Commit | Line | Data |
---|---|---|
a4bd217b JG |
1 | /* |
2 | * Copyright (C) 2016 CNEX Labs | |
3 | * Initial release: Javier Gonzalez <javier@cnexlabs.com> | |
4 | * | |
5 | * Based upon the circular ringbuffer. | |
6 | * | |
7 | * This program is free software; you can redistribute it and/or | |
8 | * modify it under the terms of the GNU General Public License version | |
9 | * 2 as published by the Free Software Foundation. | |
10 | * | |
11 | * This program is distributed in the hope that it will be useful, but | |
12 | * WITHOUT ANY WARRANTY; without even the implied warranty of | |
13 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | |
14 | * General Public License for more details. | |
15 | * | |
16 | * pblk-rb.c - pblk's write buffer | |
17 | */ | |
18 | ||
19 | #include <linux/circ_buf.h> | |
20 | ||
21 | #include "pblk.h" | |
22 | ||
23 | static DECLARE_RWSEM(pblk_rb_lock); | |
24 | ||
25 | void pblk_rb_data_free(struct pblk_rb *rb) | |
26 | { | |
27 | struct pblk_rb_pages *p, *t; | |
28 | ||
29 | down_write(&pblk_rb_lock); | |
30 | list_for_each_entry_safe(p, t, &rb->pages, list) { | |
31 | free_pages((unsigned long)page_address(p->pages), p->order); | |
32 | list_del(&p->list); | |
33 | kfree(p); | |
34 | } | |
35 | up_write(&pblk_rb_lock); | |
36 | } | |
37 | ||
38 | /* | |
39 | * Initialize ring buffer. The data and metadata buffers must be previously | |
40 | * allocated and their size must be a power of two | |
41 | * (Documentation/circular-buffers.txt) | |
42 | */ | |
43 | int pblk_rb_init(struct pblk_rb *rb, struct pblk_rb_entry *rb_entry_base, | |
44 | unsigned int power_size, unsigned int power_seg_sz) | |
45 | { | |
46 | struct pblk *pblk = container_of(rb, struct pblk, rwb); | |
47 | unsigned int init_entry = 0; | |
48 | unsigned int alloc_order = power_size; | |
49 | unsigned int max_order = MAX_ORDER - 1; | |
50 | unsigned int order, iter; | |
51 | ||
52 | down_write(&pblk_rb_lock); | |
53 | rb->entries = rb_entry_base; | |
54 | rb->seg_size = (1 << power_seg_sz); | |
55 | rb->nr_entries = (1 << power_size); | |
56 | rb->mem = rb->subm = rb->sync = rb->l2p_update = 0; | |
8154d296 | 57 | rb->flush_point = EMPTY_ENTRY; |
a4bd217b JG |
58 | |
59 | spin_lock_init(&rb->w_lock); | |
60 | spin_lock_init(&rb->s_lock); | |
61 | ||
62 | INIT_LIST_HEAD(&rb->pages); | |
63 | ||
64 | if (alloc_order >= max_order) { | |
65 | order = max_order; | |
66 | iter = (1 << (alloc_order - max_order)); | |
67 | } else { | |
68 | order = alloc_order; | |
69 | iter = 1; | |
70 | } | |
71 | ||
72 | do { | |
73 | struct pblk_rb_entry *entry; | |
74 | struct pblk_rb_pages *page_set; | |
75 | void *kaddr; | |
76 | unsigned long set_size; | |
77 | int i; | |
78 | ||
79 | page_set = kmalloc(sizeof(struct pblk_rb_pages), GFP_KERNEL); | |
80 | if (!page_set) { | |
81 | up_write(&pblk_rb_lock); | |
82 | return -ENOMEM; | |
83 | } | |
84 | ||
85 | page_set->order = order; | |
86 | page_set->pages = alloc_pages(GFP_KERNEL, order); | |
87 | if (!page_set->pages) { | |
88 | kfree(page_set); | |
89 | pblk_rb_data_free(rb); | |
90 | up_write(&pblk_rb_lock); | |
91 | return -ENOMEM; | |
92 | } | |
93 | kaddr = page_address(page_set->pages); | |
94 | ||
95 | entry = &rb->entries[init_entry]; | |
96 | entry->data = kaddr; | |
97 | entry->cacheline = pblk_cacheline_to_addr(init_entry++); | |
98 | entry->w_ctx.flags = PBLK_WRITABLE_ENTRY; | |
99 | ||
100 | set_size = (1 << order); | |
101 | for (i = 1; i < set_size; i++) { | |
102 | entry = &rb->entries[init_entry]; | |
103 | entry->cacheline = pblk_cacheline_to_addr(init_entry++); | |
104 | entry->data = kaddr + (i * rb->seg_size); | |
105 | entry->w_ctx.flags = PBLK_WRITABLE_ENTRY; | |
106 | bio_list_init(&entry->w_ctx.bios); | |
107 | } | |
108 | ||
109 | list_add_tail(&page_set->list, &rb->pages); | |
110 | iter--; | |
111 | } while (iter > 0); | |
112 | up_write(&pblk_rb_lock); | |
