commit fce16bc35a upstream.
In the exception return path, for both U/K cases, intr are already
disabled (for various existing reasons). So when we drop down to
@restore_regs, we need not redo that.
There was subtle issue - when intr were NOT being disabled for
ret-to-kernel-but-no-preemption case - now fixed by moving the
IRQ_DISABLE further up in @resume_kernel_mode.
So what do we gain:
* Shaves off a few insn in return path.
* Eliminates the need for IRQ_DISABLE_SAVE assembler macro for ARCv2
hence allows for entry code sharing.
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit 6c00350b57 upstream.
Some ARC SMP systems lack native atomic R-M-W (LLOCK/SCOND) insns and
can only use atomic EX insn (reg with mem) to build higher level R-M-W
primitives. This includes a SystemC based SMP simulation model.
So rwlocks need to use a protecting spinlock for atomic cmp-n-exchange
operation to update reader(s)/writer count.
The spinlock operation itself looks as follows:
mov reg, 1 ; 1=locked, 0=unlocked
retry:
EX reg, [lock] ; load existing, store 1, atomically
BREQ reg, 1, rety ; if already locked, retry
In single-threaded simulation, SystemC alternates between the 2 cores
with "N" insn each based scheduling. Additionally for insn with global
side effect, such as EX writing to shared mem, a core switch is
enforced too.
Given that, 2 cores doing a repeated EX on same location, Linux often
got into a livelock e.g. when both cores were fiddling with tasklist
lock (gdbserver / hackbench) for read/write respectively as the
sequence diagram below shows:
core1 core2
-------- --------
1. spin lock [EX r=0, w=1] - LOCKED
2. rwlock(Read) - LOCKED
3. spin unlock [ST 0] - UNLOCKED
spin lock [EX r=0,w=1] - LOCKED
-- resched core 1----
5. spin lock [EX r=1] - ALREADY-LOCKED
-- resched core 2----
6. rwlock(Write) - READER-LOCKED
7. spin unlock [ST 0]
8. rwlock failed, retry again
9. spin lock [EX r=0, w=1]
-- resched core 1----
10 spinlock locked in #9, retry #5
11. spin lock [EX gets 1]
-- resched core 2----
...
...
The fix was to unlock using the EX insn too (step 7), to trigger another
SystemC scheduling pass which would let core1 proceed, eliding the
livelock.
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit 0752adfda1 upstream.
Anton reported
| LTP tests syscalls/process_vm_readv01 and process_vm_writev01 fail
| similarly in one testcase test_iov_invalid -> lvec->iov_base.
| Testcase expects errno EFAULT and return code -1,
| but it gets return code 1 and ERRNO is 0 what means success.
Essentially test case was passing a pointer of -1 which access_ok()
was not catching. It was doing [@addr + @sz <= TASK_SIZE] which would
pass for @addr == -1
Fixed that by rewriting as [@addr <= TASK_SIZE - @sz]
Reported-by: Anton Kolesov <Anton.Kolesov@synopsys.com>
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit 7efd0da2d1 upstream.
Cast usecs to u64, to ensure that the (usecs * 4295 * HZ)
multiplication is 64 bit.
Initially, the (usecs * 4295 * HZ) part was done as a 32 bit
multiplication, with the result casted to 64 bit. This led to some bits
falling off, causing a "DMA initialization error" in the stmmac Ethernet
driver, due to a premature timeout.
Signed-off-by: Mischa Jonker <mjonker@synopsys.com>
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit c3567f8a35 upstream.
Commit 05b016ecf5 "ARC: Setup Vector Table Base in early boot" moved
the Interrupt vector Table setup out of arc_init_IRQ() which is called
for all CPUs, to entry point of boot cpu only, breaking booting of others.
Fix by adding the same to entry point of non-boot CPUs too.
read_arc_build_cfg_regs() printing IVT Base Register didn't help the
casue since it prints a synthetic value if zero which is totally bogus,
so fix that to print the exact Register.
[vgupta: Remove the now stale comment from header of arc_init_IRQ and
also added the commentary for halt-on-reset]
Cc: Gilad Ben-Yossef <gilad@benyossef.com>
Signed-off-by: Noam Camus <noamc@ezchip.com>
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
[Based on mainline commit 352c1d95e3: "ARC: stop using
pt_regs->orig_r8"]
Stop using orig_r8 as it could get clobbered by ST in trap_with_param,
and further it is semantically not needed either.
