xref: /openbmc/linux/tools/perf/arch/x86/util/perf_regs.c (revision 17e5530a)
1 // SPDX-License-Identifier: GPL-2.0
2 #include <errno.h>
3 #include <string.h>
4 #include <regex.h>
5 #include <linux/kernel.h>
6 #include <linux/zalloc.h>
7 
8 #include "perf_regs.h"
9 #include "../../../perf-sys.h"
10 #include "../../../util/perf_regs.h"
11 #include "../../../util/debug.h"
12 #include "../../../util/event.h"
13 #include "../../../util/pmu.h"
14 #include "../../../util/pmus.h"
15 
16 const struct sample_reg sample_reg_masks[] = {
17 	SMPL_REG(AX, PERF_REG_X86_AX),
18 	SMPL_REG(BX, PERF_REG_X86_BX),
19 	SMPL_REG(CX, PERF_REG_X86_CX),
20 	SMPL_REG(DX, PERF_REG_X86_DX),
21 	SMPL_REG(SI, PERF_REG_X86_SI),
22 	SMPL_REG(DI, PERF_REG_X86_DI),
23 	SMPL_REG(BP, PERF_REG_X86_BP),
24 	SMPL_REG(SP, PERF_REG_X86_SP),
25 	SMPL_REG(IP, PERF_REG_X86_IP),
26 	SMPL_REG(FLAGS, PERF_REG_X86_FLAGS),
27 	SMPL_REG(CS, PERF_REG_X86_CS),
28 	SMPL_REG(SS, PERF_REG_X86_SS),
29 #ifdef HAVE_ARCH_X86_64_SUPPORT
30 	SMPL_REG(R8, PERF_REG_X86_R8),
31 	SMPL_REG(R9, PERF_REG_X86_R9),
32 	SMPL_REG(R10, PERF_REG_X86_R10),
33 	SMPL_REG(R11, PERF_REG_X86_R11),
34 	SMPL_REG(R12, PERF_REG_X86_R12),
35 	SMPL_REG(R13, PERF_REG_X86_R13),
36 	SMPL_REG(R14, PERF_REG_X86_R14),
37 	SMPL_REG(R15, PERF_REG_X86_R15),
38 #endif
39 	SMPL_REG2(XMM0, PERF_REG_X86_XMM0),
40 	SMPL_REG2(XMM1, PERF_REG_X86_XMM1),
41 	SMPL_REG2(XMM2, PERF_REG_X86_XMM2),
42 	SMPL_REG2(XMM3, PERF_REG_X86_XMM3),
43 	SMPL_REG2(XMM4, PERF_REG_X86_XMM4),
44 	SMPL_REG2(XMM5, PERF_REG_X86_XMM5),
45 	SMPL_REG2(XMM6, PERF_REG_X86_XMM6),
46 	SMPL_REG2(XMM7, PERF_REG_X86_XMM7),
47 	SMPL_REG2(XMM8, PERF_REG_X86_XMM8),
48 	SMPL_REG2(XMM9, PERF_REG_X86_XMM9),
49 	SMPL_REG2(XMM10, PERF_REG_X86_XMM10),
50 	SMPL_REG2(XMM11, PERF_REG_X86_XMM11),
51 	SMPL_REG2(XMM12, PERF_REG_X86_XMM12),
52 	SMPL_REG2(XMM13, PERF_REG_X86_XMM13),
53 	SMPL_REG2(XMM14, PERF_REG_X86_XMM14),
54 	SMPL_REG2(XMM15, PERF_REG_X86_XMM15),
55 	SMPL_REG_END
56 };
57 
58 struct sdt_name_reg {
59 	const char *sdt_name;
60 	const char *uprobe_name;
61 };
62 #define SDT_NAME_REG(n, m) {.sdt_name = "%" #n, .uprobe_name = "%" #m}
63 #define SDT_NAME_REG_END {.sdt_name = NULL, .uprobe_name = NULL}
64 
65 static const struct sdt_name_reg sdt_reg_tbl[] = {
66 	SDT_NAME_REG(eax, ax),
67 	SDT_NAME_REG(rax, ax),
68 	SDT_NAME_REG(al,  ax),
69 	SDT_NAME_REG(ah,  ax),
70 	SDT_NAME_REG(ebx, bx),
71 	SDT_NAME_REG(rbx, bx),
72 	SDT_NAME_REG(bl,  bx),
73 	SDT_NAME_REG(bh,  bx),
74 	SDT_NAME_REG(ecx, cx),
75 	SDT_NAME_REG(rcx, cx),
76 	SDT_NAME_REG(cl,  cx),
77 	SDT_NAME_REG(ch,  cx),
78 	SDT_NAME_REG(edx, dx),
79 	SDT_NAME_REG(rdx, dx),
80 	SDT_NAME_REG(dl,  dx),
81 	SDT_NAME_REG(dh,  dx),
82 	SDT_NAME_REG(esi, si),
83 	SDT_NAME_REG(rsi, si),
84 	SDT_NAME_REG(sil, si),
85 	SDT_NAME_REG(edi, di),
86 	SDT_NAME_REG(rdi, di),
87 	SDT_NAME_REG(dil, di),
88 	SDT_NAME_REG(ebp, bp),
89 	SDT_NAME_REG(rbp, bp),
90 	SDT_NAME_REG(bpl, bp),
91 	SDT_NAME_REG(rsp, sp),
92 	SDT_NAME_REG(esp, sp),
93 	SDT_NAME_REG(spl, sp),
94 
95 	/* rNN registers */
96 	SDT_NAME_REG(r8b,  r8),
97 	SDT_NAME_REG(r8w,  r8),
98 	SDT_NAME_REG(r8d,  r8),
99 	SDT_NAME_REG(r9b,  r9),
100 	SDT_NAME_REG(r9w,  r9),
101 	SDT_NAME_REG(r9d,  r9),
