[deliverable/linux.git] / arch / powerpc / oprofile / op_model_power4.c

/*
 * Copyright (C) 2004 Anton Blanchard <anton@au.ibm.com>, IBM
 *
 * This program is free software; you can redistribute it and/or
 * modify it under the terms of the GNU General Public License
 * as published by the Free Software Foundation; either version
 * 2 of the License, or (at your option) any later version.
 */

#include <linux/oprofile.h>
#include <linux/init.h>
#include <linux/smp.h>
#include <asm/firmware.h>
#include <asm/ptrace.h>
#include <asm/system.h>
#include <asm/processor.h>
#include <asm/cputable.h>
#include <asm/rtas.h>
#include <asm/oprofile_impl.h>
#include <asm/reg.h>

#define dbg(args...)

static unsigned long reset_value[OP_MAX_COUNTER];

static int oprofile_running;
static int mmcra_has_sihv;

/* mmcr values are set in power4_reg_setup, used in power4_cpu_setup */
static u32 mmcr0_val;
static u64 mmcr1_val;
static u32 mmcra_val;

/*
 * Since we do not have an NMI, backtracing through spinlocks is
 * only a best guess. In light of this, allow it to be disabled at
 * runtime.
 */
static int backtrace_spinlocks;

static void power4_reg_setup(struct op_counter_config *ctr,
			     struct op_system_config *sys,
			     int num_ctrs)
{
	int i;

	/*
	 * SIHV / SIPR bits are only implemented on POWER4+ (GQ) and above.
	 * However we disable it on all POWER4 until we verify it works
	 * (I was seeing some strange behaviour last time I tried).
	 *
	 * It has been verified to work on POWER5 so we enable it there.
	 */
	if (cpu_has_feature(CPU_FTR_MMCRA_SIHV))
		mmcra_has_sihv = 1;

	/*
	 * The performance counter event settings are given in the mmcr0,
	 * mmcr1 and mmcra values passed from the user in the
	 * op_system_config structure (sys variable).
	 */
	mmcr0_val = sys->mmcr0;
	mmcr1_val = sys->mmcr1;
	mmcra_val = sys->mmcra;

	backtrace_spinlocks = sys->backtrace_spinlocks;

	for (i = 0; i < cur_cpu_spec->num_pmcs; ++i)
		reset_value[i] = 0x80000000UL - ctr[i].count;

	/* setup user and kernel profiling */
	if (sys->enable_kernel)
		mmcr0_val &= ~MMCR0_KERNEL_DISABLE;
	else
		mmcr0_val |= MMCR0_KERNEL_DISABLE;

	if (sys->enable_user)
		mmcr0_val &= ~MMCR0_PROBLEM_DISABLE;
	else
		mmcr0_val |= MMCR0_PROBLEM_DISABLE;
}

extern void ppc64_enable_pmcs(void);

/*
 * Older CPUs require the MMCRA sample bit to be always set, but newer 
 * CPUs only want it set for some groups. Eventually we will remove all
 * knowledge of this bit in the kernel, oprofile userspace should be
 * setting it when required.
 *
 * In order to keep current installations working we force the bit for
 * those older CPUs. Once everyone has updated their oprofile userspace we
 * can remove this hack.
 */
static inline int mmcra_must_set_sample(void)
{
	if (__is_processor(PV_POWER4) || __is_processor(PV_POWER4p) ||
	    __is_processor(PV_970) || __is_processor(PV_970FX) ||
	    __is_processor(PV_970MP))
		return 1;

	return 0;
}

static void power4_cpu_setup(void *unused)
{
	unsigned int mmcr0 = mmcr0_val;
	unsigned long mmcra = mmcra_val;

	ppc64_enable_pmcs();