113 | ||
114 | #ifdef CONFIG_NVM_DEBUG | |
8154d296 | 115 | atomic_set(&rb->inflight_flush_point, 0); |
a4bd217b JG |
116 | #endif |
117 | ||
118 | /* | |
119 | * Initialize rate-limiter, which controls access to the write buffer | |
120 | * but user and GC I/O | |
121 | */ | |
122 | pblk_rl_init(&pblk->rl, rb->nr_entries); | |
123 | ||
124 | return 0; | |
125 | } | |
126 | ||
127 | /* | |
128 | * pblk_rb_calculate_size -- calculate the size of the write buffer | |
129 | */ | |
130 | unsigned int pblk_rb_calculate_size(unsigned int nr_entries) | |
131 | { | |
132 | /* Alloc a write buffer that can at least fit 128 entries */ | |
133 | return (1 << max(get_count_order(nr_entries), 7)); | |
134 | } | |
135 | ||
136 | void *pblk_rb_entries_ref(struct pblk_rb *rb) | |
137 | { | |
138 | return rb->entries; | |
139 | } | |
140 | ||
141 | static void clean_wctx(struct pblk_w_ctx *w_ctx) | |
142 | { | |
143 | int flags; | |
144 | ||
145 | try: | |
146 | flags = READ_ONCE(w_ctx->flags); | |
147 | if (!(flags & PBLK_SUBMITTED_ENTRY)) | |
148 | goto try; | |
149 | ||
150 | /* Release flags on context. Protect from writes and reads */ | |
151 | smp_store_release(&w_ctx->flags, PBLK_WRITABLE_ENTRY); | |
152 | pblk_ppa_set_empty(&w_ctx->ppa); | |
07698466 | 153 | w_ctx->lba = ADDR_EMPTY; |
a4bd217b JG |
154 | } |
155 | ||
156 | #define pblk_rb_ring_count(head, tail, size) CIRC_CNT(head, tail, size) | |
157 | #define pblk_rb_ring_space(rb, head, tail, size) \ | |
158 | (CIRC_SPACE(head, tail, size)) | |
159 | ||
160 | /* | |
161 | * Buffer space is calculated with respect to the back pointer signaling | |
162 | * synchronized entries to the media. | |
163 | */ | |
164 | static unsigned int pblk_rb_space(struct pblk_rb *rb) | |
165 | { | |
166 | unsigned int mem = READ_ONCE(rb->mem); | |
167 | unsigned int sync = READ_ONCE(rb->sync); | |
168 | ||
169 | return pblk_rb_ring_space(rb, mem, sync, rb->nr_entries); | |
170 | } | |
171 | ||
172 | /* | |
173 | * Buffer count is calculated with respect to the submission entry signaling the | |
174 | * entries that are available to send to the media | |
175 | */ | |
176 | unsigned int pblk_rb_read_count(struct pblk_rb *rb) | |
177 | { | |
178 | unsigned int mem = READ_ONCE(rb->mem); | |
179 | unsigned int subm = READ_ONCE(rb->subm); | |
180 | ||
181 | return pblk_rb_ring_count(mem, subm, rb->nr_entries); | |
182 | } | |
183 | ||
ee8d5c1a JG |
184 | unsigned int pblk_rb_sync_count(struct pblk_rb *rb) |
185 | { | |
186 | unsigned int mem = READ_ONCE(rb->mem); | |
187 | unsigned int sync = READ_ONCE(rb->sync); | |
188 | ||
189 | return pblk_rb_ring_count(mem, sync, rb->nr_entries); | |
190 | } | |
191 | ||
a4bd217b JG |
192 | unsigned int pblk_rb_read_commit(struct pblk_rb *rb, unsigned int nr_entries) |
193 | { | |
194 | unsigned int subm; | |
195 | ||
196 | subm = READ_ONCE(rb->subm); | |
197 | /* Commit read means updating submission pointer */ | |
198 | smp_store_release(&rb->subm, | |
199 | (subm + nr_entries) & (rb->nr_entries - 1)); | |
200 | ||
201 | return subm; | |
202 | } | |
203 | ||
05ed3447 | 204 | static int __pblk_rb_update_l2p(struct pblk_rb *rb, unsigned int to_update) |
a4bd217b JG |
205 | { |
206 | struct pblk *pblk = container_of(rb, struct pblk, rwb); | |
207 | struct pblk_line *line; | |
208 | struct pblk_rb_entry *entry; | |
209 | struct pblk_w_ctx *w_ctx; | |
b20ba1bc | 210 | unsigned int user_io = 0, gc_io = 0; |
a4bd217b | 211 | unsigned int i; |
b20ba1bc | 212 | int flags; |
a4bd217b JG |
213 | |
214 | for (i = 0; i < to_update; i++) { | |
05ed3447 | 215 | entry = &rb->entries[rb->l2p_update]; |
a4bd217b JG |
216 | w_ctx = &entry->w_ctx; |
217 | ||
b20ba1bc JG |
218 | flags = READ_ONCE(entry->w_ctx.flags); |
219 | if (flags & PBLK_IOTYPE_USER) | |
220 | user_io++; | |
221 | else if (flags & PBLK_IOTYPE_GC) | |
222 | gc_io++; | |
223 | else | |
224 | WARN(1, "pblk: unknown IO type\n"); | |
