1 // SPDX-License-Identifier: GPL-2.0-or-later 2 /* 3 * printk_safe.c - Safe printk for printk-deadlock-prone contexts 4 */ 5 6 #include <linux/preempt.h> 7 #include <linux/spinlock.h> 8 #include <linux/debug_locks.h> 9 #include <linux/kdb.h> 10 #include <linux/smp.h> 11 #include <linux/cpumask.h> 12 #include <linux/irq_work.h> 13 #include <linux/printk.h> 14 #include <linux/kprobes.h> 15 16 #include "internal.h" 17 18 /* 19 * printk() could not take logbuf_lock in NMI context. Instead, 20 * it uses an alternative implementation that temporary stores 21 * the strings into a per-CPU buffer. The content of the buffer 22 * is later flushed into the main ring buffer via IRQ work. 23 * 24 * The alternative implementation is chosen transparently 25 * by examining current printk() context mask stored in @printk_context 26 * per-CPU variable. 27 * 28 * The implementation allows to flush the strings also from another CPU. 29 * There are situations when we want to make sure that all buffers 30 * were handled or when IRQs are blocked. 31 */ 32 33 #define SAFE_LOG_BUF_LEN ((1 << CONFIG_PRINTK_SAFE_LOG_BUF_SHIFT) - \ 34 sizeof(atomic_t) - \ 35 sizeof(atomic_t) - \ 36 sizeof(struct irq_work)) 37 38 struct printk_safe_seq_buf { 39 atomic_t len; /* length of written data */ 40 atomic_t message_lost; 41 struct irq_work work; /* IRQ work that flushes the buffer */ 42 unsigned char buffer[SAFE_LOG_BUF_LEN]; 43 }; 44 45 static DEFINE_PER_CPU(struct printk_safe_seq_buf, safe_print_seq); 46 static DEFINE_PER_CPU(int, printk_context); 47 48 static DEFINE_RAW_SPINLOCK(safe_read_lock); 49 50 #ifdef CONFIG_PRINTK_NMI 51 static DEFINE_PER_CPU(struct printk_safe_seq_buf, nmi_print_seq); 52 #endif 53 54 /* Get flushed in a more safe context. */ 55 static void queue_flush_work(struct printk_safe_seq_buf *s) 56 { 57 if (printk_percpu_data_ready()) 58 irq_work_queue(&s->work); 59 } 60 61 /* 62 * Add a message to per-CPU context-dependent buffer. NMI and printk-safe 63 * have dedicated buffers, because otherwise printk-safe preempted by 64 * NMI-printk would have overwritten the NMI messages. 65 * 66 * The messages are flushed from irq work (or from panic()), possibly, 67 * from other CPU, concurrently with printk_safe_log_store(). Should this 68 * happen, printk_safe_log_store() will notice the buffer->len mismatch 69 * and repeat the write. 70 */ 71 static __printf(2, 0) int printk_safe_log_store(struct printk_safe_seq_buf *s, 72 const char *fmt, va_list args) 73 { 74 int add; 75 size_t len; 76 va_list ap; 77 78 again: 79 len = atomic_read(&s->len); 80 81 /* The trailing '\0' is not counted into len. */ 82 if (len >= sizeof(s->buffer) - 1) { 83 atomic_inc(&s->message_lost); 84 queue_flush_work(s); 85 return 0; 86 } 87 88 /* 89 * Make sure that all old data have been read before the buffer 90 * was reset. This is not needed when we just append data. 91 */ 92 if (!len) 93 smp_rmb(); 94 95 va_copy(ap, args); 96 add = vscnprintf(s->buffer + len, sizeof(s->buffer) - len, fmt, ap); 97 va_end(ap); 98 if (!add) 99 return 0; 100 101 /* 102 * Do it once again if the buffer has been flushed in the meantime. 103 * Note that atomic_cmpxchg() is an implicit memory barrier that 104 * makes sure that the data were written before updating s->len. 105 */ 106 if (atomic_cmpxchg(&s->len, len, len + add) != len) 107 goto again; 108 109 queue_flush_work(s); 110 return add; 111 } 112 113 static inline void printk_safe_flush_line(const char *text, int len) 114 { 115 /* 116 * Avoid any console drivers calls from here, because we may be 117 * in NMI or printk_safe context (when in panic). The messages 118 * must go only into the ring buffer at this stage. Consoles will 119 * get explicitly called later when a crashdump is not generated. 120 */ 121 printk_deferred("%.