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;
151 msrs->counters[i].addr = 0;
154 for (i = 0; i < NUM_CONTROLS; i++) {
155 if (reserve_evntsel_nmi(MSR_K7_EVNTSEL0 + i))
156 msrs->controls[i].addr = MSR_K7_EVNTSEL0 + i;
158 msrs->controls[i].addr = 0;
162 static void op_amd_setup_ctrs(struct op_x86_model_spec const *model,
163 struct op_msrs const * const msrs)
168 /* setup reset_value */
169 for (i = 0; i < NUM_VIRT_COUNTERS; ++i) {
170 if (counter_config[i].enabled)
171 reset_value[i] = counter_config[i].count;
176 /* clear all counters */
177 for (i = 0; i < NUM_CONTROLS; ++i) {
178 if (unlikely(!msrs->controls[i].addr)) {
179 if (counter_config[i].enabled && !smp_processor_id())
181 * counter is reserved, this is on all
182 * cpus, so report only for cpu #0
184 op_x86_warn_reserved(i);
187 rdmsrl(msrs->controls[i].addr, val);
188 if (val & ARCH_PERFMON_EVENTSEL0_ENABLE)
189 op_x86_warn_in_use(i);
190 val &= model->reserved;
191 wrmsrl(msrs->controls[i].addr, val);
194 /* avoid a false detection of ctr overflows in NMI handler */
195 for (i = 0; i < NUM_COUNTERS; ++i) {
196 if (unlikely(!msrs->counters[i].addr))
198 wrmsrl(msrs->counters[i].addr, -1LL);
201 /* enable active counters */
202 for (i = 0; i < NUM_COUNTERS; ++i) {
203 int virt = op_x86_phys_to_virt(i);
204 if (!counter_config[virt].enabled)
206 if (!msrs->counters[i].addr)
209 /* setup counter registers */
210 wrmsrl(msrs->counters[i].addr, -(u64)reset_value[virt]);
212 /* setup control registers */
213 rdmsrl(msrs->controls[i].addr, val);
214 val &= model->reserved;
215 val |= op_x86_get_ctrl(model, &counter_config[virt]);
216 wrmsrl(msrs->controls[i].addr, val);
221 * 16-bit Linear Feedback Shift Register (LFSR)
224 * Feedback polynomial = X + X + X + X + 1
226 static unsigned int lfsr_random(void)
228 static unsigned int lfsr_value = 0xF00D;
231 /* Compute next bit to shift in */
232 bit = ((lfsr_value >> 0) ^
235 (lfsr_value >> 5)) & 0x0001;
237 /* Advance to next register value */
238 lfsr_value = (lfsr_value >> 1) | (bit << 15);
244 * IBS software randomization
246 * The IBS periodic op counter is randomized in software. The lower 12
247 * bits of the 20 bit counter are randomized. IbsOpCurCnt is
248 * initialized with a 12 bit random value.
250 static inline u64 op_amd_randomize_ibs_op(u64 val)
252 unsigned int random = lfsr_random();
254 if (!(ibs_caps & IBS_CAPS_RDWROPCNT))
256 * Work around if the hw can not write to IbsOpCurCnt
258 * Randomize the lower 8 bits of the 16 bit
259 * IbsOpMaxCnt [15:0] value in the range of -128 to
260 * +127 by adding/subtracting an offset to the
261 * maximum count (IbsOpMaxCnt).
