1//===-- xray_interface.cpp --------------------------------------*- C++ -*-===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9// This file is a part of XRay, a dynamic runtime instrumentation system.
10//
11// Implementation of the API functions.
12//
13//===----------------------------------------------------------------------===//
14
15#include "xray_interface_internal.h"
16
17#include <cinttypes>
18#include <cstdio>
19#include <errno.h>
20#include <limits>
21#include <string.h>
22#include <sys/mman.h>
23
24#if SANITIZER_FUCHSIA
25#include <zircon/process.h>
26#include <zircon/sanitizer.h>
27#include <zircon/status.h>
28#include <zircon/syscalls.h>
29#endif
30
31#include "sanitizer_common/sanitizer_addrhashmap.h"
32#include "sanitizer_common/sanitizer_common.h"
33
34#include "xray_defs.h"
35#include "xray_flags.h"
36
37extern __sanitizer::SpinMutex XRayInstrMapMutex;
38extern __sanitizer::atomic_uint8_t XRayInitialized;
39extern __xray::XRaySledMap XRayInstrMap;
40
41namespace __xray {
42
43#if defined(__x86_64__)
44static const int16_t cSledLength = 12;
45#elif defined(__aarch64__)
46static const int16_t cSledLength = 32;
47#elif defined(__arm__)
48static const int16_t cSledLength = 28;
49#elif SANITIZER_LOONGARCH64
50static const int16_t cSledLength = 48;
51#elif SANITIZER_MIPS32
52static const int16_t cSledLength = 48;
53#elif SANITIZER_MIPS64
54static const int16_t cSledLength = 64;
55#elif defined(__powerpc64__)
56static const int16_t cSledLength = 8;
57#elif defined(__hexagon__)
58static const int16_t cSledLength = 20;
59#else
60#error "Unsupported CPU Architecture"
61#endif /* CPU architecture */
62
63// This is the function to call when we encounter the entry or exit sleds.
64atomic_uintptr_t XRayPatchedFunction{.val_dont_use: 0};
65
66// This is the function to call from the arg1-enabled sleds/trampolines.
67atomic_uintptr_t XRayArgLogger{.val_dont_use: 0};
68
69// This is the function to call when we encounter a custom event log call.
70atomic_uintptr_t XRayPatchedCustomEvent{.val_dont_use: 0};
71
72// This is the function to call when we encounter a typed event log call.
73atomic_uintptr_t XRayPatchedTypedEvent{.val_dont_use: 0};
74
75// This is the global status to determine whether we are currently
76// patching/unpatching.
77atomic_uint8_t XRayPatching{.val_dont_use: 0};
78
79struct TypeDescription {
80 uint32_t type_id;
81 std::size_t description_string_length;
82};
83
84using TypeDescriptorMapType = AddrHashMap<TypeDescription, 11>;
85// An address map from immutable descriptors to type ids.
86TypeDescriptorMapType TypeDescriptorAddressMap{};
87
88atomic_uint32_t TypeEventDescriptorCounter{.val_dont_use: 0};
89
90// MProtectHelper is an RAII wrapper for calls to mprotect(...) that will
91// undo any successful mprotect(...) changes. This is used to make a page
92// writeable and executable, and upon destruction if it was successful in
93// doing so returns the page into a read-only and executable page.
94//
95// This is only used specifically for runtime-patching of the XRay
96// instrumentation points. This assumes that the executable pages are
97// originally read-and-execute only.
