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; | |
57 | rb->sync_point = EMPTY_ENTRY; | |
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 | |
115 | atomic_set(&rb->inflight_sync_point, 0); | |
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 | ||
204 | static int __pblk_rb_update_l2p(struct pblk_rb *rb, unsigned int *l2p_upd, | |
205 | unsigned int to_update) | |
206 | { | |
207 | struct pblk *pblk = container_of(rb, struct pblk, rwb); | |
208 | struct pblk_line *line; | |
209 | struct pblk_rb_entry *entry; | |
210 | struct pblk_w_ctx *w_ctx; | |
b20ba1bc | 211 | unsigned int user_io = 0, gc_io = 0; |
a4bd217b | 212 | unsigned int i; |
b20ba1bc | 213 | int flags; |
a4bd217b JG |
214 | |
215 | for (i = 0; i < to_update; i++) { | |
216 | entry = &rb->entries[*l2p_upd]; | |
217 | w_ctx = &entry->w_ctx; | |
218 | ||
b20ba1bc JG |
219 | flags = READ_ONCE(entry->w_ctx.flags); |
220 | if (flags & PBLK_IOTYPE_USER) | |
221 | user_io++; | |
222 | else if (flags & PBLK_IOTYPE_GC) | |
223 | gc_io++; | |
224 | else | |
225 | WARN(1, "pblk: unknown IO type\n"); | |
226 | ||
a4bd217b JG |
227 | pblk_update_map_dev(pblk, w_ctx->lba, w_ctx->ppa, |
228 | entry->cacheline); | |
229 | ||
230 | line = &pblk->lines[pblk_tgt_ppa_to_line(w_ctx->ppa)]; | |
231 | kref_put(&line->ref, pblk_line_put); | |
232 | clean_wctx(w_ctx); | |
233 | *l2p_upd = (*l2p_upd + 1) & (rb->nr_entries - 1); | |
234 | } | |
235 | ||
b20ba1bc JG |
236 | pblk_rl_out(&pblk->rl, user_io, gc_io); |
237 | ||
a4bd217b JG |
238 | return 0; |
239 | } | |
240 | ||
241 | /* | |
242 | * When we move the l2p_update pointer, we update the l2p table - lookups will | |
243 | * point to the physical address instead of to the cacheline in the write buffer | |
244 | * from this moment on. | |
245 | */ | |
246 | static int pblk_rb_update_l2p(struct pblk_rb *rb, unsigned int nr_entries, | |
247 | unsigned int mem, unsigned int sync) | |
248 | { | |
249 | unsigned int space, count; | |
250 | int ret = 0; | |
251 | ||
252 | lockdep_assert_held(&rb->w_lock); | |
253 | ||
254 | /* Update l2p only as buffer entries are being overwritten */ | |
255 | space = pblk_rb_ring_space(rb, mem, rb->l2p_update, rb->nr_entries); | |
256 | if (space > nr_entries) | |
257 | goto out; | |
258 | ||
259 | count = nr_entries - space; | |
260 | /* l2p_update used exclusively under rb->w_lock */ | |
261 | ret = __pblk_rb_update_l2p(rb, &rb->l2p_update, count); | |
262 | ||
263 | out: | |
264 | return ret; | |
265 | } | |
266 | ||
267 | /* | |
268 | * Update the l2p entry for all sectors stored on the write buffer. This means | |
269 | * that all future lookups to the l2p table will point to a device address, not | |
270 | * to the cacheline in the write buffer. | |
271 | */ | |
272 | void pblk_rb_sync_l2p(struct pblk_rb *rb) | |
273 | { | |
274 | unsigned int sync; | |
275 | unsigned int to_update; | |
276 | ||
277 | spin_lock(&rb->w_lock); | |
278 | ||
