perfmon.c

这个linux源代码是很全面的~基本完整了~使用c编译的~由于时间问题我没有亲自测试~但就算用来做参考资料也是非常好的

		} else if (reg_flags & (PFM_REGFL_OVFL_NOTIFY|PFM_REGFL_RANDOM)) {
				DBprintk(("cannot set ovfl_notify or random on pmc%u\n", cnum));
				goto error;
		}

		/*
		 * execute write checker, if any
		 */
		if (PMC_WR_FUNC(cnum)) {
			ret = PMC_WR_FUNC(cnum)(task, cnum, &value, regs);
			if (ret) goto error;
			ret = -EINVAL;
		}

		/*
		 * no error on this register
		 */
		PFM_REG_RETFLAG_SET(tmp.reg_flags, 0);

		/*
		 * update register return value, abort all if problem during copy.
		 * we only modify the reg_flags field. no check mode is fine because
		 * access has been verified upfront in sys_perfmonctl().
		 *
		 * If this fails, then the software state is not modified
		 */
		if (__put_user(tmp.reg_flags, &req->reg_flags)) return -EFAULT;

		/*
		 * Now we commit the changes to the software state
		 */

		/* 
		 * full flag update each time a register is programmed
		 */
		ctx->ctx_soft_pmds[cnum].flags = flags;

		if (PMC_IS_COUNTING(cnum)) {
			ctx->ctx_soft_pmds[cnum].reset_pmds[0] = reset_pmds;

			/* mark all PMDS to be accessed as used */
			CTX_USED_PMD(ctx, reset_pmds);
		}

		/*
		 * Needed in case the user does not initialize the equivalent
		 * PMD. Clearing is done in reset_pmu() so there is no possible
		 * leak here.
		 */
		CTX_USED_PMD(ctx, pmu_conf.pmc_desc[cnum].dep_pmd[0]);

		/* 
		 * keep copy the pmc, used for register reload
		 */
		th->pmc[cnum] = value;

		ia64_set_pmc(cnum, value);

		DBprintk(("[%d] pmc[%u]=0x%lx flags=0x%x used_pmds=0x%lx\n", 
			  task->pid, cnum, value, 
			  ctx->ctx_soft_pmds[cnum].flags, 
			  ctx->ctx_used_pmds[0]));

	}

	return 0;

error:
	PFM_REG_RETFLAG_SET(tmp.reg_flags, PFM_REG_RETFL_EINVAL);

	if (__put_user(tmp.reg_flags, &req->reg_flags)) ret = -EFAULT;

	DBprintk(("[%d] pmc[%u]=0x%lx error %d\n", task->pid, cnum, value, ret));

	return ret;
}

static int
pfm_write_pmds(struct task_struct *task, pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
{
	pfarg_reg_t tmp, *req = (pfarg_reg_t *)arg;
	unsigned long value, hw_value;
	unsigned int cnum;
	int i;
	int ret = 0;

	/* 
	 * Cannot do anything before PMU is enabled 
	 */
	if (!CTX_IS_ENABLED(ctx)) return -EINVAL;

	/* XXX: ctx locking may be required here */

	ret = -EINVAL;

	for (i = 0; i < count; i++, req++) {

		if (__copy_from_user(&tmp, req, sizeof(tmp))) return -EFAULT;

		cnum  = tmp.reg_num;
		value = tmp.reg_value;

		if (!PMD_IS_IMPL(cnum)) {
			DBprintk(("pmd[%u] is unimplemented or invalid\n", cnum));
			goto abort_mission;
		}

		/*
		 * execute write checker, if any
		 */
		if (PMD_WR_FUNC(cnum)) {
			unsigned long v = value;
			ret = PMD_WR_FUNC(cnum)(task, cnum, &v, regs);
			if (ret) goto abort_mission;
			value = v;
			ret = -EINVAL;
		}
		hw_value = value;
		/*
		 * no error on this register
		 */
		PFM_REG_RETFLAG_SET(tmp.reg_flags, 0);

		if (__put_user(tmp.reg_flags, &req->reg_flags)) return -EFAULT;

