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
15 #include <linux/oprofile.h>
16 #include <linux/device.h>
17 #include <linux/pci.h>
19 #include <asm/ptrace.h>
23 #include "op_x86_model.h"
24 #include "op_counter.h"
26 #define NUM_COUNTERS 4
27 #define NUM_CONTROLS 4
28 #define OP_EVENT_MASK 0x0FFF
29 #define OP_CTR_OVERFLOW (1ULL<<31)
31 #define MSR_AMD_EVENTSEL_RESERVED ((0xFFFFFCF0ULL<<32)|(1ULL<<21))
33 static unsigned long reset_value[NUM_COUNTERS];
35 #ifdef CONFIG_OPROFILE_IBS
37 /* IbsFetchCtl bits/masks */
38 #define IBS_FETCH_HIGH_VALID_BIT (1UL << 17) /* bit 49 */
39 #define IBS_FETCH_HIGH_ENABLE (1UL << 16) /* bit 48 */
40 #define IBS_FETCH_LOW_MAX_CNT_MASK 0x0000FFFFUL /* MaxCnt mask */
43 #define IBS_OP_LOW_VALID_BIT (1ULL<<18) /* bit 18 */
44 #define IBS_OP_LOW_ENABLE (1ULL<<17) /* bit 17 */
46 #define IBS_FETCH_SIZE 6
47 #define IBS_OP_SIZE 12
49 static int has_ibs; /* AMD Family10h and later */
51 struct op_ibs_config {
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;
60 static struct op_ibs_config ibs_config;
64 /* functions for op_amd_spec */
66 static void op_amd_fill_in_addresses(struct op_msrs * const msrs)
70 for (i = 0; i < NUM_COUNTERS; i++) {
71 if (reserve_perfctr_nmi(MSR_K7_PERFCTR0 + i))
72 msrs->counters[i].addr = MSR_K7_PERFCTR0 + i;
74 msrs->counters[i].addr = 0;
77 for (i = 0; i < NUM_CONTROLS; i++) {
78 if (reserve_evntsel_nmi(MSR_K7_EVNTSEL0 + i))
79 msrs->controls[i].addr = MSR_K7_EVNTSEL0 + i;
81 msrs->controls[i].addr = 0;
85 static void op_amd_setup_ctrs(struct op_x86_model_spec const *model,
86 struct op_msrs const * const msrs)
91 /* clear all counters */
92 for (i = 0 ; i < NUM_CONTROLS; ++i) {
93 if (unlikely(!CTRL_IS_RESERVED(msrs, i)))
95 rdmsrl(msrs->controls[i].addr, val);
96 val &= model->reserved;
97 wrmsrl(msrs->controls[i].addr, val);
100 /* avoid a false detection of ctr overflows in NMI handler */
101 for (i = 0; i < NUM_COUNTERS; ++i) {
102 if (unlikely(!CTR_IS_RESERVED(msrs, i)))
104 wrmsr(msrs->counters[i].addr, -1, -1);
107 /* enable active counters */
108 for (i = 0; i < NUM_COUNTERS; ++i) {
109 if ((counter_config[i].enabled) && (CTR_IS_RESERVED(msrs, i))) {
110 reset_value[i] = counter_config[i].count;
111 wrmsr(msrs->counters[i].addr, -(unsigned int)counter_config[i].count, -1);
112 rdmsrl(msrs->controls[i].addr, val);
113 val &= model->reserved;
114 val |= op_x86_get_ctrl(model, &counter_config[i]);
115 wrmsrl(msrs->controls[i].addr, val);
122 #ifdef CONFIG_OPROFILE_IBS
125 op_amd_handle_ibs(struct pt_regs * const regs,
126 struct op_msrs const * const msrs)
130 struct op_entry entry;
135 if (ibs_config.fetch_enabled) {
136 rdmsr(MSR_AMD64_IBSFETCHCTL, low, high);
137 if (high & IBS_FETCH_HIGH_VALID_BIT) {
138 rdmsrl(MSR_AMD64_IBSFETCHLINAD, msr);
139 oprofile_write_reserve(&entry, regs, msr,
140 IBS_FETCH_CODE, IBS_FETCH_SIZE);
141 oprofile_add_data(&entry, (u32)msr);
142 oprofile_add_data(&entry, (u32)(msr >> 32));
143 oprofile_add_data(&entry, low);
144 oprofile_add_data(&entry, high);
145 rdmsrl(MSR_AMD64_IBSFETCHPHYSAD, msr);
146 oprofile_add_data(&entry, (u32)msr);
147 oprofile_add_data(&entry, (u32)(msr >> 32));
148 oprofile_write_commit(&entry);
150 /* reenable the IRQ */
151 high &= ~IBS_FETCH_HIGH_VALID_BIT;
152 high |= IBS_FETCH_HIGH_ENABLE;
153 low &= IBS_FETCH_LOW_MAX_CNT_MASK;
154 wrmsr(MSR_AMD64_IBSFETCHCTL, low, high);
158 if (ibs_config.op_enabled) {