225 | ||
a4bd217b JG |
226 | pblk_update_map_dev(pblk, w_ctx->lba, w_ctx->ppa, |
227 | entry->cacheline); | |
228 | ||
b1bcfda1 | 229 | line = &pblk->lines[pblk_ppa_to_line(w_ctx->ppa)]; |
a4bd217b JG |
230 | kref_put(&line->ref, pblk_line_put); |
231 | clean_wctx(w_ctx); | |
05ed3447 | 232 | rb->l2p_update = (rb->l2p_update + 1) & (rb->nr_entries - 1); |
a4bd217b JG |
233 | } |
234 | ||
b20ba1bc JG |
235 | pblk_rl_out(&pblk->rl, user_io, gc_io); |
236 | ||
a4bd217b JG |
237 | return 0; |
238 | } | |
239 | ||
240 | /* | |
241 | * When we move the l2p_update pointer, we update the l2p table - lookups will | |
242 | * point to the physical address instead of to the cacheline in the write buffer | |
243 | * from this moment on. | |
244 | */ | |
245 | static int pblk_rb_update_l2p(struct pblk_rb *rb, unsigned int nr_entries, | |
246 | unsigned int mem, unsigned int sync) | |
247 | { | |
248 | unsigned int space, count; | |
249 | int ret = 0; | |
250 | ||
251 | lockdep_assert_held(&rb->w_lock); | |
252 | ||
253 | /* Update l2p only as buffer entries are being overwritten */ | |
254 | space = pblk_rb_ring_space(rb, mem, rb->l2p_update, rb->nr_entries); | |
255 | if (space > nr_entries) | |
256 | goto out; | |
257 | ||
258 | count = nr_entries - space; | |
259 | /* l2p_update used exclusively under rb->w_lock */ | |
05ed3447 | 260 | ret = __pblk_rb_update_l2p(rb, count); |
a4bd217b JG |
261 | |
262 | out: | |
263 | return ret; | |
264 | } | |
265 | ||
266 | /* | |
267 | * Update the l2p entry for all sectors stored on the write buffer. This means | |
268 | * that all future lookups to the l2p table will point to a device address, not | |
269 | * to the cacheline in the write buffer. | |
270 | */ | |
271 | void pblk_rb_sync_l2p(struct pblk_rb *rb) | |
272 | { | |
273 | unsigned int sync; | |
274 | unsigned int to_update; | |
275 | ||
276 | spin_lock(&rb->w_lock); | |
277 | ||
278 | /* Protect from reads and writes */ | |
279 | sync = smp_load_acquire(&rb->sync); | |
280 | ||
281 | to_update = pblk_rb_ring_count(sync, rb->l2p_update, rb->nr_entries); | |
05ed3447 | 282 | __pblk_rb_update_l2p(rb, to_update); |
a4bd217b JG |
283 | |
284 | spin_unlock(&rb->w_lock); | |
285 | } | |
286 | ||
287 | /* | |
288 | * Write @nr_entries to ring buffer from @data buffer if there is enough space. | |
289 | * Typically, 4KB data chunks coming from a bio will be copied to the ring | |
290 | * buffer, thus the write will fail if not all incoming data can be copied. | |
291 | * | |
292 | */ | |
293 | static void __pblk_rb_write_entry(struct pblk_rb *rb, void *data, | |
294 | struct pblk_w_ctx w_ctx, | |
295 | struct pblk_rb_entry *entry) | |
296 | { | |
297 | memcpy(entry->data, data, rb->seg_size); | |
298 | ||
299 | entry->w_ctx.lba = w_ctx.lba; | |
300 | entry->w_ctx.ppa = w_ctx.ppa; | |
301 | } | |
302 | ||
303 | void pblk_rb_write_entry_user(struct pblk_rb *rb, void *data, | |
304 | struct pblk_w_ctx w_ctx, unsigned int ring_pos) | |
305 | { | |
306 | struct pblk *pblk = container_of(rb, struct pblk, rwb); | |
307 | struct pblk_rb_entry *entry; | |
308 | int flags; | |
309 | ||
310 | entry = &rb->entries[ring_pos]; | |
311 | flags = READ_ONCE(entry->w_ctx.flags); | |
312 | #ifdef CONFIG_NVM_DEBUG | |
313 | /* Caller must guarantee that the entry is free */ | |
314 | BUG_ON(!(flags & PBLK_WRITABLE_ENTRY)); | |
315 | #endif | |
316 | ||
317 | __pblk_rb_write_entry(rb, data, w_ctx, entry); | |
318 | ||
319 | pblk_update_map_cache(pblk, w_ctx.lba, entry->cacheline); | |
320 | flags = w_ctx.flags | PBLK_WRITTEN_DATA; | |
321 | ||
322 | /* Release flags on write context. Protect from writes */ | |
323 | smp_store_release(&entry->w_ctx.flags, flags); | |
324 | } | |
325 | ||
326 | void pblk_rb_write_entry_gc(struct pblk_rb *rb, void *data, | |
d340121e JG |
327 | struct pblk_w_ctx w_ctx, struct pblk_line *line, |
