1 | /* SPDX-License-Identifier: GPL-2.0 */ |
2 | |
3 | #ifndef _KERNEL_PRINTK_RINGBUFFER_H |
4 | #define _KERNEL_PRINTK_RINGBUFFER_H |
5 | |
6 | #include <linux/atomic.h> |
7 | #include <linux/dev_printk.h> |
8 | |
9 | /* |
10 | * Meta information about each stored message. |
11 | * |
12 | * All fields are set by the printk code except for @seq, which is |
13 | * set by the ringbuffer code. |
14 | */ |
15 | struct printk_info { |
16 | u64 seq; /* sequence number */ |
17 | u64 ts_nsec; /* timestamp in nanoseconds */ |
18 | u16 text_len; /* length of text message */ |
19 | u8 facility; /* syslog facility */ |
20 | u8 flags:5; /* internal record flags */ |
21 | u8 level:3; /* syslog level */ |
22 | u32 caller_id; /* thread id or processor id */ |
23 | |
24 | struct dev_printk_info dev_info; |
25 | }; |
26 | |
27 | /* |
28 | * A structure providing the buffers, used by writers and readers. |
29 | * |
30 | * Writers: |
31 | * Using prb_rec_init_wr(), a writer sets @text_buf_size before calling |
32 | * prb_reserve(). On success, prb_reserve() sets @info and @text_buf to |
33 | * buffers reserved for that writer. |
34 | * |
35 | * Readers: |
36 | * Using prb_rec_init_rd(), a reader sets all fields before calling |
37 | * prb_read_valid(). Note that the reader provides the @info and @text_buf, |
38 | * buffers. On success, the struct pointed to by @info will be filled and |
39 | * the char array pointed to by @text_buf will be filled with text data. |
40 | */ |
41 | struct printk_record { |
42 | struct printk_info *info; |
43 | char *text_buf; |
44 | unsigned int text_buf_size; |
45 | }; |
46 | |
47 | /* Specifies the logical position and span of a data block. */ |
48 | struct prb_data_blk_lpos { |
49 | unsigned long begin; |
50 | unsigned long next; |
51 | }; |
52 | |
53 | /* |
54 | * A descriptor: the complete meta-data for a record. |
55 | * |
56 | * @state_var: A bitwise combination of descriptor ID and descriptor state. |
57 | */ |
58 | struct prb_desc { |
59 | atomic_long_t state_var; |
60 | struct prb_data_blk_lpos text_blk_lpos; |
61 | }; |
62 | |
63 | /* A ringbuffer of "ID + data" elements. */ |
64 | struct prb_data_ring { |
65 | unsigned int size_bits; |
66 | char *data; |
67 | atomic_long_t head_lpos; |
68 | atomic_long_t tail_lpos; |
69 | }; |
70 | |
71 | /* A ringbuffer of "struct prb_desc" elements. */ |
72 | struct prb_desc_ring { |
73 | unsigned int count_bits; |
74 | struct prb_desc *descs; |
75 | struct printk_info *infos; |
76 | atomic_long_t head_id; |
77 | atomic_long_t tail_id; |
78 | atomic_long_t last_finalized_seq; |
79 | }; |
80 | |
81 | /* |
82 | * The high level structure representing the printk ringbuffer. |
83 | * |
84 | * @fail: Count of failed prb_reserve() calls where not even a data-less |
85 | * record was created. |
86 | */ |
87 | struct printk_ringbuffer { |
88 | struct prb_desc_ring desc_ring; |
89 | struct prb_data_ring text_data_ring; |
90 | atomic_long_t fail; |
91 | }; |
92 | |
93 | /* |
94 | * Used by writers as a reserve/commit handle. |
95 | * |
96 | * @rb: Ringbuffer where the entry is reserved. |
97 | * @irqflags: Saved irq flags to restore on entry commit. |
98 | * @id: ID of the reserved descriptor. |
99 | * @text_space: Total occupied buffer space in the text data ring, including |
100 | * ID, alignment padding, and wrapping data blocks. |
101 | * |
102 | * This structure is an opaque handle for writers. Its contents are only |
103 | * to be used by the ringbuffer implementation. |
104 | */ |
105 | struct prb_reserved_entry { |
106 | struct printk_ringbuffer *rb; |
107 | unsigned long irqflags; |