	/* set the freeze bit */
	mmcr0 |= MMCR0_FC;
	mtspr(SPRN_MMCR0, mmcr0);

	mmcr0 |= MMCR0_FCM1|MMCR0_PMXE|MMCR0_FCECE;
	mmcr0 |= MMCR0_PMC1CE|MMCR0_PMCjCE;
	mtspr(SPRN_MMCR0, mmcr0);

	mtspr(SPRN_MMCR1, mmcr1_val);

	if (mmcra_must_set_sample())
		mmcra |= MMCRA_SAMPLE_ENABLE;
	mtspr(SPRN_MMCRA, mmcra);

	dbg("setup on cpu %d, mmcr0 %lx\n", smp_processor_id(),
	    mfspr(SPRN_MMCR0));
	dbg("setup on cpu %d, mmcr1 %lx\n", smp_processor_id(),
	    mfspr(SPRN_MMCR1));
	dbg("setup on cpu %d, mmcra %lx\n", smp_processor_id(),
	    mfspr(SPRN_MMCRA));
}

static void power4_start(struct op_counter_config *ctr)
{
	int i;
	unsigned int mmcr0;

	/* set the PMM bit (see comment below) */
	mtmsrd(mfmsr() | MSR_PMM);

	for (i = 0; i < cur_cpu_spec->num_pmcs; ++i) {
		if (ctr[i].enabled) {
			ctr_write(i, reset_value[i]);
		} else {
			ctr_write(i, 0);
		}
	}

	mmcr0 = mfspr(SPRN_MMCR0);

	/*
	 * We must clear the PMAO bit on some (GQ) chips. Just do it
	 * all the time
	 */
	mmcr0 &= ~MMCR0_PMAO;

	/*
	 * now clear the freeze bit, counting will not start until we
	 * rfid from this excetion, because only at that point will
	 * the PMM bit be cleared
	 */
	mmcr0 &= ~MMCR0_FC;
	mtspr(SPRN_MMCR0, mmcr0);

	oprofile_running = 1;

	dbg("start on cpu %d, mmcr0 %x\n", smp_processor_id(), mmcr0);
}

static void power4_stop(void)
{
	unsigned int mmcr0;

	/* freeze counters */
	mmcr0 = mfspr(SPRN_MMCR0);
	mmcr0 |= MMCR0_FC;
	mtspr(SPRN_MMCR0, mmcr0);

	oprofile_running = 0;

	dbg("stop on cpu %d, mmcr0 %x\n", smp_processor_id(), mmcr0);

	mb();
}

/* Fake functions used by canonicalize_pc */
static void __attribute_used__ hypervisor_bucket(void)
{
}

static void __attribute_used__ rtas_bucket(void)
{
}

static void __attribute_used__ kernel_unknown_bucket(void)
{
}

static unsigned long check_spinlock_pc(struct pt_regs *regs,
				       unsigned long profile_pc)
{
	unsigned long pc = instruction_pointer(regs);

	/*
	 * If both the SIAR (sampled instruction) and the perfmon exception
	 * occurred in a spinlock region then we account the sample to the
	 * calling function. This isnt 100% correct, we really need soft
	 * IRQ disable so we always get the perfmon exception at the
	 * point at which the SIAR is set.
	 */
	if (backtrace_spinlocks && in_lock_functions(pc) &&
			in_lock_functions(profile_pc))
		return regs->link;
	else
		return profile_pc;
}

/*
 * On GQ and newer the MMCRA stores the HV and PR bits at the time
 * the SIAR was sampled. We use that to work out if the SIAR was sampled in
 * the hypervisor, our exception vectors or RTAS.
 */
static unsigned long get_pc(struct pt_regs *regs)
{
	unsigned long pc = mfspr(SPRN_SIAR);
	unsigned long mmcra;

	/* Cant do much about it */
	if (!mmcra_has_sihv)
		return check_spinlock_pc(regs, pc);

	mmcra = mfspr(SPRN_MMCRA);

	/* Were we in the hypervisor? */
	if (firmware_has_feature(FW_FEATURE_LPAR) && (mmcra & MMCRA_SIHV))
		/* function descriptor madness */
		return *((unsigned long *)hypervisor_bucket);

