3 * athlon / K7 / K8 / Family 10h model-specific MSR operations
5 * @remark Copyright 2002-2009 OProfile authors
6 * @remark Read the file COPYING
9 * @author Philippe Elie
10 * @author Graydon Hoare
11 * @author Robert Richter <robert.richter@amd.com>
12 * @author Barry Kasindorf <barry.kasindorf@amd.com>
13 * @author Jason Yeh <jason.yeh@amd.com>
14 * @author Suravee Suthikulpanit <suravee.suthikulpanit@amd.com>
17 #include <linux/oprofile.h>
18 #include <linux/device.h>
19 #include <linux/pci.h>
20 #include <linux/percpu.h>
22 #include <asm/ptrace.h>
26 #include <asm/processor.h>
27 #include <asm/cpufeature.h>
29 #include "op_x86_model.h"
30 #include "op_counter.h"
32 #define NUM_COUNTERS 4
33 #define NUM_CONTROLS 4
34 #ifdef CONFIG_OPROFILE_EVENT_MULTIPLEX
35 #define NUM_VIRT_COUNTERS 32
36 #define NUM_VIRT_CONTROLS 32
38 #define NUM_VIRT_COUNTERS NUM_COUNTERS
39 #define NUM_VIRT_CONTROLS NUM_CONTROLS
42 #define OP_EVENT_MASK 0x0FFF
43 #define OP_CTR_OVERFLOW (1ULL<<31)
45 #define MSR_AMD_EVENTSEL_RESERVED ((0xFFFFFCF0ULL<<32)|(1ULL<<21))
47 static unsigned long reset_value[NUM_VIRT_COUNTERS];
49 /* IbsFetchCtl bits/masks */
50 #define IBS_FETCH_RAND_EN (1ULL<<57)
51 #define IBS_FETCH_VAL (1ULL<<49)
52 #define IBS_FETCH_ENABLE (1ULL<<48)
53 #define IBS_FETCH_CNT_MASK 0xFFFF0000ULL
56 #define IBS_OP_CNT_CTL (1ULL<<19)
57 #define IBS_OP_VAL (1ULL<<18)
58 #define IBS_OP_ENABLE (1ULL<<17)
60 #define IBS_FETCH_SIZE 6
61 #define IBS_OP_SIZE 12
65 struct op_ibs_config {
66 unsigned long op_enabled;
67 unsigned long fetch_enabled;
68 unsigned long max_cnt_fetch;
69 unsigned long max_cnt_op;
70 unsigned long rand_en;
71 unsigned long dispatched_ops;
74 static struct op_ibs_config ibs_config;
75 static u64 ibs_op_ctl;
78 * IBS cpuid feature detection
81 #define IBS_CPUID_FEATURES 0x8000001b
84 * Same bit mask as for IBS cpuid feature flags (Fn8000_001B_EAX), but
85 * bit 0 is used to indicate the existence of IBS.
