xref: /freebsd/sys/dev/hwpmc/hwpmc_x86.c (revision 2e5b60079b7d8c3ca68f1390cd90f305e651f8d3)
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 #if (__FreeBSD_version >= 1100000)
44 #include <x86/apicvar.h>
45 #else
46 #include <machine/apicvar.h>
47 #endif
48 #include <machine/pmc_mdep.h>
49 #include <machine/md_var.h>
50 
51 #include <vm/vm.h>
52 #include <vm/vm_param.h>
53 #include <vm/pmap.h>
54 
55 #include "hwpmc_soft.h"
56 
57 /*
58  * Attempt to walk a user call stack using a too-simple algorithm.
59  * In the general case we need unwind information associated with
60  * the executable to be able to walk the user stack.
61  *
62  * We are handed a trap frame laid down at the time the PMC interrupt
63  * was taken.  If the application is using frame pointers, the saved
64  * PC value could be:
65  * a. at the beginning of a function before the stack frame is laid
66  *    down,
67  * b. just before a 'ret', after the stack frame has been taken off,
68  * c. somewhere else in the function with a valid stack frame being
69  *    present,
70  *
71  * If the application is not using frame pointers, this algorithm will
72  * fail to yield an interesting call chain.
73  *
74  * TODO: figure out a way to use unwind information.
75  */
76 
77 int
78 pmc_save_user_callchain(uintptr_t *cc, int nframes, struct trapframe *tf)
79 {
80 	int n;
81 	uint32_t instr;
82 	uintptr_t fp, oldfp, pc, r, sp;
83 
84 	KASSERT(TRAPF_USERMODE(tf), ("[x86,%d] Not a user trap frame tf=%p",
85 	    __LINE__, (void *) tf));
86 
87 	pc = PMC_TRAPFRAME_TO_PC(tf);
88 	oldfp = fp = PMC_TRAPFRAME_TO_FP(tf);
89 	sp = PMC_TRAPFRAME_TO_USER_SP(tf);
90 
91 	*cc++ = pc; n = 1;
92 
93 	r = fp + sizeof(uintptr_t); /* points to return address */
94 
95 	if (!PMC_IN_USERSPACE(pc))
96 		return (n);
97 
98 	if (copyin((void *) pc, &instr, sizeof(instr)) != 0)
99 		return (n);
100 
101 	if (PMC_AT_FUNCTION_PROLOGUE_PUSH_BP(instr) ||
102 	    PMC_AT_FUNCTION_EPILOGUE_RET(instr)) { /* ret */
103 		if (copyin((void *) sp, &pc, sizeof(pc)) != 0)
104 			return (n);
105 	} else if (PMC_AT_FUNCTION_PROLOGUE_MOV_SP_BP(instr)) {
106 		sp += sizeof(uintptr_t);
107 		if (copyin((void *) sp, &pc, sizeof(pc)) != 0)
108 			return (n);
109 	} else if (copyin((void *) r, &pc, sizeof(pc)) != 0 ||
110 	    copyin((void *) fp, &fp, sizeof(fp)) != 0)
111 		return (n);
112 
113 	for (; n < nframes;) {
114 		if (pc == 0 || !PMC_IN_USERSPACE(pc))
115 			break;
116 
117 		*cc++ = pc; n++;
118 
119 		if (fp < oldfp)
120 			break;
121 
122 		r = fp + sizeof(uintptr_t); /* address of return address */
123 		oldfp = fp;
124 
125 		if (copyin((void *) r, &pc, sizeof(pc)) != 0 ||
126 		    copyin((void *) fp, &fp, sizeof(fp)) != 0)
127 			break;
128 	}
129 
130 	return (n);
131 }
132 
133 /*
134  * Walking the kernel call stack.
135  *
136  * We are handed the trap frame laid down at the time the PMC
137  * interrupt was taken.  The saved PC could be:
138  * a. in the lowlevel trap handler, meaning that there isn't a C stack
139  *    to traverse,
140  * b. at the beginning of a function before the stack frame is laid
141  *    down,
142  * c. just before a 'ret', after the stack frame has been taken off,
143  * d. somewhere else in a function with a valid stack frame being
144  *    present.
145  *
146  * In case (d), the previous frame pointer is at [%ebp]/[%rbp] and
147  * the return address is at [%ebp+4]/[%rbp+8].
148  *
149  * For cases (b) and (c), the return address is at [%esp]/[%rsp] and
150  * the frame pointer doesn't need to be changed when going up one
151  * level in the stack.
152  *
153  * For case (a), we check if the PC lies in low-level trap handling
154  * code, and if so we terminate our trace.
