afs: Provide a splice-read wrapper
[linux-block.git] / mm / kfence / report.c
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * KFENCE reporting.
4  *
5  * Copyright (C) 2020, Google LLC.
6  */
7
8 #include <linux/stdarg.h>
9
10 #include <linux/kernel.h>
11 #include <linux/lockdep.h>
12 #include <linux/math.h>
13 #include <linux/printk.h>
14 #include <linux/sched/debug.h>
15 #include <linux/seq_file.h>
16 #include <linux/stacktrace.h>
17 #include <linux/string.h>
18 #include <trace/events/error_report.h>
19
20 #include <asm/kfence.h>
21
22 #include "kfence.h"
23
24 /* May be overridden by <asm/kfence.h>. */
25 #ifndef ARCH_FUNC_PREFIX
26 #define ARCH_FUNC_PREFIX ""
27 #endif
28
29 extern bool no_hash_pointers;
30
31 /* Helper function to either print to a seq_file or to console. */
32 __printf(2, 3)
33 static void seq_con_printf(struct seq_file *seq, const char *fmt, ...)
34 {
35         va_list args;
36
37         va_start(args, fmt);
38         if (seq)
39                 seq_vprintf(seq, fmt, args);
40         else
41                 vprintk(fmt, args);
42         va_end(args);
43 }
44
45 /*
46  * Get the number of stack entries to skip to get out of MM internals. @type is
47  * optional, and if set to NULL, assumes an allocation or free stack.
48  */
49 static int get_stack_skipnr(const unsigned long stack_entries[], int num_entries,
50                             const enum kfence_error_type *type)
51 {
52         char buf[64];
53         int skipnr, fallback = 0;
54
55         if (type) {
56                 /* Depending on error type, find different stack entries. */
57                 switch (*type) {
58                 case KFENCE_ERROR_UAF:
59                 case KFENCE_ERROR_OOB:
60                 case KFENCE_ERROR_INVALID:
61                         /*
62                          * kfence_handle_page_fault() may be called with pt_regs
63                          * set to NULL; in that case we'll simply show the full
64                          * stack trace.
65                          */
66                         return 0;
67                 case KFENCE_ERROR_CORRUPTION:
68                 case KFENCE_ERROR_INVALID_FREE:
69                         break;
70                 }
71         }
72
73         for (skipnr = 0; skipnr < num_entries; skipnr++) {
74                 int len = scnprintf(buf, sizeof(buf), "%ps", (void *)stack_entries[skipnr]);
75
76                 if (str_has_prefix(buf, ARCH_FUNC_PREFIX "kfence_") ||
77                     str_has_prefix(buf, ARCH_FUNC_PREFIX "__kfence_") ||
78                     str_has_prefix(buf, ARCH_FUNC_PREFIX "__kmem_cache_free") ||
79                     !strncmp(buf, ARCH_FUNC_PREFIX "__slab_free", len)) {
80                         /*
81                          * In case of tail calls from any of the below to any of
82                          * the above, optimized by the compiler such that the
83                          * stack trace would omit the initial entry point below.
84                          */
85                         fallback = skipnr + 1;
86                 }
87
88                 /*
89                  * The below list should only include the initial entry points
90                  * into the slab allocators. Includes the *_bulk() variants by
91                  * checking prefixes.
