Print this page
9736 kmdb tortures via single-step miscellaneous trap
Reviewed by: Robert Mustacchi <rm@joyent.com>
Reviewed by: Jerry Jelinek <jerry.jelinek@joyent.com>
| Split |
Close |
| Expand all |
| Collapse all |
--- old/usr/src/uts/intel/kdi/kdi_idt.c
+++ new/usr/src/uts/intel/kdi/kdi_idt.c
1 1 /*
2 2 * CDDL HEADER START
3 3 *
4 4 * The contents of this file are subject to the terms of the
5 5 * Common Development and Distribution License (the "License").
6 6 * You may not use this file except in compliance with the License.
7 7 *
8 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 9 * or http://www.opensolaris.org/os/licensing.
10 10 * See the License for the specific language governing permissions
11 11 * and limitations under the License.
12 12 *
13 13 * When distributing Covered Code, include this CDDL HEADER in each
14 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 15 * If applicable, add the following below this CDDL HEADER, with the
16 16 * fields enclosed by brackets "[]" replaced with your own identifying
17 17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 18 *
19 19 * CDDL HEADER END
20 20 */
21 21 /*
22 22 * Copyright 2009 Sun Microsystems, Inc. All rights reserved.
23 23 * Use is subject to license terms.
24 24 *
25 25 * Copyright 2018 Joyent, Inc.
26 26 */
27 27
28 28 /*
29 29 * Management of KMDB's IDT, which is installed upon KMDB activation.
30 30 *
31 31 * Debugger activation has two flavors, which cover the cases where KMDB is
32 32 * loaded at boot, and when it is loaded after boot. In brief, in both cases,
33 33 * the KDI needs to interpose upon several handlers in the IDT. When
34 34 * mod-loaded KMDB is deactivated, we undo the IDT interposition, restoring the
35 35 * handlers to what they were before we started.
36 36 *
37 37 * We also take over the entirety of IDT (except the double-fault handler) on
38 38 * the active CPU when we're in kmdb so we can handle things like page faults
39 39 * sensibly.
40 40 *
41 41 * Boot-loaded KMDB
42 42 *
43 43 * When we're first activated, we're running on boot's IDT. We need to be able
44 44 * to function in this world, so we'll install our handlers into boot's IDT.
45 45 * This is a little complicated: we're using the fake cpu_t set up by
46 46 * boot_kdi_tmpinit(), so we can't access cpu_idt directly. Instead,
47 47 * kdi_idt_write() notices that cpu_idt is NULL, and works around this problem.
48 48 *
49 49 * Later, when we're about to switch to the kernel's IDT, it'll call us via
50 50 * kdi_idt_sync(), allowing us to add our handlers to the new IDT. While
51 51 * boot-loaded KMDB can't be unloaded, we still need to save the descriptors we
52 52 * replace so we can pass traps back to the kernel as necessary.
53 53 *
54 54 * The last phase of boot-loaded KMDB activation occurs at non-boot CPU
55 55 * startup. We will be called on each non-boot CPU, thus allowing us to set up
56 56 * any watchpoints that may have been configured on the boot CPU and interpose
57 57 * on the given CPU's IDT. We don't save the interposed descriptors in this
58 58 * case -- see kdi_cpu_init() for details.
59 59 *
60 60 * Mod-loaded KMDB
61 61 *
62 62 * This style of activation is much simpler, as the CPUs are already running,
63 63 * and are using their own copy of the kernel's IDT. We simply interpose upon
64 64 * each CPU's IDT. We save the handlers we replace, both for deactivation and
65 65 * for passing traps back to the kernel. Note that for the hypervisors'
66 66 * benefit, we need to xcall to the other CPUs to do this, since we need to
67 67 * actively set the trap entries in its virtual IDT from that vcpu's context
68 68 * rather than just modifying the IDT table from the CPU running kdi_activate().