102 	SDT_NAME_REG(r10b, r10),
103 	SDT_NAME_REG(r10w, r10),
104 	SDT_NAME_REG(r10d, r10),
105 	SDT_NAME_REG(r11b, r11),
106 	SDT_NAME_REG(r11w, r11),
107 	SDT_NAME_REG(r11d, r11),
108 	SDT_NAME_REG(r12b, r12),
109 	SDT_NAME_REG(r12w, r12),
110 	SDT_NAME_REG(r12d, r12),
111 	SDT_NAME_REG(r13b, r13),
112 	SDT_NAME_REG(r13w, r13),
113 	SDT_NAME_REG(r13d, r13),
114 	SDT_NAME_REG(r14b, r14),
115 	SDT_NAME_REG(r14w, r14),
116 	SDT_NAME_REG(r14d, r14),
117 	SDT_NAME_REG(r15b, r15),
118 	SDT_NAME_REG(r15w, r15),
119 	SDT_NAME_REG(r15d, r15),
120 	SDT_NAME_REG_END,
121 };
122 
123 /*
124  * Perf only supports OP which is in  +/-NUM(REG)  form.
125  * Here plus-minus sign, NUM and parenthesis are optional,
126  * only REG is mandatory.
127  *
128  * SDT events also supports indirect addressing mode with a
129  * symbol as offset, scaled mode and constants in OP. But
130  * perf does not support them yet. Below are few examples.
131  *
132  * OP with scaled mode:
133  *     (%rax,%rsi,8)
134  *     10(%ras,%rsi,8)
135  *
136  * OP with indirect addressing mode:
137  *     check_action(%rip)
138  *     mp_+52(%rip)
139  *     44+mp_(%rip)
140  *
141  * OP with constant values:
142  *     $0
143  *     $123
144  *     $-1
145  */
146 #define SDT_OP_REGEX  "^([+\\-]?)([0-9]*)(\\(?)(%[a-z][a-z0-9]+)(\\)?)$"
147 
148 static regex_t sdt_op_regex;
149 
150 static int sdt_init_op_regex(void)
151 {
152 	static int initialized;
153 	int ret = 0;
154 
155 	if (initialized)
156 		return 0;
157 
158 	ret = regcomp(&sdt_op_regex, SDT_OP_REGEX, REG_EXTENDED);
159 	if (ret < 0) {
160 		pr_debug4("Regex compilation error.\n");
161 		return ret;
162 	}
163 
164 	initialized = 1;
165 	return 0;
166 }
167 
168 /*
169  * Max x86 register name length is 5(ex: %r15d). So, 6th char
170  * should always contain NULL. This helps to find register name
171  * length using strlen, instead of maintaining one more variable.
172  */
173 #define SDT_REG_NAME_SIZE  6
174 
175 /*
176  * The uprobe parser does not support all gas register names;
177  * so, we have to replace them (ex. for x86_64: %rax -> %ax).
178  * Note: If register does not require renaming, just copy
179  * paste as it is, but don't leave it empty.
180  */
181 static void sdt_rename_register(char *sdt_reg, int sdt_len, char *uprobe_reg)
182 {
183 	int i = 0;
184 
185 	for (i = 0; sdt_reg_tbl[i].sdt_name != NULL; i++) {
186 		if (!strncmp(sdt_reg_tbl[i].sdt_name, sdt_reg, sdt_len)) {
187 			strcpy(uprobe_reg, sdt_reg_tbl[i].uprobe_name);
188 			return;
189 		}
190 	}
191 
192 	strncpy(uprobe_reg, sdt_reg, sdt_len);
193 }
194 
195 int arch_sdt_arg_parse_op(char *old_op, char **new_op)
196 {
197 	char new_reg[SDT_REG_NAME_SIZE] = {0};
198 	int new_len = 0, ret;
199 	/*
200 	 * rm[0]:  +/-NUM(REG)
201 	 * rm[1]:  +/-
202 	 * rm[2]:  NUM
203 	 * rm[3]:  (
204 	 * rm[4]:  REG
205 	 * rm[5]:  )
206 	 */
207 	regmatch_t rm[6];
208 	/*
209 	 * Max prefix length is 2 as it may contains sign(+/-)
210 	 * and displacement 0 (Both sign and displacement 0 are
211 	 * optional so it may be empty). Use one more character
212 	 * to hold last NULL so that strlen can be used to find
213 	 * prefix length, instead of maintaining one more variable.
