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_COUNTERS_F15H 6
34 #ifdef CONFIG_OPROFILE_EVENT_MULTIPLEX
35 #define NUM_VIRT_COUNTERS 32
37 #define NUM_VIRT_COUNTERS 0
40 #define OP_EVENT_MASK 0x0FFF
41 #define OP_CTR_OVERFLOW (1ULL<<31)
43 #define MSR_AMD_EVENTSEL_RESERVED ((0xFFFFFCF0ULL<<32)|(1ULL<<21))
45 static int num_counters
;
46 static unsigned long reset_value
[OP_MAX_COUNTER
];
48 #define IBS_FETCH_SIZE 6
49 #define IBS_OP_SIZE 12
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
;
60 unsigned long branch_target
;
66 unsigned long sample_size
;
69 static struct ibs_config ibs_config
;
70 static struct ibs_state ibs_state
;
73 * IBS cpuid feature detection
76 #define IBS_CPUID_FEATURES 0x8000001b
79 * Same bit mask as for IBS cpuid feature flags (Fn8000_001B_EAX), but
80 * bit 0 is used to indicate the existence of IBS.
82 #define IBS_CAPS_AVAIL (1U<<0)
83 #define IBS_CAPS_FETCHSAM (1U<<1)
84 #define IBS_CAPS_OPSAM (1U<<2)
85 #define IBS_CAPS_RDWROPCNT (1U<<3)
86 #define IBS_CAPS_OPCNT (1U<<4)
87 #define IBS_CAPS_BRNTRGT (1U<<5)
88 #define IBS_CAPS_OPCNTEXT (1U<<6)
90 #define IBS_CAPS_DEFAULT (IBS_CAPS_AVAIL \
98 #define IBSCTL_LVT_OFFSET_VALID (1ULL<<8)
99 #define IBSCTL_LVT_OFFSET_MASK 0x0F
102 * IBS randomization macros
104 #define IBS_RANDOM_BITS 12
105 #define IBS_RANDOM_MASK ((1ULL << IBS_RANDOM_BITS) - 1)
106 #define IBS_RANDOM_MAXCNT_OFFSET (1ULL << (IBS_RANDOM_BITS - 5))
108 static u32
get_ibs_caps(void)
111 unsigned int max_level
;
113 if (!boot_cpu_has(X86_FEATURE_IBS
))
116 /* check IBS cpuid feature flags */
117 max_level
= cpuid_eax(0x80000000);
118 if (max_level
< IBS_CPUID_FEATURES
)
119 return IBS_CAPS_DEFAULT
;
121 ibs_caps
= cpuid_eax(IBS_CPUID_FEATURES
);
122 if (!(ibs_caps
& IBS_CAPS_AVAIL
))
123 /* cpuid flags not valid */
124 return IBS_CAPS_DEFAULT
;
130 * 16-bit Linear Feedback Shift Register (LFSR)
133 * Feedback polynomial = X + X + X + X + 1
135 static unsigned int lfsr_random(void)
137 static unsigned int lfsr_value
= 0xF00D;
140 /* Compute next bit to shift in */
141 bit
= ((lfsr_value
>> 0) ^
144 (lfsr_value
>> 5)) & 0x0001;
146 /* Advance to next register value */
147 lfsr_value
= (lfsr_value
>> 1) | (bit
<< 15);
153 * IBS software randomization
155 * The IBS periodic op counter is randomized in software. The lower 12
156 * bits of the 20 bit counter are randomized. IbsOpCurCnt is
157 * initialized with a 12 bit random value.
159 static inline u64
op_amd_randomize_ibs_op(u64 val
)
161 unsigned int random
= lfsr_random();
163 if (!(ibs_caps
& IBS_CAPS_RDWROPCNT
))
165 * Work around if the hw can not write to IbsOpCurCnt
167 * Randomize the lower 8 bits of the 16 bit
168 * IbsOpMaxCnt [15:0] value in the range of -128 to
169 * +127 by adding/subtracting an offset to the
170 * maximum count (IbsOpMaxCnt).