108 | unsigned long id; |
109 | unsigned int text_space; |
110 | }; |
111 | |
112 | /* The possible responses of a descriptor state-query. */ |
113 | enum desc_state { |
114 | desc_miss = -1, /* ID mismatch (pseudo state) */ |
115 | desc_reserved = 0x0, /* reserved, in use by writer */ |
116 | desc_committed = 0x1, /* committed by writer, could get reopened */ |
117 | desc_finalized = 0x2, /* committed, no further modification allowed */ |
118 | desc_reusable = 0x3, /* free, not yet used by any writer */ |
119 | }; |
120 | |
121 | #define _DATA_SIZE(sz_bits) (1UL << (sz_bits)) |
122 | #define _DESCS_COUNT(ct_bits) (1U << (ct_bits)) |
123 | #define DESC_SV_BITS (sizeof(unsigned long) * 8) |
124 | #define DESC_FLAGS_SHIFT (DESC_SV_BITS - 2) |
125 | #define DESC_FLAGS_MASK (3UL << DESC_FLAGS_SHIFT) |
126 | #define DESC_STATE(sv) (3UL & (sv >> DESC_FLAGS_SHIFT)) |
127 | #define DESC_SV(id, state) (((unsigned long)state << DESC_FLAGS_SHIFT) | id) |
128 | #define DESC_ID_MASK (~DESC_FLAGS_MASK) |
129 | #define DESC_ID(sv) ((sv) & DESC_ID_MASK) |
130 | |
131 | /* |
132 | * Special data block logical position values (for fields of |
133 | * @prb_desc.text_blk_lpos). |
134 | * |
135 | * - Bit0 is used to identify if the record has no data block. (Implemented in |
136 | * the LPOS_DATALESS() macro.) |
137 | * |
138 | * - Bit1 specifies the reason for not having a data block. |
139 | * |
140 | * These special values could never be real lpos values because of the |
141 | * meta data and alignment padding of data blocks. (See to_blk_size() for |
142 | * details.) |
143 | */ |
144 | #define FAILED_LPOS 0x1 |
145 | #define EMPTY_LINE_LPOS 0x3 |
146 | |
147 | #define FAILED_BLK_LPOS \ |
148 | { \ |
149 | .begin = FAILED_LPOS, \ |
150 | .next = FAILED_LPOS, \ |
151 | } |
152 | |
153 | /* |
154 | * Descriptor Bootstrap |
155 | * |
156 | * The descriptor array is minimally initialized to allow immediate usage |
157 | * by readers and writers. The requirements that the descriptor array |
158 | * initialization must satisfy: |
159 | * |
160 | * Req1 |
161 | * The tail must point to an existing (committed or reusable) descriptor. |
162 | * This is required by the implementation of prb_first_seq(). |
163 | * |
164 | * Req2 |
165 | * Readers must see that the ringbuffer is initially empty. |
166 | * |
167 | * Req3 |
168 | * The first record reserved by a writer is assigned sequence number 0. |
169 | * |
170 | * To satisfy Req1, the tail initially points to a descriptor that is |
171 | * minimally initialized (having no data block, i.e. data-less with the |
172 | * data block's lpos @begin and @next values set to FAILED_LPOS). |
173 | * |
174 | * To satisfy Req2, the initial tail descriptor is initialized to the |
175 | * reusable state. Readers recognize reusable descriptors as existing |
176 | * records, but skip over them. |
177 | * |
178 | * To satisfy Req3, the last descriptor in the array is used as the initial |
179 | * head (and tail) descriptor. This allows the first record reserved by a |
180 | * writer (head + 1) to be the first descriptor in the array. (Only the first |
181 | * descriptor in the array could have a valid sequence number of 0.) |
182 | * |
183 | * The first time a descriptor is reserved, it is assigned a sequence number |
184 | * with the value of the array index. A "first time reserved" descriptor can |
185 | * be recognized because it has a sequence number of 0 but does not have an |
186 | * index of 0. (Only the first descriptor in the array could have a valid |
187 | * sequence number of 0.) After the first reservation, all future reservations |