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
[Based on mainline commit 502a0c775c: "ARC: pt_regs update #5"]
gdbserver needs @stop_pc, served by ptrace, but fetched from pt_regs
differently, based on in_brkpt_traps(), which in turn relies on
additional machine state in pt_regs->event bitfield.
unsigned long orig_r8:16, event:16;
For big endian config, this macro was returning false, despite being in
breakpoint Trap exception, causing wrong @stop_pc to be returned to gdb.
Issue #1: In BE, @event above is at offset 2 in word, while a STW insn
at offset 0 was used to update it. Resort to using ST insn
which updates the half-word at right location.
Issue #2: The union involving bitfields causes all the members to be
laid out at offset 0. So with fix#1 above, ASM was now
updating at offset 2, "C" code was still referencing at
offset 0. Fixed by wrapping bitfield in a struct.
Reported-by: Noam Camus <noamc@ezchip.com>
Tested-by: Anton Kolesov <akolesov@synopsys.com>
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
gdbserver inserting a breakpoint ends up calling copy_user_page() for a
code page. The generic version of which (non-aliasing config) didn't set
the PG_arch_1 bit hence update_mmu_cache() didn't sync dcache/icache for
corresponding dynamic loader code page - causing garbade to be executed.
So now aliasing versions of copy_user_highpage()/clear_page() are made
default. There is no significant overhead since all of special alias
handling code is compiled out for non-aliasing build
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
The current code uses 2 bits for determining page's dcache color, thus
sorting pages into 4 bins, whereas the aliasing dcache really has 2 bins
(8k page, 64k dcache - 4 way-set-assoc).
This can cause extraneous flushes - e.g. color 0 and 2.
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
The VM_EXEC check in update_mmu_cache() was getting optimized away
because of a stupid error in definition of macro addr_not_cache_congruent()
The intention was to have the equivalent of following:
if (a || (1 ? b : 0))
but we ended up with following:
if (a || 1 ? b : 0)
And because precedence of '||' is more that that of '?', gcc was optimizing
away evaluation of <a>
Nasty Repercussions:
1. For non-aliasing configs it would mean some extraneous dcache flushes
for non-code pages if U/K mappings were not congruent.
2. For aliasing config, some needed dcache flush for code pages might
be missed if U/K mappings were congruent.
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
This manifested as grep failing psuedo-randomly:
-------------->8---------------------
[ARCLinux]$ ip address show lo | grep inet
[ARCLinux]$ ip address show lo | grep inet
[ARCLinux]$ ip address show lo | grep inet
[ARCLinux]$
[ARCLinux]$ ip address show lo | grep inet
inet 127.0.0.1/8 scope host lo
-------------->8---------------------
ARC700 MMU provides fully orthogonal permission bits per page:
Ur, Uw, Ux, Kr, Kw, Kx
The user mode page permission templates used to have all Kernel mode
access bits enabled.
This caused a tricky race condition observed with uClibc buffered file
read and UNIX pipes.
1. Read access to an anon mapped page in libc .bss: write-protected
zero_page mapped: TLB Entry installed with Ur + K[rwx]
2. grep calls libc:getc() -> buffered read layer calls read(2) with the
internal read buffer in same .bss page.
The read() call is on STDIN which has been redirected to a pipe.
read(2) => sys_read() => pipe_read() => copy_to_user()
3. Since page has Kernel-write permission (despite being user-mode
write-protected), copy_to_user() suceeds w/o taking a MMU TLB-Miss
Exception (page-fault for ARC). core-MM is unaware that kernel
erroneously wrote to the reserved read-only zero-page (BUG #1)
4. Control returns to userspace which now does a write to same .bss page
Since Linux MM is not aware that page has been modified by kernel, it
simply reassigns a new writable zero-init page to mapping, loosing the
prior write by kernel - effectively zero'ing out the libc read buffer
under the hood - hence grep doesn't see right data (BUG #2)
The fix is to make all kernel-mode access permissions mirror the
user-mode ones. Note that the kernel still has full access to pages,
when accessed directly (w/o MMU) - this fix ensures that kernel-mode
access in copy_to_from() path uses the same faulting access model as for
pure user accesses to keep MM fully aware of page state.