279 | /* Protect from reads and writes */ | |
280 | sync = smp_load_acquire(&rb->sync); | |
281 | ||
282 | to_update = pblk_rb_ring_count(sync, rb->l2p_update, rb->nr_entries); | |
283 | __pblk_rb_update_l2p(rb, &rb->l2p_update, to_update); | |
284 | ||
285 | spin_unlock(&rb->w_lock); | |
286 | } | |
287 | ||
288 | /* | |
289 | * Write @nr_entries to ring buffer from @data buffer if there is enough space. | |
290 | * Typically, 4KB data chunks coming from a bio will be copied to the ring | |
291 | * buffer, thus the write will fail if not all incoming data can be copied. | |
292 | * | |
293 | */ | |
294 | static void __pblk_rb_write_entry(struct pblk_rb *rb, void *data, | |
295 | struct pblk_w_ctx w_ctx, | |
296 | struct pblk_rb_entry *entry) | |
297 | { | |
298 | memcpy(entry->data, data, rb->seg_size); | |
299 | ||
300 | entry->w_ctx.lba = w_ctx.lba; | |
301 | entry->w_ctx.ppa = w_ctx.ppa; | |
302 | } | |
303 | ||
304 | void pblk_rb_write_entry_user(struct pblk_rb *rb, void *data, | |
305 | struct pblk_w_ctx w_ctx, unsigned int ring_pos) | |
306 | { | |
307 | struct pblk *pblk = container_of(rb, struct pblk, rwb); | |
308 | struct pblk_rb_entry *entry; | |
309 | int flags; | |
310 | ||
311 | entry = &rb->entries[ring_pos]; | |
312 | flags = READ_ONCE(entry->w_ctx.flags); | |
313 | #ifdef CONFIG_NVM_DEBUG | |
314 | /* Caller must guarantee that the entry is free */ | |
315 | BUG_ON(!(flags & PBLK_WRITABLE_ENTRY)); | |
316 | #endif | |
317 | ||
318 | __pblk_rb_write_entry(rb, data, w_ctx, entry); | |
319 | ||
320 | pblk_update_map_cache(pblk, w_ctx.lba, entry->cacheline); | |
321 | flags = w_ctx.flags | PBLK_WRITTEN_DATA; | |
322 | ||
323 | /* Release flags on write context. Protect from writes */ | |
324 | smp_store_release(&entry->w_ctx.flags, flags); | |
325 | } | |
326 | ||
327 | void pblk_rb_write_entry_gc(struct pblk_rb *rb, void *data, | |
328 | struct pblk_w_ctx w_ctx, struct pblk_line *gc_line, | |
329 | unsigned int ring_pos) | |
330 | { | |
331 | struct pblk *pblk = container_of(rb, struct pblk, rwb); | |
332 | struct pblk_rb_entry *entry; | |
333 | int flags; | |
334 | ||
335 | entry = &rb->entries[ring_pos]; | |
336 | flags = READ_ONCE(entry->w_ctx.flags); | |
337 | #ifdef CONFIG_NVM_DEBUG | |
338 | /* Caller must guarantee that the entry is free */ | |
339 | BUG_ON(!(flags & PBLK_WRITABLE_ENTRY)); | |
340 | #endif | |
341 | ||
342 | __pblk_rb_write_entry(rb, data, w_ctx, entry); | |
343 | ||
344 | if (!pblk_update_map_gc(pblk, w_ctx.lba, entry->cacheline, gc_line)) | |
345 | entry->w_ctx.lba = ADDR_EMPTY; | |
346 | ||
347 | flags = w_ctx.flags | PBLK_WRITTEN_DATA; | |
348 | ||
349 | /* Release flags on write context. Protect from writes */ | |
350 | smp_store_release(&entry->w_ctx.flags, flags); | |
351 | } | |
352 | ||
353 | static int pblk_rb_sync_point_set(struct pblk_rb *rb, struct bio *bio, | |
354 | unsigned int pos) | |
355 | { | |
356 | struct pblk_rb_entry *entry; | |