188 | * (recycling) simply involve incrementing the sequence number by the array |
189 | * count. |
190 | * |
191 | * Hack #1 |
192 | * Only the first descriptor in the array is allowed to have the sequence |
193 | * number 0. In this case it is not possible to recognize if it is being |
194 | * reserved the first time (set to index value) or has been reserved |
195 | * previously (increment by the array count). This is handled by _always_ |
196 | * incrementing the sequence number by the array count when reserving the |
197 | * first descriptor in the array. In order to satisfy Req3, the sequence |
198 | * number of the first descriptor in the array is initialized to minus |
199 | * the array count. Then, upon the first reservation, it is incremented |
200 | * to 0, thus satisfying Req3. |
201 | * |
202 | * Hack #2 |
203 | * prb_first_seq() can be called at any time by readers to retrieve the |
204 | * sequence number of the tail descriptor. However, due to Req2 and Req3, |
205 | * initially there are no records to report the sequence number of |
206 | * (sequence numbers are u64 and there is nothing less than 0). To handle |
207 | * this, the sequence number of the initial tail descriptor is initialized |
208 | * to 0. Technically this is incorrect, because there is no record with |
209 | * sequence number 0 (yet) and the tail descriptor is not the first |
210 | * descriptor in the array. But it allows prb_read_valid() to correctly |
211 | * report the existence of a record for _any_ given sequence number at all |
212 | * times. Bootstrapping is complete when the tail is pushed the first |
213 | * time, thus finally pointing to the first descriptor reserved by a |
214 | * writer, which has the assigned sequence number 0. |
215 | */ |
216 | |
217 | /* |
218 | * Initiating Logical Value Overflows |
219 | * |
220 | * Both logical position (lpos) and ID values can be mapped to array indexes |
221 | * but may experience overflows during the lifetime of the system. To ensure |
222 | * that printk_ringbuffer can handle the overflows for these types, initial |
223 | * values are chosen that map to the correct initial array indexes, but will |
224 | * result in overflows soon. |
225 | * |
226 | * BLK0_LPOS |
227 | * The initial @head_lpos and @tail_lpos for data rings. It is at index |
228 | * 0 and the lpos value is such that it will overflow on the first wrap. |
229 | * |
230 | * DESC0_ID |
231 | * The initial @head_id and @tail_id for the desc ring. It is at the last |
232 | * index of the descriptor array (see Req3 above) and the ID value is such |
233 | * that it will overflow on the second wrap. |
234 | */ |
235 | #define BLK0_LPOS(sz_bits) (-(_DATA_SIZE(sz_bits))) |
236 | #define DESC0_ID(ct_bits) DESC_ID(-(_DESCS_COUNT(ct_bits) + 1)) |
237 | #define DESC0_SV(ct_bits) DESC_SV(DESC0_ID(ct_bits), desc_reusable) |
238 | |
239 | /* |
240 | * Define a ringbuffer with an external text data buffer. The same as |
241 | * DEFINE_PRINTKRB() but requires specifying an external buffer for the |
242 | * text data. |
243 | * |
244 | * Note: The specified external buffer must be of the size: |
245 | * 2 ^ (descbits + avgtextbits) |
246 | */ |
247 | #define _DEFINE_PRINTKRB(name, descbits, avgtextbits, text_buf) \ |
248 | static struct prb_desc _##name##_descs[_DESCS_COUNT(descbits)] = { \ |
249 | /* the initial head and tail */ \ |
250 | [_DESCS_COUNT(descbits) - 1] = { \ |
251 | /* reusable */ \ |
252 | .state_var = ATOMIC_INIT(DESC0_SV(descbits)), \ |
253 | /* no associated data block */ \ |
254 | .text_blk_lpos = FAILED_BLK_LPOS, \ |
255 | }, \ |
256 | }; \ |