The issue is peudo-random because it only shows up if the TLB entry
installed in #1 is present at the time of #3. If it is evicted out, due
to TLB pressure or some-such, then copy_to_user() does take a TLB Miss
Exception, with a routine write-to-anon COW processing installing a
fresh page for kernel writes and also usable as it is in userspace.
Further the issue was dormant for so long as it depends on where the
libc internal read buffer (in .bss) is mapped at runtime.
If it happens to reside in file-backed data mapping of libc (in the
page-aligned slack space trailing the file backed data), loader zero
padding the slack space, does the early cow page replacement, setting
things up at the very beginning itself.
With gcc 4.8 based builds, the libc buffer got pushed out to a real
anon mapping which triggers the issue.
Reported-by: Anton Kolesov <akolesov@synopsys.com>
Cc: <stable@vger.kernel.org> # 3.9
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
This is the meat of the series which prevents any dcache alias creation
by always keeping the U and K mapping of a page congruent.
If a mapping already exists, and other tries to access the page, prev
one is flushed to physical page (wback+inv)
Essentially flush_dcache_page()/copy_user_highpage() create K-mapping
of a page, but try to defer flushing, unless U-mapping exist.
When page is actually mapped to userspace, update_mmu_cache() flushes
the K-mapping (in certain cases this can be optimised out)
Additonally flush_cache_mm(), flush_cache_range(), flush_cache_page()
handle the puring of stale userspace mappings on exit/munmap...
flush_anon_page() handles the existing U-mapping for anon page before
kernel reads it via the GUP path.
Note that while not complete, this is enough to boot a simple
dynamically linked Busybox based rootfs
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
This preps the low level dcache flush helpers to take vaddr argument in
addition to the existing paddr to properly flush the VIPT dcache
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
flush_dcache_page( ) is MM hook to ensure that a page has consistent
views between kernel and userspace. Thus it is called when
* kernel writes to a page which at some later point could get mapped to
userspace (so kernel mapping needs to be flushed-n-inv)
* kernel is about to read from a page with possible userspace mappings
(so userspace mappings needs to be made coherent with kernel ones)
However for Non aliasing VIPT dcache, any userspace mapping will always
be congruent to kernel mapping. Thus d-cache need need not be flushed at
all (or delayed indefinitely).
The only reason it does need to be flushed is when mapping code pages.
Since icache doesn't snoop dcache, those dirty dcache lines need to be
written back to memory and icache line invalidated so that icache lines
fetch will get the right data.
Decent gains on LMBench fork/exec/sh and File I/O micro-benchmarks.
(1) FPGA @ 80 MHZ
Processor, Processes - times in microseconds - smaller is better
------------------------------------------------------------------------------
Host OS Mhz null null open slct sig sig fork exec sh
call I/O stat clos TCP inst hndl proc proc proc
--------- ------------- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ----
3.9-rc6-a Linux 3.9.0-r 80 4.79 8.72 66.7 116. 239. 8.39 30.4 4798 14.K 34.K
3.9-rc6-b Linux 3.9.0-r 80 4.79 8.62 65.4 111. 239. 8.35 29.0 3995 12.K 30.K
3.9-rc7-c Linux 3.9.0-r 80 4.79 9.00 66.1 106. 239. 8.61 30.4 2858 10.K 24.K
^^^^ ^^^^ ^^^
File & VM system latencies in microseconds - smaller is better
-------------------------------------------------------------------------------
Host OS 0K File 10K File Mmap Prot Page 100fd
Create Delete Create Delete Latency Fault Fault selct
--------- ------------- ------ ------ ------ ------ ------- ----- ------- -----
3.9-rc6-a Linux 3.9.0-r 317.8 204.2 1122.3 375.1 3522.0 4.288 20.7 126.8
3.9-rc6-b Linux 3.9.0-r 298.7 223.0 1141.6 367.8 3531.0 4.866 20.9 126.4
3.9-rc7-c Linux 3.9.0-r 278.4 179.2 862.1 339.3 3705.0 3.223 20.3 126.6
^^^^^ ^^^^^ ^^^^^ ^^^^
(2) Customer Silicon @ 500 MHz (166 MHz mem)
------------------------------------------------------------------------------
Host OS Mhz null null open slct sig sig fork exec sh
call I/O stat clos TCP inst hndl proc proc proc
--------- ------------- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ----
abilis-ba Linux 3.9.0-r 497 0.71 1.38 4.58 12.0 35.5 1.40 3.89 2070 5525 13.K
abilis-ca Linux 3.9.0-r 497 0.71 1.40 4.61 11.8 35.6 1.37 3.92 1411 4317 10.K
^^^^ ^^^^ ^^^
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
Now that we have same helper used for all icache invalidates (i.e.