87 #define IBS_CAPS_AVAIL (1LL<<0)
88 #define IBS_CAPS_RDWROPCNT (1LL<<3)
89 #define IBS_CAPS_OPCNT (1LL<<4)
92 * IBS randomization macros
94 #define IBS_RANDOM_BITS 12
95 #define IBS_RANDOM_MASK ((1ULL << IBS_RANDOM_BITS) - 1)
96 #define IBS_RANDOM_MAXCNT_OFFSET (1ULL << (IBS_RANDOM_BITS - 5))
98 static u32 get_ibs_caps(void)
101 unsigned int max_level;
103 if (!boot_cpu_has(X86_FEATURE_IBS))
106 /* check IBS cpuid feature flags */
107 max_level = cpuid_eax(0x80000000);
108 if (max_level < IBS_CPUID_FEATURES)
109 return IBS_CAPS_AVAIL;
111 ibs_caps = cpuid_eax(IBS_CPUID_FEATURES);
112 if (!(ibs_caps & IBS_CAPS_AVAIL))
113 /* cpuid flags not valid */
114 return IBS_CAPS_AVAIL;
119 #ifdef CONFIG_OPROFILE_EVENT_MULTIPLEX
121 static void op_mux_fill_in_addresses(struct op_msrs * const msrs)
125 for (i = 0; i < NUM_VIRT_COUNTERS; i++) {
126 int hw_counter = op_x86_virt_to_phys(i);
127 if (reserve_perfctr_nmi(MSR_K7_PERFCTR0 + i))
128 msrs->multiplex[i].addr = MSR_K7_PERFCTR0 + hw_counter;
130 msrs->multiplex[i].addr = 0;
134 static void op_mux_switch_ctrl(struct op_x86_model_spec const *model,
135 struct op_msrs const * const msrs)
140 /* enable active counters */
141 for (i = 0; i < NUM_COUNTERS; ++i) {
142 int virt = op_x86_phys_to_virt(i);
143 if (!counter_config[virt].enabled)
145 rdmsrl(msrs->controls[i].addr, val);
146 val &= model->reserved;
147 val |= op_x86_get_ctrl(model, &counter_config[virt]);
148 wrmsrl(msrs->controls[i].addr, val);
154 static inline void op_mux_fill_in_addresses(struct op_msrs * const msrs) { }
158 /* functions for op_amd_spec */
160 static void op_amd_fill_in_addresses(struct op_msrs * const msrs)
164 for (i = 0; i < NUM_COUNTERS; i++) {
165 if (reserve_perfctr_nmi(MSR_K7_PERFCTR0 + i))
166 msrs->counters[i].addr = MSR_K7_PERFCTR0 + i;
168 msrs->counters[i].addr = 0;
171 for (i = 0; i < NUM_CONTROLS; i++) {
172 if (reserve_evntsel_nmi(MSR_K7_EVNTSEL0 + i))
173 msrs->controls[i].addr = MSR_K7_EVNTSEL0 + i;
175 msrs->controls[i].addr = 0;
178 op_mux_fill_in_addresses(msrs);
181 static void op_amd_setup_ctrs(struct op_x86_model_spec const *model,
182 struct op_msrs const * const msrs)
187 /* setup reset_value */
188 for (i = 0; i < NUM_VIRT_COUNTERS; ++i) {
189 if (counter_config[i].enabled)
190 reset_value[i] = counter_config[i].count;
195 /* clear all counters */
196 for (i = 0; i < NUM_CONTROLS; ++i) {
197 if (unlikely(!msrs->controls[i].addr)) {
198 if (counter_config[i].enabled && !smp_processor_id())
200 * counter is reserved, this is on all
201 * cpus, so report only for cpu #0
203 op_x86_warn_reserved(i);
206 rdmsrl(msrs->controls[i].addr, val);
207 if (val & ARCH_PERFMON_EVENTSEL0_ENABLE)
208 op_x86_warn_in_use(i);
209 val &= model->reserved;
210 wrmsrl(msrs->controls[i].addr, val);
213 /* avoid a false detection of ctr overflows in NMI handler */
214 for (i = 0; i < NUM_COUNTERS; ++i) {
215 if (unlikely(!msrs->counters[i].addr))
217 wrmsrl(msrs->counters[i].addr, -1LL);
220 /* enable active counters */
221 for (i = 0; i < NUM_COUNTERS; ++i) {
222 int virt = op_x86_phys_to_virt(i);
223 if (!counter_config[virt].enabled)
225 if (!msrs->counters[i].addr)
228 /* setup counter registers */
229 wrmsrl(msrs->counters[i].addr, -(u64)reset_value[virt]);
231 /* setup control registers */
232 rdmsrl(msrs->controls[i].addr, val);
233 val &= model->reserved;
234 val |= op_x86_get_ctrl(model, &counter_config[virt]);
235 wrmsrl(msrs->controls[i].addr, val);
240 * 16-bit Linear Feedback Shift Register (LFSR)
243 * Feedback polynomial = X + X + X + X + 1
245 static unsigned int lfsr_random(void)
247 static unsigned int lfsr_value = 0xF00D;
250 /* Compute next bit to shift in */
251 bit = ((lfsr_value >> 0) ^
254 (lfsr_value >> 5)) & 0x0001;
256 /* Advance to next register value */
257 lfsr_value = (lfsr_value >> 1) | (bit << 15);
263 * IBS software randomization
265 * The IBS periodic op counter is randomized in software. The lower 12
266 * bits of the 20 bit counter are randomized. IbsOpCurCnt is
267 * initialized with a 12 bit random value.
