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 #ifdef CONFIG_OPROFILE_EVENT_MULTIPLEX
34 #define NUM_VIRT_COUNTERS 32
36 #define NUM_VIRT_COUNTERS NUM_COUNTERS
39 #define OP_EVENT_MASK 0x0FFF
40 #define OP_CTR_OVERFLOW (1ULL<<31)
42 #define MSR_AMD_EVENTSEL_RESERVED ((0xFFFFFCF0ULL<<32)|(1ULL<<21))
44 static unsigned long reset_value[NUM_VIRT_COUNTERS];
46 #define IBS_FETCH_SIZE 6
47 #define IBS_OP_SIZE 12
52 unsigned long op_enabled;
53 unsigned long fetch_enabled;
54 unsigned long max_cnt_fetch;
55 unsigned long max_cnt_op;
56 unsigned long rand_en;
57 unsigned long dispatched_ops;
58 unsigned long branch_target;
64 unsigned long sample_size;
67 static struct ibs_config ibs_config;
68 static struct ibs_state ibs_state;
71 * IBS cpuid feature detection
74 #define IBS_CPUID_FEATURES 0x8000001b
77 * Same bit mask as for IBS cpuid feature flags (Fn8000_001B_EAX), but
78 * bit 0 is used to indicate the existence of IBS.
80 #define IBS_CAPS_AVAIL (1U<<0)
81 #define IBS_CAPS_FETCHSAM (1U<<1)
82 #define IBS_CAPS_OPSAM (1U<<2)
83 #define IBS_CAPS_RDWROPCNT (1U<<3)
84 #define IBS_CAPS_OPCNT (1U<<4)
85 #define IBS_CAPS_BRNTRGT (1U<<5)
86 #define IBS_CAPS_OPCNTEXT (1U<<6)
88 #define IBS_CAPS_DEFAULT (IBS_CAPS_AVAIL \
96 #define IBSCTL_LVT_OFFSET_VALID (1ULL<<8)
97 #define IBSCTL_LVT_OFFSET_MASK 0x0F
100 * IBS randomization macros
102 #define IBS_RANDOM_BITS 12
103 #define IBS_RANDOM_MASK ((1ULL << IBS_RANDOM_BITS) - 1)
104 #define IBS_RANDOM_MAXCNT_OFFSET (1ULL << (IBS_RANDOM_BITS - 5))
106 static u32 get_ibs_caps(void)
109 unsigned int max_level;
111 if (!boot_cpu_has(X86_FEATURE_IBS))
114 /* check IBS cpuid feature flags */
115 max_level = cpuid_eax(0x80000000);
116 if (max_level < IBS_CPUID_FEATURES)
117 return IBS_CAPS_DEFAULT;
119 ibs_caps = cpuid_eax(IBS_CPUID_FEATURES);
120 if (!(ibs_caps & IBS_CAPS_AVAIL))
121 /* cpuid flags not valid */
122 return IBS_CAPS_DEFAULT;
128 * 16-bit Linear Feedback Shift Register (LFSR)
131 * Feedback polynomial = X + X + X + X + 1
133 static unsigned int lfsr_random(void)
135 static unsigned int lfsr_value = 0xF00D;
138 /* Compute next bit to shift in */
139 bit = ((lfsr_value >> 0) ^
142 (lfsr_value >> 5)) & 0x0001;
144 /* Advance to next register value */
145 lfsr_value = (lfsr_value >> 1) | (bit << 15);
151 * IBS software randomization
153 * The IBS periodic op counter is randomized in software. The lower 12
154 * bits of the 20 bit counter are randomized. IbsOpCurCnt is
155 * initialized with a 12 bit random value.
157 static inline u64 op_amd_randomize_ibs_op(u64 val)
159 unsigned int random = lfsr_random();
161 if (!(ibs_caps & IBS_CAPS_RDWROPCNT))
163 * Work around if the hw can not write to IbsOpCurCnt
165 * Randomize the lower 8 bits of the 16 bit
166 * IbsOpMaxCnt [15:0] value in the range of -128 to
167 * +127 by adding/subtracting an offset to the
168 * maximum count (IbsOpMaxCnt).
