kvm: take srcu lock around kvm_steal_time_set_preempted()
[linux/fpc-iii.git] / arch / powerpc / oprofile / op_model_power4.c
blob4b32e9404bbef60acb1639cfe613dff51543bcec
1 /*
2 * Copyright (C) 2004 Anton Blanchard <anton@au.ibm.com>, IBM
3 * Added mmcra[slot] support:
4 * Copyright (C) 2006-2007 Will Schmidt <willschm@us.ibm.com>, IBM
6 * This program is free software; you can redistribute it and/or
7 * modify it under the terms of the GNU General Public License
8 * as published by the Free Software Foundation; either version
9 * 2 of the License, or (at your option) any later version.
12 #include <linux/oprofile.h>
13 #include <linux/smp.h>
14 #include <asm/firmware.h>
15 #include <asm/ptrace.h>
16 #include <asm/processor.h>
17 #include <asm/cputable.h>
18 #include <asm/rtas.h>
19 #include <asm/oprofile_impl.h>
20 #include <asm/reg.h>
22 #define dbg(args...)
23 #define OPROFILE_PM_PMCSEL_MSK 0xffULL
24 #define OPROFILE_PM_UNIT_SHIFT 60
25 #define OPROFILE_PM_UNIT_MSK 0xfULL
26 #define OPROFILE_MAX_PMC_NUM 3
27 #define OPROFILE_PMSEL_FIELD_WIDTH 8
28 #define OPROFILE_UNIT_FIELD_WIDTH 4
29 #define MMCRA_SIAR_VALID_MASK 0x10000000ULL
31 static unsigned long reset_value[OP_MAX_COUNTER];
33 static int oprofile_running;
34 static int use_slot_nums;
36 /* mmcr values are set in power4_reg_setup, used in power4_cpu_setup */
37 static u32 mmcr0_val;
38 static u64 mmcr1_val;
39 static u64 mmcra_val;
40 static u32 cntr_marked_events;
42 static int power7_marked_instr_event(u64 mmcr1)
44 u64 psel, unit;
45 int pmc, cntr_marked_events = 0;
47 /* Given the MMCR1 value, look at the field for each counter to
48 * determine if it is a marked event. Code based on the function
49 * power7_marked_instr_event() in file arch/powerpc/perf/power7-pmu.c.
51 for (pmc = 0; pmc < 4; pmc++) {
52 psel = mmcr1 & (OPROFILE_PM_PMCSEL_MSK
53 << (OPROFILE_MAX_PMC_NUM - pmc)
54 * OPROFILE_PMSEL_FIELD_WIDTH);
55 psel = (psel >> ((OPROFILE_MAX_PMC_NUM - pmc)
56 * OPROFILE_PMSEL_FIELD_WIDTH)) & ~1ULL;
57 unit = mmcr1 & (OPROFILE_PM_UNIT_MSK
58 << (OPROFILE_PM_UNIT_SHIFT
59 - (pmc * OPROFILE_PMSEL_FIELD_WIDTH )));
60 unit = unit >> (OPROFILE_PM_UNIT_SHIFT
61 - (pmc * OPROFILE_PMSEL_FIELD_WIDTH));
63 switch (psel >> 4) {
64 case 2:
65 cntr_marked_events |= (pmc == 1 || pmc == 3) << pmc;
66 break;
67 case 3:
68 if (psel == 0x3c) {
69 cntr_marked_events |= (pmc == 0) << pmc;
70 break;
73 if (psel == 0x3e) {
74 cntr_marked_events |= (pmc != 1) << pmc;
75 break;
78 cntr_marked_events |= 1 << pmc;
79 break;
80 case 4:
81 case 5:
82 cntr_marked_events |= (unit == 0xd) << pmc;
83 break;
84 case 6:
85 if (psel == 0x64)
86 cntr_marked_events |= (pmc >= 2) << pmc;
87 break;
88 case 8:
89 cntr_marked_events |= (unit == 0xd) << pmc;
90 break;
93 return cntr_marked_events;
96 static int power4_reg_setup(struct op_counter_config *ctr,
97 struct op_system_config *sys,
98 int num_ctrs)
100 int i;
103 * The performance counter event settings are given in the mmcr0,
104 * mmcr1 and mmcra values passed from the user in the
105 * op_system_config structure (sys variable).