263 * To avoid over or underflows and protect upper bits
264 * starting at bit 16, the initial value for
265 * IbsOpMaxCnt must fit in the range from 0x0081 to
268 val += (s8)(random >> 4);
270 val |= (u64)(random & IBS_RANDOM_MASK) << 32;
276 op_amd_handle_ibs(struct pt_regs * const regs,
277 struct op_msrs const * const msrs)
280 struct op_entry entry;
285 if (ibs_config.fetch_enabled) {
286 rdmsrl(MSR_AMD64_IBSFETCHCTL, ctl);
287 if (ctl & IBS_FETCH_VAL) {
288 rdmsrl(MSR_AMD64_IBSFETCHLINAD, val);
289 oprofile_write_reserve(&entry, regs, val,
290 IBS_FETCH_CODE, IBS_FETCH_SIZE);
291 oprofile_add_data64(&entry, val);
292 oprofile_add_data64(&entry, ctl);
293 rdmsrl(MSR_AMD64_IBSFETCHPHYSAD, val);
294 oprofile_add_data64(&entry, val);
295 oprofile_write_commit(&entry);
297 /* reenable the IRQ */
298 ctl &= ~(IBS_FETCH_VAL | IBS_FETCH_CNT_MASK);
299 ctl |= IBS_FETCH_ENABLE;
300 wrmsrl(MSR_AMD64_IBSFETCHCTL, ctl);
304 if (ibs_config.op_enabled) {
305 rdmsrl(MSR_AMD64_IBSOPCTL, ctl);
306 if (ctl & IBS_OP_VAL) {
307 rdmsrl(MSR_AMD64_IBSOPRIP, val);
308 oprofile_write_reserve(&entry, regs, val,
309 IBS_OP_CODE, IBS_OP_SIZE);
310 oprofile_add_data64(&entry, val);
311 rdmsrl(MSR_AMD64_IBSOPDATA, val);
312 oprofile_add_data64(&entry, val);
313 rdmsrl(MSR_AMD64_IBSOPDATA2, val);
314 oprofile_add_data64(&entry, val);
315 rdmsrl(MSR_AMD64_IBSOPDATA3, val);
316 oprofile_add_data64(&entry, val);
317 rdmsrl(MSR_AMD64_IBSDCLINAD, val);
318 oprofile_add_data64(&entry, val);
319 rdmsrl(MSR_AMD64_IBSDCPHYSAD, val);
320 oprofile_add_data64(&entry, val);
321 oprofile_write_commit(&entry);
323 /* reenable the IRQ */
324 ctl = op_amd_randomize_ibs_op(ibs_op_ctl);
325 wrmsrl(MSR_AMD64_IBSOPCTL, ctl);
330 static inline void op_amd_start_ibs(void)
337 if (ibs_config.fetch_enabled) {
338 val = (ibs_config.max_cnt_fetch >> 4) & 0xFFFF;
339 val |= ibs_config.rand_en ? IBS_FETCH_RAND_EN : 0;
340 val |= IBS_FETCH_ENABLE;
341 wrmsrl(MSR_AMD64_IBSFETCHCTL, val);
344 if (ibs_config.op_enabled) {
345 ibs_op_ctl = ibs_config.max_cnt_op >> 4;
346 if (!(ibs_caps & IBS_CAPS_RDWROPCNT)) {
348 * IbsOpCurCnt not supported. See
349 * op_amd_randomize_ibs_op() for details.
351 ibs_op_ctl = clamp(ibs_op_ctl, 0x0081ULL, 0xFF80ULL);
354 * The start value is randomized with a
355 * positive offset, we need to compensate it
356 * with the half of the randomized range. Also
359 ibs_op_ctl = min(ibs_op_ctl + IBS_RANDOM_MAXCNT_OFFSET,
362 if (ibs_caps & IBS_CAPS_OPCNT && ibs_config.dispatched_ops)
363 ibs_op_ctl |= IBS_OP_CNT_CTL;
364 ibs_op_ctl |= IBS_OP_ENABLE;
365 val = op_amd_randomize_ibs_op(ibs_op_ctl);
366 wrmsrl(MSR_AMD64_IBSOPCTL, val);
370 static void op_amd_stop_ibs(void)
375 if (ibs_config.fetch_enabled)
376 /* clear max count and enable */
377 wrmsrl(MSR_AMD64_IBSFETCHCTL, 0);
379 if (ibs_config.op_enabled)
380 /* clear max count and enable */
381 wrmsrl(MSR_AMD64_IBSOPCTL, 0);
384 static int op_amd_check_ctrs(struct pt_regs * const regs,
385 struct op_msrs const * const msrs)
390 for (i = 0; i < NUM_COUNTERS; ++i) {
391 int virt = op_x86_phys_to_virt(i);
392 if (!reset_value[virt])
394 rdmsrl(msrs->counters[i].addr, val);
395 /* bit is clear if overflowed: */
396 if (val & OP_CTR_OVERFLOW)
398 oprofile_add_sample(regs, virt);
399 wrmsrl(msrs->counters[i].addr, -(u64)reset_value[virt]);
402 op_amd_handle_ibs(regs, msrs);
404 /* See op_model_ppro.c */
408 static void op_amd_start(struct op_msrs const * const msrs)
413 for (i = 0; i < NUM_COUNTERS; ++i) {
414 if (!reset_value[op_x86_phys_to_virt(i)])
416 rdmsrl(msrs->controls[i].addr, val);
417 val |= ARCH_PERFMON_EVENTSEL0_ENABLE;
418 wrmsrl(msrs->controls[i].addr, val);
424 static void op_amd_stop(struct op_msrs const * const msrs)
