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_switch_ctrl(struct op_x86_model_spec const *model,
122 struct op_msrs const * const msrs)
127 /* enable active counters */
128 for (i = 0; i < NUM_COUNTERS; ++i) {
129 int virt = op_x86_phys_to_virt(i);
130 if (!counter_config[virt].enabled)
132 rdmsrl(msrs->controls[i].addr, val);
133 val &= model->reserved;
134 val |= op_x86_get_ctrl(model, &counter_config[virt]);
135 wrmsrl(msrs->controls[i].addr, val);
141 /* functions for op_amd_spec */
143 static void op_amd_fill_in_addresses(struct op_msrs * const msrs)
147 for (i = 0; i < NUM_COUNTERS; i++) {
148 if (reserve_perfctr_nmi(MSR_K7_PERFCTR0 + i))
149 msrs->counters[i].addr = MSR_K7_PERFCTR0 + i;
152 for (i = 0; i < NUM_CONTROLS; i++) {
153 if (reserve_evntsel_nmi(MSR_K7_EVNTSEL0 + i))
154 msrs->controls[i].addr = MSR_K7_EVNTSEL0 + i;
158 static void op_amd_setup_ctrs(struct op_x86_model_spec const *model,
159 struct op_msrs const * const msrs)
164 /* setup reset_value */
165 for (i = 0; i < NUM_VIRT_COUNTERS; ++i) {
166 if (counter_config[i].enabled)
167 reset_value[i] = counter_config[i].count;
172 /* clear all counters */
173 for (i = 0; i < NUM_CONTROLS; ++i) {
174 if (unlikely(!msrs->controls[i].addr)) {
175 if (counter_config[i].enabled && !smp_processor_id())
177 * counter is reserved, this is on all
178 * cpus, so report only for cpu #0
180 op_x86_warn_reserved(i);
183 rdmsrl(msrs->controls[i].addr, val);
184 if (val & ARCH_PERFMON_EVENTSEL0_ENABLE)
185 op_x86_warn_in_use(i);
186 val &= model->reserved;
187 wrmsrl(msrs->controls[i].addr, val);
190 /* avoid a false detection of ctr overflows in NMI handler */
191 for (i = 0; i < NUM_COUNTERS; ++i) {
192 if (unlikely(!msrs->counters[i].addr))
194 wrmsrl(msrs->counters[i].addr, -1LL);
197 /* enable active counters */
198 for (i = 0; i < NUM_COUNTERS; ++i) {
199 int virt = op_x86_phys_to_virt(i);
200 if (!counter_config[virt].enabled)
202 if (!msrs->counters[i].addr)
205 /* setup counter registers */
206 wrmsrl(msrs->counters[i].addr, -(u64)reset_value[virt]);
208 /* setup control registers */
209 rdmsrl(msrs->controls[i].addr, val);
210 val &= model->reserved;
211 val |= op_x86_get_ctrl(model, &counter_config[virt]);
212 wrmsrl(msrs->controls[i].addr, val);
217 * 16-bit Linear Feedback Shift Register (LFSR)
220 * Feedback polynomial = X + X + X + X + 1
222 static unsigned int lfsr_random(void)
224 static unsigned int lfsr_value = 0xF00D;
227 /* Compute next bit to shift in */
228 bit = ((lfsr_value >> 0) ^
231 (lfsr_value >> 5)) & 0x0001;
233 /* Advance to next register value */
234 lfsr_value = (lfsr_value >> 1) | (bit << 15);
240 * IBS software randomization
242 * The IBS periodic op counter is randomized in software. The lower 12
243 * bits of the 20 bit counter are randomized. IbsOpCurCnt is
244 * initialized with a 12 bit random value.
246 static inline u64 op_amd_randomize_ibs_op(u64 val)
248 unsigned int random = lfsr_random();
250 if (!(ibs_caps & IBS_CAPS_RDWROPCNT))
252 * Work around if the hw can not write to IbsOpCurCnt
254 * Randomize the lower 8 bits of the 16 bit
255 * IbsOpMaxCnt [15:0] value in the range of -128 to
256 * +127 by adding/subtracting an offset to the
257 * maximum count (IbsOpMaxCnt).
