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
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
;
61 static u64 ibs_op_ctl
;
64 * IBS cpuid feature detection
67 #define IBS_CPUID_FEATURES 0x8000001b
70 * Same bit mask as for IBS cpuid feature flags (Fn8000_001B_EAX), but
71 * bit 0 is used to indicate the existence of IBS.
73 #define IBS_CAPS_AVAIL (1U<<0)
74 #define IBS_CAPS_RDWROPCNT (1U<<3)
75 #define IBS_CAPS_OPCNT (1U<<4)
81 #define IBSCTL_LVT_OFFSET_VALID (1ULL<<8)
82 #define IBSCTL_LVT_OFFSET_MASK 0x0F
85 * IBS randomization macros
87 #define IBS_RANDOM_BITS 12
88 #define IBS_RANDOM_MASK ((1ULL << IBS_RANDOM_BITS) - 1)
89 #define IBS_RANDOM_MAXCNT_OFFSET (1ULL << (IBS_RANDOM_BITS - 5))
91 static u32
get_ibs_caps(void)
94 unsigned int max_level
;
96 if (!boot_cpu_has(X86_FEATURE_IBS
))
99 /* check IBS cpuid feature flags */
100 max_level
= cpuid_eax(0x80000000);
101 if (max_level
< IBS_CPUID_FEATURES
)
102 return IBS_CAPS_AVAIL
;
104 ibs_caps
= cpuid_eax(IBS_CPUID_FEATURES
);
105 if (!(ibs_caps
& IBS_CAPS_AVAIL
))
106 /* cpuid flags not valid */
107 return IBS_CAPS_AVAIL
;
113 * 16-bit Linear Feedback Shift Register (LFSR)
116 * Feedback polynomial = X + X + X + X + 1
118 static unsigned int lfsr_random(void)
120 static unsigned int lfsr_value
= 0xF00D;
123 /* Compute next bit to shift in */
124 bit
= ((lfsr_value
>> 0) ^
127 (lfsr_value
>> 5)) & 0x0001;
129 /* Advance to next register value */
130 lfsr_value
= (lfsr_value
>> 1) | (bit
<< 15);
136 * IBS software randomization
138 * The IBS periodic op counter is randomized in software. The lower 12
139 * bits of the 20 bit counter are randomized. IbsOpCurCnt is
140 * initialized with a 12 bit random value.
142 static inline u64
op_amd_randomize_ibs_op(u64 val
)
144 unsigned int random
= lfsr_random();
146 if (!(ibs_caps
& IBS_CAPS_RDWROPCNT
))
148 * Work around if the hw can not write to IbsOpCurCnt
150 * Randomize the lower 8 bits of the 16 bit
151 * IbsOpMaxCnt [15:0] value in the range of -128 to
152 * +127 by adding/subtracting an offset to the
153 * maximum count (IbsOpMaxCnt).