	/* We were in userspace, nothing to do */
	if (mmcra & MMCRA_SIPR)
		return pc;

#ifdef CONFIG_PPC_RTAS
	/* Were we in RTAS? */
	if (pc >= rtas.base && pc < (rtas.base + rtas.size))
		/* function descriptor madness */
		return *((unsigned long *)rtas_bucket);
#endif

	/* Were we in our exception vectors or SLB real mode miss handler? */
	if (pc < 0x1000000UL)
		return (unsigned long)__va(pc);

	/* Not sure where we were */
	if (!is_kernel_addr(pc))
		/* function descriptor madness */
		return *((unsigned long *)kernel_unknown_bucket);

	return check_spinlock_pc(regs, pc);
}

static int get_kernel(unsigned long pc)
{
	int is_kernel;

	if (!mmcra_has_sihv) {
		is_kernel = is_kernel_addr(pc);
	} else {
		unsigned long mmcra = mfspr(SPRN_MMCRA);
		is_kernel = ((mmcra & MMCRA_SIPR) == 0);
	}

	return is_kernel;
}

static void power4_handle_interrupt(struct pt_regs *regs,
				    struct op_counter_config *ctr)
{
	unsigned long pc;
	int is_kernel;
	int val;
	int i;
	unsigned int mmcr0;

	pc = get_pc(regs);
	is_kernel = get_kernel(pc);

	/* set the PMM bit (see comment below) */
	mtmsrd(mfmsr() | MSR_PMM);

	for (i = 0; i < cur_cpu_spec->num_pmcs; ++i) {
		val = ctr_read(i);
		if (val < 0) {
			if (oprofile_running && ctr[i].enabled) {
				oprofile_add_pc(pc, is_kernel, i);
				ctr_write(i, reset_value[i]);
			} else {
				ctr_write(i, 0);
			}
		}
	}

	mmcr0 = mfspr(SPRN_MMCR0);

	/* reset the perfmon trigger */
	mmcr0 |= MMCR0_PMXE;

	/*
	 * We must clear the PMAO bit on some (GQ) chips. Just do it
	 * all the time
	 */
	mmcr0 &= ~MMCR0_PMAO;

	/*
	 * now clear the freeze bit, counting will not start until we
	 * rfid from this exception, because only at that point will
	 * the PMM bit be cleared
	 */
	mmcr0 &= ~MMCR0_FC;
	mtspr(SPRN_MMCR0, mmcr0);
}

