The Intel Architecture Optimization Reference Manual states that a short
load that follows a long store to the same object will suffer a store
forwading penalty, particularly if the two accesses use different addresses.
Trivially, a long load that follows a short store will also suffer a penalty.
__downgrade_write() in rwsem incurs both penalties: the increment operation
will not be able to reuse a recently-loaded rwsem value, and its result will
not be reused by any recently-following rwsem operation.
A comment in the code states that this is because 64-bit immediates are
special and expensive; but while they are slightly special (only a single
instruction allows them), they aren't expensive: a test shows that two loops,
one loading a 32-bit immediate and one loading a 64-bit immediate, both take
1.5 cycles per iteration.
Fix this by changing __downgrade_write to use the same add instruction on
i386 and on x86_64, so that it uses the same operand size as all the other
rwsem functions.
Signed-off-by: Avi Kivity <avi@redhat.com>
LKML-Reference: <1266049992-17419-1-git-send-email-avi@redhat.com>
Signed-off-by: H. Peter Anvin <hpa@zytor.com>
For x86-64, 32767 threads really is not enough. Change rwsem_count_t
to a signed long, so that it is 64 bits on x86-64.
This required the following changes to the assembly code:
a) %z0 doesn't work on all versions of gcc! At least gcc 4.4.2 as
shipped with Fedora 12 emits "ll" not "q" for 64 bits, even for
integer operands. Newer gccs apparently do this correctly, but
avoid this problem by using the _ASM_ macros instead of %z.
b) 64 bits immediates are only allowed in "movq $imm,%reg"
constructs... no others. Change some of the constraints to "e",
and fix the one case where we would have had to use an invalid
immediate -- in that case, we only care about the upper half
anyway, so just access the upper half.
Signed-off-by: H. Peter Anvin <hpa@zytor.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
LKML-Reference: <tip-bafaecd11df15ad5b1e598adc7736afcd38ee13d@git.kernel.org>
The fast version of the rwsems (the code that uses xadd) has
traditionally only worked on x86-32, and as a result it mixes different
kinds of types wildly - they just all happen to be 32-bit. We have
"long", we have "__s32", and we have "int".
To make it work on x86-64, the types suddenly matter a lot more. It can
be either a 32-bit or 64-bit signed type, and both work (with the caveat
that a 32-bit counter will only have 15 bits of effective write
counters, so it's limited to 32767 users). But whatever type you
choose, it needs to be used consistently.
This makes a new 'rwsem_counter_t', that is a 32-bit signed type. For a
64-bit type, you'd need to also update the BIAS values.
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
LKML-Reference: <alpine.LFD.2.00.1001121755220.17145@localhost.localdomain>
Signed-off-by: H. Peter Anvin <hpa@zytor.com>
This makes gcc use the right register names and instruction operand sizes
automatically for the rwsem inline asm statements.
So instead of using "(%%eax)" to specify the memory address that is the
semaphore, we use "(%1)" or similar. And instead of forcing the operation
to always be 32-bit, we use "%z0", taking the size from the actual
semaphore data structure itself.
This doesn't actually matter on x86-32, but if we want to use the same
inline asm for x86-64, we'll need to have the compiler generate the proper
64-bit names for the registers (%rax instead of %eax), and if we want to
use a 64-bit counter too (in order to avoid the 15-bit limit on the
write counter that limits concurrent users to 32767 threads), we'll need
to be able to generate instructions with "q" accesses rather than "l".
Since this header currently isn't enabled on x86-64, none of that matters,
but we do want to use the xadd version of the semaphores rather than have
to take spinlocks to do a rwsem. The mm->mmap_sem can be heavily contended
when you have lots of threads all taking page faults, and the fallback
rwsem code that uses a spinlock performs abysmally badly in that case.
[ hpa: modified the patch to skip size suffixes entirely when they are
redundant due to register operands. ]
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
LKML-Reference: <alpine.LFD.2.00.1001121613560.17145@localhost.localdomain>
Signed-off-by: H. Peter Anvin <hpa@zytor.com>
Change header guards named "ASM_X86__*" to "_ASM_X86_*" since:
a. the double underscore is ugly and pointless.
b. no leading underscore violates namespace constraints.
Signed-off-by: H. Peter Anvin <hpa@zytor.com>
Avi Kivity