155 * To avoid over or underflows and protect upper bits
156 * starting at bit 16, the initial value for
157 * IbsOpMaxCnt must fit in the range from 0x0081 to
160 val
+= (s8
)(random
>> 4);
162 val
|= (u64
)(random
& IBS_RANDOM_MASK
) << 32;
168 op_amd_handle_ibs(struct pt_regs
* const regs
,
169 struct op_msrs
const * const msrs
)
172 struct op_entry entry
;
177 if (ibs_config
.fetch_enabled
) {
178 rdmsrl(MSR_AMD64_IBSFETCHCTL
, ctl
);
179 if (ctl
& IBS_FETCH_VAL
) {
180 rdmsrl(MSR_AMD64_IBSFETCHLINAD
, val
);
181 oprofile_write_reserve(&entry
, regs
, val
,
182 IBS_FETCH_CODE
, IBS_FETCH_SIZE
);
183 oprofile_add_data64(&entry
, val
);
184 oprofile_add_data64(&entry
, ctl
);
185 rdmsrl(MSR_AMD64_IBSFETCHPHYSAD
, val
);
186 oprofile_add_data64(&entry
, val
);
187 oprofile_write_commit(&entry
);
189 /* reenable the IRQ */
190 ctl
&= ~(IBS_FETCH_VAL
| IBS_FETCH_CNT
);
191 ctl
|= IBS_FETCH_ENABLE
;
192 wrmsrl(MSR_AMD64_IBSFETCHCTL
, ctl
);
196 if (ibs_config
.op_enabled
) {
197 rdmsrl(MSR_AMD64_IBSOPCTL
, ctl
);
198 if (ctl
& IBS_OP_VAL
) {
199 rdmsrl(MSR_AMD64_IBSOPRIP
, val
);
200 oprofile_write_reserve(&entry
, regs
, val
,
201 IBS_OP_CODE
, IBS_OP_SIZE
);
202 oprofile_add_data64(&entry
, val
);
203 rdmsrl(MSR_AMD64_IBSOPDATA
, val
);
204 oprofile_add_data64(&entry
, val
);
205 rdmsrl(MSR_AMD64_IBSOPDATA2
, val
);
206 oprofile_add_data64(&entry
, val
);
207 rdmsrl(MSR_AMD64_IBSOPDATA3
, val
);
208 oprofile_add_data64(&entry
, val
);
209 rdmsrl(MSR_AMD64_IBSDCLINAD
, val
);
210 oprofile_add_data64(&entry
, val
);
211 rdmsrl(MSR_AMD64_IBSDCPHYSAD
, val
);
212 oprofile_add_data64(&entry
, val
);
213 oprofile_write_commit(&entry
);
215 /* reenable the IRQ */
216 ctl
= op_amd_randomize_ibs_op(ibs_op_ctl
);
217 wrmsrl(MSR_AMD64_IBSOPCTL
, ctl
);
222 static inline void op_amd_start_ibs(void)
229 if (ibs_config
.fetch_enabled
) {
230 val
= (ibs_config
.max_cnt_fetch
>> 4) & IBS_FETCH_MAX_CNT
;
231 val
|= ibs_config
.rand_en
? IBS_FETCH_RAND_EN
: 0;
232 val
|= IBS_FETCH_ENABLE
;
233 wrmsrl(MSR_AMD64_IBSFETCHCTL
, val
);
236 if (ibs_config
.op_enabled
) {
237 ibs_op_ctl
= ibs_config
.max_cnt_op
>> 4;
238 if (!(ibs_caps
& IBS_CAPS_RDWROPCNT
)) {
240 * IbsOpCurCnt not supported. See
241 * op_amd_randomize_ibs_op() for details.
243 ibs_op_ctl
= clamp(ibs_op_ctl
, 0x0081ULL
, 0xFF80ULL
);
246 * The start value is randomized with a
247 * positive offset, we need to compensate it
248 * with the half of the randomized range. Also
251 ibs_op_ctl
= min(ibs_op_ctl
+ IBS_RANDOM_MAXCNT_OFFSET
,
254 if (ibs_caps
& IBS_CAPS_OPCNT
&& ibs_config
.dispatched_ops
)
255 ibs_op_ctl
|= IBS_OP_CNT_CTL
;
256 ibs_op_ctl
|= IBS_OP_ENABLE
;
257 val
= op_amd_randomize_ibs_op(ibs_op_ctl
);
258 wrmsrl(MSR_AMD64_IBSOPCTL