357 | unsigned int subm, sync_point; | |
358 | int flags; | |
359 | ||
360 | subm = READ_ONCE(rb->subm); | |
361 | ||
362 | #ifdef CONFIG_NVM_DEBUG | |
363 | atomic_inc(&rb->inflight_sync_point); | |
364 | #endif | |
365 | ||
366 | if (pos == subm) | |
367 | return 0; | |
368 | ||
369 | sync_point = (pos == 0) ? (rb->nr_entries - 1) : (pos - 1); | |
370 | entry = &rb->entries[sync_point]; | |
371 | ||
372 | flags = READ_ONCE(entry->w_ctx.flags); | |
373 | flags |= PBLK_FLUSH_ENTRY; | |
374 | ||
375 | /* Release flags on context. Protect from writes */ | |
376 | smp_store_release(&entry->w_ctx.flags, flags); | |
377 | ||
378 | /* Protect syncs */ | |
379 | smp_store_release(&rb->sync_point, sync_point); | |
380 | ||
588726d3 JG |
381 | if (!bio) |
382 | return 0; | |
383 | ||
a4bd217b JG |
384 | spin_lock_irq(&rb->s_lock); |
385 | bio_list_add(&entry->w_ctx.bios, bio); | |
386 | spin_unlock_irq(&rb->s_lock); | |
387 | ||
388 | return 1; | |
389 | } | |
390 | ||
391 | static int __pblk_rb_may_write(struct pblk_rb *rb, unsigned int nr_entries, | |
392 | unsigned int *pos) | |
393 | { | |
394 | unsigned int mem; | |
395 | unsigned int sync; | |
396 | ||
397 | sync = READ_ONCE(rb->sync); | |
398 | mem = READ_ONCE(rb->mem); | |
399 | ||
400 | if (pblk_rb_ring_space(rb, mem, sync, rb->nr_entries) < nr_entries) | |
401 | return 0; | |
402 | ||
403 | if (pblk_rb_update_l2p(rb, nr_entries, mem, sync)) | |
404 | return 0; | |
405 | ||
406 | *pos = mem; | |
407 | ||
408 | return 1; | |
409 | } | |
410 | ||
411 | static int pblk_rb_may_write(struct pblk_rb *rb, unsigned int nr_entries, | |
412 | unsigned int *pos) | |
413 | { | |
414 | if (!__pblk_rb_may_write(rb, nr_entries, pos)) | |
415 | return 0; | |
416 | ||
417 | /* Protect from read count */ | |
418 | smp_store_release(&rb->mem, (*pos + nr_entries) & (rb->nr_entries - 1)); | |
419 | return 1; | |
420 | } | |
421 | ||
588726d3 JG |
422 | void pblk_rb_flush(struct pblk_rb *rb) |
423 | { | |
424 | struct pblk *pblk = container_of(rb, struct pblk, rwb); | |
425 | unsigned int mem = READ_ONCE(rb->mem); | |
426 | ||
427 | if (pblk_rb_sync_point_set(rb, NULL, mem)) | |
428 | return; | |
429 | ||
430 | pblk_write_should_kick(pblk); | |
431 | } | |
432 | ||
a4bd217b JG |
433 | static int pblk_rb_may_write_flush(struct pblk_rb *rb, unsigned int nr_entries, |
434 | unsigned int *pos, struct bio *bio, | |
435 | int *io_ret) | |
436 | { | |
437 | unsigned int mem; | |
438 | ||
439 | if (!__pblk_rb_may_write(rb, nr_entries, pos)) | |
440 | return 0; | |
441 | ||
442 | mem = (*pos + nr_entries) & (rb->nr_entries - 1); | |
443 | *io_ret = NVM_IO_DONE; | |
444 | ||
445 | if (bio->bi_opf & REQ_PREFLUSH) { | |
446 | struct pblk *pblk = container_of(rb, struct pblk, rwb); | |
447 | ||
448 | #ifdef CONFIG_NVM_DEBUG | |
449 | atomic_long_inc(&pblk->nr_flush); | |
450 | #endif | |
451 | if (pblk_rb_sync_point_set(&pblk->rwb, bio, mem)) | |