vaddr+paddr based exact line invalidate), consolidate the open coded
calls into one place.
Also rename flush_icache_range_vaddr => __sync_icache_dcache
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
ARC icache doesn't snoop dcache thus executable pages need to be made
coherent before mapping into userspace in flush_icache_page().
However ARC700 CDU (hardware cache flush module) requires both vaddr
(index in cache) as well as paddr (tag match) to correctly identify a
line in the VIPT cache. A typical ARC700 SoC has aliasing icache, thus
the paddr only based flush_icache_page() API couldn't be implemented
efficiently. It had to loop thru all possible alias indexes and perform
the invalidate operation (ofcourse the cache op would only succeed at
the index(es) where tag matches - typically only 1, but the cost of
visiting all the cache-bins needs to paid nevertheless).
Turns out however that the vaddr (along with paddr) is available in
update_mmu_cache() hence better suits ARC icache flush semantics.
With both vaddr+paddr, exactly one flush operation per line is done.
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
munmap ends up calling tlb_flush() which for ARC was flushing the entire
TLB unconditionally (by moving the MMU to a new ASID)
do_munmap
unmap_region
unmap_vmas
unmap_single_vma
unmap_page_range
tlb_start_vma
zap_pud_range
tlb_end_vma()
tlb_finish_mmu
tlb_flush() ---> unconditional flush_tlb_mm()
So even a single page munmap, a frequent operation when uClibc dynamic
linker (ldso) is loading the dependent shared libraries, would move the
the ASID multiple times - needlessly invalidating the pre-faulted TLB
entries (and increasing the rate of ASID wraparound + full TLB flush).
This is now optimised to only be called if tlb->full_mm (which means
for exit/execve) cases only. And for those cases, flush_tlb_mm() is
already optimised to be a no-op for mm->mm_users == 0.
So essentially there are no mmore full mm flushes - except for fork which
anyhow needs it for properly COW'ing parent address space.
munmap now needs to do TLB range flush, which is implemented with
tlb_end_vma()
Results
-------
1. ASID now consistenly moves by 4 during a simple ls (as opposed to 5 or
7 before).
2. LMBench microbenchmark also shows improvements
Basic system parameters
------------------------------------------------------------------------------
Host OS Description Mhz tlb cache mem scal
pages line par load
bytes
--------- ------------- ----------------------- ---- ----- ----- ------ ----
3.9-rc5-0 Linux 3.9.0-r 3.9-rc5-0404-gcc-4.4-ba 80 8 64 1.1000 1
3.9-rc5-0 Linux 3.9.0-r 3.9-rc5-0405-avoid-full 80 8 64 1.1200 1
Processor, Processes - times in microseconds - smaller is better
------------------------------------------------------------------------------
Host OS Mhz null null open slct sig sig fork exec sh
call I/O stat clos TCP inst hndl proc proc proc
--------- ------------- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ----
3.9-rc5-0 Linux 3.9.0-r 80 4.81 8.69 68.6 118. 239. 8.53 31.6 4839 13.K 34.K
3.9-rc5-0 Linux 3.9.0-r 80 4.46 8.36 53.8 91.3 223. 8.12 24.2 4725 13.K 33.K
File & VM system latencies in microseconds - smaller is better
-------------------------------------------------------------------------------
Host OS 0K File 10K File Mmap Prot Page 100fd
Create Delete Create Delete Latency Fault Fault selct
--------- ------------- ------ ------ ------ ------ ------- ----- ------- -----
3.9-rc5-0 Linux 3.9.0-r 314.7 223.2 1054.9 390.2 3615.0 1.590 20.1 126.6
3.9-rc5-0 Linux 3.9.0-r 265.8 183.8 1014.2 314.1 3193.0 6.910 18.8 110.4
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
Infrastructure required to make the Linux kernel compile and boot on the
Abilis Systems TB10x series of SOCs based on ARC700 CPUs:
- Kmake related files (Kconfig, Makefile, tb10x_defconfig)
- TB10x platform initialisation
Signed-off-by: Christian Ruppert <christian.ruppert@abilis.com>
Signed-off-by: Pierrick Hascoet <pierrick.hascoet@abilis.com>
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
This patch adds some room for CPU-external interrupt controllers in the
Linux interrupt space. Until now, only the 32 CPU internal interrupt lines
were supported which does not allow for external interrupt controllers such
as GPIO modules etc.