*s", len, text); 122 } 123 124 /* printk part of the temporary buffer line by line */ 125 static int printk_safe_flush_buffer(const char *start, size_t len) 126 { 127 const char *c, *end; 128 bool header; 129 130 c = start; 131 end = start + len; 132 header = true; 133 134 /* Print line by line. */ 135 while (c < end) { 136 if (*c == '\n') { 137 printk_safe_flush_line(start, c - start + 1); 138 start = ++c; 139 header = true; 140 continue; 141 } 142 143 /* Handle continuous lines or missing new line. */ 144 if ((c + 1 < end) && printk_get_level(c)) { 145 if (header) { 146 c = printk_skip_level(c); 147 continue; 148 } 149 150 printk_safe_flush_line(start, c - start); 151 start = c++; 152 header = true; 153 continue; 154 } 155 156 header = false; 157 c++; 158 } 159 160 /* Check if there was a partial line. Ignore pure header. */ 161 if (start < end && !header) { 162 static const char newline[] = KERN_CONT "\n"; 163 164 printk_safe_flush_line(start, end - start); 165 printk_safe_flush_line(newline, strlen(newline)); 166 } 167 168 return len; 169 } 170 171 static void report_message_lost(struct printk_safe_seq_buf *s) 172 { 173 int lost = atomic_xchg(&s->message_lost, 0); 174 175 if (lost) 176 printk_deferred("Lost %d message(s)!\n", lost); 177 } 178 179 /* 180 * Flush data from the associated per-CPU buffer. The function 181 * can be called either via IRQ work or independently. 182 */ 183 static void __printk_safe_flush(struct irq_work *work) 184 { 185 struct printk_safe_seq_buf *s = 186 container_of(work, struct printk_safe_seq_buf, work); 187 unsigned long flags; 188 size_t len; 189 int i; 190 191 /* 192 * The lock has two functions. First, one reader has to flush all 193 * available message to make the lockless synchronization with 194 * writers easier. Second, we do not want to mix messages from 195 * different CPUs. This is especially important when printing 196 * a backtrace. 197 */ 198 raw_spin_lock_irqsave(&safe_read_lock, flags); 199 200 i = 0; 201 more: 202 len = atomic_read(&s->len); 203 204 /* 205 * This is just a paranoid check that nobody has manipulated 206 * the buffer an unexpected way. If we printed something then 207 * @len must only increase. Also it should never overflow the 208 * buffer size. 209 */ 210 if ((i && i >= len) || len > sizeof(s->buffer)) { 211 const char *msg = "printk_safe_flush: internal error\n"; 212 213 printk_safe_flush_line(msg, strlen(msg)); 214 len = 0; 215 } 216 217 if (!len) 218 goto out; /* Someone else has already flushed the buffer. */ 219 220 /* Make sure that data has been written up to the @len */ 221 smp_rmb(); 222 i += printk_safe_flush_buffer(s->buffer + i, len - i); 223 224 /* 225 * Check that nothing has got added in the meantime and truncate 226 * the buffer. Note that atomic_cmpxchg() is an implicit memory 227 * barrier that makes sure that the data were copied before 228 * updating s->len. 229 */ 230 if (atomic_cmpxchg(&s->len, len, 0) != len) 231 goto more; 232 233 out: 234 report_message_lost(s); 235 raw_spin_unlock_irqrestore(&safe_read_lock, flags); 236 } 237 238 /** 239 * printk_safe_flush - flush all per-cpu nmi buffers. 240 * 241 * The buffers are flushed automatically via IRQ work. This function 242 * is useful only when someone wants to be sure that all buffers have 243 * been flushed at some point. 244 */ 245 void printk_safe_flush(void) 246 { 247 int cpu; 248 249 for_each_possible_cpu(cpu) { 250 #ifdef CONFIG_PRINTK_NMI 251 __printk_safe_flush(&per_cpu(nmi_print_seq, cpu).work); 252 #endif 253 __printk_safe_flush(&per_cpu(safe_print_seq, cpu).work); 254 } 255 } 256 257 /** 258 * printk_safe_flush_on_panic - flush all per-cpu nmi buffers when the system 259 * goes down. 260 * 261 * Similar to printk_safe_flush() but it can be called even in NMI context when 262 * the system goes down. It does the best effort to get NMI messages into 263 * the main ring buffer. 264 * 265 * Note that it could try harder when there is only one CPU online. 266 */ 267 void printk_safe_flush_on_panic(void) 268 { 269 /* 270 * Make sure that we could access the main ring buffer. 