430 * Subtle: stop on all counters to avoid race with setting our
433 for (i = 0; i < NUM_COUNTERS; ++i) {
434 if (!reset_value[op_x86_phys_to_virt(i)])
436 rdmsrl(msrs->controls[i].addr, val);
437 val &= ~ARCH_PERFMON_EVENTSEL0_ENABLE;
438 wrmsrl(msrs->controls[i].addr, val);
444 static void op_amd_shutdown(struct op_msrs const * const msrs)
448 for (i = 0; i < NUM_COUNTERS; ++i) {
449 if (msrs->counters[i].addr)
450 release_perfctr_nmi(MSR_K7_PERFCTR0 + i);
452 for (i = 0; i < NUM_CONTROLS; ++i) {
453 if (msrs->controls[i].addr)
454 release_evntsel_nmi(MSR_K7_EVNTSEL0 + i);
458 static u8 ibs_eilvt_off;
460 static inline void apic_init_ibs_nmi_per_cpu(void *arg)
462 ibs_eilvt_off = setup_APIC_eilvt_ibs(0, APIC_EILVT_MSG_NMI, 0);
465 static inline void apic_clear_ibs_nmi_per_cpu(void *arg)
467 setup_APIC_eilvt_ibs(0, APIC_EILVT_MSG_FIX, 1);
470 static int init_ibs_nmi(void)
472 #define IBSCTL_LVTOFFSETVAL (1 << 8)
474 struct pci_dev *cpu_cfg;
479 on_each_cpu(apic_init_ibs_nmi_per_cpu, NULL, 1);
484 cpu_cfg = pci_get_device(PCI_VENDOR_ID_AMD,
485 PCI_DEVICE_ID_AMD_10H_NB_MISC,
490 pci_write_config_dword(cpu_cfg, IBSCTL, ibs_eilvt_off
491 | IBSCTL_LVTOFFSETVAL);
492 pci_read_config_dword(cpu_cfg, IBSCTL, &value);
493 if (value != (ibs_eilvt_off | IBSCTL_LVTOFFSETVAL)) {
494 pci_dev_put(cpu_cfg);
495 printk(KERN_DEBUG "Failed to setup IBS LVT offset, "
496 "IBSCTL = 0x%08x", value);
502 printk(KERN_DEBUG "No CPU node configured for IBS");
509 /* uninitialize the APIC for the IBS interrupts if needed */
510 static void clear_ibs_nmi(void)
513 on_each_cpu(apic_clear_ibs_nmi_per_cpu, NULL, 1);
516 /* initialize the APIC for the IBS interrupts if available */
517 static void ibs_init(void)
519 ibs_caps = get_ibs_caps();
524 if (init_ibs_nmi()) {
529 printk(KERN_INFO "oprofile: AMD IBS detected (0x%08x)\n",
533 static void ibs_exit(void)
541 static int (*create_arch_files)(struct super_block *sb, struct dentry *root);
543 static int setup_ibs_files(struct super_block *sb, struct dentry *root)
548 /* architecture specific files */
549 if (create_arch_files)
550 ret = create_arch_files(sb, root);
558 /* model specific files */
560 /* setup some reasonable defaults */
561 ibs_config.max_cnt_fetch = 250000;
562 ibs_config.fetch_enabled = 0;
563 ibs_config.max_cnt_op = 250000;
564 ibs_config.op_enabled = 0;
565 ibs_config.dispatched_ops = 0;
567 dir = oprofilefs_mkdir(sb, root, "ibs_fetch");
568 oprofilefs_create_ulong(sb, dir, "enable",
569 &ibs_config.fetch_enabled);
570 oprofilefs_create_ulong(sb, dir, "max_count",
571 &ibs_config.max_cnt_fetch);
572 oprofilefs_create_ulong(sb, dir, "rand_enable",
573 &ibs_config.rand_en);
575 dir = oprofilefs_mkdir(sb, root, "ibs_op");
576 oprofilefs_create_ulong(sb, dir, "enable",
577 &ibs_config.op_enabled);
578 oprofilefs_create_ulong(sb, dir, "max_count",
579 &ibs_config.max_cnt_op);
580 if (ibs_caps & IBS_CAPS_OPCNT)
581 oprofilefs_create_ulong(sb, dir, "dispatched_ops",
582 &ibs_config.dispatched_ops);
587 static int op_amd_init(struct oprofile_operations *ops)
590 create_arch_files = ops->create_files;
591 ops->create_files = setup_ibs_files;
595 static void op_amd_exit(void)
600 struct op_x86_model_spec op_amd_spec = {
601 .num_counters = NUM_COUNTERS,
602 .num_controls = NUM_CONTROLS,
603 .num_virt_counters = NUM_VIRT_COUNTERS,
604 .reserved = MSR_AMD_EVENTSEL_RESERVED,
605 .event_mask = OP_EVENT_MASK,
608 .fill_in_addresses = &op_amd_fill_in_addresses,
609 .setup_ctrs = &op_amd_setup_ctrs,
610 .check_ctrs = &op_amd_check_ctrs,
611 .start = &op_amd_start,
612 .stop = &op_amd_stop,
613 .shutdown = &op_amd_shutdown,
614 #ifdef CONFIG_OPROFILE_EVENT_MULTIPLEX
615 .switch_ctrl = &op_mux_switch_ctrl,