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))
199 rdmsrl(msrs->controls[i].addr, val);
200 val &= model->reserved;
201 wrmsrl(msrs->controls[i].addr, val);
204 /* avoid a false detection of ctr overflows in NMI handler */
205 for (i = 0; i < NUM_COUNTERS; ++i) {
206 if (unlikely(!msrs->counters[i].addr))
208 wrmsrl(msrs->counters[i].addr, -1LL);
211 /* enable active counters */
212 for (i = 0; i < NUM_COUNTERS; ++i) {
213 int virt = op_x86_phys_to_virt(i);
214 if (!counter_config[virt].enabled)
216 if (!msrs->counters[i].addr)
219 /* setup counter registers */
220 wrmsrl(msrs->counters[i].addr, -(u64)reset_value[virt]);
222 /* setup control registers */
223 rdmsrl(msrs->controls[i].addr, val);
224 val &= model->reserved;
225 val |= op_x86_get_ctrl(model, &counter_config[virt]);
226 wrmsrl(msrs->controls[i].addr, val);
231 * 16-bit Linear Feedback Shift Register (LFSR)
234 * Feedback polynomial = X + X + X + X + 1
236 static unsigned int lfsr_random(void)
238 static unsigned int lfsr_value = 0xF00D;
241 /* Compute next bit to shift in */
242 bit = ((lfsr_value >> 0) ^
245 (lfsr_value >> 5)) & 0x0001;
247 /* Advance to next register value */
248 lfsr_value = (lfsr_value >> 1) | (bit << 15);
254 * IBS software randomization
256 * The IBS periodic op counter is randomized in software. The lower 12
257 * bits of the 20 bit counter are randomized. IbsOpCurCnt is
258 * initialized with a 12 bit random value.
260 static inline u64 op_amd_randomize_ibs_op(u64 val)
262 unsigned int random = lfsr_random();
264 if (!(ibs_caps & IBS_CAPS_RDWROPCNT))
266 * Work around if the hw can not write to IbsOpCurCnt
268 * Randomize the lower 8 bits of the 16 bit
269 * IbsOpMaxCnt [15:0] value in the range of -128 to
270 * +127 by adding/subtracting an offset to the
271 * maximum count (IbsOpMaxCnt).
273 * To avoid over or underflows and protect upper bits
274 * starting at bit 16, the initial value for
275 * IbsOpMaxCnt must fit in the range from 0x0081 to
278 val += (s8)(random >> 4);
280 val |= (u64)(random & IBS_RANDOM_MASK) << 32;
286 op_amd_handle_ibs(struct pt_regs * const regs,
287 struct op_msrs const * const msrs)
290 struct op_entry entry;
295 if (ibs_config.fetch_enabled) {
296 rdmsrl(MSR_AMD64_IBSFETCHCTL, ctl);
297 if (ctl & IBS_FETCH_VAL) {
298 rdmsrl(MSR_AMD64_IBSFETCHLINAD, val);
299 oprofile_write_reserve(&entry, regs, val,
300 IBS_FETCH_CODE, IBS_FETCH_SIZE);
301 oprofile_add_data64(&entry, val);
302 oprofile_add_data64(&entry, ctl);
303 rdmsrl(MSR_AMD64_IBSFETCHPHYSAD, val);
304 oprofile_add_data64(&entry, val);
305 oprofile_write_commit(&entry);
307 /* reenable the IRQ */
308 ctl &= ~(IBS_FETCH_VAL | IBS_FETCH_CNT_MASK);
309 ctl |= IBS_FETCH_ENABLE;
310 wrmsrl(MSR_AMD64_IBSFETCHCTL, ctl);
314 if (ibs_config.op_enabled) {
315 rdmsrl(MSR_AMD64_IBSOPCTL, ctl);
316 if (ctl & IBS_OP_VAL) {
317 rdmsrl(MSR_AMD64_IBSOPRIP, val);
318 oprofile_write_reserve(&entry, regs, val,
319 IBS_OP_CODE, IBS_OP_SIZE);
320 oprofile_add_data64(&entry, val);
321 rdmsrl(MSR_AMD64_IBSOPDATA, val);
322 oprofile_add_data64(&entry, val);
323 rdmsrl(MSR_AMD64_IBSOPDATA2, val);
324 oprofile_add_data64(&entry, val);
325 rdmsrl(MSR_AMD64_IBSOPDATA3, val);
326 oprofile_add_data64(&entry, val);
327 rdmsrl(MSR_AMD64_IBSDCLINAD, val);
328 oprofile_add_data64(&entry, val);
329 rdmsrl(MSR_AMD64_IBSDCPHYSAD, val);
330 oprofile_add_data64(&entry, val);
331 oprofile_write_commit(&entry);
333 /* reenable the IRQ */
334 ctl = op_amd_randomize_ibs_op(ibs_op_ctl);
335 wrmsrl(MSR_AMD64_IBSOPCTL, ctl);
340 static inline void op_amd_start_ibs(void)
347 if (ibs_config.fetch_enabled) {
348 val = (ibs_config.max_cnt_fetch >> 4) & 0xFFFF;
349 val |= ibs_config.rand_en ? IBS_FETCH_RAND_EN : 0;
350 val |= IBS_FETCH_ENABLE;
351 wrmsrl(MSR_AMD64_IBSFETCHCTL, val);
354 if (ibs_config.op_enabled) {
355 ibs_op_ctl = ibs_config.max_cnt_op >> 4;
356 if (!(ibs_caps & IBS_CAPS_RDWROPCNT)) {
358 * IbsOpCurCnt not supported. See
359 * op_amd_randomize_ibs_op() for details.