328 | u64 paddr, unsigned int ring_pos) | |
a4bd217b JG |
329 | { |
330 | struct pblk *pblk = container_of(rb, struct pblk, rwb); | |
331 | struct pblk_rb_entry *entry; | |
332 | int flags; | |
333 | ||
334 | entry = &rb->entries[ring_pos]; | |
335 | flags = READ_ONCE(entry->w_ctx.flags); | |
336 | #ifdef CONFIG_NVM_DEBUG | |
337 | /* Caller must guarantee that the entry is free */ | |
338 | BUG_ON(!(flags & PBLK_WRITABLE_ENTRY)); | |
339 | #endif | |
340 | ||
341 | __pblk_rb_write_entry(rb, data, w_ctx, entry); | |
342 | ||
d340121e | 343 | if (!pblk_update_map_gc(pblk, w_ctx.lba, entry->cacheline, line, paddr)) |
a4bd217b JG |
344 | entry->w_ctx.lba = ADDR_EMPTY; |
345 | ||
346 | flags = w_ctx.flags | PBLK_WRITTEN_DATA; | |
347 | ||
348 | /* Release flags on write context. Protect from writes */ | |
349 | smp_store_release(&entry->w_ctx.flags, flags); | |
350 | } | |
351 | ||
8154d296 | 352 | static int pblk_rb_flush_point_set(struct pblk_rb *rb, struct bio *bio, |
a4bd217b JG |
353 | unsigned int pos) |
354 | { | |
355 | struct pblk_rb_entry *entry; | |
533657c1 | 356 | unsigned int sync, flush_point; |
a4bd217b | 357 | |
533657c1 HH |
358 | sync = READ_ONCE(rb->sync); |
359 | ||
360 | if (pos == sync) | |
361 | return 0; | |
a4bd217b JG |
362 | |
363 | #ifdef CONFIG_NVM_DEBUG | |
8154d296 | 364 | atomic_inc(&rb->inflight_flush_point); |
a4bd217b JG |
365 | #endif |
366 | ||
8154d296 HH |
367 | flush_point = (pos == 0) ? (rb->nr_entries - 1) : (pos - 1); |
368 | entry = &rb->entries[flush_point]; | |
a4bd217b | 369 | |
b36bbf9d HH |
370 | pblk_rb_sync_init(rb, NULL); |
371 | ||
8154d296 HH |
372 | /* Protect flush points */ |
373 | smp_store_release(&rb->flush_point, flush_point); | |
a4bd217b | 374 | |
b36bbf9d HH |
375 | if (bio) |
376 | bio_list_add(&entry->w_ctx.bios, bio); | |
588726d3 | 377 | |
b36bbf9d | 378 | pblk_rb_sync_end(rb, NULL); |
a4bd217b | 379 | |
b36bbf9d | 380 | return bio ? 1 : 0; |
a4bd217b JG |
381 | } |
382 | ||
383 | static int __pblk_rb_may_write(struct pblk_rb *rb, unsigned int nr_entries, | |
384 | unsigned int *pos) | |
385 | { | |
386 | unsigned int mem; | |
387 | unsigned int sync; | |
388 | ||
389 | sync = READ_ONCE(rb->sync); | |
390 | mem = READ_ONCE(rb->mem); | |
391 | ||
392 | if (pblk_rb_ring_space(rb, mem, sync, rb->nr_entries) < nr_entries) | |
393 | return 0; | |
394 | ||
395 | if (pblk_rb_update_l2p(rb, nr_entries, mem, sync)) | |
396 | return 0; | |
397 | ||
398 | *pos = mem; | |
399 | ||
400 | return 1; | |
401 | } | |
402 | ||
403 | static int pblk_rb_may_write(struct pblk_rb *rb, unsigned int nr_entries, | |
404 | unsigned int *pos) | |
405 | { | |
406 | if (!__pblk_rb_may_write(rb, nr_entries, pos)) | |
407 | return 0; | |
408 | ||
409 | /* Protect from read count */ | |
410 | smp_store_release(&rb->mem, (*pos + nr_entries) & (rb->nr_entries - 1)); | |
411 | return 1; | |
412 | } | |
413 | ||
588726d3 JG |
414 | void pblk_rb_flush(struct pblk_rb *rb) |
415 | { | |
416 | struct pblk *pblk = container_of(rb, struct pblk, rwb); | |
417 | unsigned int mem = READ_ONCE(rb->mem); | |
418 | ||
8154d296 | 419 | if (pblk_rb_flush_point_set(rb, NULL, mem)) |
588726d3 JG |
420 | return; |
421 | ||
422 | pblk_write_should_kick(pblk); | |
423 | } | |
424 | ||
a4bd217b JG |
425 | static int pblk_rb_may_write_flush(struct pblk_rb *rb, unsigned int nr_entries, |
426 | unsigned int *pos, struct bio *bio, | |
427 | int *io_ret) | |
428 | { | |
429 | unsigned int mem; | |
430 | ||
431 | if (!__pblk_rb_may_write(rb, nr_entries, pos)) | |
432 | return 0; | |
433 | ||
434 | mem = (*pos + nr_entries) & (rb->nr_entries - 1); | |
435 | *io_ret = NVM_IO_DONE; | |
436 | ||
437 | if (bio->bi_opf & REQ_PREFLUSH) { | |
438 | struct pblk *pblk = container_of(rb, struct pblk, rwb); | |
439 | ||
5d149bfa | 440 | atomic64_inc(&pblk->nr_flush); |
8154d296 | 441 | if (pblk_rb_flush_point_set(&pblk->rwb, bio, mem)) |
a4bd217b JG |