170 * To avoid over or underflows and protect upper bits
171 * starting at bit 16, the initial value for
172 * IbsOpMaxCnt must fit in the range from 0x0081 to
175 val += (s8)(random >> 4);
177 val |= (u64)(random & IBS_RANDOM_MASK) << 32;
183 op_amd_handle_ibs(struct pt_regs * const regs,
184 struct op_msrs const * const msrs)
187 struct op_entry entry;
192 if (ibs_config.fetch_enabled) {
193 rdmsrl(MSR_AMD64_IBSFETCHCTL, ctl);
194 if (ctl & IBS_FETCH_VAL) {
195 rdmsrl(MSR_AMD64_IBSFETCHLINAD, val);
196 oprofile_write_reserve(&entry, regs, val,
197 IBS_FETCH_CODE, IBS_FETCH_SIZE);
198 oprofile_add_data64(&entry, val);
199 oprofile_add_data64(&entry, ctl);
200 rdmsrl(MSR_AMD64_IBSFETCHPHYSAD, val);
201 oprofile_add_data64(&entry, val);
202 oprofile_write_commit(&entry);
204 /* reenable the IRQ */
205 ctl &= ~(IBS_FETCH_VAL | IBS_FETCH_CNT);
206 ctl |= IBS_FETCH_ENABLE;
207 wrmsrl(MSR_AMD64_IBSFETCHCTL, ctl);
211 if (ibs_config.op_enabled) {
212 rdmsrl(MSR_AMD64_IBSOPCTL, ctl);
213 if (ctl & IBS_OP_VAL) {
214 rdmsrl(MSR_AMD64_IBSOPRIP, val);
215 oprofile_write_reserve(&entry, regs, val, IBS_OP_CODE,
216 ibs_state.sample_size);
217 oprofile_add_data64(&entry, val);
218 rdmsrl(MSR_AMD64_IBSOPDATA, val);
219 oprofile_add_data64(&entry, val);
220 rdmsrl(MSR_AMD64_IBSOPDATA2, val);
221 oprofile_add_data64(&entry, val);
222 rdmsrl(MSR_AMD64_IBSOPDATA3, val);
223 oprofile_add_data64(&entry, val);
224 rdmsrl(MSR_AMD64_IBSDCLINAD, val);
225 oprofile_add_data64(&entry, val);
226 rdmsrl(MSR_AMD64_IBSDCPHYSAD, val);
227 oprofile_add_data64(&entry, val);
228 if (ibs_state.branch_target) {
229 rdmsrl(MSR_AMD64_IBSBRTARGET, val);
230 oprofile_add_data(&entry, (unsigned long)val);
232 oprofile_write_commit(&entry);
234 /* reenable the IRQ */
235 ctl = op_amd_randomize_ibs_op(ibs_state.ibs_op_ctl);
236 wrmsrl(MSR_AMD64_IBSOPCTL, ctl);
241 static inline void op_amd_start_ibs(void)
248 memset(&ibs_state, 0, sizeof(ibs_state));
251 * Note: Since the max count settings may out of range we
252 * write back the actual used values so that userland can read
256 if (ibs_config.fetch_enabled) {
257 val = ibs_config.max_cnt_fetch >> 4;
258 val = min(val, IBS_FETCH_MAX_CNT);
259 ibs_config.max_cnt_fetch = val << 4;
260 val |= ibs_config.rand_en ? IBS_FETCH_RAND_EN : 0;
261 val |= IBS_FETCH_ENABLE;
262 wrmsrl(MSR_AMD64_IBSFETCHCTL, val);
265 if (ibs_config.op_enabled) {
266 val = ibs_config.max_cnt_op >> 4;
267 if (!(ibs_caps & IBS_CAPS_RDWROPCNT)) {
269 * IbsOpCurCnt not supported. See
270 * op_amd_randomize_ibs_op() for details.