Signed-off-by: Christian Ruppert <christian.ruppert@abilis.com>
Signed-off-by: Pierrick Hascoet <pierrick.hascoet@abilis.com>
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
ARC irqsave/restore macros were missing the compiler barrier, causing a
stale load in irq-enabled region be used in irq-safe region, despite
being changed, because the register holding the value was still live.
The problem manifested as random crashes in timer code when stress
testing ARCLinux (3.9-rc3) on a !SMP && !PREEMPT_COUNT
Here's the exact sequence which caused this:
(0). tv1[x] <----> t1 <---> t2
(1). mod_timer(t1) interrupted after it calls timer_pending()
(2). mod_timer(t2) completes
(3). mod_timer(t1) resumes but messes up the list
(4). __runt_timers( ) uses bogus timer_list entry / crashes in
timer->function
Essentially mod_timer() was racing against itself and while the spinlock
serialized the tv1[] timer link list, timer_pending() called outside the
spinlock, cached timer link list element in a register.
With low register pressure (and a deep register file), lack of barrier
in raw_local_irqsave() as well as preempt_disable (!PREEMPT_COUNT
version), there was nothing to force gcc to reload across the spinlock,
causing a stale value in reg be used for link list manipulation - ensuing
a corruption.
ARcompact disassembly which shows the culprit generated code:
mod_timer:
push_s blink
mov_s r13,r0 # timer, timer
..
###### timer_pending( )
ld_s r3,[r13] # <------ <variable>.entry.next LOADED
brne r3, 0, @.L163
.L163:
..
###### spin_lock_irq( )
lr r5, [status32] # flags
bic r4, r5, 6 # temp, flags,
and.f 0, r5, 6 # flags,
flag.nz r4
###### detach_if_pending( ) begins
tst_s r3,r3 <--------------
# timer_pending( ) checks timer->entry.next
# r3 is NOT reloaded by gcc, using stale value
beq.d @.L169
mov.eq r0,0
##### detach_timer( ): __list_del( )
ld r4,[r13,4] # <variable>.entry.prev, D.31439
st r4,[r3,4] # <variable>.prev, D.31439
st r3,[r4] # <variable>.next, D.30246
We initially tried to fix this by adding barrier() to preempt_* macros
for !PREEMPT_COUNT but Linus clarified that it was anything but wrong.
http://www.spinics.net/lists/kernel/msg1512709.html
[vgupta: updated commitlog]
Reported-by/Signed-off-by: Christian Ruppert <christian.ruppert@abilis.com>
Cc: Christian Ruppert <christian.ruppert@abilis.com>
Cc: Pierrick Hascoet <pierrick.hascoet@abilis.com>
Debugged-by/Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
orig_r8_IS_EXCPN and orig_r8_IS_BRKPT were same values due to a
copy/paste error. Although it looks bad and is wrong, it really doesn't
affect gdb working.
orig_r8_IS_BRKPT is the one relevant to debugging (breakpoints), since
it is used to provide EFA vs. ERET to a ptrace "stop_pc" request.
So when gdb has inserted a breakpoint, orig_r8_IS_BRKPT is already set,
and anything else (i.e. orig_r8_IS_EXCPN) becoming same as it, really
doesn't hurt gdb. The corollary case, could be nasty but nobody uses the
ptrace "stop_pc" request in that case
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