442 | *io_ret = NVM_IO_OK; |
443 | } | |
444 | ||
445 | /* Protect from read count */ | |
446 | smp_store_release(&rb->mem, mem); | |
6ca2f71f | 447 | |
a4bd217b JG |
448 | return 1; |
449 | } | |
450 | ||
451 | /* | |
452 | * Atomically check that (i) there is space on the write buffer for the | |
453 | * incoming I/O, and (ii) the current I/O type has enough budget in the write | |
454 | * buffer (rate-limiter). | |
455 | */ | |
456 | int pblk_rb_may_write_user(struct pblk_rb *rb, struct bio *bio, | |
457 | unsigned int nr_entries, unsigned int *pos) | |
458 | { | |
459 | struct pblk *pblk = container_of(rb, struct pblk, rwb); | |
588726d3 | 460 | int io_ret; |
a4bd217b JG |
461 | |
462 | spin_lock(&rb->w_lock); | |
588726d3 JG |
463 | io_ret = pblk_rl_user_may_insert(&pblk->rl, nr_entries); |
464 | if (io_ret) { | |
a4bd217b | 465 | spin_unlock(&rb->w_lock); |
588726d3 | 466 | return io_ret; |
a4bd217b JG |
467 | } |
468 | ||
588726d3 | 469 | if (!pblk_rb_may_write_flush(rb, nr_entries, pos, bio, &io_ret)) { |
a4bd217b JG |
470 | spin_unlock(&rb->w_lock); |
471 | return NVM_IO_REQUEUE; | |
472 | } | |
473 | ||
474 | pblk_rl_user_in(&pblk->rl, nr_entries); | |
475 | spin_unlock(&rb->w_lock); | |
476 | ||
588726d3 | 477 | return io_ret; |
a4bd217b JG |
478 | } |
479 | ||
480 | /* | |
481 | * Look at pblk_rb_may_write_user comment | |
482 | */ | |
483 | int pblk_rb_may_write_gc(struct pblk_rb *rb, unsigned int nr_entries, | |
484 | unsigned int *pos) | |
485 | { | |
486 | struct pblk *pblk = container_of(rb, struct pblk, rwb); | |
487 | ||
488 | spin_lock(&rb->w_lock); | |
489 | if (!pblk_rl_gc_may_insert(&pblk->rl, nr_entries)) { | |
490 | spin_unlock(&rb->w_lock); | |
491 | return 0; | |
492 | } | |
493 | ||
494 | if (!pblk_rb_may_write(rb, nr_entries, pos)) { | |
495 | spin_unlock(&rb->w_lock); | |
496 | return 0; | |
497 | } | |
498 | ||
499 | pblk_rl_gc_in(&pblk->rl, nr_entries); | |
500 | spin_unlock(&rb->w_lock); | |
501 | ||
502 | return 1; | |
503 | } | |
504 | ||
505 | /* | |
506 | * The caller of this function must ensure that the backpointer will not | |
507 | * overwrite the entries passed on the list. | |
508 | */ | |
509 | unsigned int pblk_rb_read_to_bio_list(struct pblk_rb *rb, struct bio *bio, | |
510 | struct list_head *list, | |
511 | unsigned int max) | |
512 | { | |
513 | struct pblk_rb_entry *entry, *tentry; | |
514 | struct page *page; | |
515 | unsigned int read = 0; | |
516 | int ret; | |
517 | ||
518 | list_for_each_entry_safe(entry, tentry, list, index) { | |
519 | if (read > max) { | |
520 | pr_err("pblk: too many entries on list\n"); | |
521 | goto out; | |
522 | } | |
523 | ||
524 | page = virt_to_page(entry->data); | |
525 | if (!page) { | |
526 | pr_err("pblk: could not allocate write bio page\n"); | |
527 | goto out; | |
528 | } | |
529 | ||
530 | ret = bio_add_page(bio, page, rb->seg_size, 0); | |
531 | if (ret != rb->seg_size) { | |
532 | pr_err("pblk: could not add page to write bio\n"); | |
533 | goto out; | |
534 | } | |
535 | ||
536 | list_del(&entry->index); | |
537 | read++; | |
538 | } | |
539 | ||
540 | out: | |
541 | return read; | |
542 | } | |
543 | ||
544 | /* | |
545 | * Read available entries on rb and add them to the given bio. To avoid a memory | |
546 | * copy, a page reference to the write buffer is used to be added to the bio. | |
547 | * | |
548 | * This function is used by the write thread to form the write bio that will | |
549 | * persist data on the write buffer to the media. | |
550 | */ | |
d624f371 | 551 | unsigned int pblk_rb_read_to_bio(struct pblk_rb *rb, struct nvm_rq *rqd, |
875d94f3 JG |
552 | unsigned int pos, unsigned int nr_entries, |
553 | unsigned int count) | |
a4bd217b JG |
554 | { |
555 | struct pblk *pblk = container_of(rb, struct pblk, rwb); | |
d624f371 JG |
556 | struct request_queue *q = pblk->dev->q; |
557 | struct pblk_c_ctx *c_ctx = nvm_rq_to_pdu(rqd); | |
