diff options
Diffstat (limited to 'arch/powerpc/platforms/pseries')
-rw-r--r-- | arch/powerpc/platforms/pseries/hvCall.S | 132 | ||||
-rw-r--r-- | arch/powerpc/platforms/pseries/hvCall_inst.c | 38 | ||||
-rw-r--r-- | arch/powerpc/platforms/pseries/lpar.c | 33 |
3 files changed, 150 insertions, 53 deletions
diff --git a/arch/powerpc/platforms/pseries/hvCall.S b/arch/powerpc/platforms/pseries/hvCall.S index c1427b3634ec..383a5d0e9818 100644 --- a/arch/powerpc/platforms/pseries/hvCall.S +++ b/arch/powerpc/platforms/pseries/hvCall.S | |||
@@ -14,68 +14,94 @@ | |||
14 | 14 | ||
15 | #define STK_PARM(i) (48 + ((i)-3)*8) | 15 | #define STK_PARM(i) (48 + ((i)-3)*8) |
16 | 16 | ||
17 | #ifdef CONFIG_HCALL_STATS | 17 | #ifdef CONFIG_TRACEPOINTS |
18 | |||
19 | .section ".toc","aw" | ||
20 | |||
21 | .globl hcall_tracepoint_refcount | ||
22 | hcall_tracepoint_refcount: | ||
23 | .llong 0 | ||
24 | |||
25 | .section ".text" | ||
26 | |||
18 | /* | 27 | /* |
19 | * precall must preserve all registers. use unused STK_PARM() | 28 | * precall must preserve all registers. use unused STK_PARM() |
20 | * areas to save snapshots and opcode. | 29 | * areas to save snapshots and opcode. We branch around this |
30 | * in early init (eg when populating the MMU hashtable) by using an | ||
31 | * unconditional cpu feature. | ||
21 | */ | 32 | */ |
22 | #define HCALL_INST_PRECALL \ | 33 | #define HCALL_INST_PRECALL(FIRST_REG) \ |
23 | std r3,STK_PARM(r3)(r1); /* save opcode */ \ | ||
24 | mftb r0; /* get timebase and */ \ | ||
25 | std r0,STK_PARM(r5)(r1); /* save for later */ \ | ||
26 | BEGIN_FTR_SECTION; \ | 34 | BEGIN_FTR_SECTION; \ |
27 | mfspr r0,SPRN_PURR; /* get PURR and */ \ | 35 | b 1f; \ |
28 | std r0,STK_PARM(r6)(r1); /* save for later */ \ | 36 | END_FTR_SECTION(0, 1); \ |
29 | END_FTR_SECTION_IFSET(CPU_FTR_PURR); | 37 | ld r12,hcall_tracepoint_refcount@toc(r2); \ |
30 | 38 | cmpdi r12,0; \ | |
39 | beq+ 1f; \ | ||
40 | mflr r0; \ | ||
41 | std r3,STK_PARM(r3)(r1); \ | ||
42 | std r4,STK_PARM(r4)(r1); \ | ||
43 | std r5,STK_PARM(r5)(r1); \ | ||
44 | std r6,STK_PARM(r6)(r1); \ | ||
45 | std r7,STK_PARM(r7)(r1); \ | ||
46 | std r8,STK_PARM(r8)(r1); \ | ||
47 | std r9,STK_PARM(r9)(r1); \ | ||
48 | std r10,STK_PARM(r10)(r1); \ | ||
49 | std r0,16(r1); \ | ||
50 | addi r4,r1,STK_PARM(FIRST_REG); \ | ||
51 | stdu r1,-STACK_FRAME_OVERHEAD(r1); \ | ||
52 | bl .__trace_hcall_entry; \ | ||
53 | addi r1,r1,STACK_FRAME_OVERHEAD; \ | ||
54 | ld r0,16(r1); \ | ||
55 | ld r3,STK_PARM(r3)(r1); \ | ||
56 | ld r4,STK_PARM(r4)(r1); \ | ||
57 | ld r5,STK_PARM(r5)(r1); \ | ||
58 | ld r6,STK_PARM(r6)(r1); \ | ||
59 | ld r7,STK_PARM(r7)(r1); \ | ||
60 | ld r8,STK_PARM(r8)(r1); \ | ||
61 | ld r9,STK_PARM(r9)(r1); \ | ||
62 | ld r10,STK_PARM(r10)(r1); \ | ||
63 | mtlr r0; \ | ||
64 | 1: | ||
65 | |||
31 | /* | 66 | /* |
32 | * postcall is performed immediately before function return which | 67 | * postcall is performed immediately before function return which |
33 | * allows liberal use of volatile registers. We branch around this | 68 | * allows liberal use of volatile registers. We branch around this |