		/*
		 * now commit changes to software state
		 */

		/* update virtualized (64bits) counter */
		if (PMD_IS_COUNTING(cnum)) {
			ctx->ctx_soft_pmds[cnum].lval = value;
			ctx->ctx_soft_pmds[cnum].val  = value & ~pmu_conf.ovfl_val;

			hw_value = value & pmu_conf.ovfl_val;

			ctx->ctx_soft_pmds[cnum].long_reset  = tmp.reg_long_reset;
			ctx->ctx_soft_pmds[cnum].short_reset = tmp.reg_short_reset;

			ctx->ctx_soft_pmds[cnum].seed = tmp.reg_random_seed;
			ctx->ctx_soft_pmds[cnum].mask = tmp.reg_random_mask;
		}

		/* keep track of what we use */
		CTX_USED_PMD(ctx, pmu_conf.pmd_desc[(cnum)].dep_pmd[0]);

		/* mark this register as used as well */
		CTX_USED_PMD(ctx, RDEP(cnum));

		/* writes to unimplemented part is ignored, so this is safe */
		ia64_set_pmd(cnum, hw_value);

		/* to go away */
		ia64_srlz_d();

		DBprintk(("[%d] pmd[%u]: value=0x%lx hw_value=0x%lx soft_pmd=0x%lx  short_reset=0x%lx "
			  "long_reset=0x%lx hw_pmd=%lx notify=%c used_pmds=0x%lx reset_pmds=0x%lx\n",
				task->pid, cnum,
				value, hw_value,
				ctx->ctx_soft_pmds[cnum].val,
				ctx->ctx_soft_pmds[cnum].short_reset,
				ctx->ctx_soft_pmds[cnum].long_reset,
				ia64_get_pmd(cnum) & pmu_conf.ovfl_val,
				PMC_OVFL_NOTIFY(ctx, cnum) ? 'Y':'N',
				ctx->ctx_used_pmds[0],
				ctx->ctx_soft_pmds[cnum].reset_pmds[0]));
	}

	return 0;

abort_mission:
	/*
	 * for now, we have only one possibility for error
	 */
	PFM_REG_RETFLAG_SET(tmp.reg_flags, PFM_REG_RETFL_EINVAL);

	/*
	 * we change the return value to EFAULT in case we cannot write register return code.
	 * The caller first must correct this error, then a resubmission of the request will
	 * eventually yield the EINVAL.
	 */
	if (__put_user(tmp.reg_flags, &req->reg_flags)) ret = -EFAULT;

	DBprintk(("[%d] pmc[%u]=0x%lx ret %d\n", task->pid, cnum, value, ret));

	return ret;
}

static int
pfm_read_pmds(struct task_struct *task, pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
{
	struct thread_struct *th = &task->thread;
	unsigned long val, lval;
	pfarg_reg_t *req = (pfarg_reg_t *)arg;
	unsigned int cnum, reg_flags = 0;
	int i, ret = 0;
#if __GNUC__ < 3
	int foo;
#endif

	if (!CTX_IS_ENABLED(ctx)) return -EINVAL;

	/*
	 * XXX: MUST MAKE SURE WE DON"T HAVE ANY PENDING OVERFLOW BEFORE READING
	 * This is required when the monitoring has been stoppped by user or kernel.
	 * If it is still going on, then that's fine because we a re not guaranteed
	 * to return an accurate value in this case.
	 */

	/* XXX: ctx locking may be required here */

	DBprintk(("ctx_last_cpu=%d for [%d]\n", atomic_read(&ctx->ctx_last_cpu), task->pid));

	for (i = 0; i < count; i++, req++) {

#if __GNUC__ < 3
		foo = __get_user(cnum, &req->reg_num);
		if (foo) return -EFAULT;
		foo = __get_user(reg_flags, &req->reg_flags);
		if (foo) return -EFAULT;
#else
		if (__get_user(cnum, &req->reg_num)) return -EFAULT;
		if (__get_user(reg_flags, &req->reg_flags)) return -EFAULT;
#endif
		lval = 0UL;

		if (!PMD_IS_IMPL(cnum)) goto abort_mission;
		/*
		 * we can only read the register that we use. That includes
		 * the one we explicitely initialize AND the one we want included
		 * in the sampling buffer (smpl_regs).
		 *
		 * Having this restriction allows optimization in the ctxsw routine
		 * without compromising security (leaks)
		 */
		if (!CTX_IS_USED_PMD(ctx, cnum)) goto abort_mission;