107 mmcr0_val = sys->mmcr0;
108 mmcr1_val = sys->mmcr1;
109 mmcra_val = sys->mmcra;
111 /* Power 7+ and newer architectures:
112 * Determine which counter events in the group (the group of events is
113 * specified by the bit settings in the MMCR1 register) are marked
114 * events for use in the interrupt handler. Do the calculation once
115 * before OProfile starts. Information is used in the interrupt
116 * handler. Starting with Power 7+ we only record the sample for
117 * marked events if the SIAR valid bit is set. For non marked events
118 * the sample is always recorded.
120 if (pvr_version_is(PVR_POWER7p))
121 cntr_marked_events = power7_marked_instr_event(mmcr1_val);
122 else
123 cntr_marked_events = 0; /* For older processors, set the bit map
124 * to zero so the sample will always be
125 * be recorded.
128 for (i = 0; i < cur_cpu_spec->num_pmcs; ++i)
129 reset_value[i] = 0x80000000UL - ctr[i].count;
131 /* setup user and kernel profiling */
132 if (sys->enable_kernel)
133 mmcr0_val &= ~MMCR0_KERNEL_DISABLE;
134 else
135 mmcr0_val |= MMCR0_KERNEL_DISABLE;
137 if (sys->enable_user)
138 mmcr0_val &= ~MMCR0_PROBLEM_DISABLE;
139 else
140 mmcr0_val |= MMCR0_PROBLEM_DISABLE;
142 if (pvr_version_is(PVR_POWER4) || pvr_version_is(PVR_POWER4p) ||
143 pvr_version_is(PVR_970) || pvr_version_is(PVR_970FX) ||
144 pvr_version_is(PVR_970MP) || pvr_version_is(PVR_970GX) ||
145 pvr_version_is(PVR_POWER5) || pvr_version_is(PVR_POWER5p))
146 use_slot_nums = 1;
148 return 0;
151 extern void ppc_enable_pmcs(void);
154 * Older CPUs require the MMCRA sample bit to be always set, but newer
155 * CPUs only want it set for some groups. Eventually we will remove all
156 * knowledge of this bit in the kernel, oprofile userspace should be
157 * setting it when required.
159 * In order to keep current installations working we force the bit for
160 * those older CPUs. Once everyone has updated their oprofile userspace we
161 * can remove this hack.
163 static inline int mmcra_must_set_sample(void)
165 if (pvr_version_is(PVR_POWER4) || pvr_version_is(PVR_POWER4p) ||
166 pvr_version_is(PVR_970) || pvr_version_is(PVR_970FX) ||
167 pvr_version_is(PVR_970MP) || pvr_version_is(PVR_970GX))
168 return 1;
170 return 0;
173 static int power4_cpu_setup(struct op_counter_config *ctr)
175 unsigned int mmcr0 = mmcr0_val;
176 unsigned long mmcra = mmcra_val;
178 ppc_enable_pmcs();
180 /* set the freeze bit */
181 mmcr0 |= MMCR0_FC;
182 mtspr(SPRN_MMCR0, mmcr0);
184 mmcr0 |= MMCR0_FCM1|MMCR0_PMXE|MMCR0_FCECE;
185 mmcr0 |= MMCR0_PMC1CE|MMCR0_PMCjCE;
186 mtspr(SPRN_MMCR0, mmcr0);
188 mtspr(SPRN_MMCR1, mmcr1_val);
190 if (mmcra_must_set_sample())
191 mmcra |= MMCRA_SAMPLE_ENABLE;
192 mtspr(SPRN_MMCRA, mmcra);
194 dbg("setup on cpu %d, mmcr0 %lx\n", smp_processor_id(),
195 mfspr(SPRN_MMCR0));
196 dbg("setup on cpu %d, mmcr1 %lx\n", smp_processor_id(),
197 mfspr(SPRN_MMCR1));
198 dbg("setup on cpu %d, mmcra %lx\n", smp_processor_id(),
199 mfspr(SPRN_MMCRA));
201 return 0;
204 static int power4_start(struct op_counter_config *ctr)
206 int i;
207 unsigned int mmcr0;
209 /* set the PMM bit (see comment below) */
210 mtmsr(mfmsr() | MSR_PMM);
212 for (i = 0; i < cur_cpu_spec->num_pmcs; ++i) {
213 if (ctr[i].enabled) {
214 classic_ctr_write(i, reset_value[i]);
215 } else {
216 classic_ctr_write(i, 0);
220 mmcr0 = mfspr(SPRN_MMCR0);