34 | * in early init (eg when populating the MMU hashtable) by using an | 69 | * in early init (eg when populating the MMU hashtable) by using an |
35 | * unconditional cpu feature. | 70 | * unconditional cpu feature. |
36 | */ | 71 | */ |
37 | #define HCALL_INST_POSTCALL \ | 72 | #define __HCALL_INST_POSTCALL \ |
38 | BEGIN_FTR_SECTION; \ | 73 | BEGIN_FTR_SECTION; \ |
39 | b 1f; \ | 74 | b 1f; \ |
40 | END_FTR_SECTION(0, 1); \ | 75 | END_FTR_SECTION(0, 1); \ |
41 | ld r4,STK_PARM(r3)(r1); /* validate opcode */ \ | 76 | ld r12,hcall_tracepoint_refcount@toc(r2); \ |
42 | cmpldi cr7,r4,MAX_HCALL_OPCODE; \ | 77 | cmpdi r12,0; \ |
43 | bgt- cr7,1f; \ | 78 | beq+ 1f; \ |
44 | \ | 79 | mflr r0; \ |
45 | /* get time and PURR snapshots after hcall */ \ | 80 | ld r6,STK_PARM(r3)(r1); \ |
46 | mftb r7; /* timebase after */ \ | 81 | std r3,STK_PARM(r3)(r1); \ |
47 | BEGIN_FTR_SECTION; \ | 82 | mr r4,r3; \ |
48 | mfspr r8,SPRN_PURR; /* PURR after */ \ | 83 | mr r3,r6; \ |
49 | ld r6,STK_PARM(r6)(r1); /* PURR before */ \ | 84 | std r0,16(r1); \ |
50 | subf r6,r6,r8; /* delta */ \ | 85 | stdu r1,-STACK_FRAME_OVERHEAD(r1); \ |
51 | END_FTR_SECTION_IFSET(CPU_FTR_PURR); \ | 86 | bl .__trace_hcall_exit; \ |
52 | ld r5,STK_PARM(r5)(r1); /* timebase before */ \ | 87 | addi r1,r1,STACK_FRAME_OVERHEAD; \ |
53 | subf r5,r5,r7; /* time delta */ \ | 88 | ld r0,16(r1); \ |
54 | \ | 89 | ld r3,STK_PARM(r3)(r1); \ |
55 | /* calculate address of stat structure r4 = opcode */ \ | 90 | mtlr r0; \ |
56 | srdi r4,r4,2; /* index into array */ \ | ||
57 | mulli r4,r4,HCALL_STAT_SIZE; \ | ||
58 | LOAD_REG_ADDR(r7, per_cpu__hcall_stats); \ | ||
59 | add r4,r4,r7; \ | ||
60 | ld r7,PACA_DATA_OFFSET(r13); /* per cpu offset */ \ | ||
61 | add r4,r4,r7; \ | ||
62 | \ | ||
63 | /* update stats */ \ | ||
64 | ld r7,HCALL_STAT_CALLS(r4); /* count */ \ | ||
65 | addi r7,r7,1; \ | ||
66 | std r7,HCALL_STAT_CALLS(r4); \ | ||
67 | ld r7,HCALL_STAT_TB(r4); /* timebase */ \ | ||
68 | add r7,r7,r5; \ | ||
69 | std r7,HCALL_STAT_TB(r4); \ | ||
70 | BEGIN_FTR_SECTION; \ | ||
71 | ld r7,HCALL_STAT_PURR(r4); /* PURR */ \ | ||
72 | add r7,r7,r6; \ | ||
73 | std r7,HCALL_STAT_PURR(r4); \ | ||
74 | END_FTR_SECTION_IFSET(CPU_FTR_PURR); \ | ||
75 | 1: | 91 | 1: |
92 | |||
93 | #define HCALL_INST_POSTCALL_NORETS \ | ||
94 | li r5,0; \ | ||
95 | __HCALL_INST_POSTCALL | ||
96 | |||
97 | #define HCALL_INST_POSTCALL(BUFREG) \ | ||
98 | mr r5,BUFREG; \ | ||
99 | __HCALL_INST_POSTCALL | ||
100 | |||
76 | #else | 101 | #else |
77 | #define HCALL_INST_PRECALL | 102 | #define HCALL_INST_PRECALL(FIRST_ARG) |
78 | #define HCALL_INST_POSTCALL | 103 | #define HCALL_INST_POSTCALL_NORETS |
104 | #define HCALL_INST_POSTCALL(BUFREG) | ||
79 | #endif | 105 | #endif |
80 | 106 | ||
81 | .text | 107 | .text |
@@ -86,11 +112,11 @@ _GLOBAL(plpar_hcall_norets) | |||
86 | mfcr r0 | 112 | mfcr r0 |
87 | stw r0,8(r1) | 113 | stw r0,8(r1) |
88 | 114 | ||
89 | HCALL_INST_PRECALL | 115 | HCALL_INST_PRECALL(r4) |
90 | 116 | ||