272 val = clamp(val, 0x0081ULL, 0xFF80ULL);
273 ibs_config.max_cnt_op = val << 4;
276 * The start value is randomized with a
277 * positive offset, we need to compensate it
278 * with the half of the randomized range. Also
281 val += IBS_RANDOM_MAXCNT_OFFSET;
282 if (ibs_caps & IBS_CAPS_OPCNTEXT)
283 val = min(val, IBS_OP_MAX_CNT_EXT);
285 val = min(val, IBS_OP_MAX_CNT);
286 ibs_config.max_cnt_op =
287 (val - IBS_RANDOM_MAXCNT_OFFSET) << 4;
289 val = ((val & ~IBS_OP_MAX_CNT) << 4) | (val & IBS_OP_MAX_CNT);
290 val |= ibs_config.dispatched_ops ? IBS_OP_CNT_CTL : 0;
291 val |= IBS_OP_ENABLE;
292 ibs_state.ibs_op_ctl = val;
293 ibs_state.sample_size = IBS_OP_SIZE;
294 if (ibs_config.branch_target) {
295 ibs_state.branch_target = 1;
296 ibs_state.sample_size++;
298 val = op_amd_randomize_ibs_op(ibs_state.ibs_op_ctl);
299 wrmsrl(MSR_AMD64_IBSOPCTL, val);
303 static void op_amd_stop_ibs(void)
308 if (ibs_config.fetch_enabled)
309 /* clear max count and enable */
310 wrmsrl(MSR_AMD64_IBSFETCHCTL, 0);
312 if (ibs_config.op_enabled)
313 /* clear max count and enable */
314 wrmsrl(MSR_AMD64_IBSOPCTL, 0);
317 #ifdef CONFIG_OPROFILE_EVENT_MULTIPLEX
319 static void op_mux_switch_ctrl(struct op_x86_model_spec const *model,
320 struct op_msrs const * const msrs)
325 /* enable active counters */
326 for (i = 0; i < NUM_COUNTERS; ++i) {
327 int virt = op_x86_phys_to_virt(i);
328 if (!reset_value[virt])
330 rdmsrl(msrs->controls[i].addr, val);
331 val &= model->reserved;
332 val |= op_x86_get_ctrl(model, &counter_config[virt]);
333 wrmsrl(msrs->controls[i].addr, val);
339 /* functions for op_amd_spec */
341 static void op_amd_shutdown(struct op_msrs const * const msrs)
345 for (i = 0; i < NUM_COUNTERS; ++i) {
346 if (!msrs->counters[i].addr)
348 release_perfctr_nmi(MSR_K7_PERFCTR0 + i);
349 release_evntsel_nmi(MSR_K7_EVNTSEL0 + i);
353 static int op_amd_fill_in_addresses(struct op_msrs * const msrs)
357 for (i = 0; i < NUM_COUNTERS; i++) {
358 if (!reserve_perfctr_nmi(MSR_K7_PERFCTR0 + i))
360 if (!reserve_evntsel_nmi(MSR_K7_EVNTSEL0 + i)) {
361 release_perfctr_nmi(MSR_K7_PERFCTR0 + i);
364 /* both registers must be reserved */
365 msrs->counters[i].addr = MSR_K7_PERFCTR0 + i;
366 msrs->controls[i].addr = MSR_K7_EVNTSEL0 + i;
369 if (!counter_config[i].enabled)
371 op_x86_warn_reserved(i);
372 op_amd_shutdown(msrs);
379 static void op_amd_setup_ctrs(struct op_x86_model_spec const *model,
380 struct op_msrs const * const msrs)
385 /* setup reset_value */
386 for (i = 0; i < NUM_VIRT_COUNTERS; ++i) {
387 if (counter_config[i].enabled
388 && msrs->counters[op_x86_virt_to_phys(i)].addr)
389 reset_value[i] = counter_config[i].count;
394 /* clear all counters */
395 for (i = 0; i < NUM_COUNTERS; ++i) {
396 if (!msrs->controls[i].addr)
398 rdmsrl(msrs->controls[i].addr, val);
399 if (val & ARCH_PERFMON_EVENTSEL_ENABLE)
400 op_x86_warn_in_use(i);
401 val &= model->reserved;
402 wrmsrl(msrs->controls[i].addr, val);
404 * avoid a false detection of ctr overflows in NMI
407 wrmsrl(msrs->counters[i].addr, -1LL);
410 /* enable active counters */
411 for (i = 0; i < NUM_COUNTERS; ++i) {
412 int virt = op_x86_phys_to_virt(i);
413 if (!reset_value[virt])
416 /* setup counter registers */
417 wrmsrl(msrs->counters[i].addr, -(u64)reset_value[virt]);
419 /* setup control registers */
420 rdmsrl(msrs->controls[i].addr, val);
421 val &= model->reserved;
422 val |= op_x86_get_ctrl(model, &counter_config[virt]);
423 wrmsrl(msrs->controls[i].addr, val);
427 setup_APIC_eilvt_ibs(0, APIC_EILVT_MSG_NMI, 0);
430 static void op_amd_cpu_shutdown(void)
433 setup_APIC_eilvt_ibs(0, APIC_EILVT_MSG_FIX, 1);
