x86/oprofile: use 64 bit values in IBS functions
[linux-2.6/linux-acpi-2.6/ibm-acpi-2.6.git] / arch / x86 / oprofile / op_model_amd.c
blob6493ef7ae9adc6e0c95fd5b6a3eee2582f02bc8b
1 /*
2 * @file op_model_amd.c
3 * athlon / K7 / K8 / Family 10h model-specific MSR operations
5 * @remark Copyright 2002-2009 OProfile authors
6 * @remark Read the file COPYING
8 * @author John Levon
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>
20 #include <asm/msr.h>
21 #include <asm/nmi.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_RAND_EN (1ULL<<57)
39 #define IBS_FETCH_VAL (1ULL<<49)
40 #define IBS_FETCH_ENABLE (1ULL<<48)
41 #define IBS_FETCH_CNT_MASK 0xFFFF0000ULL
43 /*IbsOpCtl bits */
44 #define IBS_OP_CNT_CTL (1ULL<<19)
45 #define IBS_OP_VAL (1ULL<<18)
46 #define IBS_OP_ENABLE (1ULL<<17)
48 #define IBS_FETCH_SIZE 6
49 #define IBS_OP_SIZE 12
51 static int has_ibs; /* AMD Family10h and later */
53 struct op_ibs_config {
54 unsigned long op_enabled;
55 unsigned long fetch_enabled;
56 unsigned long max_cnt_fetch;
57 unsigned long max_cnt_op;
58 unsigned long rand_en;
59 unsigned long dispatched_ops;
62 static struct op_ibs_config ibs_config;
64 #endif
66 /* functions for op_amd_spec */
68 static void op_amd_fill_in_addresses(struct op_msrs * const msrs)
70 int i;
72 for (i = 0; i < NUM_COUNTERS; i++) {
73 if (reserve_perfctr_nmi(MSR_K7_PERFCTR0 + i))
74 msrs->counters[i].addr = MSR_K7_PERFCTR0 + i;
75 else
76 msrs->counters[i].addr = 0;
79 for (i = 0; i < NUM_CONTROLS; i++) {
80 if (reserve_evntsel_nmi(MSR_K7_EVNTSEL0 + i))
81 msrs->controls[i].addr = MSR_K7_EVNTSEL0 + i;
82 else
83 msrs->controls[i].addr = 0;
87 static void op_amd_setup_ctrs(struct op_x86_model_spec const *model,
88 struct op_msrs const * const msrs)
90 u64 val;
91 int i;
93 /* clear all counters */
94 for (i = 0 ; i < NUM_CONTROLS; ++i) {
95 if (unlikely(!msrs->controls[i].addr))
96 continue;
97 rdmsrl(msrs->controls[i].addr, val);
98 val &= model->reserved;
99 wrmsrl(msrs->controls[i].addr, val);
102 /* avoid a false detection of ctr overflows in NMI handler */
103 for (i = 0; i < NUM_COUNTERS; ++i) {
104 if (unlikely(!msrs->counters[i].addr))
105 continue;
106 wrmsrl(msrs->counters[i].addr, -1LL);
109 /* enable active counters */
110 for (i = 0; i < NUM_COUNTERS; ++i) {
111 if (counter_config[i].enabled && msrs->counters[i].addr) {
112 reset_value[i] = counter_config[i].count;
113 wrmsrl(msrs->counters[i].addr,
114 -(s64)counter_config[i].count);
115 rdmsrl(msrs->controls[i].addr, val);
116 val &= model->reserved;
117 val |= op_x86_get_ctrl(model, &counter_config[i]);
118 wrmsrl(msrs->controls[i].addr, val);
119 } else {
120 reset_value[i] = 0;
125 #ifdef CONFIG_OPROFILE_IBS
127 static inline int
128 op_amd_handle_ibs(struct pt_regs * const regs,
129 struct op_msrs const * const msrs)
131 u64 val, ctl;
132 struct op_entry entry;