223 * We must clear the PMAO bit on some (GQ) chips. Just do it
224 * all the time
226 mmcr0 &= ~MMCR0_PMAO;
229 * now clear the freeze bit, counting will not start until we
230 * rfid from this excetion, because only at that point will
231 * the PMM bit be cleared
233 mmcr0 &= ~MMCR0_FC;
234 mtspr(SPRN_MMCR0, mmcr0);
236 oprofile_running = 1;
238 dbg("start on cpu %d, mmcr0 %x\n", smp_processor_id(), mmcr0);
239 return 0;
242 static void power4_stop(void)
244 unsigned int mmcr0;
246 /* freeze counters */
247 mmcr0 = mfspr(SPRN_MMCR0);
248 mmcr0 |= MMCR0_FC;
249 mtspr(SPRN_MMCR0, mmcr0);
251 oprofile_running = 0;
253 dbg("stop on cpu %d, mmcr0 %x\n", smp_processor_id(), mmcr0);
255 mb();
258 /* Fake functions used by canonicalize_pc */
259 static void __used hypervisor_bucket(void)
263 static void __used rtas_bucket(void)
267 static void __used kernel_unknown_bucket(void)
272 * On GQ and newer the MMCRA stores the HV and PR bits at the time
273 * the SIAR was sampled. We use that to work out if the SIAR was sampled in
274 * the hypervisor, our exception vectors or RTAS.
275 * If the MMCRA_SAMPLE_ENABLE bit is set, we can use the MMCRA[slot] bits
276 * to more accurately identify the address of the sampled instruction. The
277 * mmcra[slot] bits represent the slot number of a sampled instruction
278 * within an instruction group. The slot will contain a value between 1
279 * and 5 if MMCRA_SAMPLE_ENABLE is set, otherwise 0.
281 static unsigned long get_pc(struct pt_regs *regs)
283 unsigned long pc = mfspr(SPRN_SIAR);
284 unsigned long mmcra;
285 unsigned long slot;
287 /* Can't do much about it */
288 if (!cur_cpu_spec->oprofile_mmcra_sihv)
289 return pc;
291 mmcra = mfspr(SPRN_MMCRA);
293 if (use_slot_nums && (mmcra & MMCRA_SAMPLE_ENABLE)) {
294 slot = ((mmcra & MMCRA_SLOT) >> MMCRA_SLOT_SHIFT);
295 if (slot > 1)
296 pc += 4 * (slot - 1);
299 /* Were we in the hypervisor? */
300 if (firmware_has_feature(FW_FEATURE_LPAR) &&
301 (mmcra & cur_cpu_spec->oprofile_mmcra_sihv))
302 /* function descriptor madness */
303 return *((unsigned long *)hypervisor_bucket);
305 /* We were in userspace, nothing to do */
306 if (mmcra & cur_cpu_spec->oprofile_mmcra_sipr)
307 return pc;
309 #ifdef CONFIG_PPC_RTAS
310 /* Were we in RTAS? */
311 if (pc >= rtas.base && pc < (rtas.base + rtas.size))
312 /* function descriptor madness */
313 return *((unsigned long *)rtas_bucket);
314 #endif
316 /* Were we in our exception vectors or SLB real mode miss handler? */
317 if (pc < 0x1000000UL)
318 return (unsigned long)__va(pc);
320 /* Not sure where we were */
321 if (!is_kernel_addr(pc))
322 /* function descriptor madness */
323 return *((unsigned long *)kernel_unknown_bucket);
325 return pc;
328 static int get_kernel(unsigned long pc, unsigned long mmcra)
330 int is_kernel;
332 if (!cur_cpu_spec->oprofile_mmcra_sihv) {
333 is_kernel = is_kernel_addr(pc);
334 } else {
335 is_kernel = ((mmcra & cur_cpu_spec->oprofile_mmcra_sipr) == 0);
338 return is_kernel;
341 static bool pmc_overflow(unsigned long val)
343 if ((int)val < 0)
344 return true;
347 * Events on POWER7 can roll back if a speculative event doesn't
348 * eventually complete. Unfortunately in some rare cases they will
349 * raise a performance monitor exception. We need to catch this to
350 * ensure we reset the PMC. In all cases the PMC will be 256 or less
351 * cycles from overflow.