, val
);
262 static void op_amd_stop_ibs(void)
267 if (ibs_config
.fetch_enabled
)
268 /* clear max count and enable */
269 wrmsrl(MSR_AMD64_IBSFETCHCTL
, 0);
271 if (ibs_config
.op_enabled
)
272 /* clear max count and enable */
273 wrmsrl(MSR_AMD64_IBSOPCTL
, 0);
276 static inline int eilvt_is_available(int offset
)
278 /* check if we may assign a vector */
279 return !setup_APIC_eilvt(offset
, 0, APIC_EILVT_MSG_NMI
, 1);
282 static inline int ibs_eilvt_valid(void)
287 rdmsrl(MSR_AMD64_IBSCTL
, val
);
288 if (!(val
& IBSCTL_LVT_OFFSET_VALID
)) {
289 pr_err(FW_BUG
"cpu %d, invalid IBS "
290 "interrupt offset %d (MSR%08X=0x%016llx)",
291 smp_processor_id(), offset
,
292 MSR_AMD64_IBSCTL
, val
);
296 offset
= val
& IBSCTL_LVT_OFFSET_MASK
;
298 if (eilvt_is_available(offset
))
301 pr_err(FW_BUG
"cpu %d, IBS interrupt offset %d "
302 "not available (MSR%08X=0x%016llx)",
303 smp_processor_id(), offset
,
304 MSR_AMD64_IBSCTL
, val
);
309 static inline int get_ibs_offset(void)
313 rdmsrl(MSR_AMD64_IBSCTL
, val
);
314 if (!(val
& IBSCTL_LVT_OFFSET_VALID
))
317 return val
& IBSCTL_LVT_OFFSET_MASK
;
320 static void setup_APIC_ibs(void)
324 offset
= get_ibs_offset();
328 if (!setup_APIC_eilvt(offset
, 0, APIC_EILVT_MSG_NMI
, 0))
331 pr_warn("oprofile: IBS APIC setup failed on cpu #%d\n",
335 static void clear_APIC_ibs(void)
339 offset
= get_ibs_offset();
341 setup_APIC_eilvt(offset
, 0, APIC_EILVT_MSG_FIX
, 1);
344 #ifdef CONFIG_OPROFILE_EVENT_MULTIPLEX
346 static void op_mux_switch_ctrl(struct op_x86_model_spec
const *model
,
347 struct op_msrs
const * const msrs
)
352 /* enable active counters */
353 for (i
= 0; i
< NUM_COUNTERS
; ++i
) {
354 int virt
= op_x86_phys_to_virt(i
);
355 if (!reset_value
[virt
])
357 rdmsrl(msrs
->controls
[i
].addr
, val
);
358 val
&= model
->reserved
;
359 val
|= op_x86_get_ctrl(model
, &counter_config
[virt
]);
360 wrmsrl(msrs
->controls
[i
].addr
, val
);
366 /* functions for op_amd_spec */
368 static void op_amd_shutdown(struct op_msrs
const * const msrs
)
372 for (i
= 0; i
< NUM_COUNTERS
; ++i
) {
373 if (!msrs
->counters
[i
].addr
)
375 release_perfctr_nmi(MSR_K7_PERFCTR0
+ i
);
376 release_evntsel_nmi(MSR_K7_EVNTSEL0
+ i
);
380 static int op_amd_fill_in_addresses(struct op_msrs
* const msrs
)
384 for (i
= 0; i
< NUM_COUNTERS
; i
++) {
385 if (!reserve_perfctr_nmi(MSR_K7_PERFCTR0
+ i
))
387 if (!reserve_evntsel_nmi(MSR_K7_EVNTSEL0
+ i
)) {
388 release_perfctr_nmi(MSR_K7_PERFCTR0
+ i
);
391 /* both registers must be reserved */
392 msrs
->counters
[i
].addr
= MSR_K7_PERFCTR0
+ i
;
393 msrs
->controls
[i
].addr
= MSR_K7_EVNTSEL0
+ i
;
396 if (!counter_config
[i
].enabled
)
398 op_x86_warn_reserved(i
);
399 op_amd_shutdown(msrs
);