134 if (!has_ibs)
135 return 1;
137 if (ibs_config.fetch_enabled) {
138 rdmsrl(MSR_AMD64_IBSFETCHCTL, ctl);
139 if (ctl & IBS_FETCH_VAL) {
140 rdmsrl(MSR_AMD64_IBSFETCHLINAD, val);
141 oprofile_write_reserve(&entry, regs, val,
142 IBS_FETCH_CODE, IBS_FETCH_SIZE);
143 oprofile_add_data(&entry, (u32)val);
144 oprofile_add_data(&entry, (u32)(val >> 32));
145 oprofile_add_data(&entry, (u32)ctl);
146 oprofile_add_data(&entry, (u32)(ctl >> 32));
147 rdmsrl(MSR_AMD64_IBSFETCHPHYSAD, val);
148 oprofile_add_data(&entry, (u32)val);
149 oprofile_add_data(&entry, (u32)(val >> 32));
150 oprofile_write_commit(&entry);
152 /* reenable the IRQ */
153 ctl &= ~(IBS_FETCH_VAL | IBS_FETCH_CNT_MASK);
154 ctl |= IBS_FETCH_ENABLE;
155 wrmsrl(MSR_AMD64_IBSFETCHCTL, ctl);
159 if (ibs_config.op_enabled) {
160 rdmsrl(MSR_AMD64_IBSOPCTL, ctl);
161 if (ctl & IBS_OP_VAL) {
162 rdmsrl(MSR_AMD64_IBSOPRIP, val);
163 oprofile_write_reserve(&entry, regs, val,
164 IBS_OP_CODE, IBS_OP_SIZE);
165 oprofile_add_data(&entry, (u32)val);
166 oprofile_add_data(&entry, (u32)(val >> 32));
167 rdmsrl(MSR_AMD64_IBSOPDATA, val);
168 oprofile_add_data(&entry, (u32)val);
169 oprofile_add_data(&entry, (u32)(val >> 32));
170 rdmsrl(MSR_AMD64_IBSOPDATA2, val);
171 oprofile_add_data(&entry, (u32)val);
172 oprofile_add_data(&entry, (u32)(val >> 32));
173 rdmsrl(MSR_AMD64_IBSOPDATA3, val);
174 oprofile_add_data(&entry, (u32)val);
175 oprofile_add_data(&entry, (u32)(val >> 32));
176 rdmsrl(MSR_AMD64_IBSDCLINAD, val);
177 oprofile_add_data(&entry, (u32)val);
178 oprofile_add_data(&entry, (u32)(val >> 32));
179 rdmsrl(MSR_AMD64_IBSDCPHYSAD, val);
180 oprofile_add_data(&entry, (u32)val);
181 oprofile_add_data(&entry, (u32)(val >> 32));
182 oprofile_write_commit(&entry);
184 /* reenable the IRQ */
185 ctl &= ~IBS_OP_VAL & 0xFFFFFFFF;
186 ctl |= IBS_OP_ENABLE;
187 wrmsrl(MSR_AMD64_IBSOPCTL, ctl);
191 return 1;
194 static inline void op_amd_start_ibs(void)
196 u64 val;
197 if (has_ibs && ibs_config.fetch_enabled) {
198 val = (ibs_config.max_cnt_fetch >> 4) & 0xFFFF;
199 val |= ibs_config.rand_en ? IBS_FETCH_RAND_EN : 0;
200 val |= IBS_FETCH_ENABLE;
201 wrmsrl(MSR_AMD64_IBSFETCHCTL, val);
204 if (has_ibs && ibs_config.op_enabled) {
205 val = (ibs_config.max_cnt_op >> 4) & 0xFFFF;
206 val |= ibs_config.dispatched_ops ? IBS_OP_CNT_CTL : 0;
207 val |= IBS_OP_ENABLE;
208 wrmsrl(MSR_AMD64_IBSOPCTL, val);
212 static void op_amd_stop_ibs(void)
214 if (has_ibs && ibs_config.fetch_enabled)
215 /* clear max count and enable */
216 wrmsrl(MSR_AMD64_IBSFETCHCTL, 0);
218 if (has_ibs && ibs_config.op_enabled)
219 /* clear max count and enable */
220 wrmsrl(MSR_AMD64_IBSOPCTL, 0);
223 #else
225 static inline int op_amd_handle_ibs(struct pt_regs * const regs,
226 struct op_msrs const * const msrs) { }
227 static inline void op_amd_start_ibs(void) { }