		/*
		 * If the task is not the current one, then we check if the
		 * PMU state is still in the local live register due to lazy ctxsw.
		 * If true, then we read directly from the registers.
		 */
		if (atomic_read(&ctx->ctx_last_cpu) == smp_processor_id()){
			ia64_srlz_d();
			val = ia64_get_pmd(cnum);
			DBprintk(("reading pmd[%u]=0x%lx from hw\n", cnum, val));
		} else {
#ifdef CONFIG_SMP
			int cpu;
			/*
			 * for SMP system, the context may still be live on another
			 * CPU so we need to fetch it before proceeding with the read
			 * This call we only be made once for the whole loop because
			 * of ctx_last_cpu becoming == -1.
			 *
			 * We cannot reuse ctx_last_cpu as it may change before we get to the
			 * actual IPI call. In this case, we will do the call for nothing but
			 * there is no way around it. The receiving side will simply do nothing.
			 */
			cpu = atomic_read(&ctx->ctx_last_cpu);
			if (cpu != -1) {
				DBprintk(("must fetch on CPU%d for [%d]\n", cpu, task->pid));
				pfm_fetch_regs(cpu, task, ctx);
			}
#endif
			/* context has been saved */
			val = th->pmd[cnum];
		}
		if (PMD_IS_COUNTING(cnum)) {
			/*
			 * XXX: need to check for overflow
			 */
			val &= pmu_conf.ovfl_val;
			val += ctx->ctx_soft_pmds[cnum].val;

			lval = ctx->ctx_soft_pmds[cnum].lval;
		} 

		/*
		 * execute read checker, if any
		 */
		if (PMD_RD_FUNC(cnum)) {
			unsigned long v = val;
			ret = PMD_RD_FUNC(cnum)(task, cnum, &v, regs);
			val = v;
		}

		PFM_REG_RETFLAG_SET(reg_flags, ret);

		DBprintk(("read pmd[%u] ret=%d value=0x%lx pmc=0x%lx\n", 
					cnum, ret, val, ia64_get_pmc(cnum)));

		/*
		 * update register return value, abort all if problem during copy.
		 * we only modify the reg_flags field. no check mode is fine because
		 * access has been verified upfront in sys_perfmonctl().
		 */
		if (__put_user(cnum, &req->reg_num)) return -EFAULT;
		if (__put_user(val, &req->reg_value)) return -EFAULT;
		if (__put_user(reg_flags, &req->reg_flags)) return -EFAULT;
		if (__put_user(lval, &req->reg_last_reset_value)) return -EFAULT;
	}

	return 0;

abort_mission:
	PFM_REG_RETFLAG_SET(reg_flags, PFM_REG_RETFL_EINVAL);
	/* 
	 * XXX: if this fails, we stick with the original failure, flag not updated!
	 */
	__put_user(reg_flags, &req->reg_flags);

	return -EINVAL;
}

#ifdef PFM_PMU_USES_DBR
/*
 * Only call this function when a process it trying to
 * write the debug registers (reading is always allowed)
 */
int
pfm_use_debug_registers(struct task_struct *task)
{
	pfm_context_t *ctx = task->thread.pfm_context;
	int ret = 0;

	DBprintk(("called for [%d]\n", task->pid));

	/*
	 * do it only once
	 */
	if (task->thread.flags & IA64_THREAD_DBG_VALID) return 0;

	/*
	 * Even on SMP, we do not need to use an atomic here because
	 * the only way in is via ptrace() and this is possible only when the
	 * process is stopped. Even in the case where the ctxsw out is not totally
	 * completed by the time we come here, there is no way the 'stopped' process
	 * could be in the middle of fiddling with the pfm_write_ibr_dbr() routine.
	 * So this is always safe.
	 */
	if (ctx && ctx->ctx_fl_using_dbreg == 1) return -1;