69 69 */
70 70
71 71 #include <sys/types.h>
72 72 #include <sys/segments.h>
73 73 #include <sys/trap.h>
74 74 #include <sys/cpuvar.h>
75 75 #include <sys/reboot.h>
76 76 #include <sys/sunddi.h>
77 77 #include <sys/archsystm.h>
78 78 #include <sys/kdi_impl.h>
79 79 #include <sys/x_call.h>
80 80 #include <ia32/sys/psw.h>
|
↓ open down ↓ |
80 lines elided |
↑ open up ↑ |
81 81 #include <vm/hat_i86.h>
82 82
83 83 #define KDI_GATE_NVECS 3
84 84
85 85 #define KDI_IDT_NOSAVE 0
86 86 #define KDI_IDT_SAVE 1
87 87
88 88 #define KDI_IDT_DTYPE_KERNEL 0
89 89 #define KDI_IDT_DTYPE_BOOT 1
90 90
91 +/* Solely to keep kdiregs_t in the CTF, otherwise unused. */
92 +kdiregs_t kdi_regs;
93 +
91 94 kdi_cpusave_t *kdi_cpusave;
92 95 int kdi_ncpusave;
93 96
94 97 static kdi_main_t kdi_kmdb_main;
95 98
96 99 kdi_drreg_t kdi_drreg;
97 100
98 101 #ifndef __amd64
99 102 /* Used to track the current set of valid kernel selectors. */
100 103 uint32_t kdi_cs;
101 104 uint32_t kdi_ds;
102 105 uint32_t kdi_fs;
103 106 uint32_t kdi_gs;
104 107 #endif
105 108
106 109 uintptr_t kdi_kernel_handler;
107 110
108 111 int kdi_trap_switch;
109 112
110 113 #define KDI_MEMRANGES_MAX 2
111 114
112 115 kdi_memrange_t kdi_memranges[KDI_MEMRANGES_MAX];
113 116 int kdi_nmemranges;
114 117
115 118 typedef void idt_hdlr_f(void);
116 119
117 120 extern idt_hdlr_f kdi_trap0, kdi_trap1, kdi_int2, kdi_trap3, kdi_trap4;
118 121 extern idt_hdlr_f kdi_trap5, kdi_trap6, kdi_trap7, kdi_trap9;
119 122 extern idt_hdlr_f kdi_traperr10, kdi_traperr11, kdi_traperr12;
120 123 extern idt_hdlr_f kdi_traperr13, kdi_traperr14, kdi_trap16, kdi_traperr17;
121 124 extern idt_hdlr_f kdi_trap18, kdi_trap19, kdi_trap20, kdi_ivct32;
122 125 extern idt_hdlr_f kdi_invaltrap;
123 126 extern size_t kdi_ivct_size;
124 127
125 128 typedef struct kdi_gate_spec {
126 129 uint_t kgs_vec;
127 130 uint_t kgs_dpl;
128 131 } kdi_gate_spec_t;
129 132
130 133 /*
131 134 * Beware: kdi_pass_to_kernel() has unpleasant knowledge of this list.
132 135 */
133 136 static const kdi_gate_spec_t kdi_gate_specs[KDI_GATE_NVECS] = {
134 137 { T_SGLSTP, TRP_KPL },
135 138 { T_BPTFLT, TRP_UPL },
136 139 { T_DBGENTR, TRP_KPL }
137 140 };
138 141
139 142 static gate_desc_t kdi_kgates[KDI_GATE_NVECS];
140 143
141 144 extern gate_desc_t kdi_idt[NIDT];
142 145
143 146 struct idt_description {
144 147 uint_t id_low;
145 148 uint_t id_high;
146 149 idt_hdlr_f *id_basehdlr;
147 150 size_t *id_incrp;
148 151 } idt_description[] = {
149 152 { T_ZERODIV, 0, kdi_trap0, NULL },
150 153 { T_SGLSTP, 0, kdi_trap1, NULL },
151 154 { T_NMIFLT, 0, kdi_int2, NULL },
152 155 { T_BPTFLT, 0, kdi_trap3, NULL },
153 156 { T_OVFLW, 0, kdi_trap4, NULL },
154 157 { T_BOUNDFLT, 0, kdi_trap5, NULL },