875d94f3 | 558 | struct bio *bio = rqd->bio; |
a4bd217b JG |
559 | struct pblk_rb_entry *entry; |
560 | struct page *page; | |
d624f371 | 561 | unsigned int pad = 0, to_read = nr_entries; |
a4bd217b JG |
562 | unsigned int i; |
563 | int flags; | |
a4bd217b JG |
564 | |
565 | if (count < nr_entries) { | |
566 | pad = nr_entries - count; | |
567 | to_read = count; | |
568 | } | |
569 | ||
570 | c_ctx->sentry = pos; | |
571 | c_ctx->nr_valid = to_read; | |
572 | c_ctx->nr_padded = pad; | |
573 | ||
574 | for (i = 0; i < to_read; i++) { | |
575 | entry = &rb->entries[pos]; | |
576 | ||
577 | /* A write has been allowed into the buffer, but data is still | |
578 | * being copied to it. It is ok to busy wait. | |
579 | */ | |
580 | try: | |
581 | flags = READ_ONCE(entry->w_ctx.flags); | |
10888129 JG |
582 | if (!(flags & PBLK_WRITTEN_DATA)) { |
583 | io_schedule(); | |
a4bd217b | 584 | goto try; |
10888129 | 585 | } |
a4bd217b | 586 | |
a4bd217b JG |
587 | page = virt_to_page(entry->data); |
588 | if (!page) { | |
589 | pr_err("pblk: could not allocate write bio page\n"); | |
590 | flags &= ~PBLK_WRITTEN_DATA; | |
591 | flags |= PBLK_SUBMITTED_ENTRY; | |
592 | /* Release flags on context. Protect from writes */ | |
593 | smp_store_release(&entry->w_ctx.flags, flags); | |
d624f371 | 594 | return NVM_IO_ERR; |
a4bd217b JG |
595 | } |
596 | ||
d624f371 JG |
597 | if (bio_add_pc_page(q, bio, page, rb->seg_size, 0) != |
598 | rb->seg_size) { | |
a4bd217b JG |
599 | pr_err("pblk: could not add page to write bio\n"); |
600 | flags &= ~PBLK_WRITTEN_DATA; | |
601 | flags |= PBLK_SUBMITTED_ENTRY; | |
602 | /* Release flags on context. Protect from writes */ | |
603 | smp_store_release(&entry->w_ctx.flags, flags); | |
d624f371 | 604 | return NVM_IO_ERR; |
a4bd217b JG |
605 | } |
606 | ||
a4bd217b JG |
607 | flags &= ~PBLK_WRITTEN_DATA; |
608 | flags |= PBLK_SUBMITTED_ENTRY; | |
609 | ||
610 | /* Release flags on context. Protect from writes */ | |
611 | smp_store_release(&entry->w_ctx.flags, flags); | |
612 | ||
613 | pos = (pos + 1) & (rb->nr_entries - 1); | |
614 | } | |
615 | ||
d624f371 JG |
616 | if (pad) { |
617 | if (pblk_bio_add_pages(pblk, bio, GFP_KERNEL, pad)) { | |
618 | pr_err("pblk: could not pad page in write bio\n"); | |
619 | return NVM_IO_ERR; | |
620 | } | |
d624f371 | 621 | |
5d149bfa HH |
622 | if (pad < pblk->min_write_pgs) |
623 | atomic64_inc(&pblk->pad_dist[pad - 1]); | |
624 | else | |
625 | pr_warn("pblk: padding more than min. sectors\n"); | |
626 | ||
627 | atomic64_add(pad, &pblk->pad_wa); | |
628 | } | |
76758390 | 629 | |
a4bd217b | 630 | #ifdef CONFIG_NVM_DEBUG |
5d149bfa | 631 | atomic_long_add(pad, &pblk->padded_writes); |
a4bd217b | 632 | #endif |
d624f371 JG |
633 | |
634 | return NVM_IO_OK; | |
a4bd217b JG |
635 | } |
636 | ||
637 | /* | |
638 | * Copy to bio only if the lba matches the one on the given cache entry. | |
639 | * Otherwise, it means that the entry has been overwritten, and the bio should | |
640 | * be directed to disk. | |
641 | */ | |
642 | int pblk_rb_copy_to_bio(struct pblk_rb *rb, struct bio *bio, sector_t lba, | |
75cb8e93 | 643 | struct ppa_addr ppa, int bio_iter, bool advanced_bio) |
a4bd217b | 644 | { |
07698466 | 645 | struct pblk *pblk = container_of(rb, struct pblk, rwb); |
a4bd217b JG |
646 | struct pblk_rb_entry *entry; |
647 | struct pblk_w_ctx *w_ctx; | |
07698466 JG |
648 | struct ppa_addr l2p_ppa; |
649 | u64 pos = pblk_addr_to_cacheline(ppa); | |
a4bd217b JG |
650 | void *data; |
651 | int flags; | |
652 | int ret = 1; | |
653 | ||
a4bd217b JG |
654 | |
655 | #ifdef CONFIG_NVM_DEBUG | |
656 | /* Caller must ensure that the access will not cause an overflow */ | |
657 | BUG_ON(pos >= rb->nr_entries); | |
658 | #endif | |
659 | entry = &rb->entries[pos]; | |
660 | w_ctx = &entry->w_ctx; | |
661 | flags = READ_ONCE(w_ctx->flags); | |
662 | ||