259 * To avoid over or underflows and protect upper bits
260 * starting at bit 16, the initial value for
261 * IbsOpMaxCnt must fit in the range from 0x0081 to
264 val += (s8)(random >> 4);
266 val |= (u64)(random & IBS_RANDOM_MASK) << 32;
272 op_amd_handle_ibs(struct pt_regs * const regs,
273 struct op_msrs const * const msrs)
276 struct op_entry entry;
281 if (ibs_config.fetch_enabled) {
282 rdmsrl(MSR_AMD64_IBSFETCHCTL, ctl);
283 if (ctl & IBS_FETCH_VAL) {
284 rdmsrl(MSR_AMD64_IBSFETCHLINAD, val);
285 oprofile_write_reserve(&entry, regs, val,
286 IBS_FETCH_CODE, IBS_FETCH_SIZE);
287 oprofile_add_data64(&entry, val);
288 oprofile_add_data64(&entry, ctl);
289 rdmsrl(MSR_AMD64_IBSFETCHPHYSAD, val);
290 oprofile_add_data64(&entry, val);
291 oprofile_write_commit(&entry);
293 /* reenable the IRQ */
294 ctl &= ~(IBS_FETCH_VAL | IBS_FETCH_CNT_MASK);
295 ctl |= IBS_FETCH_ENABLE;
296 wrmsrl(MSR_AMD64_IBSFETCHCTL, ctl);
300 if (ibs_config.op_enabled) {
301 rdmsrl(MSR_AMD64_IBSOPCTL, ctl);
302 if (ctl & IBS_OP_VAL) {
303 rdmsrl(MSR_AMD64_IBSOPRIP, val);
304 oprofile_write_reserve(&entry, regs, val,
305 IBS_OP_CODE, IBS_OP_SIZE);
306 oprofile_add_data64(&entry, val);
307 rdmsrl(MSR_AMD64_IBSOPDATA, val);
308 oprofile_add_data64(&entry, val);
309 rdmsrl(MSR_AMD64_IBSOPDATA2, val);
310 oprofile_add_data64(&entry, val);
311 rdmsrl(MSR_AMD64_IBSOPDATA3, val);
312 oprofile_add_data64(&entry, val);
313 rdmsrl(MSR_AMD64_IBSDCLINAD, val);
314 oprofile_add_data64(&entry, val);
315 rdmsrl(MSR_AMD64_IBSDCPHYSAD, val);
316 oprofile_add_data64(&entry, val);
317 oprofile_write_commit(&entry);
319 /* reenable the IRQ */
320 ctl = op_amd_randomize_ibs_op(ibs_op_ctl);
321 wrmsrl(MSR_AMD64_IBSOPCTL, ctl);
326 static inline void op_amd_start_ibs(void)
333 if (ibs_config.fetch_enabled) {
334 val = (ibs_config.max_cnt_fetch >> 4) & 0xFFFF;
335 val |= ibs_config.rand_en ? IBS_FETCH_RAND_EN : 0;
336 val |= IBS_FETCH_ENABLE;
337 wrmsrl(MSR_AMD64_IBSFETCHCTL, val);
340 if (ibs_config.op_enabled) {
341 ibs_op_ctl = ibs_config.max_cnt_op >> 4;
342 if (!(ibs_caps & IBS_CAPS_RDWROPCNT)) {
344 * IbsOpCurCnt not supported. See
345 * op_amd_randomize_ibs_op() for details.
347 ibs_op_ctl = clamp(ibs_op_ctl, 0x0081ULL, 0xFF80ULL);
350 * The start value is randomized with a
351 * positive offset, we need to compensate it
352 * with the half of the randomized range. Also
355 ibs_op_ctl = min(ibs_op_ctl + IBS_RANDOM_MAXCNT_OFFSET,
358 if (ibs_caps & IBS_CAPS_OPCNT && ibs_config.dispatched_ops)
359 ibs_op_ctl |= IBS_OP_CNT_CTL;
360 ibs_op_ctl |= IBS_OP_ENABLE;
361 val = op_amd_randomize_ibs_op(ibs_op_ctl);
362 wrmsrl(MSR_AMD64_IBSOPCTL, val);
366 static void op_amd_stop_ibs(void)
371 if (ibs_config.fetch_enabled)
372 /* clear max count and enable */
373 wrmsrl(MSR_AMD64_IBSFETCHCTL, 0);
375 if (ibs_config.op_enabled)
376 /* clear max count and enable */
377 wrmsrl(MSR_AMD64_IBSOPCTL, 0);
380 static int op_amd_check_ctrs(struct pt_regs * const regs,
381 struct op_msrs const * const msrs)
386 for (i = 0; i < NUM_COUNTERS; ++i) {
387 int virt = op_x86_phys_to_virt(i);
388 if (!reset_value[virt])
390 rdmsrl(msrs->counters[i].addr, val);
391 /* bit is clear if overflowed: */
392 if (val & OP_CTR_OVERFLOW)
394 oprofile_add_sample(regs, virt);
395 wrmsrl(msrs->counters[i].addr, -(u64)reset_value[virt]);
398 op_amd_handle_ibs(regs, msrs);
400 /* See op_model_ppro.c */
404 static void op_amd_start(struct op_msrs const * const msrs)
409 for (i = 0; i < NUM_COUNTERS; ++i) {
410 if (!reset_value[op_x86_phys_to_virt(i)])
412 rdmsrl(msrs->controls[i].addr, val);
413 val |= ARCH_PERFMON_EVENTSEL0_ENABLE;
414 wrmsrl(msrs->controls[i].addr, val);
420 static void op_amd_stop(struct op_msrs const * const msrs)