155 158 { T_ILLINST, 0, kdi_trap6, NULL },
156 159 { T_NOEXTFLT, 0, kdi_trap7, NULL },
157 160 #if !defined(__xpv)
158 161 { T_DBLFLT, 0, syserrtrap, NULL },
159 162 #endif
160 163 { T_EXTOVRFLT, 0, kdi_trap9, NULL },
161 164 { T_TSSFLT, 0, kdi_traperr10, NULL },
162 165 { T_SEGFLT, 0, kdi_traperr11, NULL },
163 166 { T_STKFLT, 0, kdi_traperr12, NULL },
164 167 { T_GPFLT, 0, kdi_traperr13, NULL },
165 168 { T_PGFLT, 0, kdi_traperr14, NULL },
166 169 { 15, 0, kdi_invaltrap, NULL },
167 170 { T_EXTERRFLT, 0, kdi_trap16, NULL },
168 171 { T_ALIGNMENT, 0, kdi_traperr17, NULL },
169 172 { T_MCE, 0, kdi_trap18, NULL },
170 173 { T_SIMDFPE, 0, kdi_trap19, NULL },
171 174 { T_DBGENTR, 0, kdi_trap20, NULL },
172 175 { 21, 31, kdi_invaltrap, NULL },
173 176 { 32, 255, kdi_ivct32, &kdi_ivct_size },
174 177 { 0, 0, NULL },
175 178 };
176 179
177 180 void
178 181 kdi_idt_init(selector_t sel)
179 182 {
180 183 struct idt_description *id;
181 184 int i;
182 185
183 186 for (id = idt_description; id->id_basehdlr != NULL; id++) {
184 187 uint_t high = id->id_high != 0 ? id->id_high : id->id_low;
185 188 size_t incr = id->id_incrp != NULL ? *id->id_incrp : 0;
186 189
187 190 #if !defined(__xpv)
188 191 if (kpti_enable && sel == KCS_SEL && id->id_low == T_DBLFLT)
189 192 id->id_basehdlr = tr_syserrtrap;
190 193 #endif
191 194
192 195 for (i = id->id_low; i <= high; i++) {
193 196 caddr_t hdlr = (caddr_t)id->id_basehdlr +
194 197 incr * (i - id->id_low);
195 198 set_gatesegd(&kdi_idt[i], (void (*)())hdlr, sel,
196 199 SDT_SYSIGT, TRP_KPL, IST_DBG);
197 200 }
198 201 }
199 202 }
200 203
201 204 static void
202 205 kdi_idt_gates_install(selector_t sel, int saveold)
203 206 {
204 207 gate_desc_t gates[KDI_GATE_NVECS];
205 208 int i;
206 209
207 210 bzero(gates, sizeof (*gates));
208 211
209 212 for (i = 0; i < KDI_GATE_NVECS; i++) {
210 213 const kdi_gate_spec_t *gs = &kdi_gate_specs[i];
211 214 uintptr_t func = GATESEG_GETOFFSET(&kdi_idt[gs->kgs_vec]);
212 215 set_gatesegd(&gates[i], (void (*)())func, sel, SDT_SYSIGT,
213 216 gs->kgs_dpl, IST_DBG);
214 217 }
215 218
216 219 for (i = 0; i < KDI_GATE_NVECS; i++) {
217 220 uint_t vec = kdi_gate_specs[i].kgs_vec;
218 221
219 222 if (saveold)
220 223 kdi_kgates[i] = CPU->cpu_m.mcpu_idt[vec];
221 224
222 225 kdi_idt_write(&gates[i], vec);
223 226 }
224 227 }
225 228
226 229 static void
227 230 kdi_idt_gates_restore(void)
228 231 {
229 232 int i;
230 233
231 234 for (i = 0; i < KDI_GATE_NVECS; i++)
232 235 kdi_idt_write(&kdi_kgates[i], kdi_gate_specs[i].kgs_vec);
233 236 }
234 237
235 238 /*
236 239 * Called when we switch to the kernel's IDT. We need to interpose on the
237 240 * kernel's IDT entries and stop using KMDBCODE_SEL.
