the very end just before returning
to Trap handler
ARCv2 EXCEPTION_PROLOGUE already did that, so same for ARcompact and the
common trap handler adjusted to use cached ECR
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
Alexey Brodkin
Vineet Gupta
Yuriy Kolerov
Laurent Dufour
Linus Torvalds
Akinobu Mita