159 rdmsr(MSR_AMD64_IBSOPCTL, low, high);
160 if (low & IBS_OP_LOW_VALID_BIT) {
161 rdmsrl(MSR_AMD64_IBSOPRIP, msr);
162 oprofile_write_reserve(&entry, regs, msr,
163 IBS_OP_CODE, IBS_OP_SIZE);
164 oprofile_add_data(&entry, (u32)msr);
165 oprofile_add_data(&entry, (u32)(msr >> 32));
166 rdmsrl(MSR_AMD64_IBSOPDATA, msr);
167 oprofile_add_data(&entry, (u32)msr);
168 oprofile_add_data(&entry, (u32)(msr >> 32));
169 rdmsrl(MSR_AMD64_IBSOPDATA2, msr);
170 oprofile_add_data(&entry, (u32)msr);
171 oprofile_add_data(&entry, (u32)(msr >> 32));
172 rdmsrl(MSR_AMD64_IBSOPDATA3, msr);
173 oprofile_add_data(&entry, (u32)msr);
174 oprofile_add_data(&entry, (u32)(msr >> 32));
175 rdmsrl(MSR_AMD64_IBSDCLINAD, msr);
176 oprofile_add_data(&entry, (u32)msr);
177 oprofile_add_data(&entry, (u32)(msr >> 32));
178 rdmsrl(MSR_AMD64_IBSDCPHYSAD, msr);
179 oprofile_add_data(&entry, (u32)msr);
180 oprofile_add_data(&entry, (u32)(msr >> 32));
181 oprofile_write_commit(&entry);
183 /* reenable the IRQ */
185 low &= ~IBS_OP_LOW_VALID_BIT;
186 low |= IBS_OP_LOW_ENABLE;
187 wrmsr(MSR_AMD64_IBSOPCTL, low, high);
194 static inline void op_amd_start_ibs(void)
196 unsigned int low, high;
197 if (has_ibs && ibs_config.fetch_enabled) {
198 low = (ibs_config.max_cnt_fetch >> 4) & 0xFFFF;
199 high = ((ibs_config.rand_en & 0x1) << 25) /* bit 57 */
200 + IBS_FETCH_HIGH_ENABLE;
201 wrmsr(MSR_AMD64_IBSFETCHCTL, low, high);
204 if (has_ibs && ibs_config.op_enabled) {
205 low = ((ibs_config.max_cnt_op >> 4) & 0xFFFF)
206 + ((ibs_config.dispatched_ops & 0x1) << 19) /* bit 19 */
209 wrmsr(MSR_AMD64_IBSOPCTL, low, high);
213 static void op_amd_stop_ibs(void)
215 unsigned int low, high;
216 if (has_ibs && ibs_config.fetch_enabled) {
217 /* clear max count and enable */
220 wrmsr(MSR_AMD64_IBSFETCHCTL, low, high);
223 if (has_ibs && ibs_config.op_enabled) {
224 /* clear max count and enable */
227 wrmsr(MSR_AMD64_IBSOPCTL, low, high);
233 static inline int op_amd_handle_ibs(struct pt_regs * const regs,
234 struct op_msrs const * const msrs) { }
235 static inline void op_amd_start_ibs(void) { }
236 static inline void op_amd_stop_ibs(void) { }
240 static int op_amd_check_ctrs(struct pt_regs * const regs,
241 struct op_msrs const * const msrs)
246 for (i = 0 ; i < NUM_COUNTERS; ++i) {
249 rdmsrl(msrs->counters[i].addr, val);
250 /* bit is clear if overflowed: */
251 if (val & OP_CTR_OVERFLOW)
253 oprofile_add_sample(regs, i);
254 wrmsr(msrs->counters[i].addr, -(unsigned int)reset_value[i], -1);
257 op_amd_handle_ibs(regs, msrs);
259 /* See op_model_ppro.c */
263 static void op_amd_start(struct op_msrs const * const msrs)
267 for (i = 0 ; i < NUM_COUNTERS ; ++i) {
268 if (reset_value[i]) {
269 rdmsrl(msrs->controls[i].addr, val);
270 val |= ARCH_PERFMON_EVENTSEL0_ENABLE;
271 wrmsrl(msrs->controls[i].addr, val);
278 static void op_amd_stop(struct op_msrs const * const msrs)
284 * Subtle: stop on all counters to avoid race with setting our
287 for (i = 0 ; i < NUM_COUNTERS ; ++i) {
290 rdmsrl(msrs->controls[i].addr, val);
291 val &= ~ARCH_PERFMON_EVENTSEL0_ENABLE;
292 wrmsrl(msrs->controls[i].addr, val);
298 static void op_amd_shutdown(struct op_msrs const * const msrs)
302 for (i = 0 ; i < NUM_COUNTERS ; ++i) {
303 if (CTR_IS_RESERVED(msrs, i))
304 release_perfctr_nmi(MSR_K7_PERFCTR0 + i);
306 for (i = 0 ; i < NUM_CONTROLS ; ++i) {
307 if (CTRL_IS_RESERVED(msrs, i))