92                  */
93                 if (str_has_prefix(buf, ARCH_FUNC_PREFIX "kfree") ||
94                     str_has_prefix(buf, ARCH_FUNC_PREFIX "kmem_cache_free") ||
95                     str_has_prefix(buf, ARCH_FUNC_PREFIX "__kmalloc") ||
96                     str_has_prefix(buf, ARCH_FUNC_PREFIX "kmem_cache_alloc"))
97                         goto found;
98         }
99         if (fallback < num_entries)
100                 return fallback;
101 found:
102         skipnr++;
103         return skipnr < num_entries ? skipnr : 0;
104 }
105
106 static void kfence_print_stack(struct seq_file *seq, const struct kfence_metadata *meta,
107                                bool show_alloc)
108 {
109         const struct kfence_track *track = show_alloc ? &meta->alloc_track : &meta->free_track;
110         u64 ts_sec = track->ts_nsec;
111         unsigned long rem_nsec = do_div(ts_sec, NSEC_PER_SEC);
112
113         /* Timestamp matches printk timestamp format. */
114         seq_con_printf(seq, "%s by task %d on cpu %d at %lu.%06lus:\n",
115                        show_alloc ? "allocated" : "freed", track->pid,
116                        track->cpu, (unsigned long)ts_sec, rem_nsec / 1000);
117
118         if (track->num_stack_entries) {
119                 /* Skip allocation/free internals stack. */
120                 int i = get_stack_skipnr(track->stack_entries, track->num_stack_entries, NULL);
121
122                 /* stack_trace_seq_print() does not exist; open code our own. */
123                 for (; i < track->num_stack_entries; i++)
124                         seq_con_printf(seq, " %pS\n", (void *)track->stack_entries[i]);
125         } else {
126                 seq_con_printf(seq, " no %s stack\n", show_alloc ? "allocation" : "deallocation");
127         }
128 }
129
130 void kfence_print_object(struct seq_file *seq, const struct kfence_metadata *meta)
131 {
132         const int size = abs(meta->size);
133         const unsigned long start = meta->addr;
134         const struct kmem_cache *const cache = meta->cache;
135
136         lockdep_assert_held(&meta->lock);
137
138         if (meta->state == KFENCE_OBJECT_UNUSED) {
139                 seq_con_printf(seq, "kfence-#%td unused\n", meta - kfence_metadata);
140                 return;
141         }
142
143         seq_con_printf(seq, "kfence-#%td: 0x%p-0x%p, size=%d, cache=%s\n\n",
144                        meta - kfence_metadata, (void *)start, (void *)(start + size - 1),
145                        size, (cache && cache->name) ? cache->name : "<destroyed>");
146
147         kfence_print_stack(seq, meta, true);
148
149         if (meta->state == KFENCE_OBJECT_FREED) {
150                 seq_con_printf(seq, "\n");
151                 kfence_print_stack(seq, meta, false);
152         }
153 }
154
155 /*
156  * Show bytes at @addr that are different from the expected canary values, up to
157  * @max_bytes.
158  */
159 static void print_diff_canary(unsigned long address, size_t bytes_to_show,
160                               const struct kfence_metadata *meta)
161 {
162         const unsigned long show_until_addr = address + bytes_to_show;
163         const u8 *cur, *end;
164
165         /* Do not show contents of object nor read into following guard page. */
166         end = (const u8 *)(address < meta->addr ? min(show_until_addr, meta->addr)
167                                                 : min(show_until_addr, PAGE_ALIGN(address)));
168
169         pr_cont("[");
170         for (cur = (const u8 *)address; cur < end; cur++) {
171                 if (*cur == KFENCE_CANARY_PATTERN_U8(cur))
172                         pr_cont(" .");
173                 else if (no_hash_pointers)
174                         pr_cont(" 0x%02x", *cur);
175                 else /* Do not leak kernel memory in non-debug builds. */
176                         pr_cont(" !");
177         }
178         pr_cont(" ]");
179 }
180
181 static const char *get_access_type(bool is_write)
182 {
183         return is_write ? "write" : "read";
184 }
185
186 void kfence_report_error(unsigned long address, bool is_write, struct pt_regs *regs,
187                          const struct kfence_metadata *meta, enum kfence_error_type type)
188 {
189         unsigned long stack_entries[KFENCE_STACK_DEPTH] = { 0 };
190         const ptrdiff_t object_index = meta ? meta - kfence_metadata : -1;
191         int num_stack_entries;
192         int skipnr = 0;
193
194         if (regs) {
195                 num_stack_entries = stack_trace_save_regs(regs, stack_entries, KFENCE_STACK_DEPTH, 0);
196         } else {
197                 num_stack_entries = stack_trace_save(stack_entries, KFENCE_STACK_DEPTH, 1);
198                 skipnr = get_stack_skipnr(stack_entries, num_stack_entries, &type);
199         }
200
201         /* Require non-NULL meta, except if KFENCE_ERROR_INVALID. */
202         if (WARN_ON(type != KFENCE_ERROR_INVALID && !meta))
203                 return;
204
205         if (meta)
206                 lockdep_assert_held(&meta->lock);
207         /*
208          * Because we may generate reports in printk-unfriendly parts of the
209          * kernel, such as scheduler code, the use of printk() could deadlock.
210          * Until such time that all printing code here is safe in all parts of
211          * the kernel, accept the risk, and just get our message out (given the
212          * system might already behave unpredictably due to the memory error).
213          * As such, also disable lockdep to hide warnings, and avoid disabling
214          * lockdep for the rest of the kernel.