228 static inline void op_amd_stop_ibs(void) { }
230 #endif
232 static int op_amd_check_ctrs(struct pt_regs * const regs,
233 struct op_msrs const * const msrs)
235 u64 val;
236 int i;
238 for (i = 0 ; i < NUM_COUNTERS; ++i) {
239 if (!reset_value[i])
240 continue;
241 rdmsrl(msrs->counters[i].addr, val);
242 /* bit is clear if overflowed: */
243 if (val & OP_CTR_OVERFLOW)
244 continue;
245 oprofile_add_sample(regs, i);
246 wrmsrl(msrs->counters[i].addr, -(s64)reset_value[i]);
249 op_amd_handle_ibs(regs, msrs);
251 /* See op_model_ppro.c */
252 return 1;
255 static void op_amd_start(struct op_msrs const * const msrs)
257 u64 val;
258 int i;
259 for (i = 0 ; i < NUM_COUNTERS ; ++i) {
260 if (reset_value[i]) {
261 rdmsrl(msrs->controls[i].addr, val);
262 val |= ARCH_PERFMON_EVENTSEL0_ENABLE;
263 wrmsrl(msrs->controls[i].addr, val);
267 op_amd_start_ibs();
270 static void op_amd_stop(struct op_msrs const * const msrs)
272 u64 val;
273 int i;
276 * Subtle: stop on all counters to avoid race with setting our
277 * pm callback
279 for (i = 0 ; i < NUM_COUNTERS ; ++i) {
280 if (!reset_value[i])
281 continue;
282 rdmsrl(msrs->controls[i].addr, val);
283 val &= ~ARCH_PERFMON_EVENTSEL0_ENABLE;
284 wrmsrl(msrs->controls[i].addr, val);
287 op_amd_stop_ibs();
290 static void op_amd_shutdown(struct op_msrs const * const msrs)
292 int i;
294 for (i = 0 ; i < NUM_COUNTERS ; ++i) {
295 if (msrs->counters[i].addr)
296 release_perfctr_nmi(MSR_K7_PERFCTR0 + i);
298 for (i = 0 ; i < NUM_CONTROLS ; ++i) {
299 if (msrs->controls[i].addr)
300 release_evntsel_nmi(MSR_K7_EVNTSEL0 + i);
304 #ifdef CONFIG_OPROFILE_IBS
306 static u8 ibs_eilvt_off;
308 static inline void apic_init_ibs_nmi_per_cpu(void *arg)
310 ibs_eilvt_off = setup_APIC_eilvt_ibs(0, APIC_EILVT_MSG_NMI, 0);
313 static inline void apic_clear_ibs_nmi_per_cpu(void *arg)
315 setup_APIC_eilvt_ibs(0, APIC_EILVT_MSG_FIX, 1);
318 static int init_ibs_nmi(void)
320 #define IBSCTL_LVTOFFSETVAL (1 << 8)
321 #define IBSCTL 0x1cc
322 struct pci_dev *cpu_cfg;
323 int nodes;
324 u32 value = 0;
326 /* per CPU setup */
327 on_each_cpu(apic_init_ibs_nmi_per_cpu, NULL, 1);
329 nodes = 0;
330 cpu_cfg = NULL;
331 do {
332 cpu_cfg = pci_get_device(PCI_VENDOR_ID_AMD,
333 PCI_DEVICE_ID_AMD_10H_NB_MISC,
334 cpu_cfg);
335 if (!cpu_cfg)
336 break;
337 ++nodes;
338 pci_write_config_dword(cpu_cfg, IBSCTL, ibs_eilvt_off
339 | IBSCTL_LVTOFFSETVAL);
340 pci_read_config_dword(cpu_cfg, IBSCTL, &value);
341 if (value != (ibs_eilvt_off | IBSCTL_LVTOFFSETVAL)) {
342 pci_dev_put(cpu_cfg);
343 printk(KERN_DEBUG "Failed to setup IBS LVT offset, "
344 "IBSCTL = 0x%08x", value);
345 return 1;
347 } while (1);
349 if (!nodes) {
350 printk(KERN_DEBUG "No CPU node configured for IBS");
351 return 1;
354 #ifdef CONFIG_NUMA
355 /* Sanity check */
356 /* Works only for 64bit with proper numa implementation. */