struct op_powerpc_model op_model_power4 = {
	.reg_setup		= power4_reg_setup,
	.cpu_setup		= power4_cpu_setup,
	.start			= power4_start,
	.stop			= power4_stop,
	.handle_interrupt	= power4_handle_interrupt,
};
Commit	Line	Data
1da177e4 LT	1	/*
	2	* Copyright (C) 2004 Anton Blanchard <anton@au.ibm.com>, IBM
	3	*
	4	* This program is free software; you can redistribute it and/or
	5	* modify it under the terms of the GNU General Public License
	6	* as published by the Free Software Foundation; either version
	7	* 2 of the License, or (at your option) any later version.
	8	*/
	9
	10	#include <linux/oprofile.h>
	11	#include <linux/init.h>
	12	#include <linux/smp.h>
57cfb814	13	#include <asm/firmware.h>
1da177e4 LT	14	#include <asm/ptrace.h>
	15	#include <asm/system.h>
	16	#include <asm/processor.h>
	17	#include <asm/cputable.h>
1da177e4	18	#include <asm/rtas.h>
dca85932	19	#include <asm/oprofile_impl.h>
cb09cff3	20	#include <asm/reg.h>
1da177e4 LT	21
	22	#define dbg(args...)
	23
1da177e4 LT	24	static unsigned long reset_value[OP_MAX_COUNTER];
1da177e4 LT	25
1da177e4 LT	26	static int oprofile_running;
	27	static int mmcra_has_sihv;
	28
	29	/* mmcr values are set in power4_reg_setup, used in power4_cpu_setup */
	30	static u32 mmcr0_val;
	31	static u64 mmcr1_val;
	32	static u32 mmcra_val;
	33
	34	/*
	35	* Since we do not have an NMI, backtracing through spinlocks is
	36	* only a best guess. In light of this, allow it to be disabled at
	37	* runtime.
	38	*/
	39	static int backtrace_spinlocks;
	40
	41	static void power4_reg_setup(struct op_counter_config *ctr,
	42	struct op_system_config *sys,
	43	int num_ctrs)
	44	{
	45	int i;
	46
1da177e4 LT	47	/*
	48	* SIHV / SIPR bits are only implemented on POWER4+ (GQ) and above.
	49	* However we disable it on all POWER4 until we verify it works
	50	* (I was seeing some strange behaviour last time I tried).
	51	*
	52	* It has been verified to work on POWER5 so we enable it there.
	53	*/
	54	if (cpu_has_feature(CPU_FTR_MMCRA_SIHV))
	55	mmcra_has_sihv = 1;
	56
	57	/*
	58	* The performance counter event settings are given in the mmcr0,
	59	* mmcr1 and mmcra values passed from the user in the
	60	* op_system_config structure (sys variable).
	61	*/
	62	mmcr0_val = sys->mmcr0;
	63	mmcr1_val = sys->mmcr1;
	64	mmcra_val = sys->mmcra;
	65
	66	backtrace_spinlocks = sys->backtrace_spinlocks;
	67
a6908cd0	68	for (i = 0; i < cur_cpu_spec->num_pmcs; ++i)
1da177e4 LT	69	reset_value[i] = 0x80000000UL - ctr[i].count;
	70
	71	/* setup user and kernel profiling */
	72	if (sys->enable_kernel)
	73	mmcr0_val &= ~MMCR0_KERNEL_DISABLE;
	74	else
	75	mmcr0_val \|= MMCR0_KERNEL_DISABLE;
	76
	77	if (sys->enable_user)
	78	mmcr0_val &= ~MMCR0_PROBLEM_DISABLE;
	79	else
	80	mmcr0_val \|= MMCR0_PROBLEM_DISABLE;
	81	}
	82
	83	extern void ppc64_enable_pmcs(void);
	84
cb09cff3 AB	85	/*
	86	* Older CPUs require the MMCRA sample bit to be always set, but newer
	87	* CPUs only want it set for some groups. Eventually we will remove all
	88	* knowledge of this bit in the kernel, oprofile userspace should be
	89	* setting it when required.
	90	*
	91	* In order to keep current installations working we force the bit for
	92	* those older CPUs. Once everyone has updated their oprofile userspace we