426 * Subtle: stop on all counters to avoid race with setting our
429 for (i = 0; i < NUM_COUNTERS; ++i) {
430 if (!reset_value[op_x86_phys_to_virt(i)])
432 rdmsrl(msrs->controls[i].addr, val);
433 val &= ~ARCH_PERFMON_EVENTSEL0_ENABLE;
434 wrmsrl(msrs->controls[i].addr, val);
440 static void op_amd_shutdown(struct op_msrs const * const msrs)
444 for (i = 0; i < NUM_COUNTERS; ++i) {
445 if (msrs->counters[i].addr)
446 release_perfctr_nmi(MSR_K7_PERFCTR0 + i);
448 for (i = 0; i < NUM_CONTROLS; ++i) {
449 if (msrs->controls[i].addr)
450 release_evntsel_nmi(MSR_K7_EVNTSEL0 + i);
454 static u8 ibs_eilvt_off;
456 static inline void apic_init_ibs_nmi_per_cpu(void *arg)
458 ibs_eilvt_off = setup_APIC_eilvt_ibs(0, APIC_EILVT_MSG_NMI, 0);
461 static inline void apic_clear_ibs_nmi_per_cpu(void *arg)
463 setup_APIC_eilvt_ibs(0, APIC_EILVT_MSG_FIX, 1);
466 static int init_ibs_nmi(void)
468 #define IBSCTL_LVTOFFSETVAL (1 << 8)
470 struct pci_dev *cpu_cfg;
475 on_each_cpu(apic_init_ibs_nmi_per_cpu, NULL, 1);
480 cpu_cfg = pci_get_device(PCI_VENDOR_ID_AMD,
481 PCI_DEVICE_ID_AMD_10H_NB_MISC,
486 pci_write_config_dword(cpu_cfg, IBSCTL, ibs_eilvt_off
487 | IBSCTL_LVTOFFSETVAL);
488 pci_read_config_dword(cpu_cfg, IBSCTL, &value);
489 if (value != (ibs_eilvt_off | IBSCTL_LVTOFFSETVAL)) {
490 pci_dev_put(cpu_cfg);
491 printk(KERN_DEBUG "Failed to setup IBS LVT offset, "
492 "IBSCTL = 0x%08x", value);
498 printk(KERN_DEBUG "No CPU node configured for IBS");
505 /* uninitialize the APIC for the IBS interrupts if needed */
506 static void clear_ibs_nmi(void)
509 on_each_cpu(apic_clear_ibs_nmi_per_cpu, NULL, 1);
512 /* initialize the APIC for the IBS interrupts if available */
513 static void ibs_init(void)
515 ibs_caps = get_ibs_caps();
520 if (init_ibs_nmi()) {
525 printk(KERN_INFO "oprofile: AMD IBS detected (0x%08x)\n",
529 static void ibs_exit(void)
537 static int (*create_arch_files)(struct super_block *sb, struct dentry *root);
539 static int setup_ibs_files(struct super_block *sb, struct dentry *root)
544 /* architecture specific files */
545 if (create_arch_files)
546 ret = create_arch_files(sb, root);
554 /* model specific files */
556 /* setup some reasonable defaults */
557 ibs_config.max_cnt_fetch = 250000;
558 ibs_config.fetch_enabled = 0;
559 ibs_config.max_cnt_op = 250000;
560 ibs_config.op_enabled = 0;
561 ibs_config.dispatched_ops = 0;
563 dir = oprofilefs_mkdir(sb, root, "ibs_fetch");
564 oprofilefs_create_ulong(sb, dir, "enable",
565 &ibs_config.fetch_enabled);
566 oprofilefs_create_ulong(sb, dir, "max_count",
567 &ibs_config.max_cnt_fetch);
568 oprofilefs_create_ulong(sb, dir, "rand_enable",
569 &ibs_config.rand_en);
571 dir = oprofilefs_mkdir(sb, root, "ibs_op");
572 oprofilefs_create_ulong(sb, dir, "enable",
573 &ibs_config.op_enabled);
574 oprofilefs_create_ulong(sb, dir, "max_count",
575 &ibs_config.max_cnt_op);
576 if (ibs_caps & IBS_CAPS_OPCNT)
577 oprofilefs_create_ulong(sb, dir, "dispatched_ops",
578 &ibs_config.dispatched_ops);
583 static int op_amd_init(struct oprofile_operations *ops)
586 create_arch_files = ops->create_files;
587 ops->create_files = setup_ibs_files;
591 static void op_amd_exit(void)
596 struct op_x86_model_spec op_amd_spec = {
597 .num_counters = NUM_COUNTERS,
598 .num_controls = NUM_CONTROLS,
599 .num_virt_counters = NUM_VIRT_COUNTERS,
600 .reserved = MSR_AMD_EVENTSEL_RESERVED,
601 .event_mask = OP_EVENT_MASK,
604 .fill_in_addresses = &op_amd_fill_in_addresses,
605 .setup_ctrs = &op_amd_setup_ctrs,
606 .check_ctrs = &op_amd_check_ctrs,
607 .start = &op_amd_start,
608 .stop = &op_amd_stop,
609 .shutdown = &op_amd_shutdown,
610 #ifdef CONFIG_OPROFILE_EVENT_MULTIPLEX
611 .switch_ctrl = &op_mux_switch_ctrl,