214 	 */
215 	char prefix[3] = {0};
216 
217 	ret = sdt_init_op_regex();
218 	if (ret < 0)
219 		return ret;
220 
221 	/*
222 	 * If unsupported OR does not match with regex OR
223 	 * register name too long, skip it.
224 	 */
225 	if (strchr(old_op, ',') || strchr(old_op, '$') ||
226 	    regexec(&sdt_op_regex, old_op, 6, rm, 0)   ||
227 	    rm[4].rm_eo - rm[4].rm_so > SDT_REG_NAME_SIZE) {
228 		pr_debug4("Skipping unsupported SDT argument: %s\n", old_op);
229 		return SDT_ARG_SKIP;
230 	}
231 
232 	/*
233 	 * Prepare prefix.
234 	 * If SDT OP has parenthesis but does not provide
235 	 * displacement, add 0 for displacement.
236 	 *     SDT         Uprobe     Prefix
237 	 *     -----------------------------
238 	 *     +24(%rdi)   +24(%di)   +
239 	 *     24(%rdi)    +24(%di)   +
240 	 *     %rdi        %di
241 	 *     (%rdi)      +0(%di)    +0
242 	 *     -80(%rbx)   -80(%bx)   -
243 	 */
244 	if (rm[3].rm_so != rm[3].rm_eo) {
245 		if (rm[1].rm_so != rm[1].rm_eo)
246 			prefix[0] = *(old_op + rm[1].rm_so);
247 		else if (rm[2].rm_so != rm[2].rm_eo)
248 			prefix[0] = '+';
249 		else
250 			scnprintf(prefix, sizeof(prefix), "+0");
251 	}
252 
253 	/* Rename register */
254 	sdt_rename_register(old_op + rm[4].rm_so, rm[4].rm_eo - rm[4].rm_so,
255 			    new_reg);
256 
257 	/* Prepare final OP which should be valid for uprobe_events */
258 	new_len = strlen(prefix)              +
259 		  (rm[2].rm_eo - rm[2].rm_so) +
260 		  (rm[3].rm_eo - rm[3].rm_so) +
261 		  strlen(new_reg)             +
262 		  (rm[5].rm_eo - rm[5].rm_so) +
263 		  1;					/* NULL */
264 
265 	*new_op = zalloc(new_len);
266 	if (!*new_op)
267 		return -ENOMEM;
268 
269 	scnprintf(*new_op, new_len, "%.*s%.*s%.*s%.*s%.*s",
270 		  strlen(prefix), prefix,
271 		  (int)(rm[2].rm_eo - rm[2].rm_so), old_op + rm[2].rm_so,
272 		  (int)(rm[3].rm_eo - rm[3].rm_so), old_op + rm[3].rm_so,
273 		  strlen(new_reg), new_reg,
274 		  (int)(rm[5].rm_eo - rm[5].rm_so), old_op + rm[5].rm_so);
275 
276 	return SDT_ARG_VALID;
277 }
278 
279 uint64_t arch__intr_reg_mask(void)
280 {
281 	struct perf_event_attr attr = {
282 		.type			= PERF_TYPE_HARDWARE,
283 		.config			= PERF_COUNT_HW_CPU_CYCLES,
284 		.sample_type		= PERF_SAMPLE_REGS_INTR,
285 		.sample_regs_intr	= PERF_REG_EXTENDED_MASK,
286 		.precise_ip		= 1,
287 		.disabled 		= 1,
288 		.exclude_kernel		= 1,
289 	};
290 	int fd;
291 	/*
292 	 * In an unnamed union, init it here to build on older gcc versions
293 	 */
294 	attr.sample_period = 1;
295 
296 	if (perf_pmus__num_core_pmus() > 1) {
297 		struct perf_pmu *pmu = NULL;
298 		__u64 type = PERF_TYPE_RAW;
299 
300 		/*
301 		 * The same register set is supported among different hybrid PMUs.
302 		 * Only check the first available one.
303 		 */
304 		while ((pmu = perf_pmus__scan_core(pmu)) != NULL) {
305 			type = pmu->type;
306 			break;
307 		}
308 		attr.config |= type << PERF_PMU_TYPE_SHIFT;
309 	}
310 
311 	event_attr_init(&attr);
312 
313 	fd = sys_perf_event_open(&attr, 0, -1, -1, 0);
314 	if (fd != -1) {
315 		close(fd);
316 		return (PERF_REG_EXTENDED_MASK | PERF_REGS_MASK);
317 	}
318 
319 	return PERF_REGS_MASK;
320 }
321 
322 uint64_t arch__user_reg_mask(void)
323 {
324 	return PERF_REGS_MASK;
325 }
326