361 ibs_op_ctl = clamp(ibs_op_ctl, 0x0081ULL, 0xFF80ULL);
364 * The start value is randomized with a
365 * positive offset, we need to compensate it
366 * with the half of the randomized range. Also
369 ibs_op_ctl = min(ibs_op_ctl + IBS_RANDOM_MAXCNT_OFFSET,
372 if (ibs_caps & IBS_CAPS_OPCNT && ibs_config.dispatched_ops)
373 ibs_op_ctl |= IBS_OP_CNT_CTL;
374 ibs_op_ctl |= IBS_OP_ENABLE;
375 val = op_amd_randomize_ibs_op(ibs_op_ctl);
376 wrmsrl(MSR_AMD64_IBSOPCTL, val);
380 static void op_amd_stop_ibs(void)
385 if (ibs_config.fetch_enabled)
386 /* clear max count and enable */
387 wrmsrl(MSR_AMD64_IBSFETCHCTL, 0);
389 if (ibs_config.op_enabled)
390 /* clear max count and enable */
391 wrmsrl(MSR_AMD64_IBSOPCTL, 0);
394 static int op_amd_check_ctrs(struct pt_regs * const regs,
395 struct op_msrs const * const msrs)
400 for (i = 0; i < NUM_COUNTERS; ++i) {
401 int virt = op_x86_phys_to_virt(i);
402 if (!reset_value[virt])
404 rdmsrl(msrs->counters[i].addr, val);
405 /* bit is clear if overflowed: */
406 if (val & OP_CTR_OVERFLOW)
408 oprofile_add_sample(regs, virt);
409 wrmsrl(msrs->counters[i].addr, -(u64)reset_value[virt]);
412 op_amd_handle_ibs(regs, msrs);
414 /* See op_model_ppro.c */
418 static void op_amd_start(struct op_msrs const * const msrs)
423 for (i = 0; i < NUM_COUNTERS; ++i) {
424 if (!reset_value[op_x86_phys_to_virt(i)])
426 rdmsrl(msrs->controls[i].addr, val);
427 val |= ARCH_PERFMON_EVENTSEL0_ENABLE;
428 wrmsrl(msrs->controls[i].addr, val);
434 static void op_amd_stop(struct op_msrs const * const msrs)
440 * Subtle: stop on all counters to avoid race with setting our
443 for (i = 0; i < NUM_COUNTERS; ++i) {
444 if (!reset_value[op_x86_phys_to_virt(i)])
446 rdmsrl(msrs->controls[i].addr, val);
447 val &= ~ARCH_PERFMON_EVENTSEL0_ENABLE;
448 wrmsrl(msrs->controls[i].addr, val);
454 static void op_amd_shutdown(struct op_msrs const * const msrs)
458 for (i = 0; i < NUM_COUNTERS; ++i) {
459 if (msrs->counters[i].addr)
460 release_perfctr_nmi(MSR_K7_PERFCTR0 + i);
462 for (i = 0; i < NUM_CONTROLS; ++i) {
463 if (msrs->controls[i].addr)
464 release_evntsel_nmi(MSR_K7_EVNTSEL0 + i);
468 static u8 ibs_eilvt_off;
470 static inline void apic_init_ibs_nmi_per_cpu(void *arg)
472 ibs_eilvt_off = setup_APIC_eilvt_ibs(0, APIC_EILVT_MSG_NMI, 0);
475 static inline void apic_clear_ibs_nmi_per_cpu(void *arg)
477 setup_APIC_eilvt_ibs(0, APIC_EILVT_MSG_FIX, 1);
480 static int init_ibs_nmi(void)
482 #define IBSCTL_LVTOFFSETVAL (1 << 8)
484 struct pci_dev *cpu_cfg;
489 on_each_cpu(apic_init_ibs_nmi_per_cpu, NULL, 1);
494 cpu_cfg = pci_get_device(PCI_VENDOR_ID_AMD,
495 PCI_DEVICE_ID_AMD_10H_NB_MISC,