257 | static struct printk_info _##name##_infos[_DESCS_COUNT(descbits)] = { \ |
258 | /* this will be the first record reserved by a writer */ \ |
259 | [0] = { \ |
260 | /* will be incremented to 0 on the first reservation */ \ |
261 | .seq = -(u64)_DESCS_COUNT(descbits), \ |
262 | }, \ |
263 | /* the initial head and tail */ \ |
264 | [_DESCS_COUNT(descbits) - 1] = { \ |
265 | /* reports the first seq value during the bootstrap phase */ \ |
266 | .seq = 0, \ |
267 | }, \ |
268 | }; \ |
269 | static struct printk_ringbuffer name = { \ |
270 | .desc_ring = { \ |
271 | .count_bits = descbits, \ |
272 | .descs = &_##name##_descs[0], \ |
273 | .infos = &_##name##_infos[0], \ |
274 | .head_id = ATOMIC_INIT(DESC0_ID(descbits)), \ |
275 | .tail_id = ATOMIC_INIT(DESC0_ID(descbits)), \ |
276 | .last_finalized_seq = ATOMIC_INIT(0), \ |
277 | }, \ |
278 | .text_data_ring = { \ |
279 | .size_bits = (avgtextbits) + (descbits), \ |
280 | .data = text_buf, \ |
281 | .head_lpos = ATOMIC_LONG_INIT(BLK0_LPOS((avgtextbits) + (descbits))), \ |
282 | .tail_lpos = ATOMIC_LONG_INIT(BLK0_LPOS((avgtextbits) + (descbits))), \ |
283 | }, \ |
284 | .fail = ATOMIC_LONG_INIT(0), \ |
285 | } |
286 | |
287 | /** |
288 | * DEFINE_PRINTKRB() - Define a ringbuffer. |
289 | * |
290 | * @name: The name of the ringbuffer variable. |
291 | * @descbits: The number of descriptors as a power-of-2 value. |
292 | * @avgtextbits: The average text data size per record as a power-of-2 value. |
293 | * |
294 | * This is a macro for defining a ringbuffer and all internal structures |
295 | * such that it is ready for immediate use. See _DEFINE_PRINTKRB() for a |
296 | * variant where the text data buffer can be specified externally. |
297 | */ |
298 | #define DEFINE_PRINTKRB(name, descbits, avgtextbits) \ |
299 | static char _##name##_text[1U << ((avgtextbits) + (descbits))] \ |
300 | __aligned(__alignof__(unsigned long)); \ |
301 | _DEFINE_PRINTKRB(name, descbits, avgtextbits, &_##name##_text[0]) |
302 | |
303 | /* Writer Interface */ |
304 | |
305 | /** |
306 | * prb_rec_init_wr() - Initialize a buffer for writing records. |
307 | * |
308 | * @r: The record to initialize. |
309 | * @text_buf_size: The needed text buffer size. |
310 | */ |
311 | static inline void prb_rec_init_wr(struct printk_record *r, |
312 | unsigned int text_buf_size) |
313 | { |
314 | r->info = NULL; |
315 | r->text_buf = NULL; |
316 | r->text_buf_size = text_buf_size; |
317 | } |
318 | |
319 | bool prb_reserve(struct prb_reserved_entry *e, struct printk_ringbuffer *rb, |
320 | struct printk_record *r); |
321 | bool prb_reserve_in_last(struct prb_reserved_entry *e, struct printk_ringbuffer *rb, |
322 | struct printk_record *r, u32 caller_id, unsigned int max_size); |
323 | void prb_commit(struct prb_reserved_entry *e); |
324 | void prb_final_commit(struct prb_reserved_entry *e); |
325 | |
326 | void prb_init(struct printk_ringbuffer *rb, |
327 | char *text_buf, unsigned int text_buf_size, |
328 | struct prb_desc *descs, unsigned int descs_count_bits, |
329 | struct printk_info *infos); |
330 | unsigned int prb_record_text_space(struct prb_reserved_entry *e); |
331 | |
332 | /* Reader Interface */ |
333 | |
334 | /** |
335 | * prb_rec_init_rd() - Initialize a buffer for reading records. |
336 | * |
337 | * @r: The record to initialize. |
338 | * @info: A buffer to store record meta-data. |
339 | * @text_buf: A buffer to store text data. |
340 | * @text_buf_size: The size of @text_buf. |
341 | * |
342 | * Initialize all the fields that a reader is interested in. All arguments |