271 * Do not risk a double release when more CPUs are up. 272 */ 273 if (raw_spin_is_locked(&logbuf_lock)) { 274 if (num_online_cpus() > 1) 275 return; 276 277 debug_locks_off(); 278 raw_spin_lock_init(&logbuf_lock); 279 } 280 281 if (raw_spin_is_locked(&safe_read_lock)) { 282 if (num_online_cpus() > 1) 283 return; 284 285 debug_locks_off(); 286 raw_spin_lock_init(&safe_read_lock); 287 } 288 289 printk_safe_flush(); 290 } 291 292 #ifdef CONFIG_PRINTK_NMI 293 /* 294 * Safe printk() for NMI context. It uses a per-CPU buffer to 295 * store the message. NMIs are not nested, so there is always only 296 * one writer running. But the buffer might get flushed from another 297 * CPU, so we need to be careful. 298 */ 299 static __printf(1, 0) int vprintk_nmi(const char *fmt, va_list args) 300 { 301 struct printk_safe_seq_buf *s = this_cpu_ptr(&nmi_print_seq); 302 303 return printk_safe_log_store(s, fmt, args); 304 } 305 306 void noinstr printk_nmi_enter(void) 307 { 308 this_cpu_add(printk_context, PRINTK_NMI_CONTEXT_OFFSET); 309 } 310 311 void noinstr printk_nmi_exit(void) 312 { 313 this_cpu_sub(printk_context, PRINTK_NMI_CONTEXT_OFFSET); 314 } 315 316 /* 317 * Marks a code that might produce many messages in NMI context 318 * and the risk of losing them is more critical than eventual 319 * reordering. 320 * 321 * It has effect only when called in NMI context. Then printk() 322 * will try to store the messages into the main logbuf directly 323 * and use the per-CPU buffers only as a fallback when the lock 324 * is not available. 325 */ 326 void printk_nmi_direct_enter(void) 327 { 328 if (this_cpu_read(printk_context) & PRINTK_NMI_CONTEXT_MASK) 329 this_cpu_or(printk_context, PRINTK_NMI_DIRECT_CONTEXT_MASK); 330 } 331 332 void printk_nmi_direct_exit(void) 333 { 334 this_cpu_and(printk_context, ~PRINTK_NMI_DIRECT_CONTEXT_MASK); 335 } 336 337 #else 338 339 static __printf(1, 0) int vprintk_nmi(const char *fmt, va_list args) 340 { 341 return 0; 342 } 343 344 #endif /* CONFIG_PRINTK_NMI */ 345 346 /* 347 * Lock-less printk(), to avoid deadlocks should the printk() recurse 348 * into itself. It uses a per-CPU buffer to store the message, just like 349 * NMI. 350 */ 351 static __printf(1, 0) int vprintk_safe(const char *fmt, va_list args) 352 { 353 struct printk_safe_seq_buf *s = this_cpu_ptr(&safe_print_seq); 354 355 return printk_safe_log_store(s, fmt, args); 356 } 357 358 /* Can be preempted by NMI. */ 359 void __printk_safe_enter(void) 360 { 361 this_cpu_inc(printk_context); 362 } 363 364 /* Can be preempted by NMI. */ 365 void __printk_safe_exit(void) 366 { 367 this_cpu_dec(printk_context); 368 } 369 370 __printf(1, 0) int vprintk_func(const char *fmt, va_list args) 371 { 372 #ifdef CONFIG_KGDB_KDB 373 /* Allow to pass printk() to kdb but avoid a recursion. */ 374 if (unlikely(kdb_trap_printk && kdb_printf_cpu < 0)) 375 return vkdb_printf(KDB_MSGSRC_PRINTK, fmt, args); 376 #endif 377 378 /* 379 * Try to use the main logbuf even in NMI. But avoid calling console 380 * drivers that might have their own locks. 381 */ 382 if ((this_cpu_read(printk_context) & PRINTK_NMI_DIRECT_CONTEXT_MASK) && 383 raw_spin_trylock(&logbuf_lock)) { 384 int len; 385 386 len = vprintk_store(0, LOGLEVEL_DEFAULT, NULL, fmt, args); 387 raw_spin_unlock(&logbuf_lock); 388 defer_console_output(); 389 return len; 390 } 391 392 /* Use extra buffer in NMI when logbuf_lock is taken or in safe mode. */ 393 if (this_cpu_read(printk_context) & PRINTK_NMI_CONTEXT_MASK) 394 return vprintk_nmi(fmt, args); 395 396 /* Use extra buffer to prevent a recursion deadlock in safe mode. */ 397 if (this_cpu_read(printk_context) & PRINTK_SAFE_CONTEXT_MASK) 398 return vprintk_safe(fmt, args); 399 400 /* No obstacles. */ 401 return vprintk_default(fmt, args); 402 } 403 404 void __init printk_safe_init(void) 405 { 406 int cpu; 407 408 for_each_possible_cpu(cpu) { 409 struct printk_safe_seq_buf *s; 410 411 s = &per_cpu(safe_print_seq, cpu); 412 init_irq_work(&s->work, __printk_safe_flush); 413 414 #ifdef CONFIG_PRINTK_NMI 415 s = &per_cpu(nmi_print_seq, cpu); 416 init_irq_work(&s->work, __printk_safe_flush); 417 #endif 418 } 419 420 /* Flush pending messages that did not have scheduled IRQ works. */ 421 printk_safe_flush(); 422 } 423