452 | *io_ret = NVM_IO_OK; | |
453 | } | |
454 | ||
455 | /* Protect from read count */ | |
456 | smp_store_release(&rb->mem, mem); | |
457 | return 1; | |
458 | } | |
459 | ||
460 | /* | |
461 | * Atomically check that (i) there is space on the write buffer for the | |
462 | * incoming I/O, and (ii) the current I/O type has enough budget in the write | |
463 | * buffer (rate-limiter). | |
464 | */ | |
465 | int pblk_rb_may_write_user(struct pblk_rb *rb, struct bio *bio, | |
466 | unsigned int nr_entries, unsigned int *pos) | |
467 | { | |
468 | struct pblk *pblk = container_of(rb, struct pblk, rwb); | |
588726d3 | 469 | int io_ret; |
a4bd217b JG |
470 | |
471 | spin_lock(&rb->w_lock); | |
588726d3 JG |
472 | io_ret = pblk_rl_user_may_insert(&pblk->rl, nr_entries); |
473 | if (io_ret) { | |
a4bd217b | 474 | spin_unlock(&rb->w_lock); |
588726d3 | 475 | return io_ret; |
a4bd217b JG |
476 | } |
477 | ||
588726d3 | 478 | if (!pblk_rb_may_write_flush(rb, nr_entries, pos, bio, &io_ret)) { |
a4bd217b JG |
479 | spin_unlock(&rb->w_lock); |
480 | return NVM_IO_REQUEUE; | |
481 | } | |
482 | ||
483 | pblk_rl_user_in(&pblk->rl, nr_entries); | |
484 | spin_unlock(&rb->w_lock); | |
485 | ||
588726d3 | 486 | return io_ret; |
a4bd217b JG |
487 | } |
488 | ||
489 | /* | |
490 | * Look at pblk_rb_may_write_user comment | |
491 | */ | |
492 | int pblk_rb_may_write_gc(struct pblk_rb *rb, unsigned int nr_entries, | |
493 | unsigned int *pos) | |
494 | { | |
495 | struct pblk *pblk = container_of(rb, struct pblk, rwb); | |
496 | ||
497 | spin_lock(&rb->w_lock); | |
498 | if (!pblk_rl_gc_may_insert(&pblk->rl, nr_entries)) { | |
499 | spin_unlock(&rb->w_lock); | |
500 | return 0; | |
501 | } | |
502 | ||
503 | if (!pblk_rb_may_write(rb, nr_entries, pos)) { | |
504 | spin_unlock(&rb->w_lock); | |
505 | return 0; | |
506 | } | |
507 | ||
508 | pblk_rl_gc_in(&pblk->rl, nr_entries); | |
509 | spin_unlock(&rb->w_lock); | |
510 | ||
511 | return 1; | |
512 | } | |
513 | ||
514 | /* | |
515 | * The caller of this function must ensure that the backpointer will not | |
516 | * overwrite the entries passed on the list. | |
517 | */ | |
518 | unsigned int pblk_rb_read_to_bio_list(struct pblk_rb *rb, struct bio *bio, | |
519 | struct list_head *list, | |
520 | unsigned int max) | |
521 | { | |
522 | struct pblk_rb_entry *entry, *tentry; | |
523 | struct page *page; | |
524 | unsigned int read = 0; | |
525 | int ret; | |
526 | ||
527 | list_for_each_entry_safe(entry, tentry, list, index) { | |
528 | if (read > max) { | |
529 | pr_err("pblk: too many entries on list\n"); | |
530 | goto out; | |
531 | } | |
532 | ||
533 | page = virt_to_page(entry->data); | |
534 | if (!page) { | |
535 | pr_err("pblk: could not allocate write bio page\n"); | |
536 | goto out; | |
537 | } | |
538 | ||
539 | ret = bio_add_page(bio, page, rb->seg_size, 0); | |
540 | if (ret != rb->seg_size) { | |
541 | pr_err("pblk: could not add page to write bio\n"); | |