269 static inline u64 op_amd_randomize_ibs_op(u64 val)
271 unsigned int random = lfsr_random();
273 if (!(ibs_caps & IBS_CAPS_RDWROPCNT))
275 * Work around if the hw can not write to IbsOpCurCnt
277 * Randomize the lower 8 bits of the 16 bit
278 * IbsOpMaxCnt [15:0] value in the range of -128 to
279 * +127 by adding/subtracting an offset to the
280 * maximum count (IbsOpMaxCnt).
282 * To avoid over or underflows and protect upper bits
283 * starting at bit 16, the initial value for
284 * IbsOpMaxCnt must fit in the range from 0x0081 to
287 val += (s8)(random >> 4);
289 val |= (u64)(random & IBS_RANDOM_MASK) << 32;
295 op_amd_handle_ibs(struct pt_regs * const regs,
296 struct op_msrs const * const msrs)
299 struct op_entry entry;
304 if (ibs_config.fetch_enabled) {
305 rdmsrl(MSR_AMD64_IBSFETCHCTL, ctl);
306 if (ctl & IBS_FETCH_VAL) {
307 rdmsrl(MSR_AMD64_IBSFETCHLINAD, val);
308 oprofile_write_reserve(&entry, regs, val,
309 IBS_FETCH_CODE, IBS_FETCH_SIZE);
310 oprofile_add_data64(&entry, val);
311 oprofile_add_data64(&entry, ctl);
312 rdmsrl(MSR_AMD64_IBSFETCHPHYSAD, val);
313 oprofile_add_data64(&entry, val);
314 oprofile_write_commit(&entry);
316 /* reenable the IRQ */
317 ctl &= ~(IBS_FETCH_VAL | IBS_FETCH_CNT_MASK);
318 ctl |= IBS_FETCH_ENABLE;
319 wrmsrl(MSR_AMD64_IBSFETCHCTL, ctl);
323 if (ibs_config.op_enabled) {
324 rdmsrl(MSR_AMD64_IBSOPCTL, ctl);
325 if (ctl & IBS_OP_VAL) {
326 rdmsrl(MSR_AMD64_IBSOPRIP, val);
327 oprofile_write_reserve(&entry, regs, val,
328 IBS_OP_CODE, IBS_OP_SIZE);
329 oprofile_add_data64(&entry, val);
330 rdmsrl(MSR_AMD64_IBSOPDATA, val);
331 oprofile_add_data64(&entry, val);
332 rdmsrl(MSR_AMD64_IBSOPDATA2, val);
333 oprofile_add_data64(&entry, val);
334 rdmsrl(MSR_AMD64_IBSOPDATA3, val);
335 oprofile_add_data64(&entry, val);
336 rdmsrl(MSR_AMD64_IBSDCLINAD, val);
337 oprofile_add_data64(&entry, val);
338 rdmsrl(MSR_AMD64_IBSDCPHYSAD, val);
339 oprofile_add_data64(&entry, val);
340 oprofile_write_commit(&entry);
342 /* reenable the IRQ */
343 ctl = op_amd_randomize_ibs_op(ibs_op_ctl);
344 wrmsrl(MSR_AMD64_IBSOPCTL, ctl);
349 static inline void op_amd_start_ibs(void)
356 if (ibs_config.fetch_enabled) {
357 val = (ibs_config.max_cnt_fetch >> 4) & 0xFFFF;
358 val |= ibs_config.rand_en ? IBS_FETCH_RAND_EN : 0;
359 val |= IBS_FETCH_ENABLE;
360 wrmsrl(MSR_AMD64_IBSFETCHCTL, val);
363 if (ibs_config.op_enabled) {
364 ibs_op_ctl = ibs_config.max_cnt_op >> 4;
365 if (!(ibs_caps & IBS_CAPS_RDWROPCNT)) {
367 * IbsOpCurCnt not supported. See
368 * op_amd_randomize_ibs_op() for details.
