linux-toradex.git/arch/s390/include/asm/cmpxchg.h, branch v4.4.97

s390/cmpxchg: remove dead code

2015-10-14T12:32:15+00:00

With the removal of 31 bit support a couple of defines became unused.
Remove them.

Signed-off-by: Heiko Carstens 
Signed-off-by: Martin Schwidefsky

arch: Remove __ARCH_HAVE_CMPXCHG

2015-05-13T08:55:42+00:00

We removed the only user of this define in the rtmutex code. Get rid
of it.

Signed-off-by: Thomas Gleixner 
Cc: Sebastian Andrzej Siewior

s390/cmpxchg: simplify cmpxchg_double

2015-03-25T10:49:36+00:00

Since sizeof(long) == 4 is always false now, simplify cmpxchg_double a bit.

Signed-off-by: Heiko Carstens 
Signed-off-by: Martin Schwidefsky

s390/cmpxchg: use compiler builtins

2014-11-03T12:29:47+00:00

The kernel build for s390 fails for gcc compilers with version 3.x,
set the minimum required version of gcc to version 4.3.

As the atomic builtins are available with all gcc 4.x compilers,
use the __sync_val_compare_and_swap and __sync_bool_compare_and_swap
functions to replace the complex macro and inline assembler magic
in include/asm/cmpxchg.h. The compiler can just-do-it and generates
better code with the builtins.

While we are at it use __sync_bool_compare_and_swap for the
_raw_compare_and_swap function in the spinlock code as well.

Signed-off-by: Martin Schwidefsky

s390/32bit: fix cmpxchg64

2014-01-22T13:02:15+00:00

Fix broken inline assembly contraints for cmpxchg64 on 32bit.

Fixes this crash:

specification exception: 0006 [#1] SMP
CPU: 0 PID: 0 Comm: swapper/0 Not tainted 3.13.0 #4
task: 005a16c8 ti: 00592000 task.ti: 00592000
Krnl PSW : 070ce000 8029abd6 (lockref_get+0x3e/0x9c)
...
Krnl Code: 8029abcc: a71a0001           ahi     %r1,1
           8029abd0: 1852               lr      %r5,%r2
          #8029abd2: bb40f064           cds     %r4,%r0,100(%r15)
          >8029abd6: 1943               cr      %r4,%r3
           8029abd8: 1815               lr      %r1,%r5
Call Trace:
([<0000000078e01870>] 0x78e01870)
 [<000000000021105a>] sysfs_mount+0xd2/0x1c8
 [<00000000001b551e>] mount_fs+0x3a/0x134
 [<00000000001ce768>] vfs_kern_mount+0x44/0x11c
 [<00000000001ce864>] kern_mount_data+0x24/0x3c
 [<00000000005cc4b8>] sysfs_init+0x74/0xd4
 [<00000000005cb5b4>] mnt_init+0xe0/0x1fc
 [<00000000005cb16a>] vfs_caches_init+0xb6/0x14c
 [<00000000005be794>] start_kernel+0x318/0x33c
 [<000000000010001c>] _stext+0x1c/0x80

Reported-by: Mike Frysinger 
Signed-off-by: Heiko Carstens 
Signed-off-by: Martin Schwidefsky

s390/cmpxchg,percpu: implement cmpxchg_double()

2012-09-26T13:45:25+00:00

Signed-off-by: Heiko Carstens 
Signed-off-by: Martin Schwidefsky

s390/cmpxchg: fix sign extension bugs

2012-05-30T07:07:58+00:00

For 1 and 2 byte operands for xchg and cmpxchg the old and new values
get or'ed into the larger 4 byte old value before the compare and swap
instruction gets executed. This is done without using the proper byte
mask before or'ing the values.
If the caller passed in negative old or new values these got sign
extended by the caller. Which in turn means that either the old value
never matches, or, even worse, unrelated bytes would be changed in memory.

Luckily there don't seem to be any callers around yet, since that would
have resulted in the specification exception fixed in an earlies patch.

Signed-off-by: Heiko Carstens 
Signed-off-by: Martin Schwidefsky

s390/cmpxchg: fix 1 and 2 byte memory accesses

2012-05-30T07:07:57+00:00

When accessing a 1 or 2 byte memory operand we cannot use the
passed address since the compare and swap instruction only works
for 4 byte aligned memory operands.
Hence we calculate an aligned address so that compare and swap works
correctly. However we don't pass the calculated address to the inline
assembly. This results in incorrect memory accesses and in a
specification exception if used on non 4 byte aligned memory operands.

Since this didn't happen until now, there don't seem to be
too many users of cmpxchg on unaligned addresses.

Signed-off-by: Heiko Carstens 
Signed-off-by: Martin Schwidefsky

s390/cmpxchg: fix compile warnings specific to s390

2012-05-30T07:07:56+00:00

The cmpxchg macros and functions are a bit different than on other
architectures. In particular the macros do not store the return
value of a __cmpxchg function call in a variable before returning the
value.

This causes compile warnings that only occur on s390 like this one:

net/ipv4/af_inet.c: In function 'build_ehash_secret':
net/ipv4/af_inet.c:241:2: warning: value computed is not used [-Wunused-value]

To get rid of these warnings use the same construct that we already use
for the xchg macro, which was introduced for the same reason.

Signed-off-by: Heiko Carstens 
Signed-off-by: Martin Schwidefsky

s390/cmpxchg: add missing memory barrier to cmpxchg64

2012-05-30T07:07:55+00:00

All cmpxchg functions imply a memory barrier.
cmpxch64 did not have one for 31 bit code, so add it.

Signed-off-by: Heiko Carstens 
Signed-off-by: Martin Schwidefsky