308 release_evntsel_nmi(MSR_K7_EVNTSEL0 + i);
312 #ifdef CONFIG_OPROFILE_IBS
314 static u8 ibs_eilvt_off;
316 static inline void apic_init_ibs_nmi_per_cpu(void *arg)
318 ibs_eilvt_off = setup_APIC_eilvt_ibs(0, APIC_EILVT_MSG_NMI, 0);
321 static inline void apic_clear_ibs_nmi_per_cpu(void *arg)
323 setup_APIC_eilvt_ibs(0, APIC_EILVT_MSG_FIX, 1);
326 static int init_ibs_nmi(void)
328 #define IBSCTL_LVTOFFSETVAL (1 << 8)
330 struct pci_dev *cpu_cfg;
335 on_each_cpu(apic_init_ibs_nmi_per_cpu, NULL, 1);
340 cpu_cfg = pci_get_device(PCI_VENDOR_ID_AMD,
341 PCI_DEVICE_ID_AMD_10H_NB_MISC,
346 pci_write_config_dword(cpu_cfg, IBSCTL, ibs_eilvt_off
347 | IBSCTL_LVTOFFSETVAL);
348 pci_read_config_dword(cpu_cfg, IBSCTL, &value);
349 if (value != (ibs_eilvt_off | IBSCTL_LVTOFFSETVAL)) {
350 pci_dev_put(cpu_cfg);
351 printk(KERN_DEBUG "Failed to setup IBS LVT offset, "
352 "IBSCTL = 0x%08x", value);
358 printk(KERN_DEBUG "No CPU node configured for IBS");
364 /* Works only for 64bit with proper numa implementation. */
365 if (nodes != num_possible_nodes()) {
366 printk(KERN_DEBUG "Failed to setup CPU node(s) for IBS, "
367 "found: %d, expected %d",
368 nodes, num_possible_nodes());
375 /* uninitialize the APIC for the IBS interrupts if needed */
376 static void clear_ibs_nmi(void)
379 on_each_cpu(apic_clear_ibs_nmi_per_cpu, NULL, 1);
382 /* initialize the APIC for the IBS interrupts if available */
383 static void ibs_init(void)
385 has_ibs = boot_cpu_has(X86_FEATURE_IBS);
390 if (init_ibs_nmi()) {
395 printk(KERN_INFO "oprofile: AMD IBS detected\n");
398 static void ibs_exit(void)
406 static int (*create_arch_files)(struct super_block *sb, struct dentry *root);
408 static int setup_ibs_files(struct super_block *sb, struct dentry *root)
413 /* architecture specific files */
414 if (create_arch_files)
415 ret = create_arch_files(sb, root);
423 /* model specific files */
425 /* setup some reasonable defaults */
426 ibs_config.max_cnt_fetch = 250000;
427 ibs_config.fetch_enabled = 0;
428 ibs_config.max_cnt_op = 250000;
429 ibs_config.op_enabled = 0;
430 ibs_config.dispatched_ops = 1;
432 dir = oprofilefs_mkdir(sb, root, "ibs_fetch");
433 oprofilefs_create_ulong(sb, dir, "enable",
434 &ibs_config.fetch_enabled);
435 oprofilefs_create_ulong(sb, dir, "max_count",
436 &ibs_config.max_cnt_fetch);
437 oprofilefs_create_ulong(sb, dir, "rand_enable",
438 &ibs_config.rand_en);
440 dir = oprofilefs_mkdir(sb, root, "ibs_op");
441 oprofilefs_create_ulong(sb, dir, "enable",
442 &ibs_config.op_enabled);
443 oprofilefs_create_ulong(sb, dir, "max_count",
444 &ibs_config.max_cnt_op);
445 oprofilefs_create_ulong(sb, dir, "dispatched_ops",
446 &ibs_config.dispatched_ops);
451 static int op_amd_init(struct oprofile_operations *ops)
454 create_arch_files = ops->create_files;
455 ops->create_files = setup_ibs_files;
459 static void op_amd_exit(void)
468 static int op_amd_init(struct oprofile_operations *ops)
473 static void op_amd_exit(void) {}
475 #endif /* CONFIG_OPROFILE_IBS */
477 struct op_x86_model_spec const op_amd_spec = {
478 .num_counters = NUM_COUNTERS,
479 .num_controls = NUM_CONTROLS,
480 .reserved = MSR_AMD_EVENTSEL_RESERVED,
481 .event_mask = OP_EVENT_MASK,
484 .fill_in_addresses = &op_amd_fill_in_addresses,
485 .setup_ctrs = &op_amd_setup_ctrs,
486 .check_ctrs = &op_amd_check_ctrs,
487 .start = &op_amd_start,
488 .stop = &op_amd_stop,
489 .shutdown = &op_amd_shutdown,