357 if (nodes != num_possible_nodes()) {
358 printk(KERN_DEBUG "Failed to setup CPU node(s) for IBS, "
359 "found: %d, expected %d",
360 nodes, num_possible_nodes());
361 return 1;
363 #endif
364 return 0;
367 /* uninitialize the APIC for the IBS interrupts if needed */
368 static void clear_ibs_nmi(void)
370 if (has_ibs)
371 on_each_cpu(apic_clear_ibs_nmi_per_cpu, NULL, 1);
374 /* initialize the APIC for the IBS interrupts if available */
375 static void ibs_init(void)
377 has_ibs = boot_cpu_has(X86_FEATURE_IBS);
379 if (!has_ibs)
380 return;
382 if (init_ibs_nmi()) {
383 has_ibs = 0;
384 return;
387 printk(KERN_INFO "oprofile: AMD IBS detected\n");
390 static void ibs_exit(void)
392 if (!has_ibs)
393 return;
395 clear_ibs_nmi();
398 static int (*create_arch_files)(struct super_block *sb, struct dentry *root);
400 static int setup_ibs_files(struct super_block *sb, struct dentry *root)
402 struct dentry *dir;
403 int ret = 0;
405 /* architecture specific files */
406 if (create_arch_files)
407 ret = create_arch_files(sb, root);
409 if (ret)
410 return ret;
412 if (!has_ibs)
413 return ret;
415 /* model specific files */
417 /* setup some reasonable defaults */
418 ibs_config.max_cnt_fetch = 250000;
419 ibs_config.fetch_enabled = 0;
420 ibs_config.max_cnt_op = 250000;
421 ibs_config.op_enabled = 0;
422 ibs_config.dispatched_ops = 1;
424 dir = oprofilefs_mkdir(sb, root, "ibs_fetch");
425 oprofilefs_create_ulong(sb, dir, "enable",
426 &ibs_config.fetch_enabled);
427 oprofilefs_create_ulong(sb, dir, "max_count",
428 &ibs_config.max_cnt_fetch);
429 oprofilefs_create_ulong(sb, dir, "rand_enable",
430 &ibs_config.rand_en);
432 dir = oprofilefs_mkdir(sb, root, "ibs_op");
433 oprofilefs_create_ulong(sb, dir, "enable",
434 &ibs_config.op_enabled);
435 oprofilefs_create_ulong(sb, dir, "max_count",
436 &ibs_config.max_cnt_op);
437 oprofilefs_create_ulong(sb, dir, "dispatched_ops",
438 &ibs_config.dispatched_ops);
440 return 0;
443 static int op_amd_init(struct oprofile_operations *ops)
445 ibs_init();
446 create_arch_files = ops->create_files;
447 ops->create_files = setup_ibs_files;
448 return 0;
451 static void op_amd_exit(void)
453 ibs_exit();
456 #else
458 /* no IBS support */
460 static int op_amd_init(struct oprofile_operations *ops)
462 return 0;
465 static void op_amd_exit(void) {}
467 #endif /* CONFIG_OPROFILE_IBS */
469 struct op_x86_model_spec const op_amd_spec = {
470 .num_counters = NUM_COUNTERS,
471 .num_controls = NUM_CONTROLS,
472 .reserved = MSR_AMD_EVENTSEL_RESERVED,
473 .event_mask = OP_EVENT_MASK,
474 .init = op_amd_init,
475 .exit = op_amd_exit,
476 .fill_in_addresses = &op_amd_fill_in_addresses,
477 .setup_ctrs = &op_amd_setup_ctrs,
478 .check_ctrs = &op_amd_check_ctrs,
479 .start = &op_amd_start,
480 .stop = &op_amd_stop,
481 .shutdown = &op_amd_shutdown,