xref: /openbmc/linux/kernel/printk/printk_safe.c (revision 6cd70754)
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