[perfmon2] exposing syscall base number in /sys

[stephane e. <er...@go...>]

Hello,

Last week, I pushed a patch to expose the perfmon base syscall number
in /sys/kernel/perfmon/syscall.
This is a temporary measure to work around the issue that with each
new kernel release the perfmon
syscall numbers change because new syscalls are added. Libpfm (CVS)
has been modified to look in this file
if it exists, otherwise it defaults to hardcoded constants for each
kernel version (starting with 2.6.24).

Phil Mucci pointed out that 64-bit kernels often also support running
32-bit native binaries, i.e., a 32-bit ABI.
On those kernels, the perfmon interface is also supported for 32-bit
binaries. This is the case on X86, where
an i386 binary using perfmon can run unmodified. Oftentimes, though,
the syscall numbers for 32 vs. 64 bit
ABI are different, this is true on X86 for instance. Thus, there is a
problem for exporting the number via a single
entry in /sys.

I have now just pushed a patch which makes the content of
/sys/kernel/perfmon/syscall dependent on the ABI
of the process which reads the file. Thus, with a single entry in /sys
we can service any kind of binary.

On SPARC and PPC, it seems the same syscall numbers are used, so there
should not be any difficulty.
On IA-64, there is no native 32-bit ABI, so the problem does not exist.

On X86, the kernel detects the ABI of the process (via existing TIF
flags) and returns the corresponding syscall
base number.

The only architecture for which we do not have yet the full solution
is MIPS where they support 3 ABIs.
Of those, two 32-bit ABI which are not easy to distinguish. I am
hopeful we will have a solution in the next
few days.

Nevertheless, I have decided to release this patch so that you could
start testing it on your architecture and report
any problems. Note that you will need to use the libpfm from CVS.

Thanks.