172 * To avoid over or underflows and protect upper bits
173 * starting at bit 16, the initial value for
174 * IbsOpMaxCnt must fit in the range from 0x0081 to
177 val
+= (s8
)(random
>> 4);
179 val
|= (u64
)(random
& IBS_RANDOM_MASK
) << 32;
185 op_amd_handle_ibs(struct pt_regs
* const regs
,
186 struct op_msrs
const * const msrs
)
189 struct op_entry entry
;
194 if (ibs_config
.fetch_enabled
) {
195 rdmsrl(MSR_AMD64_IBSFETCHCTL
, ctl
);
196 if (ctl
& IBS_FETCH_VAL
) {
197 rdmsrl(MSR_AMD64_IBSFETCHLINAD
, val
);
198 oprofile_write_reserve(&entry
, regs
, val
,
199 IBS_FETCH_CODE
, IBS_FETCH_SIZE
);
200 oprofile_add_data64(&entry
, val
);
201 oprofile_add_data64(&entry
, ctl
);
202 rdmsrl(MSR_AMD64_IBSFETCHPHYSAD
, val
);
203 oprofile_add_data64(&entry
, val
);
204 oprofile_write_commit(&entry
);
206 /* reenable the IRQ */
207 ctl
&= ~(IBS_FETCH_VAL
| IBS_FETCH_CNT
);
208 ctl
|= IBS_FETCH_ENABLE
;
209 wrmsrl(MSR_AMD64_IBSFETCHCTL
, ctl
);
213 if (ibs_config
.op_enabled
) {
214 rdmsrl(MSR_AMD64_IBSOPCTL
, ctl
);
215 if (ctl
& IBS_OP_VAL
) {
216 rdmsrl(MSR_AMD64_IBSOPRIP
, val
);
217 oprofile_write_reserve(&entry
, regs
, val
, IBS_OP_CODE
,
218 ibs_state
.sample_size
);
219 oprofile_add_data64(&entry
, val
);
220 rdmsrl(MSR_AMD64_IBSOPDATA
, val
);
221 oprofile_add_data64(&entry
, val
);
222 rdmsrl(MSR_AMD64_IBSOPDATA2
, val
);
223 oprofile_add_data64(&entry
, val
);
224 rdmsrl(MSR_AMD64_IBSOPDATA3
, val
);
225 oprofile_add_data64(&entry
, val
);
226 rdmsrl(MSR_AMD64_IBSDCLINAD
, val
);
227 oprofile_add_data64(&entry
, val
);
228 rdmsrl(MSR_AMD64_IBSDCPHYSAD
, val
);
229 oprofile_add_data64(&entry
, val
);
230 if (ibs_state
.branch_target
) {
231 rdmsrl(MSR_AMD64_IBSBRTARGET
, val
);
232 oprofile_add_data(&entry
, (unsigned long)val
);
234 oprofile_write_commit(&entry
);
236 /* reenable the IRQ */
237 ctl
= op_amd_randomize_ibs_op(ibs_state
.ibs_op_ctl
);
238 wrmsrl(MSR_AMD64_IBSOPCTL
, ctl
);
243 static inline void op_amd_start_ibs(void)
250 memset(&ibs_state
, 0, sizeof(ibs_state
));
253 * Note: Since the max count settings may out of range we
254 * write back the actual used values so that userland can read
258 if (ibs_config
.fetch_enabled
) {
259 val
= ibs_config
.max_cnt_fetch
>> 4;
260 val
= min(val
, IBS_FETCH_MAX_CNT
);
261 ibs_config
.max_cnt_fetch
= val
<< 4;
262 val
|= ibs_config
.rand_en
? IBS_FETCH_RAND_EN
: 0;
263 val
|= IBS_FETCH_ENABLE
;
264 wrmsrl(MSR_AMD64_IBSFETCHCTL
, val
);
267 if (ibs_config
.op_enabled
) {
268 val
= ibs_config
.max_cnt_op
>> 4;
269 if (!(ibs_caps
& IBS_CAPS_RDWROPCNT
)) {
271 * IbsOpCurCnt not supported. See
272 * op_amd_randomize_ibs_op() for details.