406 static void op_amd_setup_ctrs(struct op_x86_model_spec
const *model
,
407 struct op_msrs
const * const msrs
)
412 /* setup reset_value */
413 for (i
= 0; i
< NUM_VIRT_COUNTERS
; ++i
) {
414 if (counter_config
[i
].enabled
415 && msrs
->counters
[op_x86_virt_to_phys(i
)].addr
)
416 reset_value
[i
] = counter_config
[i
].count
;
421 /* clear all counters */
422 for (i
= 0; i
< NUM_COUNTERS
; ++i
) {
423 if (!msrs
->controls
[i
].addr
)
425 rdmsrl(msrs
->controls
[i
].addr
, val
);
426 if (val
& ARCH_PERFMON_EVENTSEL_ENABLE
)
427 op_x86_warn_in_use(i
);
428 val
&= model
->reserved
;
429 wrmsrl(msrs
->controls
[i
].addr
, val
);
431 * avoid a false detection of ctr overflows in NMI
434 wrmsrl(msrs
->counters
[i
].addr
, -1LL);
437 /* enable active counters */
438 for (i
= 0; i
< NUM_COUNTERS
; ++i
) {
439 int virt
= op_x86_phys_to_virt(i
);
440 if (!reset_value
[virt
])
443 /* setup counter registers */
444 wrmsrl(msrs
->counters
[i
].addr
, -(u64
)reset_value
[virt
]);
446 /* setup control registers */
447 rdmsrl(msrs
->controls
[i
].addr
, val
);
448 val
&= model
->reserved
;
449 val
|= op_x86_get_ctrl(model
, &counter_config
[virt
]);
450 wrmsrl(msrs
->controls
[i
].addr
, val
);
457 static void op_amd_cpu_shutdown(void)
463 static int op_amd_check_ctrs(struct pt_regs
* const regs
,
464 struct op_msrs
const * const msrs
)
469 for (i
= 0; i
< NUM_COUNTERS
; ++i
) {
470 int virt
= op_x86_phys_to_virt(i
);
471 if (!reset_value
[virt
])
473 rdmsrl(msrs
->counters
[i
].addr
, val
);
474 /* bit is clear if overflowed: */
475 if (val
& OP_CTR_OVERFLOW
)
477 oprofile_add_sample(regs
, virt
);
478 wrmsrl(msrs
->counters
[i
].addr
, -(u64
)reset_value
[virt
]);
481 op_amd_handle_ibs(regs
, msrs
);
483 /* See op_model_ppro.c */
487 static void op_amd_start(struct op_msrs
const * const msrs
)
492 for (i
= 0; i
< NUM_COUNTERS
; ++i
) {
493 if (!reset_value
[op_x86_phys_to_virt(i
)])
495 rdmsrl(msrs
->controls
[i
].addr
, val
);
496 val
|= ARCH_PERFMON_EVENTSEL_ENABLE
;
497 wrmsrl(msrs
->controls
[i
].addr
, val
);
503 static void op_amd_stop(struct op_msrs
const * const msrs
)
509 * Subtle: stop on all counters to avoid race with setting our
512 for (i
= 0; i
< NUM_COUNTERS
; ++i
) {
513 if (!reset_value
[op_x86_phys_to_virt(i
)])
515 rdmsrl(msrs
->controls
[i
].addr
, val
);
516 val
&= ~ARCH_PERFMON_EVENTSEL_ENABLE
;
517 wrmsrl(msrs
->controls
[i
].addr
, val
);
523 static int setup_ibs_ctl(int ibs_eilvt_off
)
525 struct pci_dev
*cpu_cfg
;
532 cpu_cfg
= pci_get_device(PCI_VENDOR_ID_AMD
,
533 PCI_DEVICE_ID_AMD_10H_NB_MISC
,
538 pci_write_config_dword(cpu_cfg
, IBSCTL
, ibs_eilvt_off
539 | IBSCTL_LVT_OFFSET_VALID
);
540 pci_read_config_dword(cpu_cfg
, IBSCTL
, &value
);
541 if (value
!= (ibs_eilvt_off
| IBSCTL_LVT_OFFSET_VALID