	LOCK_PFS();

	/*
	 * We cannot allow setting breakpoints when system wide monitoring
	 * sessions are using the debug registers.
	 */
	if (pfm_sessions.pfs_sys_use_dbregs> 0)
		ret = -1;
	else
		pfm_sessions.pfs_ptrace_use_dbregs++;

	DBprintk(("ptrace_use_dbregs=%u  sys_use_dbregs=%u by [%d] ret = %d\n", 
		  pfm_sessions.pfs_ptrace_use_dbregs, 
		  pfm_sessions.pfs_sys_use_dbregs, 
		  task->pid, ret));

	UNLOCK_PFS();

	return ret;
}

/*
 * This function is called for every task that exits with the
 * IA64_THREAD_DBG_VALID set. This indicates a task which was
 * able to use the debug registers for debugging purposes via
 * ptrace(). Therefore we know it was not using them for
 * perfmormance monitoring, so we only decrement the number
 * of "ptraced" debug register users to keep the count up to date
 */
int
pfm_release_debug_registers(struct task_struct *task)
{
	int ret;

	LOCK_PFS();
	if (pfm_sessions.pfs_ptrace_use_dbregs == 0) {
		printk("perfmon: invalid release for [%d] ptrace_use_dbregs=0\n", task->pid);
		ret = -1;
	}  else {
		pfm_sessions.pfs_ptrace_use_dbregs--;
		ret = 0;
	}
	UNLOCK_PFS();

	return ret;
}
#else /* PFM_PMU_USES_DBR is true */
/*
 * in case, the PMU does not use the debug registers, these two functions are nops.
 * The first function is called from arch/ia64/kernel/ptrace.c.
 * The second function is called from arch/ia64/kernel/process.c.
 */
int
pfm_use_debug_registers(struct task_struct *task)
{
	return 0;
}

int
pfm_release_debug_registers(struct task_struct *task)
{
	return 0;
}
#endif /* PFM_PMU_USES_DBR */

static int
pfm_restart(struct task_struct *task, pfm_context_t *ctx, void *arg, int count, 
	 struct pt_regs *regs)
{
	void *sem = &ctx->ctx_restart_sem;

	/* 
	 * Cannot do anything before PMU is enabled 
	 */
	if (!CTX_IS_ENABLED(ctx)) return -EINVAL;

	if (task == current) {
		DBprintk(("restarting self %d frozen=%d ovfl_regs=0x%lx\n", 
			task->pid, 
			ctx->ctx_fl_frozen,
			ctx->ctx_ovfl_regs[0]));

		pfm_reset_regs(ctx, ctx->ctx_ovfl_regs, PFM_PMD_LONG_RESET);

		ctx->ctx_ovfl_regs[0] = 0UL;

		/*
		 * We ignore block/don't block because we never block
		 * for a self-monitoring process.
		 */
		ctx->ctx_fl_frozen = 0;

		if (CTX_HAS_SMPL(ctx)) {
			ctx->ctx_psb->psb_hdr->hdr_count = 0;
			ctx->ctx_psb->psb_index = 0;
		}

		/* simply unfreeze */
		ia64_set_pmc(0, 0);
		ia64_srlz_d();

		return 0;
	} 
	/* restart on another task */

	/*
	 * if blocking, then post the semaphore.
	 * if non-blocking, then we ensure that the task will go into
	 * pfm_overflow_must_block() before returning to user mode. 
	 * We cannot explicitely reset another task, it MUST always
	 * be done by the task itself. This works for system wide because
	 * the tool that is controlling the session is doing "self-monitoring".
	 *
	 * XXX: what if the task never goes back to user?
	 *
	 */
	if (CTX_OVFL_NOBLOCK(ctx) == 0) {
		DBprintk(("unblocking %d \n", task->pid));
		up(sem);
	} else {
		task->thread.pfm_ovfl_block_reset = 1;
	}