98class MProtectHelper {
99 void *PageAlignedAddr;
100 std::size_t MProtectLen;
101 bool MustCleanup;
102
103public:
104 explicit MProtectHelper(void *PageAlignedAddr,
105 std::size_t MProtectLen,
106 std::size_t PageSize) XRAY_NEVER_INSTRUMENT
107 : PageAlignedAddr(PageAlignedAddr),
108 MProtectLen(MProtectLen),
109 MustCleanup(false) {
110#if SANITIZER_FUCHSIA
111 MProtectLen = RoundUpTo(MProtectLen, PageSize);
112#endif
113 }
114
115 int MakeWriteable() XRAY_NEVER_INSTRUMENT {
116#if SANITIZER_FUCHSIA
117 auto R = __sanitizer_change_code_protection(
118 reinterpret_cast<uintptr_t>(PageAlignedAddr), MProtectLen, true);
119 if (R != ZX_OK) {
120 Report("XRay: cannot change code protection: %s\n",
121 _zx_status_get_string(R));
122 return -1;
123 }
124 MustCleanup = true;
125 return 0;
126#else
127 auto R = mprotect(addr: PageAlignedAddr, len: MProtectLen,
128 PROT_READ | PROT_WRITE | PROT_EXEC);
129 if (R != -1)
130 MustCleanup = true;
131 return R;
132#endif
133 }
134
135 ~MProtectHelper() XRAY_NEVER_INSTRUMENT {
136 if (MustCleanup) {
137#if SANITIZER_FUCHSIA
138 auto R = __sanitizer_change_code_protection(
139 reinterpret_cast<uintptr_t>(PageAlignedAddr), MProtectLen, false);
140 if (R != ZX_OK) {
141 Report("XRay: cannot change code protection: %s\n",
142 _zx_status_get_string(R));
143 }
144#else
145 mprotect(addr: PageAlignedAddr, len: MProtectLen, PROT_READ | PROT_EXEC);
146#endif
147 }
148 }
149};
150
151namespace {
152
153bool patchSled(const XRaySledEntry &Sled, bool Enable,
154 int32_t FuncId) XRAY_NEVER_INSTRUMENT {
155 bool Success = false;
156 switch (Sled.Kind) {
157 case XRayEntryType::ENTRY:
158 Success = patchFunctionEntry(Enable, FuncId, Sled, Trampoline: __xray_FunctionEntry);
159 break;
160 case XRayEntryType::EXIT:
161 Success = patchFunctionExit(Enable, FuncId, Sled);
162 break;
163 case XRayEntryType::TAIL:
164 Success = patchFunctionTailExit(Enable, FuncId, Sled);
165 break;
166 case XRayEntryType::LOG_ARGS_ENTRY:
167 Success = patchFunctionEntry(Enable, FuncId, Sled, Trampoline: __xray_ArgLoggerEntry);
168 break;
169 case XRayEntryType::CUSTOM_EVENT:
170 Success = patchCustomEvent(Enable, FuncId, Sled);
171 break;
172 case XRayEntryType::TYPED_EVENT:
173 Success = patchTypedEvent(Enable, FuncId, Sled);
174 break;
175 default:
176 Report(format: "Unsupported sled kind '%" PRIu64 "' @%04x\n", Sled.Address,
177 int(Sled.Kind));
178 return false;
179 }
180 return Success;
181}
182
183const XRayFunctionSledIndex
184findFunctionSleds(int32_t FuncId,
185 const XRaySledMap &InstrMap) XRAY_NEVER_INSTRUMENT {
186 int32_t CurFn = 0;
187 uint64_t LastFnAddr = 0;
188 XRayFunctionSledIndex Index = {.Begin: nullptr, .Size: 0};
189
190 for (std::size_t I = 0; I < InstrMap.Entries && CurFn <= FuncId; I++) {
191 const auto &Sled = InstrMap.Sleds[I];
192 const auto Function = Sled.function();
193 if (Function != LastFnAddr) {
194 CurFn++;
195 LastFnAddr = Function;
196 }
197
198 if (CurFn == FuncId) {
199 if (Index.Begin == nullptr)
200 Index.Begin = &Sled;
201 Index.Size = &Sled - Index.Begin + 1;
202 }
203 }
204
205 return Index;
206}
207
208XRayPatchingStatus patchFunction(int32_t FuncId,
209 bool Enable) XRAY_NEVER_INSTRUMENT {
210 if (!atomic_load(a: &XRayInitialized,
211 mo: memory_order_acquire))
212 return XRayPatchingStatus::NOT_INITIALIZED; // Not initialized.
213
214 uint8_t NotPatching = false;
215 if (!atomic_compare_exchange_strong(
216 a: &XRayPatching, cmp: &NotPatching, xchg: true, mo: memory_order_acq_rel))
217 return XRayPatchingStatus::ONGOING; // Already patching.
218
219 // Next, we look for the function index.
220 XRaySledMap InstrMap;
221 {
222 SpinMutexLock Guard(&XRayInstrMapMutex);
223 InstrMap = XRayInstrMap;
224 }
225
226 // If we don't have an index, we can't patch individual functions.