343 | * (except @r) are optional. Only record data for arguments that are |
344 | * non-NULL or non-zero will be read. |
345 | */ |
346 | static inline void prb_rec_init_rd(struct printk_record *r, |
347 | struct printk_info *info, |
348 | char *text_buf, unsigned int text_buf_size) |
349 | { |
350 | r->info = info; |
351 | r->text_buf = text_buf; |
352 | r->text_buf_size = text_buf_size; |
353 | } |
354 | |
355 | /** |
356 | * prb_for_each_record() - Iterate over the records of a ringbuffer. |
357 | * |
358 | * @from: The sequence number to begin with. |
359 | * @rb: The ringbuffer to iterate over. |
360 | * @s: A u64 to store the sequence number on each iteration. |
361 | * @r: A printk_record to store the record on each iteration. |
362 | * |
363 | * This is a macro for conveniently iterating over a ringbuffer. |
364 | * Note that @s may not be the sequence number of the record on each |
365 | * iteration. For the sequence number, @r->info->seq should be checked. |
366 | * |
367 | * Context: Any context. |
368 | */ |
369 | #define prb_for_each_record(from, rb, s, r) \ |
370 | for ((s) = from; prb_read_valid(rb, s, r); (s) = (r)->info->seq + 1) |
371 | |
372 | /** |
373 | * prb_for_each_info() - Iterate over the meta data of a ringbuffer. |
374 | * |
375 | * @from: The sequence number to begin with. |
376 | * @rb: The ringbuffer to iterate over. |
377 | * @s: A u64 to store the sequence number on each iteration. |
378 | * @i: A printk_info to store the record meta data on each iteration. |
379 | * @lc: An unsigned int to store the text line count of each record. |
380 | * |
381 | * This is a macro for conveniently iterating over a ringbuffer. |
382 | * Note that @s may not be the sequence number of the record on each |
383 | * iteration. For the sequence number, @r->info->seq should be checked. |
384 | * |
385 | * Context: Any context. |
386 | */ |
387 | #define prb_for_each_info(from, rb, s, i, lc) \ |
388 | for ((s) = from; prb_read_valid_info(rb, s, i, lc); (s) = (i)->seq + 1) |
389 | |
390 | bool prb_read_valid(struct printk_ringbuffer *rb, u64 seq, |
391 | struct printk_record *r); |
392 | bool prb_read_valid_info(struct printk_ringbuffer *rb, u64 seq, |
393 | struct printk_info *info, unsigned int *line_count); |
394 | |
395 | u64 prb_first_seq(struct printk_ringbuffer *rb); |
396 | u64 prb_first_valid_seq(struct printk_ringbuffer *rb); |
397 | u64 prb_next_seq(struct printk_ringbuffer *rb); |
398 | u64 prb_next_reserve_seq(struct printk_ringbuffer *rb); |
399 | |
400 | #ifdef CONFIG_64BIT |
401 | |
402 | #define __u64seq_to_ulseq(u64seq) (u64seq) |
403 | #define __ulseq_to_u64seq(rb, ulseq) (ulseq) |
404 | |
405 | #else /* CONFIG_64BIT */ |
406 | |
407 | #define __u64seq_to_ulseq(u64seq) ((u32)u64seq) |
408 | |
409 | static inline u64 __ulseq_to_u64seq(struct printk_ringbuffer *rb, u32 ulseq) |
410 | { |
411 | u64 rb_first_seq = prb_first_seq(rb); |
412 | u64 seq; |
413 | |
414 | /* |
415 | * The provided sequence is only the lower 32 bits of the ringbuffer |
416 | * sequence. It needs to be expanded to 64bit. Get the first sequence |
417 | * number from the ringbuffer and fold it. |
418 | * |
419 | * Having a 32bit representation in the console is sufficient. |
420 | * If a console ever gets more than 2^31 records behind |
421 | * the ringbuffer then this is the least of the problems. |
422 | * |
423 | * Also the access to the ring buffer is always safe. |
424 | */ |
425 | seq = rb_first_seq - (s32)((u32)rb_first_seq - ulseq); |
426 | |
427 | return seq; |
428 | } |
429 | |
430 | #endif /* CONFIG_64BIT */ |
431 | |
432 | #endif /* _KERNEL_PRINTK_RINGBUFFER_H */ |
433 | |