xref: /freebsd/sys/dev/hwpmc/hwpmc_x86.c (revision 8d20be1e22095c27faf8fe8b2f0d089739cc742e)
1 /*-
2  * Copyright (c) 2005,2008 Joseph Koshy
3  * Copyright (c) 2007 The FreeBSD Foundation
4  * All rights reserved.
5  *
6  * Portions of this software were developed by A. Joseph Koshy under
7  * sponsorship from the FreeBSD Foundation and Google, Inc.
8  *
9  * Redistribution and use in source and binary forms, with or without
10  * modification, are permitted provided that the following conditions
11  * are met:
12  * 1. Redistributions of source code must retain the above copyright
13  *    notice, this list of conditions and the following disclaimer.
14  * 2. Redistributions in binary form must reproduce the above copyright
15  *    notice, this list of conditions and the following disclaimer in the
16  *    documentation and/or other materials provided with the distribution.
17  *
18  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
19  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
20  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
21  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
22  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
23  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
24  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
25  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
26  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
27  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
28  * SUCH DAMAGE.
29  */
30 
31 #include <sys/cdefs.h>
32 __FBSDID("$FreeBSD$");
33 
34 #include <sys/param.h>
35 #include <sys/bus.h>
36 #include <sys/pmc.h>
37 #include <sys/proc.h>
38 #include <sys/systm.h>
39 
40 #include <machine/cpu.h>
41 #include <machine/cputypes.h>
42 #include <machine/intr_machdep.h>
43 #include <machine/apicvar.h>
44 #include <machine/pmc_mdep.h>
45 #include <machine/md_var.h>
46 
47 #include <vm/vm.h>
48 #include <vm/vm_param.h>
49 #include <vm/pmap.h>
50 
51 #include "hwpmc_soft.h"
52 
53 /*
54  * Attempt to walk a user call stack using a too-simple algorithm.
55  * In the general case we need unwind information associated with
56  * the executable to be able to walk the user stack.
57  *
58  * We are handed a trap frame laid down at the time the PMC interrupt
59  * was taken.  If the application is using frame pointers, the saved
60  * PC value could be:
61  * a. at the beginning of a function before the stack frame is laid
62  *    down,
63  * b. just before a 'ret', after the stack frame has been taken off,
64  * c. somewhere else in the function with a valid stack frame being
65  *    present,
66  *
67  * If the application is not using frame pointers, this algorithm will
68  * fail to yield an interesting call chain.
69  *
70  * TODO: figure out a way to use unwind information.
71  */
72 
73 int
74 pmc_save_user_callchain(uintptr_t *cc, int nframes, struct trapframe *tf)
75 {
76 	int n;
77 	uint32_t instr;
78 	uintptr_t fp, oldfp, pc, r, sp;
79 
80 	KASSERT(TRAPF_USERMODE(tf), ("[x86,%d] Not a user trap frame tf=%p",
81 	    __LINE__, (void *) tf));
82 
83 	pc = PMC_TRAPFRAME_TO_PC(tf);
84 	oldfp = fp = PMC_TRAPFRAME_TO_FP(tf);
85 	sp = PMC_TRAPFRAME_TO_USER_SP(tf);
86 
87 	*cc++ = pc; n = 1;
88 
89 	r = fp + sizeof(uintptr_t); /* points to return address */
90 
91 	if (!PMC_IN_USERSPACE(pc))
92 		return (n);
93 
94 	if (copyin((void *) pc, &instr, sizeof(instr)) != 0)
95 		return (n);
96 
97 	if (PMC_AT_FUNCTION_PROLOGUE_PUSH_BP(instr) ||
98 	    PMC_AT_FUNCTION_EPILOGUE_RET(instr)) { /* ret */
99 		if (copyin((void *) sp, &pc, sizeof(pc)) != 0)
100 			return (n);
101 	} else if (PMC_AT_FUNCTION_PROLOGUE_MOV_SP_BP(instr)) {
102 		sp += sizeof(uintptr_t);
103 		if (copyin((void *) sp, &pc, sizeof(pc)) != 0)
104 			return (n);
105 	} else if (copyin((void *) r, &pc, sizeof(pc)) != 0 ||
106 	    copyin((void *) fp, &fp, sizeof(fp)) != 0)
107 		return (n);
108 
109 	for (; n < nframes;) {
110 		if (pc == 0 || !PMC_IN_USERSPACE(pc))
111 			break;
112 
113 		*cc++ = pc; n++;
114 
115 		if (fp < oldfp)
116 			break;
117 
118 		r = fp + sizeof(uintptr_t); /* address of return address */
119 		oldfp = fp;
120 
121 		if (copyin((void *) r, &pc, sizeof(pc)) != 0 ||
122 		    copyin((void *) fp, &fp, sizeof(fp)) != 0)
123 			break;
124 	}
125 
126 	return (n);
127 }
128 
129 /*
130  * Walking the kernel call stack.
131  *
132  * We are handed the trap frame laid down at the time the PMC
133  * interrupt was taken.  The saved PC could be:
134  * a. in the lowlevel trap handler, meaning that there isn't a C stack
135  *    to traverse,
136  * b. at the beginning of a function before the stack frame is laid
137  *    down,
138  * c. just before a 'ret', after the stack frame has been taken off,
139  * d. somewhere else in a function with a valid stack frame being
140  *    present.