353 * We only do this if the first pass fails to find any overflowing
354 * PMCs because a user might set a period of less than 256 and we
355 * don't want to mistakenly reset them.
357 if (pvr_version_is(PVR_POWER7) && ((0x80000000 - val) <= 256))
358 return true;
360 return false;
363 static void power4_handle_interrupt(struct pt_regs *regs,
364 struct op_counter_config *ctr)
366 unsigned long pc;
367 int is_kernel;
368 int val;
369 int i;
370 unsigned int mmcr0;
371 unsigned long mmcra;
372 bool siar_valid = false;
374 mmcra = mfspr(SPRN_MMCRA);
376 pc = get_pc(regs);
377 is_kernel = get_kernel(pc, mmcra);
379 /* set the PMM bit (see comment below) */
380 mtmsr(mfmsr() | MSR_PMM);
382 /* Check that the SIAR valid bit in MMCRA is set to 1. */
383 if ((mmcra & MMCRA_SIAR_VALID_MASK) == MMCRA_SIAR_VALID_MASK)
384 siar_valid = true;
386 for (i = 0; i < cur_cpu_spec->num_pmcs; ++i) {
387 val = classic_ctr_read(i);
388 if (pmc_overflow(val)) {
389 if (oprofile_running && ctr[i].enabled) {
390 /* Power 7+ and newer architectures:
391 * If the event is a marked event, then only
392 * save the sample if the SIAR valid bit is
393 * set. If the event is not marked, then
394 * always save the sample.
395 * Note, the Sample enable bit in the MMCRA
396 * register must be set to 1 if the group
397 * contains a marked event.
399 if ((siar_valid &&
400 (cntr_marked_events & (1 << i)))
401 || !(cntr_marked_events & (1 << i)))
402 oprofile_add_ext_sample(pc, regs, i,
403 is_kernel);
405 classic_ctr_write(i, reset_value[i]);
406 } else {
407 classic_ctr_write(i, 0);
412 mmcr0 = mfspr(SPRN_MMCR0);
414 /* reset the perfmon trigger */
415 mmcr0 |= MMCR0_PMXE;
418 * We must clear the PMAO bit on some (GQ) chips. Just do it
419 * all the time
421 mmcr0 &= ~MMCR0_PMAO;
423 /* Clear the appropriate bits in the MMCRA */
424 mmcra &= ~cur_cpu_spec->oprofile_mmcra_clear;
425 mtspr(SPRN_MMCRA, mmcra);
428 * now clear the freeze bit, counting will not start until we
429 * rfid from this exception, because only at that point will
430 * the PMM bit be cleared
432 mmcr0 &= ~MMCR0_FC;
433 mtspr(SPRN_MMCR0, mmcr0);
436 struct op_powerpc_model op_model_power4 = {
437 .reg_setup = power4_reg_setup,
438 .cpu_setup = power4_cpu_setup,
439 .start = power4_start,
440 .stop = power4_stop,
441 .handle_interrupt = power4_handle_interrupt,