227 if (InstrMap.Functions == 0)
228 return XRayPatchingStatus::NOT_INITIALIZED;
229
230 // FuncId must be a positive number, less than the number of functions
231 // instrumented.
232 if (FuncId <= 0 || static_cast<size_t>(FuncId) > InstrMap.Functions) {
233 Report(format: "Invalid function id provided: %d\n", FuncId);
234 return XRayPatchingStatus::FAILED;
235 }
236
237 // Now we patch ths sleds for this specific function.
238 XRayFunctionSledIndex SledRange;
239 if (InstrMap.SledsIndex) {
240 SledRange = {.Begin: InstrMap.SledsIndex[FuncId - 1].fromPCRelative(),
241 .Size: InstrMap.SledsIndex[FuncId - 1].Size};
242 } else {
243 SledRange = findFunctionSleds(FuncId, InstrMap);
244 }
245 auto *f = SledRange.Begin;
246 bool SucceedOnce = false;
247 for (size_t i = 0; i != SledRange.Size; ++i)
248 SucceedOnce |= patchSled(Sled: f[i], Enable, FuncId);
249
250 atomic_store(a: &XRayPatching, v: false,
251 mo: memory_order_release);
252
253 if (!SucceedOnce) {
254 Report(format: "Failed patching any sled for function '%d'.", FuncId);
255 return XRayPatchingStatus::FAILED;
256 }
257
258 return XRayPatchingStatus::SUCCESS;
259}
260
261// controlPatching implements the common internals of the patching/unpatching
262// implementation. |Enable| defines whether we're enabling or disabling the
263// runtime XRay instrumentation.
264XRayPatchingStatus controlPatching(bool Enable) XRAY_NEVER_INSTRUMENT {
265 if (!atomic_load(a: &XRayInitialized,
266 mo: memory_order_acquire))
267 return XRayPatchingStatus::NOT_INITIALIZED; // Not initialized.
268
269 uint8_t NotPatching = false;
270 if (!atomic_compare_exchange_strong(
271 a: &XRayPatching, cmp: &NotPatching, xchg: true, mo: memory_order_acq_rel))
272 return XRayPatchingStatus::ONGOING; // Already patching.
273
274 uint8_t PatchingSuccess = false;
275 auto XRayPatchingStatusResetter =
276 at_scope_exit(fn: [&PatchingSuccess] {
277 if (!PatchingSuccess)
278 atomic_store(a: &XRayPatching, v: false,
279 mo: memory_order_release);
280 });
281
282 XRaySledMap InstrMap;
283 {
284 SpinMutexLock Guard(&XRayInstrMapMutex);
285 InstrMap = XRayInstrMap;
286 }
287 if (InstrMap.Entries == 0)
288 return XRayPatchingStatus::NOT_INITIALIZED;
289
290 uint32_t FuncId = 1;
291 uint64_t CurFun = 0;
292
293 // First we want to find the bounds for which we have instrumentation points,
294 // and try to get as few calls to mprotect(...) as possible. We're assuming
295 // that all the sleds for the instrumentation map are contiguous as a single
296 // set of pages. When we do support dynamic shared object instrumentation,
297 // we'll need to do this for each set of page load offsets per DSO loaded. For
298 // now we're assuming we can mprotect the whole section of text between the
299 // minimum sled address and the maximum sled address (+ the largest sled
300 // size).