500 pci_write_config_dword(cpu_cfg, IBSCTL, ibs_eilvt_off
501 | IBSCTL_LVTOFFSETVAL);
502 pci_read_config_dword(cpu_cfg, IBSCTL, &value);
503 if (value != (ibs_eilvt_off | IBSCTL_LVTOFFSETVAL)) {
504 pci_dev_put(cpu_cfg);
505 printk(KERN_DEBUG "Failed to setup IBS LVT offset, "
506 "IBSCTL = 0x%08x", value);
512 printk(KERN_DEBUG "No CPU node configured for IBS");
519 /* uninitialize the APIC for the IBS interrupts if needed */
520 static void clear_ibs_nmi(void)
523 on_each_cpu(apic_clear_ibs_nmi_per_cpu, NULL, 1);
526 /* initialize the APIC for the IBS interrupts if available */
527 static void ibs_init(void)
529 ibs_caps = get_ibs_caps();
534 if (init_ibs_nmi()) {
539 printk(KERN_INFO "oprofile: AMD IBS detected (0x%08x)\n",
543 static void ibs_exit(void)
551 static int (*create_arch_files)(struct super_block *sb, struct dentry *root);
553 static int setup_ibs_files(struct super_block *sb, struct dentry *root)
558 /* architecture specific files */
559 if (create_arch_files)
560 ret = create_arch_files(sb, root);
568 /* model specific files */
570 /* setup some reasonable defaults */
571 ibs_config.max_cnt_fetch = 250000;
572 ibs_config.fetch_enabled = 0;
573 ibs_config.max_cnt_op = 250000;
574 ibs_config.op_enabled = 0;
575 ibs_config.dispatched_ops = 0;
577 dir = oprofilefs_mkdir(sb, root, "ibs_fetch");
578 oprofilefs_create_ulong(sb, dir, "enable",
579 &ibs_config.fetch_enabled);
580 oprofilefs_create_ulong(sb, dir, "max_count",
581 &ibs_config.max_cnt_fetch);
582 oprofilefs_create_ulong(sb, dir, "rand_enable",
583 &ibs_config.rand_en);
585 dir = oprofilefs_mkdir(sb, root, "ibs_op");
586 oprofilefs_create_ulong(sb, dir, "enable",
587 &ibs_config.op_enabled);
588 oprofilefs_create_ulong(sb, dir, "max_count",
589 &ibs_config.max_cnt_op);
590 if (ibs_caps & IBS_CAPS_OPCNT)
591 oprofilefs_create_ulong(sb, dir, "dispatched_ops",
592 &ibs_config.dispatched_ops);
597 static int op_amd_init(struct oprofile_operations *ops)
600 create_arch_files = ops->create_files;
601 ops->create_files = setup_ibs_files;
605 static void op_amd_exit(void)
610 struct op_x86_model_spec op_amd_spec = {
611 .num_counters = NUM_COUNTERS,
612 .num_controls = NUM_CONTROLS,
613 .num_virt_counters = NUM_VIRT_COUNTERS,
614 .reserved = MSR_AMD_EVENTSEL_RESERVED,
615 .event_mask = OP_EVENT_MASK,
618 .fill_in_addresses = &op_amd_fill_in_addresses,
619 .setup_ctrs = &op_amd_setup_ctrs,
620 .check_ctrs = &op_amd_check_ctrs,
621 .start = &op_amd_start,
622 .stop = &op_amd_stop,
623 .shutdown = &op_amd_shutdown,
624 #ifdef CONFIG_OPROFILE_EVENT_MULTIPLEX
625 .switch_ctrl = &op_mux_switch_ctrl,