542 | goto out; | |
543 | } | |
544 | ||
545 | list_del(&entry->index); | |
546 | read++; | |
547 | } | |
548 | ||
549 | out: | |
550 | return read; | |
551 | } | |
552 | ||
553 | /* | |
554 | * Read available entries on rb and add them to the given bio. To avoid a memory | |
555 | * copy, a page reference to the write buffer is used to be added to the bio. | |
556 | * | |
557 | * This function is used by the write thread to form the write bio that will | |
558 | * persist data on the write buffer to the media. | |
559 | */ | |
d624f371 JG |
560 | unsigned int pblk_rb_read_to_bio(struct pblk_rb *rb, struct nvm_rq *rqd, |
561 | struct bio *bio, unsigned int pos, | |
562 | unsigned int nr_entries, unsigned int count) | |
a4bd217b JG |
563 | { |
564 | struct pblk *pblk = container_of(rb, struct pblk, rwb); | |
d624f371 JG |
565 | struct request_queue *q = pblk->dev->q; |
566 | struct pblk_c_ctx *c_ctx = nvm_rq_to_pdu(rqd); | |
a4bd217b JG |
567 | struct pblk_rb_entry *entry; |
568 | struct page *page; | |
d624f371 | 569 | unsigned int pad = 0, to_read = nr_entries; |
a4bd217b JG |
570 | unsigned int i; |
571 | int flags; | |
a4bd217b JG |
572 | |
573 | if (count < nr_entries) { | |
574 | pad = nr_entries - count; | |
575 | to_read = count; | |
576 | } | |
577 | ||
578 | c_ctx->sentry = pos; | |
579 | c_ctx->nr_valid = to_read; | |
580 | c_ctx->nr_padded = pad; | |
581 | ||
582 | for (i = 0; i < to_read; i++) { | |
583 | entry = &rb->entries[pos]; | |
584 | ||
585 | /* A write has been allowed into the buffer, but data is still | |
586 | * being copied to it. It is ok to busy wait. | |
587 | */ | |
588 | try: | |
589 | flags = READ_ONCE(entry->w_ctx.flags); | |
10888129 JG |
590 | if (!(flags & PBLK_WRITTEN_DATA)) { |
591 | io_schedule(); | |
a4bd217b | 592 | goto try; |
10888129 | 593 | } |
a4bd217b | 594 | |
a4bd217b JG |
595 | page = virt_to_page(entry->data); |
596 | if (!page) { | |
597 | pr_err("pblk: could not allocate write bio page\n"); | |
598 | flags &= ~PBLK_WRITTEN_DATA; | |
599 | flags |= PBLK_SUBMITTED_ENTRY; | |
600 | /* Release flags on context. Protect from writes */ | |
601 | smp_store_release(&entry->w_ctx.flags, flags); | |
d624f371 | 602 | return NVM_IO_ERR; |
a4bd217b JG |
603 | } |
604 | ||
d624f371 JG |
605 | if (bio_add_pc_page(q, bio, page, rb->seg_size, 0) != |
606 | rb->seg_size) { | |
a4bd217b JG |
607 | pr_err("pblk: could not add page to write bio\n"); |
608 | flags &= ~PBLK_WRITTEN_DATA; | |
609 | flags |= PBLK_SUBMITTED_ENTRY; | |
610 | /* Release flags on context. Protect from writes */ | |
611 | smp_store_release(&entry->w_ctx.flags, flags); | |
d624f371 | 612 | return NVM_IO_ERR; |
a4bd217b JG |
613 | } |
614 | ||
615 | if (flags & PBLK_FLUSH_ENTRY) { | |
616 | unsigned int sync_point; | |
617 | ||
618 | sync_point = READ_ONCE(rb->sync_point); | |
619 | if (sync_point == pos) { | |
620 | /* Protect syncs */ | |