141  *
142  * In case (d), the previous frame pointer is at [%ebp]/[%rbp] and
143  * the return address is at [%ebp+4]/[%rbp+8].
144  *
145  * For cases (b) and (c), the return address is at [%esp]/[%rsp] and
146  * the frame pointer doesn't need to be changed when going up one
147  * level in the stack.
148  *
149  * For case (a), we check if the PC lies in low-level trap handling
150  * code, and if so we terminate our trace.
151  */
152 
153 int
154 pmc_save_kernel_callchain(uintptr_t *cc, int nframes, struct trapframe *tf)
155 {
156 	int n;
157 	uint32_t instr;
158 	uintptr_t fp, pc, r, sp, stackstart, stackend;
159 	struct thread *td;
160 
161 	KASSERT(TRAPF_USERMODE(tf) == 0,("[x86,%d] not a kernel backtrace",
162 	    __LINE__));
163 
164 	td = curthread;
165 	pc = PMC_TRAPFRAME_TO_PC(tf);
166 	fp = PMC_TRAPFRAME_TO_FP(tf);
167 	sp = PMC_TRAPFRAME_TO_KERNEL_SP(tf);
168 
169 	*cc++ = pc;
170 	r = fp + sizeof(uintptr_t); /* points to return address */
171 
172 	if (nframes <= 1)
173 		return (1);
174 
175 	stackstart = (uintptr_t) td->td_kstack;
176 	stackend = (uintptr_t) td->td_kstack + td->td_kstack_pages * PAGE_SIZE;
177 
178 	if (PMC_IN_TRAP_HANDLER(pc) ||
179 	    !PMC_IN_KERNEL(pc) ||
180 	    !PMC_IN_KERNEL_STACK(r, stackstart, stackend) ||
181 	    !PMC_IN_KERNEL_STACK(sp, stackstart, stackend) ||
182 	    !PMC_IN_KERNEL_STACK(fp, stackstart, stackend))
183 		return (1);
184 
185 	instr = *(uint32_t *) pc;
186 
187 	/*
188 	 * Determine whether the interrupted function was in the
189 	 * processing of either laying down its stack frame or taking
190 	 * it off.
191 	 *
192 	 * If we haven't started laying down a stack frame, or are
193 	 * just about to return, then our caller's address is at
194 	 * *sp, and we don't have a frame to unwind.
195 	 */
196 	if (PMC_AT_FUNCTION_PROLOGUE_PUSH_BP(instr) ||
197 	    PMC_AT_FUNCTION_EPILOGUE_RET(instr))
198 		pc = *(uintptr_t *) sp;
199 	else if (PMC_AT_FUNCTION_PROLOGUE_MOV_SP_BP(instr)) {
200 		/*
201 		 * The code was midway through laying down a frame.
202 		 * At this point sp[0] has a frame back pointer,
203 		 * and the caller's address is therefore at sp[1].
204 		 */
205 		sp += sizeof(uintptr_t);
206 		if (!PMC_IN_KERNEL_STACK(sp, stackstart, stackend))
207 			return (1);
208 		pc = *(uintptr_t *) sp;
209 	} else {
210 		/*
211 		 * Not in the function prologue or epilogue.
212 		 */
213 		pc = *(uintptr_t *) r;
214 		fp = *(uintptr_t *) fp;
215 	}
216 
217 	for (n = 1; n < nframes; n++) {
218 		*cc++ = pc;
219 
220 		if (PMC_IN_TRAP_HANDLER(pc))
221 			break;
222 
223 		r = fp + sizeof(uintptr_t);
224 		if (!PMC_IN_KERNEL_STACK(fp, stackstart, stackend) ||
225 		    !PMC_IN_KERNEL_STACK(r, stackstart, stackend))
226 			break;
227 		pc = *(uintptr_t *) r;
228 		fp = *(uintptr_t *) fp;
229 	}
230 
231 	return (n);
232 }
233 
234 /*
235  * Machine dependent initialization for x86 class platforms.
236  */
237 
238 struct pmc_mdep *
239 pmc_md_initialize()
240 {
241 	int i;
242 	struct pmc_mdep *md;
243 
244 	/* determine the CPU kind */
245 	if (cpu_vendor_id == CPU_VENDOR_AMD)
246 		md = pmc_amd_initialize();
247 	else if (cpu_vendor_id == CPU_VENDOR_INTEL)
248 		md = pmc_intel_initialize();
249 	else
250 		return (NULL);
251 
252 	/* disallow sampling if we do not have an LAPIC */
253 	if (md != NULL && !lapic_enable_pmc())
254 		for (i = 0; i < md->pmd_nclass; i++) {
255 			if (i == PMC_CLASS_INDEX_SOFT)
256 				continue;
257 			md->pmd_classdep[i].pcd_caps &= ~PMC_CAP_INTERRUPT;
258 		}
259 
260 	return (md);
261 }
262 
263 void
264 pmc_md_finalize(struct pmc_mdep *md)
265 {
266 
267 	lapic_disable_pmc();
268 	if (cpu_vendor_id == CPU_VENDOR_AMD)
269 		pmc_amd_finalize(md);
270 	else if (cpu_vendor_id == CPU_VENDOR_INTEL)
271 		pmc_intel_finalize(md);
272 	else
273 		KASSERT(0, ("[x86,%d] Unknown vendor", __LINE__));
274 }
275