301 auto *MinSled = &InstrMap.Sleds[0];
302 auto *MaxSled = &InstrMap.Sleds[InstrMap.Entries - 1];
303 for (std::size_t I = 0; I < InstrMap.Entries; I++) {
304 const auto &Sled = InstrMap.Sleds[I];
305 if (Sled.address() < MinSled->address())
306 MinSled = &Sled;
307 if (Sled.address() > MaxSled->address())
308 MaxSled = &Sled;
309 }
310
311 const size_t PageSize = flags()->xray_page_size_override > 0
312 ? flags()->xray_page_size_override
313 : GetPageSizeCached();
314 if ((PageSize == 0) || ((PageSize & (PageSize - 1)) != 0)) {
315 Report(format: "System page size is not a power of two: %zu\n", PageSize);
316 return XRayPatchingStatus::FAILED;
317 }
318
319 void *PageAlignedAddr =
320 reinterpret_cast<void *>(MinSled->address() & ~(PageSize - 1));
321 size_t MProtectLen =
322 (MaxSled->address() - reinterpret_cast<uptr>(PageAlignedAddr)) +
323 cSledLength;
324 MProtectHelper Protector(PageAlignedAddr, MProtectLen, PageSize);
325 if (Protector.MakeWriteable() == -1) {
326 Report(format: "Failed mprotect: %d\n", errno);
327 return XRayPatchingStatus::FAILED;
328 }
329
330 for (std::size_t I = 0; I < InstrMap.Entries; ++I) {
331 auto &Sled = InstrMap.Sleds[I];
332 auto F = Sled.function();
333 if (CurFun == 0)
334 CurFun = F;
335 if (F != CurFun) {
336 ++FuncId;
337 CurFun = F;
338 }
339 patchSled(Sled, Enable, FuncId);
340 }
341 atomic_store(a: &XRayPatching, v: false,
342 mo: memory_order_release);
343 PatchingSuccess = true;
344 return XRayPatchingStatus::SUCCESS;
345}
346
347XRayPatchingStatus mprotectAndPatchFunction(int32_t FuncId,
348 bool Enable) XRAY_NEVER_INSTRUMENT {
349 XRaySledMap InstrMap;
350 {
351 SpinMutexLock Guard(&XRayInstrMapMutex);
352 InstrMap = XRayInstrMap;
353 }
354
355 // FuncId must be a positive number, less than the number of functions
356 // instrumented.
357 if (FuncId <= 0 || static_cast<size_t>(FuncId) > InstrMap.Functions) {
358 Report(format: "Invalid function id provided: %d\n", FuncId);
359 return XRayPatchingStatus::FAILED;
360 }
361
362 const size_t PageSize = flags()->xray_page_size_override > 0
363 ? flags()->xray_page_size_override
364 : GetPageSizeCached();
365 if ((PageSize == 0) || ((PageSize & (PageSize - 1)) != 0)) {
366 Report(format: "Provided page size is not a power of two: %zu\n", PageSize);
367 return XRayPatchingStatus::FAILED;
368 }
369
370 // Here we compute the minimum sled and maximum sled associated with a
371 // particular function ID.
372 XRayFunctionSledIndex SledRange;
373 if (InstrMap.SledsIndex) {
374 SledRange = {.Begin: InstrMap.SledsIndex[FuncId - 1].fromPCRelative(),
375 .Size: InstrMap.SledsIndex[FuncId - 1].Size};
376 } else {
377 SledRange = findFunctionSleds(FuncId, InstrMap);
378 }
379 auto *f = SledRange.Begin;
380 auto *e = SledRange.Begin + SledRange.Size;
381 auto *MinSled = f;
382 auto *MaxSled = e - 1;
383 while (f != e) {
384 if (f->address() < MinSled->address())
385 MinSled = f;
386 if (f->address() > MaxSled->address())
387 MaxSled = f;
388 ++f;
389 }
390
391 void *PageAlignedAddr =
392 reinterpret_cast<void *>(MinSled->address() & ~(PageSize - 1));
393 size_t MProtectLen =
394 (MaxSled->address() - reinterpret_cast<uptr>(PageAlignedAddr)) +
395 cSledLength;
396 MProtectHelper Protector(PageAlignedAddr, MProtectLen, PageSize);
397 if (Protector.MakeWriteable() == -1) {
398 Report(format: "Failed mprotect: %d\n", errno);
399 return XRayPatchingStatus::FAILED;
400 }
401 return patchFunction(FuncId, Enable);
402}
403
404} // namespace
405
406} // namespace __xray
407
408using namespace __xray;
409
410// The following functions are declared `extern "C" {...}` in the header, hence
411// they're defined in the global namespace.