621 | smp_store_release(&rb->sync_point, EMPTY_ENTRY); | |
622 | } | |
623 | ||
624 | flags &= ~PBLK_FLUSH_ENTRY; | |
625 | #ifdef CONFIG_NVM_DEBUG | |
626 | atomic_dec(&rb->inflight_sync_point); | |
627 | #endif | |
628 | } | |
629 | ||
630 | flags &= ~PBLK_WRITTEN_DATA; | |
631 | flags |= PBLK_SUBMITTED_ENTRY; | |
632 | ||
633 | /* Release flags on context. Protect from writes */ | |
634 | smp_store_release(&entry->w_ctx.flags, flags); | |
635 | ||
636 | pos = (pos + 1) & (rb->nr_entries - 1); | |
637 | } | |
638 | ||
d624f371 JG |
639 | if (pad) { |
640 | if (pblk_bio_add_pages(pblk, bio, GFP_KERNEL, pad)) { | |
641 | pr_err("pblk: could not pad page in write bio\n"); | |
642 | return NVM_IO_ERR; | |
643 | } | |
644 | } | |
645 | ||
a4bd217b JG |
646 | #ifdef CONFIG_NVM_DEBUG |
647 | atomic_long_add(pad, &((struct pblk *) | |
648 | (container_of(rb, struct pblk, rwb)))->padded_writes); | |
649 | #endif | |
d624f371 JG |
650 | |
651 | return NVM_IO_OK; | |
a4bd217b JG |
652 | } |
653 | ||
654 | /* | |
655 | * Copy to bio only if the lba matches the one on the given cache entry. | |
656 | * Otherwise, it means that the entry has been overwritten, and the bio should | |
657 | * be directed to disk. | |
658 | */ | |
659 | int pblk_rb_copy_to_bio(struct pblk_rb *rb, struct bio *bio, sector_t lba, | |
75cb8e93 | 660 | struct ppa_addr ppa, int bio_iter, bool advanced_bio) |
a4bd217b | 661 | { |
07698466 | 662 | struct pblk *pblk = container_of(rb, struct pblk, rwb); |
a4bd217b JG |
663 | struct pblk_rb_entry *entry; |
664 | struct pblk_w_ctx *w_ctx; | |
07698466 JG |
665 | struct ppa_addr l2p_ppa; |
666 | u64 pos = pblk_addr_to_cacheline(ppa); | |
a4bd217b JG |
667 | void *data; |
668 | int flags; | |
669 | int ret = 1; | |
670 | ||
a4bd217b JG |
671 | |
672 | #ifdef CONFIG_NVM_DEBUG | |
673 | /* Caller must ensure that the access will not cause an overflow */ | |
674 | BUG_ON(pos >= rb->nr_entries); | |
675 | #endif | |
676 | entry = &rb->entries[pos]; | |
677 | w_ctx = &entry->w_ctx; | |
678 | flags = READ_ONCE(w_ctx->flags); | |
679 | ||
07698466 JG |
680 | spin_lock(&rb->w_lock); |
681 | spin_lock(&pblk->trans_lock); | |
682 | l2p_ppa = pblk_trans_map_get(pblk, lba); | |
683 | spin_unlock(&pblk->trans_lock); | |
684 | ||
a4bd217b | 685 | /* Check if the entry has been overwritten or is scheduled to be */ |
07698466 JG |
686 | if (!pblk_ppa_comp(l2p_ppa, ppa) || w_ctx->lba != lba || |
687 | flags & PBLK_WRITABLE_ENTRY) { | |
a4bd217b JG |
688 | ret = 0; |
689 | goto out; | |
690 | } | |
691 | ||
692 | /* Only advance the bio if it hasn't been advanced already. If advanced, | |
693 | * this bio is at least a partial bio (i.e., it has partially been | |
694 | * filled with data from the cache). If part of the data resides on the | |
695 | * media, we will read later on | |
696 | */ | |
75cb8e93 | 697 | if (unlikely(!advanced_bio)) |
a4bd217b JG |
698 | bio_advance(bio, bio_iter * PBLK_EXPOSED_PAGE_SIZE); |