274 val
= clamp(val
, 0x0081ULL
, 0xFF80ULL
);
275 ibs_config
.max_cnt_op
= val
<< 4;
278 * The start value is randomized with a
279 * positive offset, we need to compensate it
280 * with the half of the randomized range. Also
283 val
+= IBS_RANDOM_MAXCNT_OFFSET
;
284 if (ibs_caps
& IBS_CAPS_OPCNTEXT
)
285 val
= min(val
, IBS_OP_MAX_CNT_EXT
);
287 val
= min(val
, IBS_OP_MAX_CNT
);
288 ibs_config
.max_cnt_op
=
289 (val
- IBS_RANDOM_MAXCNT_OFFSET
) << 4;
291 val
= ((val
& ~IBS_OP_MAX_CNT
) << 4) | (val
& IBS_OP_MAX_CNT
);
292 val
|= ibs_config
.dispatched_ops
? IBS_OP_CNT_CTL
: 0;
293 val
|= IBS_OP_ENABLE
;
294 ibs_state
.ibs_op_ctl
= val
;
295 ibs_state
.sample_size
= IBS_OP_SIZE
;
296 if (ibs_config
.branch_target
) {
297 ibs_state
.branch_target
= 1;
298 ibs_state
.sample_size
++;
300 val
= op_amd_randomize_ibs_op(ibs_state
.ibs_op_ctl
);
301 wrmsrl(MSR_AMD64_IBSOPCTL
, val
);
305 static void op_amd_stop_ibs(void)
310 if (ibs_config
.fetch_enabled
)
311 /* clear max count and enable */
312 wrmsrl(MSR_AMD64_IBSFETCHCTL
, 0);
314 if (ibs_config
.op_enabled
)
315 /* clear max count and enable */
316 wrmsrl(MSR_AMD64_IBSOPCTL
, 0);
319 static inline int eilvt_is_available(int offset
)
321 /* check if we may assign a vector */
322 return !setup_APIC_eilvt(offset
, 0, APIC_EILVT_MSG_NMI
, 1);
325 static inline int ibs_eilvt_valid(void)
330 rdmsrl(MSR_AMD64_IBSCTL
, val
);
331 offset
= val
& IBSCTL_LVT_OFFSET_MASK
;
333 if (!(val
& IBSCTL_LVT_OFFSET_VALID
)) {
334 pr_err(FW_BUG
"cpu %d, invalid IBS interrupt offset %d (MSR%08X=0x%016llx)\n",
335 smp_processor_id(), offset
, MSR_AMD64_IBSCTL
, val
);
339 if (!eilvt_is_available(offset
)) {
340 pr_err(FW_BUG
"cpu %d, IBS interrupt offset %d not available (MSR%08X=0x%016llx)\n",
341 smp_processor_id(), offset
, MSR_AMD64_IBSCTL
, val
);
348 static inline int get_ibs_offset(void)
352 rdmsrl(MSR_AMD64_IBSCTL
, val
);
353 if (!(val
& IBSCTL_LVT_OFFSET_VALID
))
356 return val
& IBSCTL_LVT_OFFSET_MASK
;
359 static void setup_APIC_ibs(void)
363 offset
= get_ibs_offset();
367 if (!setup_APIC_eilvt(offset
, 0, APIC_EILVT_MSG_NMI
, 0))
370 pr_warn("oprofile: IBS APIC setup failed on cpu #%d\n",
374 static void clear_APIC_ibs(void)
378 offset
= get_ibs_offset();
380 setup_APIC_eilvt(offset
, 0, APIC_EILVT_MSG_FIX
, 1);
383 #ifdef CONFIG_OPROFILE_EVENT_MULTIPLEX
385 static void op_mux_switch_ctrl(struct op_x86_model_spec
const *model
,
386 struct op_msrs
const * const msrs
)
391 /* enable active counters */
392 for (i
= 0; i
< num_counters
; ++i
) {
393 int virt
= op_x86_phys_to_virt(i
);
394 if (!reset_value
[virt
])
396 rdmsrl(msrs
->controls
[i
].addr
, val
);
397 val
&= model
->reserved
;
398 val
|= op_x86_get_ctrl(model
, &counter_config
[virt
]);
399 wrmsrl(msrs
->controls
[i
].addr
, val
);
405 /* functions for op_amd_spec */
407 static void op_amd_shutdown(struct op_msrs
const * const msrs
)
411 for (i
= 0; i
< num_counters
; ++i
) {
412 if (!msrs
->counters
[i
].addr
)
414 release_perfctr_nmi(MSR_K7_PERFCTR0
+ i
);
415 release_evntsel_nmi(MSR_K7_EVNTSEL0
+ i
);
419 static int op_amd_fill_in_addresses(struct op_msrs
* const msrs
)
423 for (i
= 0; i
< num_counters
; i
++) {
424 if (!reserve_perfctr_nmi(MSR_K7_PERFCTR0
+ i
))
426 if (!reserve_evntsel_nmi(MSR_K7_EVNTSEL0
+ i
)) {