91 | HVSC /* invoke the hypervisor */ | 117 | HVSC /* invoke the hypervisor */ |
92 | 118 | ||
93 | HCALL_INST_POSTCALL | 119 | HCALL_INST_POSTCALL_NORETS |
94 | 120 | ||
95 | lwz r0,8(r1) | 121 | lwz r0,8(r1) |
96 | mtcrf 0xff,r0 | 122 | mtcrf 0xff,r0 |
@@ -102,7 +128,7 @@ _GLOBAL(plpar_hcall) | |||
102 | mfcr r0 | 128 | mfcr r0 |
103 | stw r0,8(r1) | 129 | stw r0,8(r1) |
104 | 130 | ||
105 | HCALL_INST_PRECALL | 131 | HCALL_INST_PRECALL(r5) |
106 | 132 | ||
107 | std r4,STK_PARM(r4)(r1) /* Save ret buffer */ | 133 | std r4,STK_PARM(r4)(r1) /* Save ret buffer */ |
108 | 134 | ||
@@ -121,7 +147,7 @@ _GLOBAL(plpar_hcall) | |||
121 | std r6, 16(r12) | 147 | std r6, 16(r12) |
122 | std r7, 24(r12) | 148 | std r7, 24(r12) |
123 | 149 | ||
124 | HCALL_INST_POSTCALL | 150 | HCALL_INST_POSTCALL(r12) |
125 | 151 | ||
126 | lwz r0,8(r1) | 152 | lwz r0,8(r1) |
127 | mtcrf 0xff,r0 | 153 | mtcrf 0xff,r0 |
@@ -168,7 +194,7 @@ _GLOBAL(plpar_hcall9) | |||
168 | mfcr r0 | 194 | mfcr r0 |
169 | stw r0,8(r1) | 195 | stw r0,8(r1) |
170 | 196 | ||
171 | HCALL_INST_PRECALL | 197 | HCALL_INST_PRECALL(r5) |
172 | 198 | ||
173 | std r4,STK_PARM(r4)(r1) /* Save ret buffer */ | 199 | std r4,STK_PARM(r4)(r1) /* Save ret buffer */ |
174 | 200 | ||
@@ -196,7 +222,7 @@ _GLOBAL(plpar_hcall9) | |||
196 | std r11,56(r12) | 222 | std r11,56(r12) |
197 | std r0, 64(r12) | 223 | std r0, 64(r12) |
198 | 224 | ||
199 | HCALL_INST_POSTCALL | 225 | HCALL_INST_POSTCALL(r12) |
200 | 226 | ||
201 | lwz r0,8(r1) | 227 | lwz r0,8(r1) |
202 | mtcrf 0xff,r0 | 228 | mtcrf 0xff,r0 |
diff --git a/arch/powerpc/platforms/pseries/hvCall_inst.c b/arch/powerpc/platforms/pseries/hvCall_inst.c index 3631a4f277eb..2f58c71b7259 100644 --- a/arch/powerpc/platforms/pseries/hvCall_inst.c +++ b/arch/powerpc/platforms/pseries/hvCall_inst.c | |||
@@ -26,6 +26,7 @@ | |||
26 | #include <asm/hvcall.h> | 26 | #include <asm/hvcall.h> |
27 | #include <asm/firmware.h> | 27 | #include <asm/firmware.h> |
28 | #include <asm/cputable.h> | 28 | #include <asm/cputable.h> |
29 | #include <asm/trace.h> | ||
29 | 30 | ||
30 | DEFINE_PER_CPU(struct hcall_stats[HCALL_STAT_ARRAY_SIZE], hcall_stats); | 31 | DEFINE_PER_CPU(struct hcall_stats[HCALL_STAT_ARRAY_SIZE], hcall_stats); |
31 | 32 | ||
@@ -100,6 +101,35 @@ static const struct file_operations hcall_inst_seq_fops = { | |||
100 | #define HCALL_ROOT_DIR "hcall_inst" | 101 | #define HCALL_ROOT_DIR "hcall_inst" |
101 | #define CPU_NAME_BUF_SIZE 32 | 102 | #define CPU_NAME_BUF_SIZE 32 |
102 | 103 | ||
104 | |||
105 | static void probe_hcall_entry(unsigned long opcode, unsigned long *args) | ||
106 | { | ||
107 | struct hcall_stats *h; | ||
108 | |||
109 | if (opcode > MAX_HCALL_OPCODE) | ||
110 | return; | ||
111 | |||
112 | h = &get_cpu_var(hcall_stats)[opcode / 4]; | ||
113 | h->tb_start = mftb(); | ||
114 | h->purr_start = mfspr(SPRN_PURR); | ||
115 | } | ||
116 | |||
117 | static void probe_hcall_exit(unsigned long opcode, unsigned long retval, | ||
118 | unsigned long *retbuf) | ||
119 | { | ||
120 | struct hcall_stats *h; | ||
121 | |||
