1
linux/arch/mips/math-emu/dsemul.c
David Daney b6ee75ed4f MIPS: Collect FPU emulator statistics per-CPU.
On SMP systems, the collection of statistics can cause cache line
bouncing in the lines associated with the counters.  Also there are
races incrementing the counters on multiple CPUs.

To fix both problems, we collect the statistics in per-CPU variables,
and add them up in the debugfs read operation.

As a test I ran the LTP float_bessel test on a 12 CPU Octeon system.

Without CONFIG_DEBUG_FS :             2602 seconds.
With CONFIG_DEBUG_FS:                 2640 seconds.
With non-cpu-local atomic statistics: 14569 seconds.

Signed-off-by: David Daney <ddaney@caviumnetworks.com>
Cc: linux-mips@linux-mips.org
Signed-off-by: Ralf Baechle <ralf@linux-mips.org>
2009-12-17 01:57:08 +00:00

168 lines
4.4 KiB
C

#include <linux/compiler.h>
#include <linux/mm.h>
#include <linux/signal.h>
#include <linux/smp.h>
#include <asm/asm.h>
#include <asm/bootinfo.h>
#include <asm/byteorder.h>
#include <asm/cpu.h>
#include <asm/inst.h>
#include <asm/processor.h>
#include <asm/uaccess.h>
#include <asm/branch.h>
#include <asm/mipsregs.h>
#include <asm/system.h>
#include <asm/cacheflush.h>
#include <asm/fpu_emulator.h>
#include "ieee754.h"
/* Strap kernel emulator for full MIPS IV emulation */
#ifdef __mips
#undef __mips
#endif
#define __mips 4
/*
* Emulate the arbritrary instruction ir at xcp->cp0_epc. Required when
* we have to emulate the instruction in a COP1 branch delay slot. Do
* not change cp0_epc due to the instruction
*
* According to the spec:
* 1) it shouldnt be a branch :-)
* 2) it can be a COP instruction :-(
* 3) if we are tring to run a protected memory space we must take
* special care on memory access instructions :-(
*/
/*
* "Trampoline" return routine to catch exception following
* execution of delay-slot instruction execution.
*/
struct emuframe {
mips_instruction emul;
mips_instruction badinst;
mips_instruction cookie;
unsigned long epc;
};
int mips_dsemul(struct pt_regs *regs, mips_instruction ir, unsigned long cpc)
{
extern asmlinkage void handle_dsemulret(void);
struct emuframe __user *fr;
int err;
if (ir == 0) { /* a nop is easy */
regs->cp0_epc = cpc;
regs->cp0_cause &= ~CAUSEF_BD;
return 0;
}
#ifdef DSEMUL_TRACE
printk("dsemul %lx %lx\n", regs->cp0_epc, cpc);
#endif
/*
* The strategy is to push the instruction onto the user stack
* and put a trap after it which we can catch and jump to
* the required address any alternative apart from full
* instruction emulation!!.
*
* Algorithmics used a system call instruction, and
* borrowed that vector. MIPS/Linux version is a bit
* more heavyweight in the interests of portability and
* multiprocessor support. For Linux we generate a
* an unaligned access and force an address error exception.
*
* For embedded systems (stand-alone) we prefer to use a
* non-existing CP1 instruction. This prevents us from emulating
* branches, but gives us a cleaner interface to the exception
* handler (single entry point).
*/
/* Ensure that the two instructions are in the same cache line */
fr = (struct emuframe __user *)
((regs->regs[29] - sizeof(struct emuframe)) & ~0x7);
/* Verify that the stack pointer is not competely insane */
if (unlikely(!access_ok(VERIFY_WRITE, fr, sizeof(struct emuframe))))
return SIGBUS;
err = __put_user(ir, &fr->emul);
err |= __put_user((mips_instruction)BREAK_MATH, &fr->badinst);
err |= __put_user((mips_instruction)BD_COOKIE, &fr->cookie);
err |= __put_user(cpc, &fr->epc);
if (unlikely(err)) {
MIPS_FPU_EMU_INC_STATS(errors);
return SIGBUS;
}
regs->cp0_epc = (unsigned long) &fr->emul;
flush_cache_sigtramp((unsigned long)&fr->badinst);
return SIGILL; /* force out of emulation loop */
}
int do_dsemulret(struct pt_regs *xcp)
{
struct emuframe __user *fr;
unsigned long epc;
u32 insn, cookie;
int err = 0;
fr = (struct emuframe __user *)
(xcp->cp0_epc - sizeof(mips_instruction));
/*
* If we can't even access the area, something is very wrong, but we'll
* leave that to the default handling
*/
if (!access_ok(VERIFY_READ, fr, sizeof(struct emuframe)))
return 0;
/*
* Do some sanity checking on the stackframe:
*
* - Is the instruction pointed to by the EPC an BREAK_MATH?
* - Is the following memory word the BD_COOKIE?
*/
err = __get_user(insn, &fr->badinst);
err |= __get_user(cookie, &fr->cookie);
if (unlikely(err || (insn != BREAK_MATH) || (cookie != BD_COOKIE))) {
MIPS_FPU_EMU_INC_STATS(errors);
return 0;
}
/*
* At this point, we are satisfied that it's a BD emulation trap. Yes,
* a user might have deliberately put two malformed and useless
* instructions in a row in his program, in which case he's in for a
* nasty surprise - the next instruction will be treated as a
* continuation address! Alas, this seems to be the only way that we
* can handle signals, recursion, and longjmps() in the context of
* emulating the branch delay instruction.
*/
#ifdef DSEMUL_TRACE
printk("dsemulret\n");
#endif
if (__get_user(epc, &fr->epc)) { /* Saved EPC */
/* This is not a good situation to be in */
force_sig(SIGBUS, current);
return 0;
}
/* Set EPC to return to post-branch instruction */
xcp->cp0_epc = epc;
return 1;
}