	93	* can remove this hack.
	94	*/
	95	static inline int mmcra_must_set_sample(void)
	96	{
	97	if (__is_processor(PV_POWER4) \|\| __is_processor(PV_POWER4p) \|\|
	98	__is_processor(PV_970) \|\| __is_processor(PV_970FX) \|\|
	99	__is_processor(PV_970MP))
	100	return 1;
	101
	102	return 0;
	103	}
	104
1da177e4 LT	105	static void power4_cpu_setup(void *unused)
	106	{
	107	unsigned int mmcr0 = mmcr0_val;
	108	unsigned long mmcra = mmcra_val;
	109
	110	ppc64_enable_pmcs();
	111
	112	/* set the freeze bit */
	113	mmcr0 \|= MMCR0_FC;
	114	mtspr(SPRN_MMCR0, mmcr0);
	115
	116	mmcr0 \|= MMCR0_FCM1\|MMCR0_PMXE\|MMCR0_FCECE;
	117	mmcr0 \|= MMCR0_PMC1CE\|MMCR0_PMCjCE;
	118	mtspr(SPRN_MMCR0, mmcr0);
	119
	120	mtspr(SPRN_MMCR1, mmcr1_val);
	121
cb09cff3 AB	122	if (mmcra_must_set_sample())
cb09cff3 AB	123	mmcra \|= MMCRA_SAMPLE_ENABLE;
1da177e4 LT	124	mtspr(SPRN_MMCRA, mmcra);
	125
	126	dbg("setup on cpu %d, mmcr0 %lx\n", smp_processor_id(),
	127	mfspr(SPRN_MMCR0));
	128	dbg("setup on cpu %d, mmcr1 %lx\n", smp_processor_id(),
	129	mfspr(SPRN_MMCR1));
	130	dbg("setup on cpu %d, mmcra %lx\n", smp_processor_id(),
	131	mfspr(SPRN_MMCRA));
	132	}
	133
	134	static void power4_start(struct op_counter_config *ctr)
	135	{
	136	int i;
	137	unsigned int mmcr0;
	138
	139	/* set the PMM bit (see comment below) */
	140	mtmsrd(mfmsr() \| MSR_PMM);
	141
a6908cd0	142	for (i = 0; i < cur_cpu_spec->num_pmcs; ++i) {
1da177e4 LT	143	if (ctr[i].enabled) {
	144	ctr_write(i, reset_value[i]);
	145	} else {
	146	ctr_write(i, 0);
	147	}
	148	}
	149
	150	mmcr0 = mfspr(SPRN_MMCR0);
	151
	152	/*
	153	* We must clear the PMAO bit on some (GQ) chips. Just do it
	154	* all the time
	155	*/
	156	mmcr0 &= ~MMCR0_PMAO;
	157
	158	/*
	159	* now clear the freeze bit, counting will not start until we
	160	* rfid from this excetion, because only at that point will
	161	* the PMM bit be cleared
	162	*/
	163	mmcr0 &= ~MMCR0_FC;
	164	mtspr(SPRN_MMCR0, mmcr0);
	165
	166	oprofile_running = 1;
	167
	168	dbg("start on cpu %d, mmcr0 %x\n", smp_processor_id(), mmcr0);
	169	}
	170
	171	static void power4_stop(void)
	172	{
	173	unsigned int mmcr0;
	174
	175	/* freeze counters */
	176	mmcr0 = mfspr(SPRN_MMCR0);
	177	mmcr0 \|= MMCR0_FC;
	178	mtspr(SPRN_MMCR0, mmcr0);
	179
	180	oprofile_running = 0;
	181
	182	dbg("stop on cpu %d, mmcr0 %x\n", smp_processor_id(), mmcr0);
	183
	184	mb();
	185	}
	186
	187	/* Fake functions used by canonicalize_pc */
	188	static void __attribute_used__ hypervisor_bucket(void)
	189	{
	190	}
	191
	192	static void __attribute_used__ rtas_bucket(void)
	193	{
	194	}
	195
	196	static void __attribute_used__ kernel_unknown_bucket(void)
	197	{
	198	}
	199
	200	static unsigned long check_spinlock_pc(struct pt_regs *regs,
	201	unsigned long profile_pc)
	202	{
	203	unsigned long pc = instruction_pointer(regs);
	204
	205	/*
	206	* If both the SIAR (sampled instruction) and the perfmon exception
207	* occurred in a spinlock region then we account the sample to the
208	* calling function. This isnt 100% correct, we really need soft
209	* IRQ disable so we always get the perfmon exception at the
210	* point at which the SIAR is set.
211	*/
212	if (backtrace_spinlocks && in_lock_functions(pc) &&