427 release_perfctr_nmi(MSR_K7_PERFCTR0
+ i
);
430 /* both registers must be reserved */
431 if (num_counters
== NUM_COUNTERS_F15H
) {
432 msrs
->counters
[i
].addr
= MSR_F15H_PERF_CTR
+ (i
<< 1);
433 msrs
->controls
[i
].addr
= MSR_F15H_PERF_CTL
+ (i
<< 1);
435 msrs
->controls
[i
].addr
= MSR_K7_EVNTSEL0
+ i
;
436 msrs
->counters
[i
].addr
= MSR_K7_PERFCTR0
+ i
;
440 if (!counter_config
[i
].enabled
)
442 op_x86_warn_reserved(i
);
443 op_amd_shutdown(msrs
);
450 static void op_amd_setup_ctrs(struct op_x86_model_spec
const *model
,
451 struct op_msrs
const * const msrs
)
456 /* setup reset_value */
457 for (i
= 0; i
< OP_MAX_COUNTER
; ++i
) {
458 if (counter_config
[i
].enabled
459 && msrs
->counters
[op_x86_virt_to_phys(i
)].addr
)
460 reset_value
[i
] = counter_config
[i
].count
;
465 /* clear all counters */
466 for (i
= 0; i
< num_counters
; ++i
) {
467 if (!msrs
->controls
[i
].addr
)
469 rdmsrl(msrs
->controls
[i
].addr
, val
);
470 if (val
& ARCH_PERFMON_EVENTSEL_ENABLE
)
471 op_x86_warn_in_use(i
);
472 val
&= model
->reserved
;
473 wrmsrl(msrs
->controls
[i
].addr
, val
);
475 * avoid a false detection of ctr overflows in NMI
478 wrmsrl(msrs
->counters
[i
].addr
, -1LL);
481 /* enable active counters */
482 for (i
= 0; i
< num_counters
; ++i
) {
483 int virt
= op_x86_phys_to_virt(i
);
484 if (!reset_value
[virt
])
487 /* setup counter registers */
488 wrmsrl(msrs
->counters
[i
].addr
, -(u64
)reset_value
[virt
]);
490 /* setup control registers */
491 rdmsrl(msrs
->controls
[i
].addr
, val
);
492 val
&= model
->reserved
;
493 val
|= op_x86_get_ctrl(model
, &counter_config
[virt
]);
494 wrmsrl(msrs
->controls
[i
].addr
, val
);
501 static void op_amd_cpu_shutdown(void)
507 static int op_amd_check_ctrs(struct pt_regs
* const regs
,
508 struct op_msrs
const * const msrs
)
513 for (i
= 0; i
< num_counters
; ++i
) {
514 int virt
= op_x86_phys_to_virt(i
);
515 if (!reset_value
[virt
])
517 rdmsrl(msrs
->counters
[i
].addr
, val
);
518 /* bit is clear if overflowed: */
519 if (val
& OP_CTR_OVERFLOW
)
521 oprofile_add_sample(regs
, virt
);
522 wrmsrl(msrs
->counters
[i
].addr
, -(u64
)reset_value
[virt
]);
525 op_amd_handle_ibs(regs
, msrs
);
527 /* See op_model_ppro.c */
531 static void op_amd_start(struct op_msrs
const * const msrs
)
536 for (i
= 0; i
< num_counters
; ++i
) {
537 if (!reset_value
[op_x86_phys_to_virt(i
)])
539 rdmsrl(msrs
->controls
[i
].addr
, val
);
540 val
|= ARCH_PERFMON_EVENTSEL_ENABLE
;
541 wrmsrl(msrs
->controls
[i
].addr
, val
);
547 static void op_amd_stop(struct op_msrs
const * const msrs
)
553 * Subtle: stop on all counters to avoid race with setting our
556 for (i
= 0; i
< num_counters
; ++i
) {
557 if (!reset_value
[op_x86_phys_to_virt(i
)])
559 rdmsrl(msrs
->controls
[i
].addr
, val
);
560 val
&= ~ARCH_PERFMON_EVENTSEL_ENABLE
;
561 wrmsrl(msrs
->controls
[i
].addr
, val
);
567 static int setup_ibs_ctl(int ibs_eilvt_off
)
569 struct pci_dev
*cpu_cfg
;
576 cpu_cfg
= pci_get_device(PCI_VENDOR_ID_AMD
,
577 PCI_DEVICE_ID_AMD_10H_NB_MISC
,
582 pci_write_config_dword(cpu_cfg
, IBSCTL
, ibs_eilvt_off
583 | IBSCTL_LVT_OFFSET_VALID
);
584 pci_read_config_dword(cpu_cfg
, IBSCTL
, &value
);
585 if (value
!= (ibs_eilvt_off
| IBSCTL_LVT_OFFSET_VALID
)) {
586 pci_dev_put(cpu_cfg
);
587 printk(KERN_DEBUG