)) {
542 pci_dev_put(cpu_cfg
);
543 printk(KERN_DEBUG
"Failed to setup IBS LVT offset, "
544 "IBSCTL = 0x%08x\n", value
);
550 printk(KERN_DEBUG
"No CPU node configured for IBS\n");
557 static int force_ibs_eilvt_setup(void)
562 /* find the next free available EILVT entry */
563 for (i
= 1; i
< 4; i
++) {
564 if (!eilvt_is_available(i
))
566 ret
= setup_ibs_ctl(i
);
572 printk(KERN_DEBUG
"No EILVT entry available\n");
577 static int __init_ibs_nmi(void)
581 if (ibs_eilvt_valid())
584 ret
= force_ibs_eilvt_setup();
588 if (!ibs_eilvt_valid())
591 pr_err(FW_BUG
"workaround enabled for IBS LVT offset\n");
596 /* initialize the APIC for the IBS interrupts if available */
597 static void init_ibs(void)
599 ibs_caps
= get_ibs_caps();
604 if (__init_ibs_nmi()) {
609 printk(KERN_INFO
"oprofile: AMD IBS detected (0x%08x)\n",
613 static int (*create_arch_files
)(struct super_block
*sb
, struct dentry
*root
);
615 static int setup_ibs_files(struct super_block
*sb
, struct dentry
*root
)
620 /* architecture specific files */
621 if (create_arch_files
)
622 ret
= create_arch_files(sb
, root
);
630 /* model specific files */
632 /* setup some reasonable defaults */
633 ibs_config
.max_cnt_fetch
= 250000;
634 ibs_config
.fetch_enabled
= 0;
635 ibs_config
.max_cnt_op
= 250000;
636 ibs_config
.op_enabled
= 0;
637 ibs_config
.dispatched_ops
= 0;
639 dir
= oprofilefs_mkdir(sb
, root
, "ibs_fetch");
640 oprofilefs_create_ulong(sb
, dir
, "enable",
641 &ibs_config
.fetch_enabled
);
642 oprofilefs_create_ulong(sb
, dir
, "max_count",
643 &ibs_config
.max_cnt_fetch
);
644 oprofilefs_create_ulong(sb
, dir
, "rand_enable",
645 &ibs_config
.rand_en
);
647 dir
= oprofilefs_mkdir(sb
, root
, "ibs_op");
648 oprofilefs_create_ulong(sb
, dir
, "enable",
649 &ibs_config
.op_enabled
);
650 oprofilefs_create_ulong(sb
, dir
, "max_count",
651 &ibs_config
.max_cnt_op
);
652 if (ibs_caps
& IBS_CAPS_OPCNT
)
653 oprofilefs_create_ulong(sb
, dir
, "dispatched_ops",
654 &ibs_config
.dispatched_ops
);
659 static int op_amd_init(struct oprofile_operations
*ops
)
662 create_arch_files
= ops
->create_files
;
663 ops
->create_files
= setup_ibs_files
;
667 struct op_x86_model_spec op_amd_spec
= {
668 .num_counters
= NUM_COUNTERS
,
669 .num_controls
= NUM_COUNTERS
,
670 .num_virt_counters
= NUM_VIRT_COUNTERS
,
671 .reserved
= MSR_AMD_EVENTSEL_RESERVED
,
672 .event_mask
= OP_EVENT_MASK
,
674 .fill_in_addresses
= &op_amd_fill_in_addresses
,
675 .setup_ctrs
= &op_amd_setup_ctrs
,
676 .cpu_down
= &op_amd_cpu_shutdown
,
677 .check_ctrs
= &op_amd_check_ctrs
,
678 .start
= &op_amd_start
,
679 .stop
= &op_amd_stop
,
680 .shutdown
= &op_amd_shutdown
,
681 #ifdef CONFIG_OPROFILE_EVENT_MULTIPLEX
682 .switch_ctrl
= &op_mux_switch_ctrl
,