412
413int __xray_set_handler(void (*entry)(int32_t,
414 XRayEntryType)) XRAY_NEVER_INSTRUMENT {
415 if (atomic_load(a: &XRayInitialized,
416 mo: memory_order_acquire)) {
417
418 atomic_store(a: &__xray::XRayPatchedFunction,
419 v: reinterpret_cast<uintptr_t>(entry),
420 mo: memory_order_release);
421 return 1;
422 }
423 return 0;
424}
425
426int __xray_set_customevent_handler(void (*entry)(void *, size_t))
427 XRAY_NEVER_INSTRUMENT {
428 if (atomic_load(a: &XRayInitialized,
429 mo: memory_order_acquire)) {
430 atomic_store(a: &__xray::XRayPatchedCustomEvent,
431 v: reinterpret_cast<uintptr_t>(entry),
432 mo: memory_order_release);
433 return 1;
434 }
435 return 0;
436}
437
438int __xray_set_typedevent_handler(void (*entry)(size_t, const void *,
439 size_t)) XRAY_NEVER_INSTRUMENT {
440 if (atomic_load(a: &XRayInitialized,
441 mo: memory_order_acquire)) {
442 atomic_store(a: &__xray::XRayPatchedTypedEvent,
443 v: reinterpret_cast<uintptr_t>(entry),
444 mo: memory_order_release);
445 return 1;
446 }
447 return 0;
448}
449
450int __xray_remove_handler() XRAY_NEVER_INSTRUMENT {
451 return __xray_set_handler(entry: nullptr);
452}
453
454int __xray_remove_customevent_handler() XRAY_NEVER_INSTRUMENT {
455 return __xray_set_customevent_handler(entry: nullptr);
456}
457
458int __xray_remove_typedevent_handler() XRAY_NEVER_INSTRUMENT {
459 return __xray_set_typedevent_handler(entry: nullptr);
460}
461
462uint16_t __xray_register_event_type(
463 const char *const event_type) XRAY_NEVER_INSTRUMENT {
464 TypeDescriptorMapType::Handle h(&TypeDescriptorAddressMap, (uptr)event_type);
465 if (h.created()) {
466 h->type_id = atomic_fetch_add(
467 a: &TypeEventDescriptorCounter, v: 1, mo: memory_order_acq_rel);
468 h->description_string_length = strnlen(string: event_type, maxlen: 1024);
469 }
470 return h->type_id;
471}
472
473XRayPatchingStatus __xray_patch() XRAY_NEVER_INSTRUMENT {
474 return controlPatching(Enable: true);
475}
476
477XRayPatchingStatus __xray_unpatch() XRAY_NEVER_INSTRUMENT {
478 return controlPatching(Enable: false);
479}
480
481XRayPatchingStatus __xray_patch_function(int32_t FuncId) XRAY_NEVER_INSTRUMENT {
482 return mprotectAndPatchFunction(FuncId, Enable: true);
483}
484
485XRayPatchingStatus
486__xray_unpatch_function(int32_t FuncId) XRAY_NEVER_INSTRUMENT {
487 return mprotectAndPatchFunction(FuncId, Enable: false);
488}
489
490int __xray_set_handler_arg1(void (*entry)(int32_t, XRayEntryType, uint64_t)) {
491 if (!atomic_load(a: &XRayInitialized,
492 mo: memory_order_acquire))
493 return 0;
494
495 // A relaxed write might not be visible even if the current thread gets
496 // scheduled on a different CPU/NUMA node. We need to wait for everyone to
497 // have this handler installed for consistency of collected data across CPUs.
498 atomic_store(a: &XRayArgLogger, v: reinterpret_cast<uint64_t>(entry),
499 mo: memory_order_release);
500 return 1;
501}
502
503int __xray_remove_handler_arg1() { return __xray_set_handler_arg1(entry: nullptr); }
504
505uintptr_t __xray_function_address(int32_t FuncId) XRAY_NEVER_INSTRUMENT {
506 XRaySledMap InstrMap;
507 {
508 SpinMutexLock Guard(&XRayInstrMapMutex);
509 InstrMap = XRayInstrMap;
510 }
511
512 if (FuncId <= 0 || static_cast<size_t>(FuncId) > InstrMap.Functions)
513 return 0;
514 const XRaySledEntry *Sled =
515 InstrMap.SledsIndex ? InstrMap.SledsIndex[FuncId - 1].fromPCRelative()
516 : findFunctionSleds(FuncId, InstrMap).Begin;
517 return Sled->function()
518// On PPC, function entries are always aligned to 16 bytes. The beginning of a
519// sled might be a local entry, which is always +8 based on the global entry.
520// Always return the global entry.
521#ifdef __PPC__
522 & ~0xf
523#endif
524 ;
525}
526
527size_t __xray_max_function_id() XRAY_NEVER_INSTRUMENT {
528 SpinMutexLock Guard(&XRayInstrMapMutex);
529 return XRayInstrMap.Functions;
530}
531