"Failed to setup IBS LVT offset, "
588 "IBSCTL = 0x%08x\n", value
);
594 printk(KERN_DEBUG
"No CPU node configured for IBS\n");
601 static int force_ibs_eilvt_setup(void)
606 /* find the next free available EILVT entry */
607 for (i
= 1; i
< 4; i
++) {
608 if (!eilvt_is_available(i
))
610 ret
= setup_ibs_ctl(i
);
613 pr_err(FW_BUG
"using offset %d for IBS interrupts\n", i
);
617 printk(KERN_DEBUG
"No EILVT entry available\n");
622 static int __init_ibs_nmi(void)
626 if (ibs_eilvt_valid())
629 ret
= force_ibs_eilvt_setup();
633 if (!ibs_eilvt_valid())
636 pr_err(FW_BUG
"workaround enabled for IBS LVT offset\n");
642 * check and reserve APIC extended interrupt LVT offset for IBS if
645 * init_ibs() preforms implicitly cpu-local operations, so pin this
646 * thread to its current CPU
649 static void init_ibs(void)
653 ibs_caps
= get_ibs_caps();
657 if (__init_ibs_nmi() < 0)
660 printk(KERN_INFO
"oprofile: AMD IBS detected (0x%08x)\n", ibs_caps
);
666 static int (*create_arch_files
)(struct super_block
*sb
, struct dentry
*root
);
668 static int setup_ibs_files(struct super_block
*sb
, struct dentry
*root
)
673 /* architecture specific files */
674 if (create_arch_files
)
675 ret
= create_arch_files(sb
, root
);
683 /* model specific files */
685 /* setup some reasonable defaults */
686 memset(&ibs_config
, 0, sizeof(ibs_config
));
687 ibs_config
.max_cnt_fetch
= 250000;
688 ibs_config
.max_cnt_op
= 250000;
690 if (ibs_caps
& IBS_CAPS_FETCHSAM
) {
691 dir
= oprofilefs_mkdir(sb
, root
, "ibs_fetch");
692 oprofilefs_create_ulong(sb
, dir
, "enable",
693 &ibs_config
.fetch_enabled
);
694 oprofilefs_create_ulong(sb
, dir
, "max_count",
695 &ibs_config
.max_cnt_fetch
);
696 oprofilefs_create_ulong(sb
, dir
, "rand_enable",
697 &ibs_config
.rand_en
);
700 if (ibs_caps
& IBS_CAPS_OPSAM
) {
701 dir
= oprofilefs_mkdir(sb
, root
, "ibs_op");
702 oprofilefs_create_ulong(sb
, dir
, "enable",
703 &ibs_config
.op_enabled
);
704 oprofilefs_create_ulong(sb
, dir
, "max_count",
705 &ibs_config
.max_cnt_op
);
706 if (ibs_caps
& IBS_CAPS_OPCNT
)
707 oprofilefs_create_ulong(sb
, dir
, "dispatched_ops",
708 &ibs_config
.dispatched_ops
);
709 if (ibs_caps
& IBS_CAPS_BRNTRGT
)
710 oprofilefs_create_ulong(sb
, dir
, "branch_target",
711 &ibs_config
.branch_target
);
717 struct op_x86_model_spec op_amd_spec
;
719 static int op_amd_init(struct oprofile_operations
*ops
)
722 create_arch_files
= ops
->create_files
;
723 ops
->create_files
= setup_ibs_files
;
725 if (boot_cpu_data
.x86
== 0x15) {
726 num_counters
= NUM_COUNTERS_F15H
;
728 num_counters
= NUM_COUNTERS
;
731 op_amd_spec
.num_counters
= num_counters
;
732 op_amd_spec
.num_controls
= num_counters
;
733 op_amd_spec
.num_virt_counters
= max(num_counters
, NUM_VIRT_COUNTERS
);
738 struct op_x86_model_spec op_amd_spec
= {
739 /* num_counters/num_controls filled in at runtime */
740 .reserved
= MSR_AMD_EVENTSEL_RESERVED
,
741 .event_mask
= OP_EVENT_MASK
,
743 .fill_in_addresses
= &op_amd_fill_in_addresses
,
744 .setup_ctrs
= &op_amd_setup_ctrs
,
745 .cpu_down
= &op_amd_cpu_shutdown
,
746 .check_ctrs
= &op_amd_check_ctrs
,
747 .start
= &op_amd_start
,
748 .stop
= &op_amd_stop
,
749 .shutdown
= &op_amd_shutdown
,
750 #ifdef CONFIG_OPROFILE_EVENT_MULTIPLEX
751 .switch_ctrl
= &op_mux_switch_ctrl
,