07698466 JG |
663 | spin_lock(&rb->w_lock); |
664 | spin_lock(&pblk->trans_lock); | |
665 | l2p_ppa = pblk_trans_map_get(pblk, lba); | |
666 | spin_unlock(&pblk->trans_lock); | |
667 | ||
a4bd217b | 668 | /* Check if the entry has been overwritten or is scheduled to be */ |
07698466 JG |
669 | if (!pblk_ppa_comp(l2p_ppa, ppa) || w_ctx->lba != lba || |
670 | flags & PBLK_WRITABLE_ENTRY) { | |
a4bd217b JG |
671 | ret = 0; |
672 | goto out; | |
673 | } | |
674 | ||
675 | /* Only advance the bio if it hasn't been advanced already. If advanced, | |
676 | * this bio is at least a partial bio (i.e., it has partially been | |
677 | * filled with data from the cache). If part of the data resides on the | |
678 | * media, we will read later on | |
679 | */ | |
75cb8e93 | 680 | if (unlikely(!advanced_bio)) |
a4bd217b JG |
681 | bio_advance(bio, bio_iter * PBLK_EXPOSED_PAGE_SIZE); |
682 | ||
683 | data = bio_data(bio); | |
684 | memcpy(data, entry->data, rb->seg_size); | |
685 | ||
686 | out: | |
687 | spin_unlock(&rb->w_lock); | |
688 | return ret; | |
689 | } | |
690 | ||
691 | struct pblk_w_ctx *pblk_rb_w_ctx(struct pblk_rb *rb, unsigned int pos) | |
692 | { | |
693 | unsigned int entry = pos & (rb->nr_entries - 1); | |
694 | ||
695 | return &rb->entries[entry].w_ctx; | |
696 | } | |
697 | ||
698 | unsigned int pblk_rb_sync_init(struct pblk_rb *rb, unsigned long *flags) | |
699 | __acquires(&rb->s_lock) | |
700 | { | |
701 | if (flags) | |
702 | spin_lock_irqsave(&rb->s_lock, *flags); | |
703 | else | |
704 | spin_lock_irq(&rb->s_lock); | |
705 | ||
706 | return rb->sync; | |
707 | } | |
708 | ||
709 | void pblk_rb_sync_end(struct pblk_rb *rb, unsigned long *flags) | |
710 | __releases(&rb->s_lock) | |
711 | { | |
712 | lockdep_assert_held(&rb->s_lock); | |
713 | ||
714 | if (flags) | |
715 | spin_unlock_irqrestore(&rb->s_lock, *flags); | |
716 | else | |
717 | spin_unlock_irq(&rb->s_lock); | |
718 | } | |
719 | ||
720 | unsigned int pblk_rb_sync_advance(struct pblk_rb *rb, unsigned int nr_entries) | |
721 | { | |
533657c1 | 722 | unsigned int sync, flush_point; |
a4bd217b JG |
723 | lockdep_assert_held(&rb->s_lock); |
724 | ||
725 | sync = READ_ONCE(rb->sync); | |
533657c1 HH |
726 | flush_point = READ_ONCE(rb->flush_point); |
727 | ||
728 | if (flush_point != EMPTY_ENTRY) { | |
729 | unsigned int secs_to_flush; | |
a4bd217b | 730 | |
533657c1 HH |
731 | secs_to_flush = pblk_rb_ring_count(flush_point, sync, |
732 | rb->nr_entries); | |
733 | if (secs_to_flush < nr_entries) { | |
734 | /* Protect flush points */ | |
735 | smp_store_release(&rb->flush_point, EMPTY_ENTRY); | |
736 | } | |
737 | } | |
738 | ||
739 | sync = (sync + nr_entries) & (rb->nr_entries - 1); | |
a4bd217b JG |
740 | |
741 | /* Protect from counts */ | |
742 | smp_store_release(&rb->sync, sync); | |
743 | ||
744 | return sync; | |
745 | } | |
746 | ||
533657c1 | 747 | /* Calculate how many sectors to submit up to the current flush point. */ |
8154d296 | 748 | unsigned int pblk_rb_flush_point_count(struct pblk_rb *rb) |
a4bd217b | 749 | { |
533657c1 HH |
750 | unsigned int subm, sync, flush_point; |
751 | unsigned int submitted, to_flush; | |
a4bd217b | 752 | |
8154d296 HH |
753 | /* Protect flush points */ |
754 | flush_point = smp_load_acquire(&rb->flush_point); | |
755 | if (flush_point == EMPTY_ENTRY) | |
a4bd217b JG |
756 | return 0; |
757 | ||
533657c1 HH |
758 | /* Protect syncs */ |
759 | sync = smp_load_acquire(&rb->sync); | |
760 | ||
a4bd217b | 761 | subm = READ_ONCE(rb->subm); |
533657c1 | 762 | submitted = pblk_rb_ring_count(subm, sync, rb->nr_entries); |
a4bd217b JG |
763 | |
764 | /* The sync point itself counts as a sector to sync */ | |
533657c1 | 765 | to_flush = pblk_rb_ring_count(flush_point, sync, rb->nr_entries) + 1; |
a4bd217b | 766 | |
533657c1 | 767 | return (submitted < to_flush) ? (to_flush - submitted) : 0; |
a4bd217b JG |
768 | } |
769 | ||