122 | if (opcode > MAX_HCALL_OPCODE) | ||
123 | return; | ||
124 | |||
125 | h = &__get_cpu_var(hcall_stats)[opcode / 4]; | ||
126 | h->num_calls++; | ||
127 | h->tb_total = mftb() - h->tb_start; | ||
128 | h->purr_total = mfspr(SPRN_PURR) - h->purr_start; | ||
129 | |||
130 | put_cpu_var(hcall_stats); | ||
131 | } | ||
132 | |||
103 | static int __init hcall_inst_init(void) | 133 | static int __init hcall_inst_init(void) |
104 | { | 134 | { |
105 | struct dentry *hcall_root; | 135 | struct dentry *hcall_root; |
@@ -110,6 +140,14 @@ static int __init hcall_inst_init(void) | |||
110 | if (!firmware_has_feature(FW_FEATURE_LPAR)) | 140 | if (!firmware_has_feature(FW_FEATURE_LPAR)) |
111 | return 0; | 141 | return 0; |
112 | 142 | ||
143 | if (register_trace_hcall_entry(probe_hcall_entry)) | ||
144 | return -EINVAL; | ||
145 | |||
146 | if (register_trace_hcall_exit(probe_hcall_exit)) { | ||
147 | unregister_trace_hcall_entry(probe_hcall_entry); | ||
148 | return -EINVAL; | ||
149 | } | ||
150 | |||
113 | hcall_root = debugfs_create_dir(HCALL_ROOT_DIR, NULL); | 151 | hcall_root = debugfs_create_dir(HCALL_ROOT_DIR, NULL); |
114 | if (!hcall_root) | 152 | if (!hcall_root) |
115 | return -ENOMEM; | 153 | return -ENOMEM; |
diff --git a/arch/powerpc/platforms/pseries/lpar.c b/arch/powerpc/platforms/pseries/lpar.c index 903eb9eec687..0707653612ba 100644 --- a/arch/powerpc/platforms/pseries/lpar.c +++ b/arch/powerpc/platforms/pseries/lpar.c | |||
@@ -39,6 +39,7 @@ | |||
39 | #include <asm/cputable.h> | 39 | #include <asm/cputable.h> |
40 | #include <asm/udbg.h> | 40 | #include <asm/udbg.h> |
41 | #include <asm/smp.h> | 41 | #include <asm/smp.h> |
42 | #include <asm/trace.h> | ||
42 | 43 | ||
43 | #include "plpar_wrappers.h" | 44 | #include "plpar_wrappers.h" |
44 | #include "pseries.h" | 45 | #include "pseries.h" |
@@ -661,3 +662,35 @@ void arch_free_page(struct page *page, int order) | |||
661 | EXPORT_SYMBOL(arch_free_page); | 662 | EXPORT_SYMBOL(arch_free_page); |
662 | 663 | ||
663 | #endif | 664 | #endif |
665 | |||
666 | #ifdef CONFIG_TRACEPOINTS | ||
667 | /* | ||
668 | * We optimise our hcall path by placing hcall_tracepoint_refcount | ||
669 | * directly in the TOC so we can check if the hcall tracepoints are | ||
670 | * enabled via a single load. | ||
671 | */ | ||
672 | |||
673 | /* NB: reg/unreg are called while guarded with the tracepoints_mutex */ | ||
674 | extern long hcall_tracepoint_refcount; | ||
675 | |||
676 | void hcall_tracepoint_regfunc(void) | ||
677 | { | ||
678 | hcall_tracepoint_refcount++; | ||
679 | } | ||
680 | |||
681 | void hcall_tracepoint_unregfunc(void) | ||
682 | { | ||
683 | hcall_tracepoint_refcount--; | ||
684 | } | ||
685 | |||
686 | void __trace_hcall_entry(unsigned long opcode, unsigned long *args) | ||
687 | { | ||
688 | trace_hcall_entry(opcode, args); | ||
689 | } | ||
690 | |||
691 | void __trace_hcall_exit(long opcode, unsigned long retval, | ||
692 | unsigned long *retbuf) | ||
693 | { | ||
694 | trace_hcall_exit(opcode, retval, retbuf); | ||
695 | } | ||
696 | #endif | ||