155  */
156 
157 int
158 pmc_save_kernel_callchain(uintptr_t *cc, int nframes, struct trapframe *tf)
159 {
160 	int n;
161 	uint32_t instr;
162 	uintptr_t fp, pc, r, sp, stackstart, stackend;
163 	struct thread *td;
164 
165 	KASSERT(TRAPF_USERMODE(tf) == 0,("[x86,%d] not a kernel backtrace",
166 	    __LINE__));
167 
168 	td = curthread;
169 	pc = PMC_TRAPFRAME_TO_PC(tf);
170 	fp = PMC_TRAPFRAME_TO_FP(tf);
171 	sp = PMC_TRAPFRAME_TO_KERNEL_SP(tf);
172 
173 	*cc++ = pc;
174 	r = fp + sizeof(uintptr_t); /* points to return address */
175 
176 	if (nframes <= 1)
177 		return (1);
178 
179 	stackstart = (uintptr_t) td->td_kstack;
180 	stackend = (uintptr_t) td->td_kstack + td->td_kstack_pages * PAGE_SIZE;
181 
182 	if (PMC_IN_TRAP_HANDLER(pc) ||
183 	    !PMC_IN_KERNEL(pc) ||
184 	    !PMC_IN_KERNEL_STACK(r, stackstart, stackend) ||
185 	    !PMC_IN_KERNEL_STACK(sp, stackstart, stackend) ||
186 	    !PMC_IN_KERNEL_STACK(fp, stackstart, stackend))
187 		return (1);
188 
189 	instr = *(uint32_t *) pc;
190 
191 	/*
192 	 * Determine whether the interrupted function was in the
193 	 * processing of either laying down its stack frame or taking
194 	 * it off.
195 	 *
196 	 * If we haven't started laying down a stack frame, or are
197 	 * just about to return, then our caller's address is at
198 	 * *sp, and we don't have a frame to unwind.
199 	 */
200 	if (PMC_AT_FUNCTION_PROLOGUE_PUSH_BP(instr) ||
201 	    PMC_AT_FUNCTION_EPILOGUE_RET(instr))
202 		pc = *(uintptr_t *) sp;
203 	else if (PMC_AT_FUNCTION_PROLOGUE_MOV_SP_BP(instr)) {
204 		/*
205 		 * The code was midway through laying down a frame.
206 		 * At this point sp[0] has a frame back pointer,
207 		 * and the caller's address is therefore at sp[1].
208 		 */
209 		sp += sizeof(uintptr_t);
210 		if (!PMC_IN_KERNEL_STACK(sp, stackstart, stackend))
211 			return (1);
212 		pc = *(uintptr_t *) sp;
213 	} else {
214 		/*
215 		 * Not in the function prologue or epilogue.
216 		 */
217 		pc = *(uintptr_t *) r;
218 		fp = *(uintptr_t *) fp;
219 	}
220 
221 	for (n = 1; n < nframes; n++) {
222 		*cc++ = pc;
223 
224 		if (PMC_IN_TRAP_HANDLER(pc))
225 			break;
226 
227 		r = fp + sizeof(uintptr_t);
228 		if (!PMC_IN_KERNEL_STACK(fp, stackstart, stackend) ||
229 		    !PMC_IN_KERNEL_STACK(r, stackstart, stackend))
230 			break;
231 		pc = *(uintptr_t *) r;
232 		fp = *(uintptr_t *) fp;
233 	}
234 
235 	return (n);
236 }
237 
238 /*
239  * Machine dependent initialization for x86 class platforms.
240  */
241 
242 struct pmc_mdep *
243 pmc_md_initialize()
244 {
245 	int i;
246 	struct pmc_mdep *md;
247 
248 	/* determine the CPU kind */
249 	if (cpu_vendor_id == CPU_VENDOR_AMD)
250 		md = pmc_amd_initialize();
251 	else if (cpu_vendor_id == CPU_VENDOR_INTEL)
252 		md = pmc_intel_initialize();
253 	else
254 		return (NULL);
255 
256 	/* disallow sampling if we do not have an LAPIC */
257 	if (md != NULL && !lapic_enable_pmc())
258 		for (i = 0; i < md->pmd_nclass; i++) {
259 			if (i == PMC_CLASS_INDEX_SOFT)
260 				continue;
261 			md->pmd_classdep[i].pcd_caps &= ~PMC_CAP_INTERRUPT;
262 		}
263 
264 	return (md);
265 }
266 
267 void
268 pmc_md_finalize(struct pmc_mdep *md)
269 {
270 
271 	lapic_disable_pmc();
272 	if (cpu_vendor_id == CPU_VENDOR_AMD)
273 		pmc_amd_finalize(md);
274 	else if (cpu_vendor_id == CPU_VENDOR_INTEL)
275 		pmc_intel_finalize(md);
276 	else
277 		KASSERT(0, ("[x86,%d] Unknown vendor", __LINE__));
278 }
279