370 ibs_op_ctl = clamp(ibs_op_ctl, 0x0081ULL, 0xFF80ULL);
373 * The start value is randomized with a
374 * positive offset, we need to compensate it
375 * with the half of the randomized range. Also
378 ibs_op_ctl = min(ibs_op_ctl + IBS_RANDOM_MAXCNT_OFFSET,
381 if (ibs_caps & IBS_CAPS_OPCNT && ibs_config.dispatched_ops)
382 ibs_op_ctl |= IBS_OP_CNT_CTL;
383 ibs_op_ctl |= IBS_OP_ENABLE;
384 val = op_amd_randomize_ibs_op(ibs_op_ctl);
385 wrmsrl(MSR_AMD64_IBSOPCTL, val);
389 static void op_amd_stop_ibs(void)
394 if (ibs_config.fetch_enabled)
395 /* clear max count and enable */
396 wrmsrl(MSR_AMD64_IBSFETCHCTL, 0);
398 if (ibs_config.op_enabled)
399 /* clear max count and enable */
400 wrmsrl(MSR_AMD64_IBSOPCTL, 0);
403 static int op_amd_check_ctrs(struct pt_regs * const regs,
404 struct op_msrs const * const msrs)
409 for (i = 0; i < NUM_COUNTERS; ++i) {
410 int virt = op_x86_phys_to_virt(i);
411 if (!reset_value[virt])
413 rdmsrl(msrs->counters[i].addr, val);
414 /* bit is clear if overflowed: */
415 if (val & OP_CTR_OVERFLOW)
417 oprofile_add_sample(regs, virt);
418 wrmsrl(msrs->counters[i].addr, -(u64)reset_value[virt]);
421 op_amd_handle_ibs(regs, msrs);
423 /* See op_model_ppro.c */
427 static void op_amd_start(struct op_msrs const * const msrs)
432 for (i = 0; i < NUM_COUNTERS; ++i) {
433 if (!reset_value[op_x86_phys_to_virt(i)])
435 rdmsrl(msrs->controls[i].addr, val);
436 val |= ARCH_PERFMON_EVENTSEL0_ENABLE;
437 wrmsrl(msrs->controls[i].addr, val);
443 static void op_amd_stop(struct op_msrs const * const msrs)
449 * Subtle: stop on all counters to avoid race with setting our
452 for (i = 0; i < NUM_COUNTERS; ++i) {
453 if (!reset_value[op_x86_phys_to_virt(i)])
455 rdmsrl(msrs->controls[i].addr, val);
456 val &= ~ARCH_PERFMON_EVENTSEL0_ENABLE;
457 wrmsrl(msrs->controls[i].addr, val);
463 static void op_amd_shutdown(struct op_msrs const * const msrs)
467 for (i = 0; i < NUM_COUNTERS; ++i) {
468 if (msrs->counters[i].addr)
469 release_perfctr_nmi(MSR_K7_PERFCTR0 + i);
471 for (i = 0; i < NUM_CONTROLS; ++i) {
472 if (msrs->controls[i].addr)
473 release_evntsel_nmi(MSR_K7_EVNTSEL0 + i);
477 static u8 ibs_eilvt_off;
479 static inline void apic_init_ibs_nmi_per_cpu(void *arg)
481 ibs_eilvt_off = setup_APIC_eilvt_ibs(0, APIC_EILVT_MSG_NMI, 0);
484 static inline void apic_clear_ibs_nmi_per_cpu(void *arg)
486 setup_APIC_eilvt_ibs(0, APIC_EILVT_MSG_FIX, 1);
489 static int init_ibs_nmi(void)
491 #define IBSCTL_LVTOFFSETVAL (1 << 8)
493 struct pci_dev *cpu_cfg;
498 on_each_cpu(apic_init_ibs_nmi_per_cpu, NULL, 1);
503 cpu_cfg = pci_get_device(PCI_VENDOR_ID_AMD,
504 PCI_DEVICE_ID_AMD_10H_NB_MISC,