436 static int op_amd_check_ctrs(struct pt_regs * const regs,
437 struct op_msrs const * const msrs)
442 for (i = 0; i < NUM_COUNTERS; ++i) {
443 int virt = op_x86_phys_to_virt(i);
444 if (!reset_value[virt])
446 rdmsrl(msrs->counters[i].addr, val);
447 /* bit is clear if overflowed: */
448 if (val & OP_CTR_OVERFLOW)
450 oprofile_add_sample(regs, virt);
451 wrmsrl(msrs->counters[i].addr, -(u64)reset_value[virt]);
454 op_amd_handle_ibs(regs, msrs);
456 /* See op_model_ppro.c */
460 static void op_amd_start(struct op_msrs const * const msrs)
465 for (i = 0; i < NUM_COUNTERS; ++i) {
466 if (!reset_value[op_x86_phys_to_virt(i)])
468 rdmsrl(msrs->controls[i].addr, val);
469 val |= ARCH_PERFMON_EVENTSEL_ENABLE;
470 wrmsrl(msrs->controls[i].addr, val);
476 static void op_amd_stop(struct op_msrs const * const msrs)
482 * Subtle: stop on all counters to avoid race with setting our
485 for (i = 0; i < NUM_COUNTERS; ++i) {
486 if (!reset_value[op_x86_phys_to_virt(i)])
488 rdmsrl(msrs->controls[i].addr, val);
489 val &= ~ARCH_PERFMON_EVENTSEL_ENABLE;
490 wrmsrl(msrs->controls[i].addr, val);
496 static int __init_ibs_nmi(void)
498 #define IBSCTL_LVTOFFSETVAL (1 << 8)
500 struct pci_dev *cpu_cfg;
505 ibs_eilvt_off = setup_APIC_eilvt_ibs(0, APIC_EILVT_MSG_FIX, 1);
510 cpu_cfg = pci_get_device(PCI_VENDOR_ID_AMD,
511 PCI_DEVICE_ID_AMD_10H_NB_MISC,
516 pci_write_config_dword(cpu_cfg, IBSCTL, ibs_eilvt_off
517 | IBSCTL_LVTOFFSETVAL);
518 pci_read_config_dword(cpu_cfg, IBSCTL, &value);
519 if (value != (ibs_eilvt_off | IBSCTL_LVTOFFSETVAL)) {
520 pci_dev_put(cpu_cfg);
521 printk(KERN_DEBUG "Failed to setup IBS LVT offset, "
522 "IBSCTL = 0x%08x", value);
528 printk(KERN_DEBUG "No CPU node configured for IBS");
535 /* initialize the APIC for the IBS interrupts if available */
536 static void init_ibs(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 int (*create_arch_files)(struct super_block *sb, struct dentry *root);
554 static int setup_ibs_files(struct super_block *sb, struct dentry *root)
559 /* architecture specific files */
560 if (create_arch_files)
561 ret = create_arch_files(sb, root);
569 /* model specific files */
571 /* setup some reasonable defaults */
572 memset(&ibs_config, 0, sizeof(ibs_config));
573 ibs_config.max_cnt_fetch = 250000;
574 ibs_config.max_cnt_op = 250000;
576 if (ibs_caps & IBS_CAPS_FETCHSAM) {
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);
586 if (ibs_caps & IBS_CAPS_OPSAM) {
587 dir = oprofilefs_mkdir(sb, root, "ibs_op");
588 oprofilefs_create_ulong(sb, dir, "enable",
589 &ibs_config.op_enabled);
590 oprofilefs_create_ulong(sb, dir, "max_count",
591 &ibs_config.max_cnt_op);
592 if (ibs_caps & IBS_CAPS_OPCNT)
593 oprofilefs_create_ulong(sb, dir, "dispatched_ops",
594 &ibs_config.dispatched_ops);
595 if (ibs_caps & IBS_CAPS_BRNTRGT)
596 oprofilefs_create_ulong(sb, dir, "branch_target",
597 &ibs_config.branch_target);
603 static int op_amd_init(struct oprofile_operations *ops)
606 create_arch_files = ops->create_files;
607 ops->create_files = setup_ibs_files;
611 struct op_x86_model_spec op_amd_spec = {
612 .num_counters = NUM_COUNTERS,
613 .num_controls = NUM_COUNTERS,
614 .num_virt_counters = NUM_VIRT_COUNTERS,
615 .reserved = MSR_AMD_EVENTSEL_RESERVED,
616 .event_mask = OP_EVENT_MASK,
618 .fill_in_addresses = &op_amd_fill_in_addresses,
619 .setup_ctrs = &op_amd_setup_ctrs,
620 .cpu_down = &op_amd_cpu_shutdown,
621 .check_ctrs = &op_amd_check_ctrs,
622 .start = &op_amd_start,
623 .stop = &op_amd_stop,
624 .shutdown = &op_amd_shutdown,
625 #ifdef CONFIG_OPROFILE_EVENT_MULTIPLEX
626 .switch_ctrl = &op_mux_switch_ctrl,