215          */
216         lockdep_off();
217
218         pr_err("==================================================================\n");
219         /* Print report header. */
220         switch (type) {
221         case KFENCE_ERROR_OOB: {
222                 const bool left_of_object = address < meta->addr;
223
224                 pr_err("BUG: KFENCE: out-of-bounds %s in %pS\n\n", get_access_type(is_write),
225                        (void *)stack_entries[skipnr]);
226                 pr_err("Out-of-bounds %s at 0x%p (%luB %s of kfence-#%td):\n",
227                        get_access_type(is_write), (void *)address,
228                        left_of_object ? meta->addr - address : address - meta->addr,
229                        left_of_object ? "left" : "right", object_index);
230                 break;
231         }
232         case KFENCE_ERROR_UAF:
233                 pr_err("BUG: KFENCE: use-after-free %s in %pS\n\n", get_access_type(is_write),
234                        (void *)stack_entries[skipnr]);
235                 pr_err("Use-after-free %s at 0x%p (in kfence-#%td):\n",
236                        get_access_type(is_write), (void *)address, object_index);
237                 break;
238         case KFENCE_ERROR_CORRUPTION:
239                 pr_err("BUG: KFENCE: memory corruption in %pS\n\n", (void *)stack_entries[skipnr]);
240                 pr_err("Corrupted memory at 0x%p ", (void *)address);
241                 print_diff_canary(address, 16, meta);
242                 pr_cont(" (in kfence-#%td):\n", object_index);
243                 break;
244         case KFENCE_ERROR_INVALID:
245                 pr_err("BUG: KFENCE: invalid %s in %pS\n\n", get_access_type(is_write),
246                        (void *)stack_entries[skipnr]);
247                 pr_err("Invalid %s at 0x%p:\n", get_access_type(is_write),
248                        (void *)address);
249                 break;
250         case KFENCE_ERROR_INVALID_FREE:
251                 pr_err("BUG: KFENCE: invalid free in %pS\n\n", (void *)stack_entries[skipnr]);
252                 pr_err("Invalid free of 0x%p (in kfence-#%td):\n", (void *)address,
253                        object_index);
254                 break;
255         }
256
257         /* Print stack trace and object info. */
258         stack_trace_print(stack_entries + skipnr, num_stack_entries - skipnr, 0);
259
260         if (meta) {
261                 pr_err("\n");
262                 kfence_print_object(NULL, meta);
263         }
264
265         /* Print report footer. */
266         pr_err("\n");
267         if (no_hash_pointers && regs)
268                 show_regs(regs);
269         else
270                 dump_stack_print_info(KERN_ERR);
271         trace_error_report_end(ERROR_DETECTOR_KFENCE, address);
272         pr_err("==================================================================\n");
273
274         lockdep_on();
275
276         check_panic_on_warn("KFENCE");
277
278         /* We encountered a memory safety error, taint the kernel! */
279         add_taint(TAINT_BAD_PAGE, LOCKDEP_STILL_OK);
280 }
281
282 #ifdef CONFIG_PRINTK
283 static void kfence_to_kp_stack(const struct kfence_track *track, void **kp_stack)
284 {
285         int i, j;
286
287         i = get_stack_skipnr(track->stack_entries, track->num_stack_entries, NULL);
288         for (j = 0; i < track->num_stack_entries && j < KS_ADDRS_COUNT; ++i, ++j)
289                 kp_stack[j] = (void *)track->stack_entries[i];
290         if (j < KS_ADDRS_COUNT)
291                 kp_stack[j] = NULL;
292 }
293
294 bool __kfence_obj_info(struct kmem_obj_info *kpp, void *object, struct slab *slab)
295 {
296         struct kfence_metadata *meta = addr_to_metadata((unsigned long)object);
297         unsigned long flags;
298
299         if (!meta)
300                 return false;
301
302         /*
303          * If state is UNUSED at least show the pointer requested; the rest
304          * would be garbage data.
305          */
306         kpp->kp_ptr = object;
307
308         /* Requesting info an a never-used object is almost certainly a bug. */
309         if (WARN_ON(meta->state == KFENCE_OBJECT_UNUSED))
310                 return true;
311
312         raw_spin_lock_irqsave(&meta->lock, flags);
313
314         kpp->kp_slab = slab;
315         kpp->kp_slab_cache = meta->cache;
316         kpp->kp_objp = (void *)meta->addr;
317         kfence_to_kp_stack(&meta->alloc_track, kpp->kp_stack);
318         if (meta->state == KFENCE_OBJECT_FREED)
319                 kfence_to_kp_stack(&meta->free_track, kpp->kp_free_stack);
320         /* get_stack_skipnr() ensures the first entry is outside allocator. */
321         kpp->kp_ret = kpp->kp_stack[0];
322
323         raw_spin_unlock_irqrestore(&meta->lock, flags);
324
325         return true;
326 }
327 #endif