xref: /openbmc/linux/kernel/printk/printk_safe.c (revision ed1666f6)
1 /*
2  * printk_safe.c - Safe printk for printk-deadlock-prone contexts
3  *
4  * This program is free software; you can redistribute it and/or
5  * modify it under the terms of the GNU General Public License
6  * as published by the Free Software Foundation; either version 2
7  * of the License, or (at your option) any later version.
8  *
9  * This program is distributed in the hope that it will be useful,
10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
12  * GNU General Public License for more details.
13  *
14  * You should have received a copy of the GNU General Public License
15  * along with this program; if not, see <http://www.gnu.org/licenses/>.
16  */
17 
18 #include <linux/preempt.h>
19 #include <linux/spinlock.h>
20 #include <linux/debug_locks.h>
21 #include <linux/smp.h>
22 #include <linux/cpumask.h>
23 #include <linux/irq_work.h>
24 #include <linux/printk.h>
25 
26 #include "internal.h"
27 
28 /*
29  * printk() could not take logbuf_lock in NMI context. Instead,
30  * it uses an alternative implementation that temporary stores
31  * the strings into a per-CPU buffer. The content of the buffer
32  * is later flushed into the main ring buffer via IRQ work.
33  *
34  * The alternative implementation is chosen transparently
35  * by examinig current printk() context mask stored in @printk_context
36  * per-CPU variable.
37  *
38  * The implementation allows to flush the strings also from another CPU.
39  * There are situations when we want to make sure that all buffers
40  * were handled or when IRQs are blocked.
41  */
42 static int printk_safe_irq_ready __read_mostly;
43 
44 #define SAFE_LOG_BUF_LEN ((1 << CONFIG_PRINTK_SAFE_LOG_BUF_SHIFT) -	\
45 				sizeof(atomic_t) -			\
46 				sizeof(atomic_t) -			\
47 				sizeof(struct irq_work))
48 
49 struct printk_safe_seq_buf {
50 	atomic_t		len;	/* length of written data */
51 	atomic_t		message_lost;
52 	struct irq_work		work;	/* IRQ work that flushes the buffer */
53 	unsigned char		buffer[SAFE_LOG_BUF_LEN];
54 };
55 
56 static DEFINE_PER_CPU(struct printk_safe_seq_buf, safe_print_seq);
57 static DEFINE_PER_CPU(int, printk_context);
58 
59 #ifdef CONFIG_PRINTK_NMI
60 static DEFINE_PER_CPU(struct printk_safe_seq_buf, nmi_print_seq);
61 #endif
62 
63 /* Get flushed in a more safe context. */
64 static void queue_flush_work(struct printk_safe_seq_buf *s)
65 {
66 	if (printk_safe_irq_ready)
67 		irq_work_queue(&s->work);
68 }
69 
70 /*
71  * Add a message to per-CPU context-dependent buffer. NMI and printk-safe
72  * have dedicated buffers, because otherwise printk-safe preempted by
73  * NMI-printk would have overwritten the NMI messages.
74  *
75  * The messages are flushed from irq work (or from panic()), possibly,
76  * from other CPU, concurrently with printk_safe_log_store(). Should this
77  * happen, printk_safe_log_store() will notice the buffer->len mismatch
78  * and repeat the write.
79  */
80 static __printf(2, 0) int printk_safe_log_store(struct printk_safe_seq_buf *s,
81 						const char *fmt, va_list args)
82 {
83 	int add;
84 	size_t len;
85 	va_list ap;
86 
87 again:
88 	len = atomic_read(&s->len);
89 
90 	/* The trailing '\0' is not counted into len. */
91 	if (len >= sizeof(s->buffer) - 1) {
92 		atomic_inc(&s->message_lost);
93 		queue_flush_work(s);
94 		return 0;
95 	}
96 
97 	/*
98 	 * Make sure that all old data have been read before the buffer
99 	 * was reset. This is not needed when we just append data.
100 	 */
101 	if (!len)
102 		smp_rmb();
103 
104 	va_copy(ap, args);
105 	add = vscnprintf(s->buffer + len, sizeof(s->buffer) - len, fmt, ap);
106 	va_end(ap);
107 	if (!add)
108 		return 0;
109 
110 	/*
111 	 * Do it once again if the buffer has been flushed in the meantime.
112 	 * Note that atomic_cmpxchg() is an implicit memory barrier that
113 	 * makes sure that the data were written before updating s->len.
114 	 */
115 	if (atomic_cmpxchg(&s->len, len, len + add) != len)
116 		goto again;
117 
118 	queue_flush_work(s);
119 	return add;
120 }
121 
122 static inline void printk_safe_flush_line(const char *text, int len)
123 {
124 	/*
125 	 * Avoid any console drivers calls from here, because we may be
126 	 * in NMI or printk_safe context (when in panic). The messages
127 	 * must go only into the ring buffer at this stage.  Consoles will
128 	 * get explicitly called later when a crashdump is not generated.
