xref: /freebsd/sys/dev/hwpmc/hwpmc_x86.c (revision b3aaa0cc21c63d388230c7ef2a80abd631ff20d5)
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/apicreg.h>
43 #include <machine/pmc_mdep.h>
44 #include <machine/md_var.h>
45 
46 #include <vm/vm.h>
47 #include <vm/vm_param.h>
48 #include <vm/pmap.h>
49 
50 extern volatile lapic_t *lapic;
51 
52 void
53 pmc_x86_lapic_enable_pmc_interrupt(void)
54 {
55 	uint32_t value;
56 
57 	value =  lapic->lvt_pcint;
58 	value &= ~APIC_LVT_M;
59 	lapic->lvt_pcint = value;
60 }
61 
62 /*
63  * Attempt to walk a user call stack using a too-simple algorithm.
64  * In the general case we need unwind information associated with
65  * the executable to be able to walk the user stack.
66  *
67  * We are handed a trap frame laid down at the time the PMC interrupt
68  * was taken.  If the application is using frame pointers, the saved
69  * PC value could be:
70  * a. at the beginning of a function before the stack frame is laid
71  *    down,
72  * b. just before a 'ret', after the stack frame has been taken off,
73  * c. somewhere else in the function with a valid stack frame being
74  *    present,
75  *
76  * If the application is not using frame pointers, this algorithm will
77  * fail to yield an interesting call chain.
78  *
79  * TODO: figure out a way to use unwind information.
80  */
81 
82 int
83 pmc_save_user_callchain(uintptr_t *cc, int nframes, struct trapframe *tf)
84 {
85 	int n;
86 	uint32_t instr;
87 	uintptr_t fp, oldfp, pc, r, sp;
88 
89 	KASSERT(TRAPF_USERMODE(tf), ("[x86,%d] Not a user trap frame tf=%p",
90 	    __LINE__, (void *) tf));
91 
92 	pc = PMC_TRAPFRAME_TO_PC(tf);
93 	oldfp = fp = PMC_TRAPFRAME_TO_FP(tf);
94 	sp = PMC_TRAPFRAME_TO_USER_SP(tf);
95 
96 	*cc++ = pc; n = 1;
97 
98 	r = fp + sizeof(uintptr_t); /* points to return address */
99 
100 	if (!PMC_IN_USERSPACE(pc))
101 		return (n);
102 
103 	if (copyin((void *) pc, &instr, sizeof(instr)) != 0)
104 		return (n);
105 
106 	if (PMC_AT_FUNCTION_PROLOGUE_PUSH_BP(instr) ||
107 	    PMC_AT_FUNCTION_EPILOGUE_RET(instr)) { /* ret */
108 		if (copyin((void *) sp, &pc, sizeof(pc)) != 0)
109 			return (n);
110 	} else if (PMC_AT_FUNCTION_PROLOGUE_MOV_SP_BP(instr)) {
111 		sp += sizeof(uintptr_t);
112 		if (copyin((void *) sp, &pc, sizeof(pc)) != 0)
113 			return (n);
114 	} else if (copyin((void *) r, &pc, sizeof(pc)) != 0 ||
115 	    copyin((void *) fp, &fp, sizeof(fp) != 0))
116 		return (n);
117 
118 	for (; n < nframes;) {
119 		if (pc == 0 || !PMC_IN_USERSPACE(pc))
120 			break;
121 
122 		*cc++ = pc; n++;
123 
124 		if (fp < oldfp)
125 			break;
126 
127 		r = fp + sizeof(uintptr_t); /* address of return address */
128 		oldfp = fp;
129 
130 		if (copyin((void *) r, &pc, sizeof(pc)) != 0 ||
131 		    copyin((void *) fp, &fp, sizeof(fp)) != 0)
132 			break;
133 	}
134 
135 	return (n);
136 }
137 
138 /*
139  * Walking the kernel call stack.
140  *
141  * We are handed the trap frame laid down at the time the PMC
142  * interrupt was taken.  The saved PC could be:
143  * a. in the lowlevel trap handler, meaning that there isn't a C stack
144  *    to traverse,
145  * b. at the beginning of a function before the stack frame is laid
146  *    down,
147  * c. just before a 'ret', after the stack frame has been taken off,
148  * d. somewhere else in a function with a valid stack frame being
149  *    present.
150  *
151  * In case (d), the previous frame pointer is at [%ebp]/[%rbp] and
152  * the return address is at [%ebp+4]/[%rbp+8].
153  *
154  * For cases (b) and (c), the return address is at [%esp]/[%rsp] and
155  * the frame pointer doesn't need to be changed when going up one
156  * level in the stack.
157  *
158  * For case (a), we check if the PC lies in low-level trap handling
159  * code, and if so we terminate our trace.
160  */
161 
162 int
163 pmc_save_kernel_callchain(uintptr_t *cc, int nframes, struct trapframe *tf)
164 {
165 	int n;
166 	uint32_t instr;
167 	uintptr_t fp, pc, r, sp, stackstart, stackend;
168 	struct thread *td;
169 
170 	KASSERT(TRAPF_USERMODE(tf) == 0,("[x86,%d] not a kernel backtrace",
171 	    __LINE__));
172 
173 	pc = PMC_TRAPFRAME_TO_PC(tf);
174 	fp = PMC_TRAPFRAME_TO_FP(tf);
175 	sp = PMC_TRAPFRAME_TO_KERNEL_SP(tf);
176 
177 	*cc++ = pc;
178 	r = fp + sizeof(uintptr_t); /* points to return address */
179 
180 	if ((td = curthread) == NULL)
181 		return (1);
182 
183 	if (nframes <= 1)
184 		return (1);
185 
186 	stackstart = (uintptr_t) td->td_kstack;
187 	stackend = (uintptr_t) td->td_kstack + td->td_kstack_pages * PAGE_SIZE;
188 
189 	if (PMC_IN_TRAP_HANDLER(pc) ||
190 	    !PMC_IN_KERNEL(pc) || !PMC_IN_KERNEL(r) ||
191 	    !PMC_IN_KERNEL_STACK(sp, stackstart, stackend) ||
192 	    !PMC_IN_KERNEL_STACK(fp, stackstart, stackend))
193 		return (1);
194 
195 	instr = *(uint32_t *) pc;
196 
197 	/*
198 	 * Determine whether the interrupted function was in the
199 	 * processing of either laying down its stack frame or taking
200 	 * it off.
201 	 *
202 	 * If we haven't started laying down a stack frame, or are
203 	 * just about to return, then our caller's address is at
204 	 * *sp, and we don't have a frame to unwind.
205 	 */
206 	if (PMC_AT_FUNCTION_PROLOGUE_PUSH_BP(instr) ||
207 	    PMC_AT_FUNCTION_EPILOGUE_RET(instr))
208 		pc = *(uintptr_t *) sp;
209 	else if (PMC_AT_FUNCTION_PROLOGUE_MOV_SP_BP(instr)) {
210 		/*
211 		 * The code was midway through laying down a frame.
212 		 * At this point sp[0] has a frame back pointer,
213 		 * and the caller's address is therefore at sp[1].
214 		 */
215 		sp += sizeof(uintptr_t);
216 		if (!PMC_IN_KERNEL_STACK(sp, stackstart, stackend))
217 			return (1);
218 		pc = *(uintptr_t *) sp;
219 	} else {
220 		/*
221 		 * Not in the function prologue or epilogue.
222 		 */
223 		pc = *(uintptr_t *) r;
224 		fp = *(uintptr_t *) fp;
225 	}
226 
227 	for (n = 1; n < nframes; n++) {
228 		*cc++ = pc;
229 
230 		if (PMC_IN_TRAP_HANDLER(pc))
231 			break;
232 
233 		r = fp + sizeof(uintptr_t);
234 		if (!PMC_IN_KERNEL_STACK(fp, stackstart, stackend) ||
235 		    !PMC_IN_KERNEL(r))
236 			break;
237 		pc = *(uintptr_t *) r;
238 		fp = *(uintptr_t *) fp;
239 	}
240 
241 	return (n);
242 }
243 
244 /*
245  * Machine dependent initialization for x86 class platforms.
246  */
247 
248 struct pmc_mdep *
249 pmc_md_initialize()
250 {
251 	int i;
252 	struct pmc_mdep *md;
253 
254 	/* determine the CPU kind */
255 	md = NULL;
256 	if (cpu_vendor_id == CPU_VENDOR_AMD)
257 		md = pmc_amd_initialize();
258 	else if (cpu_vendor_id == CPU_VENDOR_INTEL)
259 		md = pmc_intel_initialize();
260 	else
261 		KASSERT(0, ("[x86,%d] Unknown vendor", __LINE__));
262 
263 	/* disallow sampling if we do not have an LAPIC */
264 	if (md != NULL && lapic == NULL)
265 		for (i = 1; i < md->pmd_nclass; i++)
266 			md->pmd_classdep[i].pcd_caps &= ~PMC_CAP_INTERRUPT;
267 
268 	return (md);
269 }
270 
271 void
272 pmc_md_finalize(struct pmc_mdep *md)
273 {
274 	if (cpu_vendor_id == CPU_VENDOR_AMD)
275 		pmc_amd_finalize(md);
276 	else if (cpu_vendor_id == CPU_VENDOR_INTEL)
277 		pmc_intel_finalize(md);
278 	else
279 		KASSERT(0, ("[x86,%d] Unknown vendor", __LINE__));
280 }
281