770 | /* | |
771 | * Scan from the current position of the sync pointer to find the entry that | |
772 | * corresponds to the given ppa. This is necessary since write requests can be | |
773 | * completed out of order. The assumption is that the ppa is close to the sync | |
774 | * pointer thus the search will not take long. | |
775 | * | |
776 | * The caller of this function must guarantee that the sync pointer will no | |
777 | * reach the entry while it is using the metadata associated with it. With this | |
778 | * assumption in mind, there is no need to take the sync lock. | |
779 | */ | |
780 | struct pblk_rb_entry *pblk_rb_sync_scan_entry(struct pblk_rb *rb, | |
781 | struct ppa_addr *ppa) | |
782 | { | |
783 | unsigned int sync, subm, count; | |
784 | unsigned int i; | |
785 | ||
786 | sync = READ_ONCE(rb->sync); | |
787 | subm = READ_ONCE(rb->subm); | |
788 | count = pblk_rb_ring_count(subm, sync, rb->nr_entries); | |
789 | ||
790 | for (i = 0; i < count; i++) | |
791 | sync = (sync + 1) & (rb->nr_entries - 1); | |
792 | ||
793 | return NULL; | |
794 | } | |
795 | ||
796 | int pblk_rb_tear_down_check(struct pblk_rb *rb) | |
797 | { | |
798 | struct pblk_rb_entry *entry; | |
799 | int i; | |
800 | int ret = 0; | |
801 | ||
802 | spin_lock(&rb->w_lock); | |
803 | spin_lock_irq(&rb->s_lock); | |
804 | ||
805 | if ((rb->mem == rb->subm) && (rb->subm == rb->sync) && | |
806 | (rb->sync == rb->l2p_update) && | |
8154d296 | 807 | (rb->flush_point == EMPTY_ENTRY)) { |
a4bd217b JG |
808 | goto out; |
809 | } | |
810 | ||
811 | if (!rb->entries) { | |
812 | ret = 1; | |
813 | goto out; | |
814 | } | |
815 | ||
816 | for (i = 0; i < rb->nr_entries; i++) { | |
817 | entry = &rb->entries[i]; | |
818 | ||
819 | if (!entry->data) { | |
820 | ret = 1; | |
821 | goto out; | |
822 | } | |
823 | } | |
824 | ||
825 | out: | |
826 | spin_unlock(&rb->w_lock); | |
827 | spin_unlock_irq(&rb->s_lock); | |
828 | ||
829 | return ret; | |
830 | } | |
831 | ||
832 | unsigned int pblk_rb_wrap_pos(struct pblk_rb *rb, unsigned int pos) | |
833 | { | |
834 | return (pos & (rb->nr_entries - 1)); | |
835 | } | |
836 | ||
837 | int pblk_rb_pos_oob(struct pblk_rb *rb, u64 pos) | |
838 | { | |
839 | return (pos >= rb->nr_entries); | |
840 | } | |
841 | ||
842 | ssize_t pblk_rb_sysfs(struct pblk_rb *rb, char *buf) | |
843 | { | |
844 | struct pblk *pblk = container_of(rb, struct pblk, rwb); | |
845 | struct pblk_c_ctx *c; | |
846 | ssize_t offset; | |
847 | int queued_entries = 0; | |
848 | ||
849 | spin_lock_irq(&rb->s_lock); | |
850 | list_for_each_entry(c, &pblk->compl_list, list) | |
851 | queued_entries++; | |
852 | spin_unlock_irq(&rb->s_lock); | |
853 | ||
8154d296 | 854 | if (rb->flush_point != EMPTY_ENTRY) |
a4bd217b JG |
855 | offset = scnprintf(buf, PAGE_SIZE, |
856 | "%u\t%u\t%u\t%u\t%u\t%u\t%u - %u/%u/%u - %d\n", | |
857 | rb->nr_entries, | |
858 | rb->mem, | |
859 | rb->subm, | |
860 | rb->sync, | |
861 | rb->l2p_update, | |
862 | #ifdef CONFIG_NVM_DEBUG | |
8154d296 | 863 | atomic_read(&rb->inflight_flush_point), |
a4bd217b JG |
864 | #else |
865 | 0, | |
866 | #endif | |
8154d296 | 867 | rb->flush_point, |
a4bd217b JG |
868 | pblk_rb_read_count(rb), |
869 | pblk_rb_space(rb), | |
8154d296 | 870 | pblk_rb_flush_point_count(rb), |
a4bd217b JG |
871 | queued_entries); |
872 | else | |
873 | offset = scnprintf(buf, PAGE_SIZE, | |
874 | "%u\t%u\t%u\t%u\t%u\t%u\tNULL - %u/%u/%u - %d\n", | |
875 | rb->nr_entries, | |
876 | rb->mem, | |
877 | rb->subm, | |
878 | rb->sync, | |
879 | rb->l2p_update, | |
880 | #ifdef CONFIG_NVM_DEBUG | |
8154d296 | 881 | atomic_read(&rb->inflight_flush_point), |
a4bd217b JG |
882 | #else |
883 | 0, | |
884 | #endif | |
885 | pblk_rb_read_count(rb), | |
886 | pblk_rb_space(rb), | |
8154d296 | 887 | pblk_rb_flush_point_count(rb), |
a4bd217b JG |
888 | queued_entries); |
889 | ||
890 | return offset; | |
891 | } |