213	in_lock_functions(profile_pc))
214	return regs->link;
215	else
216	return profile_pc;
217	}
218
219	/*
220	* On GQ and newer the MMCRA stores the HV and PR bits at the time
221	* the SIAR was sampled. We use that to work out if the SIAR was sampled in
222	* the hypervisor, our exception vectors or RTAS.
223	*/
224	static unsigned long get_pc(struct pt_regs *regs)
225	{
226	unsigned long pc = mfspr(SPRN_SIAR);
227	unsigned long mmcra;
228
229	/* Cant do much about it */
230	if (!mmcra_has_sihv)
231	return check_spinlock_pc(regs, pc);
232
233	mmcra = mfspr(SPRN_MMCRA);
234
235	/* Were we in the hypervisor? */
57cfb814	236	if (firmware_has_feature(FW_FEATURE_LPAR) && (mmcra & MMCRA_SIHV))
1da177e4 LT	237	/* function descriptor madness */
	238	return ((unsigned long )hypervisor_bucket);
	239
	240	/* We were in userspace, nothing to do */
	241	if (mmcra & MMCRA_SIPR)
	242	return pc;
	243
	244	#ifdef CONFIG_PPC_RTAS
	245	/* Were we in RTAS? */
	246	if (pc >= rtas.base && pc < (rtas.base + rtas.size))
	247	/* function descriptor madness */
	248	return ((unsigned long )rtas_bucket);
	249	#endif
	250
	251	/* Were we in our exception vectors or SLB real mode miss handler? */
	252	if (pc < 0x1000000UL)
	253	return (unsigned long)__va(pc);
	254
	255	/* Not sure where we were */
51fae6de	256	if (!is_kernel_addr(pc))
1da177e4 LT	257	/* function descriptor madness */
	258	return ((unsigned long )kernel_unknown_bucket);
	259
	260	return check_spinlock_pc(regs, pc);
	261	}
	262
	263	static int get_kernel(unsigned long pc)
	264	{
	265	int is_kernel;
	266
	267	if (!mmcra_has_sihv) {
51fae6de	268	is_kernel = is_kernel_addr(pc);
1da177e4 LT	269	} else {
	270	unsigned long mmcra = mfspr(SPRN_MMCRA);
	271	is_kernel = ((mmcra & MMCRA_SIPR) == 0);
	272	}
	273
	274	return is_kernel;
	275	}
	276
	277	static void power4_handle_interrupt(struct pt_regs *regs,
	278	struct op_counter_config *ctr)
	279	{
	280	unsigned long pc;
	281	int is_kernel;
	282	int val;
	283	int i;
	284	unsigned int mmcr0;
	285
	286	pc = get_pc(regs);
	287	is_kernel = get_kernel(pc);
	288
	289	/* set the PMM bit (see comment below) */
	290	mtmsrd(mfmsr() \| MSR_PMM);
	291
a6908cd0	292	for (i = 0; i < cur_cpu_spec->num_pmcs; ++i) {
1da177e4 LT	293	val = ctr_read(i);
	294	if (val < 0) {
	295	if (oprofile_running && ctr[i].enabled) {
	296	oprofile_add_pc(pc, is_kernel, i);
	297	ctr_write(i, reset_value[i]);
	298	} else {
	299	ctr_write(i, 0);
	300	}
	301	}
	302	}
	303
	304	mmcr0 = mfspr(SPRN_MMCR0);
	305
	306	/* reset the perfmon trigger */
	307	mmcr0 \|= MMCR0_PMXE;
	308
	309	/*
	310	* We must clear the PMAO bit on some (GQ) chips. Just do it
	311	* all the time
	312	*/
	313	mmcr0 &= ~MMCR0_PMAO;
	314
	315	/*
	316	* now clear the freeze bit, counting will not start until we
	317	* rfid from this exception, because only at that point will
	318	* the PMM bit be cleared
	319	*/
	320	mmcr0 &= ~MMCR0_FC;
	321	mtspr(SPRN_MMCR0, mmcr0);
	322	}
	323
a3e48c10	324	struct op_powerpc_model op_model_power4 = {
1da177e4 LT	325	.reg_setup = power4_reg_setup,
	326	.cpu_setup = power4_cpu_setup,
	327	.start = power4_start,
	328	.stop = power4_stop,
	329	.handle_interrupt = power4_handle_interrupt,
	330	};