129 	 */
130 	printk_deferred("%.*s", len, text);
131 }
132 
133 /* printk part of the temporary buffer line by line */
134 static int printk_safe_flush_buffer(const char *start, size_t len)
135 {
136 	const char *c, *end;
137 	bool header;
138 
139 	c = start;
140 	end = start + len;
141 	header = true;
142 
143 	/* Print line by line. */
144 	while (c < end) {
145 		if (*c == '\n') {
146 			printk_safe_flush_line(start, c - start + 1);
147 			start = ++c;
148 			header = true;
149 			continue;
150 		}
151 
152 		/* Handle continuous lines or missing new line. */
153 		if ((c + 1 < end) && printk_get_level(c)) {
154 			if (header) {
155 				c = printk_skip_level(c);
156 				continue;
157 			}
158 
159 			printk_safe_flush_line(start, c - start);
160 			start = c++;
161 			header = true;
162 			continue;
163 		}
164 
165 		header = false;
166 		c++;
167 	}
168 
169 	/* Check if there was a partial line. Ignore pure header. */
170 	if (start < end && !header) {
171 		static const char newline[] = KERN_CONT "\n";
172 
173 		printk_safe_flush_line(start, end - start);
174 		printk_safe_flush_line(newline, strlen(newline));
175 	}
176 
177 	return len;
178 }
179 
180 static void report_message_lost(struct printk_safe_seq_buf *s)
181 {
182 	int lost = atomic_xchg(&s->message_lost, 0);
183 
184 	if (lost)
185 		printk_deferred("Lost %d message(s)!\n", lost);
186 }
187 
188 /*
189  * Flush data from the associated per-CPU buffer. The function
190  * can be called either via IRQ work or independently.
191  */
192 static void __printk_safe_flush(struct irq_work *work)
193 {
194 	static raw_spinlock_t read_lock =
195 		__RAW_SPIN_LOCK_INITIALIZER(read_lock);
196 	struct printk_safe_seq_buf *s =
197 		container_of(work, struct printk_safe_seq_buf, work);
198 	unsigned long flags;
199 	size_t len;
200 	int i;
201 
202 	/*
203 	 * The lock has two functions. First, one reader has to flush all
204 	 * available message to make the lockless synchronization with
205 	 * writers easier. Second, we do not want to mix messages from
206 	 * different CPUs. This is especially important when printing
207 	 * a backtrace.
208 	 */
209 	raw_spin_lock_irqsave(&read_lock, flags);
210 
211 	i = 0;
212 more:
213 	len = atomic_read(&s->len);
214 
215 	/*
216 	 * This is just a paranoid check that nobody has manipulated
217 	 * the buffer an unexpected way. If we printed something then
218 	 * @len must only increase. Also it should never overflow the
219 	 * buffer size.
220 	 */
221 	if ((i && i >= len) || len > sizeof(s->buffer)) {
222 		const char *msg = "printk_safe_flush: internal error\n";
223 
224 		printk_safe_flush_line(msg, strlen(msg));
225 		len = 0;
226 	}
227 
228 	if (!len)
229 		goto out; /* Someone else has already flushed the buffer. */
230 
231 	/* Make sure that data has been written up to the @len */
232 	smp_rmb();
233 	i += printk_safe_flush_buffer(s->buffer + i, len - i);
234 
235 	/*
236 	 * Check that nothing has got added in the meantime and truncate
237 	 * the buffer. Note that atomic_cmpxchg() is an implicit memory
238 	 * barrier that makes sure that the data were copied before
239 	 * updating s->len.
240 	 */
241 	if (atomic_cmpxchg(&s->len, len, 0) != len)
242 		goto more;
243 
244 out:
245 	report_message_lost(s);
246 	raw_spin_unlock_irqrestore(&read_lock, flags);
247 }
248 
249 /**
250  * printk_safe_flush - flush all per-cpu nmi buffers.
251  *
252  * The buffers are flushed automatically via IRQ work. This function
253  * is useful only when someone wants to be sure that all buffers have
254  * been flushed at some point.
255  */
256 void printk_safe_flush(void)
257 {
258 	int cpu;
259 
260 	for_each_possible_cpu(cpu) {
261 #ifdef CONFIG_PRINTK_NMI
262 		__printk_safe_flush(&per_cpu(nmi_print_seq, cpu).work);
263 #endif
264 		__printk_safe_flush(&per_cpu(safe_print_seq, cpu).work);
265 	}
266 }
267 
268 /**
269  * printk_safe_flush_on_panic - flush all per-cpu nmi buffers when the system
270  *	goes down.
271  *
272  * Similar to printk_safe_flush() but it can be called even in NMI context when
273  * the system goes down. It does the best effort to get NMI messages into
274  * the main ring buffer.
275  *
276  * Note that it could try harder when there is only one CPU online.
277  */
278 void printk_safe_flush_on_panic(void)
279 {
280 	/*
281 	 * Make sure that we could access the main ring buffer.
282 	 * Do not risk a double release when more CPUs are up.