699 | ||
700 | data = bio_data(bio); | |
701 | memcpy(data, entry->data, rb->seg_size); | |
702 | ||
703 | out: | |
704 | spin_unlock(&rb->w_lock); | |
705 | return ret; | |
706 | } | |
707 | ||
708 | struct pblk_w_ctx *pblk_rb_w_ctx(struct pblk_rb *rb, unsigned int pos) | |
709 | { | |
710 | unsigned int entry = pos & (rb->nr_entries - 1); | |
711 | ||
712 | return &rb->entries[entry].w_ctx; | |
713 | } | |
714 | ||
715 | unsigned int pblk_rb_sync_init(struct pblk_rb *rb, unsigned long *flags) | |
716 | __acquires(&rb->s_lock) | |
717 | { | |
718 | if (flags) | |
719 | spin_lock_irqsave(&rb->s_lock, *flags); | |
720 | else | |
721 | spin_lock_irq(&rb->s_lock); | |
722 | ||
723 | return rb->sync; | |
724 | } | |
725 | ||
726 | void pblk_rb_sync_end(struct pblk_rb *rb, unsigned long *flags) | |
727 | __releases(&rb->s_lock) | |
728 | { | |
729 | lockdep_assert_held(&rb->s_lock); | |
730 | ||
731 | if (flags) | |
732 | spin_unlock_irqrestore(&rb->s_lock, *flags); | |
733 | else | |
734 | spin_unlock_irq(&rb->s_lock); | |
735 | } | |
736 | ||
737 | unsigned int pblk_rb_sync_advance(struct pblk_rb *rb, unsigned int nr_entries) | |
738 | { | |
739 | unsigned int sync; | |
740 | unsigned int i; | |
741 | ||
742 | lockdep_assert_held(&rb->s_lock); | |
743 | ||
744 | sync = READ_ONCE(rb->sync); | |
745 | ||
746 | for (i = 0; i < nr_entries; i++) | |
747 | sync = (sync + 1) & (rb->nr_entries - 1); | |
748 | ||
749 | /* Protect from counts */ | |
750 | smp_store_release(&rb->sync, sync); | |
751 | ||
752 | return sync; | |
753 | } | |
754 | ||
755 | unsigned int pblk_rb_sync_point_count(struct pblk_rb *rb) | |
756 | { | |
757 | unsigned int subm, sync_point; | |
758 | unsigned int count; | |
759 | ||
760 | /* Protect syncs */ | |
761 | sync_point = smp_load_acquire(&rb->sync_point); | |
762 | if (sync_point == EMPTY_ENTRY) | |
763 | return 0; | |
764 | ||
765 | subm = READ_ONCE(rb->subm); | |
766 | ||
767 | /* The sync point itself counts as a sector to sync */ | |
768 | count = pblk_rb_ring_count(sync_point, subm, rb->nr_entries) + 1; | |
769 | ||
770 | return count; | |
771 | } | |
772 | ||
773 | /* | |
774 | * Scan from the current position of the sync pointer to find the entry that | |
775 | * corresponds to the given ppa. This is necessary since write requests can be | |
776 | * completed out of order. The assumption is that the ppa is close to the sync | |
777 | * pointer thus the search will not take long. | |
778 | * | |
779 | * The caller of this function must guarantee that the sync pointer will no | |
780 | * reach the entry while it is using the metadata associated with it. With this | |
781 | * assumption in mind, there is no need to take the sync lock. | |
782 | */ | |
783 | struct pblk_rb_entry *pblk_rb_sync_scan_entry(struct pblk_rb *rb, | |
784 | struct ppa_addr *ppa) | |
785 | { | |
786 | unsigned int sync, subm, count; | |
787 | unsigned int i; | |
788 | ||