509 pci_write_config_dword(cpu_cfg, IBSCTL, ibs_eilvt_off
510 | IBSCTL_LVTOFFSETVAL);
511 pci_read_config_dword(cpu_cfg, IBSCTL, &value);
512 if (value != (ibs_eilvt_off | IBSCTL_LVTOFFSETVAL)) {
513 pci_dev_put(cpu_cfg);
514 printk(KERN_DEBUG "Failed to setup IBS LVT offset, "
515 "IBSCTL = 0x%08x", value);
521 printk(KERN_DEBUG "No CPU node configured for IBS");
528 /* uninitialize the APIC for the IBS interrupts if needed */
529 static void clear_ibs_nmi(void)
532 on_each_cpu(apic_clear_ibs_nmi_per_cpu, NULL, 1);
535 /* initialize the APIC for the IBS interrupts if available */
536 static void ibs_init(void)
538 ibs_caps = get_ibs_caps();
543 if (init_ibs_nmi()) {
548 printk(KERN_INFO "oprofile: AMD IBS detected (0x%08x)\n",
552 static void ibs_exit(void)
560 static int (*create_arch_files)(struct super_block *sb, struct dentry *root);
562 static int setup_ibs_files(struct super_block *sb, struct dentry *root)
567 /* architecture specific files */
568 if (create_arch_files)
569 ret = create_arch_files(sb, root);
577 /* model specific files */
579 /* setup some reasonable defaults */
580 ibs_config.max_cnt_fetch = 250000;
581 ibs_config.fetch_enabled = 0;
582 ibs_config.max_cnt_op = 250000;
583 ibs_config.op_enabled = 0;
584 ibs_config.dispatched_ops = 0;
586 dir = oprofilefs_mkdir(sb, root, "ibs_fetch");
587 oprofilefs_create_ulong(sb, dir, "enable",
588 &ibs_config.fetch_enabled);
589 oprofilefs_create_ulong(sb, dir, "max_count",
590 &ibs_config.max_cnt_fetch);
591 oprofilefs_create_ulong(sb, dir, "rand_enable",
592 &ibs_config.rand_en);
594 dir = oprofilefs_mkdir(sb, root, "ibs_op");
595 oprofilefs_create_ulong(sb, dir, "enable",
596 &ibs_config.op_enabled);
597 oprofilefs_create_ulong(sb, dir, "max_count",
598 &ibs_config.max_cnt_op);
599 if (ibs_caps & IBS_CAPS_OPCNT)
600 oprofilefs_create_ulong(sb, dir, "dispatched_ops",
601 &ibs_config.dispatched_ops);
606 static int op_amd_init(struct oprofile_operations *ops)
609 create_arch_files = ops->create_files;
610 ops->create_files = setup_ibs_files;
614 static void op_amd_exit(void)
619 struct op_x86_model_spec op_amd_spec = {
620 .num_counters = NUM_COUNTERS,
621 .num_controls = NUM_CONTROLS,
622 .num_virt_counters = NUM_VIRT_COUNTERS,
623 .reserved = MSR_AMD_EVENTSEL_RESERVED,
624 .event_mask = OP_EVENT_MASK,
627 .fill_in_addresses = &op_amd_fill_in_addresses,
628 .setup_ctrs = &op_amd_setup_ctrs,
629 .check_ctrs = &op_amd_check_ctrs,
630 .start = &op_amd_start,
631 .stop = &op_amd_stop,
632 .shutdown = &op_amd_shutdown,
633 #ifdef CONFIG_OPROFILE_EVENT_MULTIPLEX
634 .switch_ctrl = &op_mux_switch_ctrl,