283 	 */
284 	if (raw_spin_is_locked(&logbuf_lock)) {
285 		if (num_online_cpus() > 1)
286 			return;
287 
288 		debug_locks_off();
289 		raw_spin_lock_init(&logbuf_lock);
290 	}
291 
292 	printk_safe_flush();
293 }
294 
295 #ifdef CONFIG_PRINTK_NMI
296 /*
297  * Safe printk() for NMI context. It uses a per-CPU buffer to
298  * store the message. NMIs are not nested, so there is always only
299  * one writer running. But the buffer might get flushed from another
300  * CPU, so we need to be careful.
301  */
302 static __printf(1, 0) int vprintk_nmi(const char *fmt, va_list args)
303 {
304 	struct printk_safe_seq_buf *s = this_cpu_ptr(&nmi_print_seq);
305 
306 	return printk_safe_log_store(s, fmt, args);
307 }
308 
309 void notrace printk_nmi_enter(void)
310 {
311 	this_cpu_or(printk_context, PRINTK_NMI_CONTEXT_MASK);
312 }
313 
314 void notrace printk_nmi_exit(void)
315 {
316 	this_cpu_and(printk_context, ~PRINTK_NMI_CONTEXT_MASK);
317 }
318 
319 /*
320  * Marks a code that might produce many messages in NMI context
321  * and the risk of losing them is more critical than eventual
322  * reordering.
323  *
324  * It has effect only when called in NMI context. Then printk()
325  * will try to store the messages into the main logbuf directly
326  * and use the per-CPU buffers only as a fallback when the lock
327  * is not available.
328  */
329 void printk_nmi_direct_enter(void)
330 {
331 	if (this_cpu_read(printk_context) & PRINTK_NMI_CONTEXT_MASK)
332 		this_cpu_or(printk_context, PRINTK_NMI_DIRECT_CONTEXT_MASK);
333 }
334 
335 void printk_nmi_direct_exit(void)
336 {
337 	this_cpu_and(printk_context, ~PRINTK_NMI_DIRECT_CONTEXT_MASK);
338 }
339 
340 #else
341 
342 static __printf(1, 0) int vprintk_nmi(const char *fmt, va_list args)
343 {
344 	return 0;
345 }
346 
347 #endif /* CONFIG_PRINTK_NMI */
348 
349 /*
350  * Lock-less printk(), to avoid deadlocks should the printk() recurse
351  * into itself. It uses a per-CPU buffer to store the message, just like
352  * NMI.
353  */
354 static __printf(1, 0) int vprintk_safe(const char *fmt, va_list args)
355 {
356 	struct printk_safe_seq_buf *s = this_cpu_ptr(&safe_print_seq);
357 
358 	return printk_safe_log_store(s, fmt, args);
359 }
360 
361 /* Can be preempted by NMI. */
362 void __printk_safe_enter(void)
363 {
364 	this_cpu_inc(printk_context);
365 }
366 
367 /* Can be preempted by NMI. */
368 void __printk_safe_exit(void)
369 {
370 	this_cpu_dec(printk_context);
371 }
372 
373 __printf(1, 0) int vprintk_func(const char *fmt, va_list args)
374 {
375 	/*
376 	 * Try to use the main logbuf even in NMI. But avoid calling console
377 	 * drivers that might have their own locks.
378 	 */
379 	if ((this_cpu_read(printk_context) & PRINTK_NMI_DIRECT_CONTEXT_MASK) &&
380 	    raw_spin_trylock(&logbuf_lock)) {
381 		int len;
382 
383 		len = vprintk_store(0, LOGLEVEL_DEFAULT, NULL, 0, fmt, args);
384 		raw_spin_unlock(&logbuf_lock);
385 		defer_console_output();
386 		return len;
387 	}
388 
389 	/* Use extra buffer in NMI when logbuf_lock is taken or in safe mode. */
390 	if (this_cpu_read(printk_context) & PRINTK_NMI_CONTEXT_MASK)
391 		return vprintk_nmi(fmt, args);
392 
393 	/* Use extra buffer to prevent a recursion deadlock in safe mode. */
394 	if (this_cpu_read(printk_context) & PRINTK_SAFE_CONTEXT_MASK)
395 		return vprintk_safe(fmt, args);
396 
397 	/* No obstacles. */
398 	return vprintk_default(fmt, args);
399 }
400 
401 void __init printk_safe_init(void)
402 {
403 	int cpu;
404 
405 	for_each_possible_cpu(cpu) {
406 		struct printk_safe_seq_buf *s;
407 
408 		s = &per_cpu(safe_print_seq, cpu);
409 		init_irq_work(&s->work, __printk_safe_flush);
410 
411 #ifdef CONFIG_PRINTK_NMI
412 		s = &per_cpu(nmi_print_seq, cpu);
413 		init_irq_work(&s->work, __printk_safe_flush);
414 #endif
415 	}
416 
417 	/*
418 	 * In the highly unlikely event that a NMI were to trigger at
419 	 * this moment. Make sure IRQ work is set up before this
420 	 * variable is set.
421 	 */
422 	barrier();
423 	printk_safe_irq_ready = 1;
424 
425 	/* Flush pending messages that did not have scheduled IRQ works. */
426 	printk_safe_flush();
427 }
428