789 | sync = READ_ONCE(rb->sync); | |
790 | subm = READ_ONCE(rb->subm); | |
791 | count = pblk_rb_ring_count(subm, sync, rb->nr_entries); | |
792 | ||
793 | for (i = 0; i < count; i++) | |
794 | sync = (sync + 1) & (rb->nr_entries - 1); | |
795 | ||
796 | return NULL; | |
797 | } | |
798 | ||
799 | int pblk_rb_tear_down_check(struct pblk_rb *rb) | |
800 | { | |
801 | struct pblk_rb_entry *entry; | |
802 | int i; | |
803 | int ret = 0; | |
804 | ||
805 | spin_lock(&rb->w_lock); | |
806 | spin_lock_irq(&rb->s_lock); | |
807 | ||
808 | if ((rb->mem == rb->subm) && (rb->subm == rb->sync) && | |
809 | (rb->sync == rb->l2p_update) && | |
810 | (rb->sync_point == EMPTY_ENTRY)) { | |
811 | goto out; | |
812 | } | |
813 | ||
814 | if (!rb->entries) { | |
815 | ret = 1; | |
816 | goto out; | |
817 | } | |
818 | ||
819 | for (i = 0; i < rb->nr_entries; i++) { | |
820 | entry = &rb->entries[i]; | |
821 | ||
822 | if (!entry->data) { | |
823 | ret = 1; | |
824 | goto out; | |
825 | } | |
826 | } | |
827 | ||
828 | out: | |
829 | spin_unlock(&rb->w_lock); | |
830 | spin_unlock_irq(&rb->s_lock); | |
831 | ||
832 | return ret; | |
833 | } | |
834 | ||
835 | unsigned int pblk_rb_wrap_pos(struct pblk_rb *rb, unsigned int pos) | |
836 | { | |
837 | return (pos & (rb->nr_entries - 1)); | |
838 | } | |
839 | ||
840 | int pblk_rb_pos_oob(struct pblk_rb *rb, u64 pos) | |
841 | { | |
842 | return (pos >= rb->nr_entries); | |
843 | } | |
844 | ||
845 | ssize_t pblk_rb_sysfs(struct pblk_rb *rb, char *buf) | |
846 | { | |
847 | struct pblk *pblk = container_of(rb, struct pblk, rwb); | |
848 | struct pblk_c_ctx *c; | |
849 | ssize_t offset; | |
850 | int queued_entries = 0; | |
851 | ||
852 | spin_lock_irq(&rb->s_lock); | |
853 | list_for_each_entry(c, &pblk->compl_list, list) | |
854 | queued_entries++; | |
855 | spin_unlock_irq(&rb->s_lock); | |
856 | ||
857 | if (rb->sync_point != EMPTY_ENTRY) | |
858 | offset = scnprintf(buf, PAGE_SIZE, | |
859 | "%u\t%u\t%u\t%u\t%u\t%u\t%u - %u/%u/%u - %d\n", | |
860 | rb->nr_entries, | |
861 | rb->mem, | |
862 | rb->subm, | |
863 | rb->sync, | |
864 | rb->l2p_update, | |
865 | #ifdef CONFIG_NVM_DEBUG | |
866 | atomic_read(&rb->inflight_sync_point), | |
867 | #else | |
868 | 0, | |
869 | #endif | |
870 | rb->sync_point, | |
871 | pblk_rb_read_count(rb), | |
872 | pblk_rb_space(rb), | |
873 | pblk_rb_sync_point_count(rb), | |
874 | queued_entries); | |
875 | else | |
876 | offset = scnprintf(buf, PAGE_SIZE, | |
877 | "%u\t%u\t%u\t%u\t%u\t%u\tNULL - %u/%u/%u - %d\n", | |
878 | rb->nr_entries, | |
879 | rb->mem, | |
880 | rb->subm, | |
881 | rb->sync, | |
882 | rb->l2p_update, | |
883 | #ifdef CONFIG_NVM_DEBUG | |
884 | atomic_read(&rb->inflight_sync_point), | |
885 | #else | |
886 | 0, | |
887 | #endif | |
888 | pblk_rb_read_count(rb), | |
889 | pblk_rb_space(rb), | |
890 | pblk_rb_sync_point_count(rb), | |
891 | queued_entries); | |
892 | ||
893 | return offset; | |
894 | } |