238 241 */
239 242 void
240 243 kdi_idt_sync(void)
241 244 {
242 245 kdi_idt_init(KCS_SEL);
243 246 kdi_idt_gates_install(KCS_SEL, KDI_IDT_SAVE);
244 247 }
245 248
246 249 void
247 250 kdi_update_drreg(kdi_drreg_t *drreg)
248 251 {
249 252 kdi_drreg = *drreg;
250 253 }
251 254
252 255 void
253 256 kdi_memrange_add(caddr_t base, size_t len)
254 257 {
255 258 kdi_memrange_t *mr = &kdi_memranges[kdi_nmemranges];
256 259
257 260 ASSERT(kdi_nmemranges != KDI_MEMRANGES_MAX);
258 261
259 262 mr->mr_base = base;
260 263 mr->mr_lim = base + len - 1;
261 264 kdi_nmemranges++;
262 265 }
263 266
264 267 void
265 268 kdi_idt_switch(kdi_cpusave_t *cpusave)
266 269 {
267 270 if (cpusave == NULL)
268 271 kdi_idtr_set(kdi_idt, sizeof (kdi_idt) - 1);
269 272 else
270 273 kdi_idtr_set(cpusave->krs_idt, (sizeof (*idt0) * NIDT) - 1);
271 274 }
272 275
273 276 /*
274 277 * Activation for CPUs other than the boot CPU, called from that CPU's
275 278 * mp_startup(). We saved the kernel's descriptors when we initialized the
276 279 * boot CPU, so we don't want to do it again. Saving the handlers from this
277 280 * CPU's IDT would actually be dangerous with the CPU initialization method in
278 281 * use at the time of this writing. With that method, the startup code creates
279 282 * the IDTs for slave CPUs by copying the one used by the boot CPU, which has
280 283 * already been interposed upon by KMDB. Were we to interpose again, we'd
281 284 * replace the kernel's descriptors with our own in the save area. By not
282 285 * saving, but still overwriting, we'll work in the current world, and in any
283 286 * future world where the IDT is generated from scratch.
284 287 */
285 288 void
286 289 kdi_cpu_init(void)
287 290 {
288 291 kdi_idt_gates_install(KCS_SEL, KDI_IDT_NOSAVE);
289 292 /* Load the debug registers. */
290 293 kdi_cpu_debug_init(&kdi_cpusave[CPU->cpu_id]);
291 294 }
292 295
293 296 /*
294 297 * Activation for all CPUs for mod-loaded kmdb, i.e. a kmdb that wasn't
295 298 * loaded at boot.
296 299 */
297 300 static int
298 301 kdi_cpu_activate(void)
299 302 {
300 303 kdi_idt_gates_install(KCS_SEL, KDI_IDT_SAVE);
301 304 return (0);
302 305 }
303 306
304 307 void
305 308 kdi_activate(kdi_main_t main, kdi_cpusave_t *cpusave, uint_t ncpusave)
306 309 {
307 310 int i;
308 311 cpuset_t cpuset;
309 312
310 313 CPUSET_ALL(cpuset);
311 314
312 315 kdi_cpusave = cpusave;
313 316 kdi_ncpusave = ncpusave;
314 317
315 318 kdi_kmdb_main = main;
316 319
317 320 for (i = 0; i < kdi_ncpusave; i++) {
318 321 kdi_cpusave[i].krs_cpu_id = i;
319 322
320 323 kdi_cpusave[i].krs_curcrumb =
321 324 &kdi_cpusave[i].krs_crumbs[KDI_NCRUMBS - 1];
322 325 kdi_cpusave[i].krs_curcrumbidx = KDI_NCRUMBS - 1;
323 326 }
324 327
325 328 if (boothowto & RB_KMDB)
326 329 kdi_idt_init(KMDBCODE_SEL);
327 330 else
328 331 kdi_idt_init(KCS_SEL);
329 332
330 333 /* The initial selector set. Updated by the debugger-entry code */
331 334 #ifndef __amd64
332 335 kdi_cs = B32CODE_SEL;
333 336 kdi_ds = kdi_fs = kdi_gs = B32DATA_SEL;
334 337 #endif
335 338
336 339 kdi_memranges[0].mr_base = kdi_segdebugbase;
337 340 kdi_memranges[0].mr_lim = kdi_segdebugbase + kdi_segdebugsize - 1;
338 341 kdi_nmemranges = 1;
339 342
340 343 kdi_drreg.dr_ctl = KDIREG_DRCTL_RESERVED;
341 344 kdi_drreg.dr_stat = KDIREG_DRSTAT_RESERVED;
342 345
343 346 if (boothowto & RB_KMDB) {
344 347 kdi_idt_gates_install(KMDBCODE_SEL, KDI_IDT_NOSAVE);
345 348 } else {
346 349 xc_call(0, 0, 0, CPUSET2BV(cpuset),
347 350 (xc_func_t)kdi_cpu_activate);
348 351 }
349 352 }
350 353
351 354 static int
352 355 kdi_cpu_deactivate(void)
353 356 {
354 357 kdi_idt_gates_restore();
355 358 return (0);
356 359 }
357 360
358 361 void
359 362 kdi_deactivate(void)
360 363 {
361 364 cpuset_t cpuset;
362 365 CPUSET_ALL(cpuset);
363 366
364 367 xc_call(0, 0, 0, CPUSET2BV(cpuset), (xc_func_t)kdi_cpu_deactivate);
365 368 kdi_nmemranges = 0;
366 369 }
367 370
368 371 /*
369 372 * We receive all breakpoints and single step traps. Some of them, including
370 373 * those from userland and those induced by DTrace providers, are intended for
371 374 * the kernel, and must be processed there. We adopt this
372 375 * ours-until-proven-otherwise position due to the painful consequences of
373 376 * sending the kernel an unexpected breakpoint or single step. Unless someone
374 377 * can prove to us that the kernel is prepared to handle the trap, we'll assume
375 378 * there's a problem and will give the user a chance to debug it.
376 379 *
377 380 * If we return 2, then the calling code should restore the trap-time %cr3: that
378 381 * is, it really is a kernel-originated trap.
379 382 */
380 383 int
381 384 kdi_trap_pass(kdi_cpusave_t *cpusave)
382 385 {
383 386 greg_t tt = cpusave->krs_gregs[KDIREG_TRAPNO];
384 387 greg_t pc = cpusave->krs_gregs[KDIREG_PC];
385 388 greg_t cs = cpusave->krs_gregs[KDIREG_CS];
386 389
387 390 if (USERMODE(cs))
388 391 return (1);
389 392
390 393 if (tt != T_BPTFLT && tt != T_SGLSTP)
391 394 return (0);
392 395
393 396 if (tt == T_BPTFLT && kdi_dtrace_get_state() ==
394 397 KDI_DTSTATE_DTRACE_ACTIVE)
395 398 return (2);
396 399
397 400 /*
398 401 * See the comments in the kernel's T_SGLSTP handler for why we need to
399 402 * do this.
400 403 */
401 404 #if !defined(__xpv)
402 405 if (tt == T_SGLSTP &&
403 406 (pc == (greg_t)sys_sysenter || pc == (greg_t)brand_sys_sysenter ||
404 407 pc == (greg_t)tr_sys_sysenter ||
405 408 pc == (greg_t)tr_brand_sys_sysenter)) {
406 409 #else
407 410 if (tt == T_SGLSTP &&
408 411 (pc == (greg_t)sys_sysenter || pc == (greg_t)brand_sys_sysenter)) {
409 412 #endif
410 413 return (1);
411 414 }
412 415
413 416 return (0);
414 417 }
415 418
416 419 /*
417 420 * State has been saved, and all CPUs are on the CPU-specific stacks. All
418 421 * CPUs enter here, and head off into the debugger proper.
419 422 */
420 423 void
421 424 kdi_debugger_entry(kdi_cpusave_t *cpusave)
422 425 {
423 426 /*
424 427 * BPTFLT gives us control with %eip set to the instruction *after*
425 428 * the int 3. Back it off, so we're looking at the instruction that
426 429 * triggered the fault.
427 430 */
428 431 if (cpusave->krs_gregs[KDIREG_TRAPNO] == T_BPTFLT)
429 432 cpusave->krs_gregs[KDIREG_PC]--;
430 433
431 434 kdi_kmdb_main(cpusave);
432 435 }
|
↓ open down ↓ |
332 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX