]>
Commit | Line | Data |
---|---|---|
02a1520d | 1 | // SPDX-License-Identifier: GPL-2.0 |
a4bd217b JG |
2 | /* |
3 | * Copyright (C) 2016 CNEX Labs | |
4 | * Initial release: Javier Gonzalez <javier@cnexlabs.com> | |
5 | * | |
6 | * Based upon the circular ringbuffer. | |
7 | * | |
8 | * This program is free software; you can redistribute it and/or | |
9 | * modify it under the terms of the GNU General Public License version | |
10 | * 2 as published by the Free Software Foundation. | |
11 | * | |
12 | * This program is distributed in the hope that it will be useful, but | |
13 | * WITHOUT ANY WARRANTY; without even the implied warranty of | |
14 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | |
15 | * General Public License for more details. | |
16 | * | |
17 | * pblk-rb.c - pblk's write buffer | |
18 | */ | |
19 | ||
20 | #include <linux/circ_buf.h> | |
21 | ||
22 | #include "pblk.h" | |
23 | ||
24 | static DECLARE_RWSEM(pblk_rb_lock); | |
25 | ||
26 | void pblk_rb_data_free(struct pblk_rb *rb) | |
27 | { | |
28 | struct pblk_rb_pages *p, *t; | |
29 | ||
30 | down_write(&pblk_rb_lock); | |
31 | list_for_each_entry_safe(p, t, &rb->pages, list) { | |
32 | free_pages((unsigned long)page_address(p->pages), p->order); | |
33 | list_del(&p->list); | |
34 | kfree(p); | |
35 | } | |
36 | up_write(&pblk_rb_lock); | |
37 | } | |
38 | ||
39 | /* | |
40 | * Initialize ring buffer. The data and metadata buffers must be previously | |
41 | * allocated and their size must be a power of two | |
5fb94e9c | 42 | * (Documentation/core-api/circular-buffers.rst) |
a4bd217b JG |
43 | */ |
44 | int pblk_rb_init(struct pblk_rb *rb, struct pblk_rb_entry *rb_entry_base, | |
45 | unsigned int power_size, unsigned int power_seg_sz) | |
46 | { | |
47 | struct pblk *pblk = container_of(rb, struct pblk, rwb); | |
48 | unsigned int init_entry = 0; | |
49 | unsigned int alloc_order = power_size; | |
50 | unsigned int max_order = MAX_ORDER - 1; | |
51 | unsigned int order, iter; | |
52 | ||
53 | down_write(&pblk_rb_lock); | |
54 | rb->entries = rb_entry_base; | |
55 | rb->seg_size = (1 << power_seg_sz); | |
56 | rb->nr_entries = (1 << power_size); | |
57 | rb->mem = rb->subm = rb->sync = rb->l2p_update = 0; | |
8154d296 | 58 | rb->flush_point = EMPTY_ENTRY; |
a4bd217b JG |
59 | |
60 | spin_lock_init(&rb->w_lock); | |
61 | spin_lock_init(&rb->s_lock); | |
62 | ||
63 | INIT_LIST_HEAD(&rb->pages); | |
64 | ||
65 | if (alloc_order >= max_order) { | |
66 | order = max_order; | |
67 | iter = (1 << (alloc_order - max_order)); | |
68 | } else { | |
69 | order = alloc_order; | |
70 | iter = 1; | |
71 | } | |
72 | ||
73 | do { | |
74 | struct pblk_rb_entry *entry; | |
75 | struct pblk_rb_pages *page_set; | |
76 | void *kaddr; | |
77 | unsigned long set_size; | |
78 | int i; | |
79 | ||
80 | page_set = kmalloc(sizeof(struct pblk_rb_pages), GFP_KERNEL); | |
81 | if (!page_set) { | |
82 | up_write(&pblk_rb_lock); | |
83 | return -ENOMEM; | |
84 | } | |
85 | ||
86 | page_set->order = order; | |
87 | page_set->pages = alloc_pages(GFP_KERNEL, order); | |
88 | if (!page_set->pages) { | |
89 | kfree(page_set); | |
90 | pblk_rb_data_free(rb); | |
91 | up_write(&pblk_rb_lock); | |
92 | return -ENOMEM; | |
93 | } | |
94 | kaddr = page_address(page_set->pages); | |
95 | ||
96 | entry = &rb->entries[init_entry]; | |
97 | entry->data = kaddr; | |
98 | entry->cacheline = pblk_cacheline_to_addr(init_entry++); | |
99 | entry->w_ctx.flags = PBLK_WRITABLE_ENTRY; | |
100 | ||
101 | set_size = (1 << order); | |
102 | for (i = 1; i < set_size; i++) { | |
103 | entry = &rb->entries[init_entry]; | |
104 | entry->cacheline = pblk_cacheline_to_addr(init_entry++); | |
105 | entry->data = kaddr + (i * rb->seg_size); | |
106 | entry->w_ctx.flags = PBLK_WRITABLE_ENTRY; | |
107 | bio_list_init(&entry->w_ctx.bios); | |
108 | } | |
109 | ||
110 | list_add_tail(&page_set->list, &rb->pages); | |
111 | iter--; | |
112 | } while (iter > 0); | |
113 | up_write(&pblk_rb_lock); | |
114 | ||
880eda54 | 115 | #ifdef CONFIG_NVM_PBLK_DEBUG |
8154d296 | 116 | atomic_set(&rb->inflight_flush_point, 0); |
a4bd217b JG |
117 | #endif |
118 | ||
119 | /* | |
120 | * Initialize rate-limiter, which controls access to the write buffer | |
121 | * but user and GC I/O | |
122 | */ | |
123 | pblk_rl_init(&pblk->rl, rb->nr_entries); | |
124 | ||
125 | return 0; | |
126 | } | |
127 | ||
128 | /* | |
129 | * pblk_rb_calculate_size -- calculate the size of the write buffer | |
130 | */ | |
131 | unsigned int pblk_rb_calculate_size(unsigned int nr_entries) | |
132 | { | |
133 | /* Alloc a write buffer that can at least fit 128 entries */ | |
134 | return (1 << max(get_count_order(nr_entries), 7)); | |
135 | } | |
136 | ||
137 | void *pblk_rb_entries_ref(struct pblk_rb *rb) | |
138 | { | |
139 | return rb->entries; | |
140 | } | |
141 | ||
142 | static void clean_wctx(struct pblk_w_ctx *w_ctx) | |
143 | { | |
144 | int flags; | |
145 | ||
a4bd217b | 146 | flags = READ_ONCE(w_ctx->flags); |
e37d0798 JG |
147 | WARN_ONCE(!(flags & PBLK_SUBMITTED_ENTRY), |
148 | "pblk: overwriting unsubmitted data\n"); | |
a4bd217b JG |
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 | ||
40b8657d JG |
172 | unsigned int pblk_rb_ptr_wrap(struct pblk_rb *rb, unsigned int p, |
173 | unsigned int nr_entries) | |
174 | { | |
175 | return (p + nr_entries) & (rb->nr_entries - 1); | |
176 | } | |
177 | ||
a4bd217b JG |
178 | /* |
179 | * Buffer count is calculated with respect to the submission entry signaling the | |
180 | * entries that are available to send to the media | |
181 | */ | |
182 | unsigned int pblk_rb_read_count(struct pblk_rb *rb) | |
183 | { | |
184 | unsigned int mem = READ_ONCE(rb->mem); | |
185 | unsigned int subm = READ_ONCE(rb->subm); | |
186 | ||
187 | return pblk_rb_ring_count(mem, subm, rb->nr_entries); | |
188 | } | |
189 | ||
ee8d5c1a JG |
190 | unsigned int pblk_rb_sync_count(struct pblk_rb *rb) |
191 | { | |
192 | unsigned int mem = READ_ONCE(rb->mem); | |
193 | unsigned int sync = READ_ONCE(rb->sync); | |
194 | ||
195 | return pblk_rb_ring_count(mem, sync, rb->nr_entries); | |
196 | } | |
197 | ||
a4bd217b JG |
198 | unsigned int pblk_rb_read_commit(struct pblk_rb *rb, unsigned int nr_entries) |
199 | { | |
200 | unsigned int subm; | |
201 | ||
202 | subm = READ_ONCE(rb->subm); | |
203 | /* Commit read means updating submission pointer */ | |
40b8657d | 204 | smp_store_release(&rb->subm, pblk_rb_ptr_wrap(rb, subm, nr_entries)); |
a4bd217b JG |
205 | |
206 | return subm; | |
207 | } | |
208 | ||
05ed3447 | 209 | static int __pblk_rb_update_l2p(struct pblk_rb *rb, unsigned int to_update) |
a4bd217b JG |
210 | { |
211 | struct pblk *pblk = container_of(rb, struct pblk, rwb); | |
212 | struct pblk_line *line; | |
213 | struct pblk_rb_entry *entry; | |
214 | struct pblk_w_ctx *w_ctx; | |
b20ba1bc | 215 | unsigned int user_io = 0, gc_io = 0; |
a4bd217b | 216 | unsigned int i; |
b20ba1bc | 217 | int flags; |
a4bd217b JG |
218 | |
219 | for (i = 0; i < to_update; i++) { | |
05ed3447 | 220 | entry = &rb->entries[rb->l2p_update]; |
a4bd217b JG |
221 | w_ctx = &entry->w_ctx; |
222 | ||
b20ba1bc JG |
223 | flags = READ_ONCE(entry->w_ctx.flags); |
224 | if (flags & PBLK_IOTYPE_USER) | |
225 | user_io++; | |
226 | else if (flags & PBLK_IOTYPE_GC) | |
227 | gc_io++; | |
228 | else | |
229 | WARN(1, "pblk: unknown IO type\n"); | |
230 | ||
a4bd217b JG |
231 | pblk_update_map_dev(pblk, w_ctx->lba, w_ctx->ppa, |
232 | entry->cacheline); | |
233 | ||
cb21665c | 234 | line = pblk_ppa_to_line(pblk, w_ctx->ppa); |
a4bd217b JG |
235 | kref_put(&line->ref, pblk_line_put); |
236 | clean_wctx(w_ctx); | |
40b8657d | 237 | rb->l2p_update = pblk_rb_ptr_wrap(rb, rb->l2p_update, 1); |
a4bd217b JG |
238 | } |
239 | ||
b20ba1bc JG |
240 | pblk_rl_out(&pblk->rl, user_io, gc_io); |
241 | ||
a4bd217b JG |
242 | return 0; |
243 | } | |
244 | ||
245 | /* | |
246 | * When we move the l2p_update pointer, we update the l2p table - lookups will | |
247 | * point to the physical address instead of to the cacheline in the write buffer | |
248 | * from this moment on. | |
249 | */ | |
250 | static int pblk_rb_update_l2p(struct pblk_rb *rb, unsigned int nr_entries, | |
251 | unsigned int mem, unsigned int sync) | |
252 | { | |
253 | unsigned int space, count; | |
254 | int ret = 0; | |
255 | ||
256 | lockdep_assert_held(&rb->w_lock); | |
257 | ||
258 | /* Update l2p only as buffer entries are being overwritten */ | |
259 | space = pblk_rb_ring_space(rb, mem, rb->l2p_update, rb->nr_entries); | |
260 | if (space > nr_entries) | |
261 | goto out; | |
262 | ||
263 | count = nr_entries - space; | |
264 | /* l2p_update used exclusively under rb->w_lock */ | |
05ed3447 | 265 | ret = __pblk_rb_update_l2p(rb, count); |
a4bd217b JG |
266 | |
267 | out: | |
268 | return ret; | |
269 | } | |
270 | ||
271 | /* | |
272 | * Update the l2p entry for all sectors stored on the write buffer. This means | |
273 | * that all future lookups to the l2p table will point to a device address, not | |
274 | * to the cacheline in the write buffer. | |
275 | */ | |
276 | void pblk_rb_sync_l2p(struct pblk_rb *rb) | |
277 | { | |
278 | unsigned int sync; | |
279 | unsigned int to_update; | |
280 | ||
281 | spin_lock(&rb->w_lock); | |
282 | ||
283 | /* Protect from reads and writes */ | |
284 | sync = smp_load_acquire(&rb->sync); | |
285 | ||
286 | to_update = pblk_rb_ring_count(sync, rb->l2p_update, rb->nr_entries); | |
05ed3447 | 287 | __pblk_rb_update_l2p(rb, to_update); |
a4bd217b JG |
288 | |
289 | spin_unlock(&rb->w_lock); | |
290 | } | |
291 | ||
292 | /* | |
293 | * Write @nr_entries to ring buffer from @data buffer if there is enough space. | |
294 | * Typically, 4KB data chunks coming from a bio will be copied to the ring | |
295 | * buffer, thus the write will fail if not all incoming data can be copied. | |
296 | * | |
297 | */ | |
298 | static void __pblk_rb_write_entry(struct pblk_rb *rb, void *data, | |
299 | struct pblk_w_ctx w_ctx, | |
300 | struct pblk_rb_entry *entry) | |
301 | { | |
302 | memcpy(entry->data, data, rb->seg_size); | |
303 | ||
304 | entry->w_ctx.lba = w_ctx.lba; | |
305 | entry->w_ctx.ppa = w_ctx.ppa; | |
306 | } | |
307 | ||
308 | void pblk_rb_write_entry_user(struct pblk_rb *rb, void *data, | |
309 | struct pblk_w_ctx w_ctx, unsigned int ring_pos) | |
310 | { | |
311 | struct pblk *pblk = container_of(rb, struct pblk, rwb); | |
312 | struct pblk_rb_entry *entry; | |
313 | int flags; | |
314 | ||
315 | entry = &rb->entries[ring_pos]; | |
316 | flags = READ_ONCE(entry->w_ctx.flags); | |
880eda54 | 317 | #ifdef CONFIG_NVM_PBLK_DEBUG |
a4bd217b JG |
318 | /* Caller must guarantee that the entry is free */ |
319 | BUG_ON(!(flags & PBLK_WRITABLE_ENTRY)); | |
320 | #endif | |
321 | ||
322 | __pblk_rb_write_entry(rb, data, w_ctx, entry); | |
323 | ||
324 | pblk_update_map_cache(pblk, w_ctx.lba, entry->cacheline); | |
325 | flags = w_ctx.flags | PBLK_WRITTEN_DATA; | |
326 | ||
327 | /* Release flags on write context. Protect from writes */ | |
328 | smp_store_release(&entry->w_ctx.flags, flags); | |
329 | } | |
330 | ||
331 | void pblk_rb_write_entry_gc(struct pblk_rb *rb, void *data, | |
d340121e JG |
332 | struct pblk_w_ctx w_ctx, struct pblk_line *line, |
333 | u64 paddr, unsigned int ring_pos) | |
a4bd217b JG |
334 | { |
335 | struct pblk *pblk = container_of(rb, struct pblk, rwb); | |
336 | struct pblk_rb_entry *entry; | |
337 | int flags; | |
338 | ||
339 | entry = &rb->entries[ring_pos]; | |
340 | flags = READ_ONCE(entry->w_ctx.flags); | |
880eda54 | 341 | #ifdef CONFIG_NVM_PBLK_DEBUG |
a4bd217b JG |
342 | /* Caller must guarantee that the entry is free */ |
343 | BUG_ON(!(flags & PBLK_WRITABLE_ENTRY)); | |
344 | #endif | |
345 | ||
346 | __pblk_rb_write_entry(rb, data, w_ctx, entry); | |
347 | ||
d340121e | 348 | if (!pblk_update_map_gc(pblk, w_ctx.lba, entry->cacheline, line, paddr)) |
a4bd217b JG |
349 | entry->w_ctx.lba = ADDR_EMPTY; |
350 | ||
351 | flags = w_ctx.flags | PBLK_WRITTEN_DATA; | |
352 | ||
353 | /* Release flags on write context. Protect from writes */ | |
354 | smp_store_release(&entry->w_ctx.flags, flags); | |
355 | } | |
356 | ||
8154d296 | 357 | static int pblk_rb_flush_point_set(struct pblk_rb *rb, struct bio *bio, |
cc9c9a00 | 358 | unsigned int pos) |
a4bd217b JG |
359 | { |
360 | struct pblk_rb_entry *entry; | |
533657c1 | 361 | unsigned int sync, flush_point; |
a4bd217b | 362 | |
3c05ef11 | 363 | pblk_rb_sync_init(rb, NULL); |
533657c1 HH |
364 | sync = READ_ONCE(rb->sync); |
365 | ||
3c05ef11 HH |
366 | if (pos == sync) { |
367 | pblk_rb_sync_end(rb, NULL); | |
533657c1 | 368 | return 0; |
3c05ef11 | 369 | } |
a4bd217b | 370 | |
880eda54 | 371 | #ifdef CONFIG_NVM_PBLK_DEBUG |
8154d296 | 372 | atomic_inc(&rb->inflight_flush_point); |
a4bd217b JG |
373 | #endif |
374 | ||
8154d296 HH |
375 | flush_point = (pos == 0) ? (rb->nr_entries - 1) : (pos - 1); |
376 | entry = &rb->entries[flush_point]; | |
a4bd217b | 377 | |
8154d296 HH |
378 | /* Protect flush points */ |
379 | smp_store_release(&rb->flush_point, flush_point); | |
a4bd217b | 380 | |
b36bbf9d HH |
381 | if (bio) |
382 | bio_list_add(&entry->w_ctx.bios, bio); | |
588726d3 | 383 | |
b36bbf9d | 384 | pblk_rb_sync_end(rb, NULL); |
a4bd217b | 385 | |
b36bbf9d | 386 | return bio ? 1 : 0; |
a4bd217b JG |
387 | } |
388 | ||
389 | static int __pblk_rb_may_write(struct pblk_rb *rb, unsigned int nr_entries, | |
390 | unsigned int *pos) | |
391 | { | |
392 | unsigned int mem; | |
393 | unsigned int sync; | |
394 | ||
395 | sync = READ_ONCE(rb->sync); | |
396 | mem = READ_ONCE(rb->mem); | |
397 | ||
398 | if (pblk_rb_ring_space(rb, mem, sync, rb->nr_entries) < nr_entries) | |
399 | return 0; | |
400 | ||
401 | if (pblk_rb_update_l2p(rb, nr_entries, mem, sync)) | |
402 | return 0; | |
403 | ||
404 | *pos = mem; | |
405 | ||
406 | return 1; | |
407 | } | |
408 | ||
409 | static int pblk_rb_may_write(struct pblk_rb *rb, unsigned int nr_entries, | |
410 | unsigned int *pos) | |
411 | { | |
412 | if (!__pblk_rb_may_write(rb, nr_entries, pos)) | |
413 | return 0; | |
414 | ||
415 | /* Protect from read count */ | |
40b8657d | 416 | smp_store_release(&rb->mem, pblk_rb_ptr_wrap(rb, *pos, nr_entries)); |
a4bd217b JG |
417 | return 1; |
418 | } | |
419 | ||
588726d3 JG |
420 | void pblk_rb_flush(struct pblk_rb *rb) |
421 | { | |
422 | struct pblk *pblk = container_of(rb, struct pblk, rwb); | |
423 | unsigned int mem = READ_ONCE(rb->mem); | |
424 | ||
8154d296 | 425 | if (pblk_rb_flush_point_set(rb, NULL, mem)) |
588726d3 JG |
426 | return; |
427 | ||
cc9c9a00 | 428 | pblk_write_kick(pblk); |
588726d3 JG |
429 | } |
430 | ||
a4bd217b JG |
431 | static int pblk_rb_may_write_flush(struct pblk_rb *rb, unsigned int nr_entries, |
432 | unsigned int *pos, struct bio *bio, | |
433 | int *io_ret) | |
434 | { | |
435 | unsigned int mem; | |
436 | ||
437 | if (!__pblk_rb_may_write(rb, nr_entries, pos)) | |
438 | return 0; | |
439 | ||
40b8657d | 440 | mem = pblk_rb_ptr_wrap(rb, *pos, nr_entries); |
a4bd217b JG |
441 | *io_ret = NVM_IO_DONE; |
442 | ||
443 | if (bio->bi_opf & REQ_PREFLUSH) { | |
444 | struct pblk *pblk = container_of(rb, struct pblk, rwb); | |
445 | ||
5d149bfa | 446 | atomic64_inc(&pblk->nr_flush); |
8154d296 | 447 | if (pblk_rb_flush_point_set(&pblk->rwb, bio, mem)) |
a4bd217b JG |
448 | *io_ret = NVM_IO_OK; |
449 | } | |
450 | ||
451 | /* Protect from read count */ | |
452 | smp_store_release(&rb->mem, mem); | |
6ca2f71f | 453 | |
a4bd217b JG |
454 | return 1; |
455 | } | |
456 | ||
457 | /* | |
458 | * Atomically check that (i) there is space on the write buffer for the | |
459 | * incoming I/O, and (ii) the current I/O type has enough budget in the write | |
460 | * buffer (rate-limiter). | |
461 | */ | |
462 | int pblk_rb_may_write_user(struct pblk_rb *rb, struct bio *bio, | |
463 | unsigned int nr_entries, unsigned int *pos) | |
464 | { | |
465 | struct pblk *pblk = container_of(rb, struct pblk, rwb); | |
588726d3 | 466 | int io_ret; |
a4bd217b JG |
467 | |
468 | spin_lock(&rb->w_lock); | |
588726d3 JG |
469 | io_ret = pblk_rl_user_may_insert(&pblk->rl, nr_entries); |
470 | if (io_ret) { | |
a4bd217b | 471 | spin_unlock(&rb->w_lock); |
588726d3 | 472 | return io_ret; |
a4bd217b JG |
473 | } |
474 | ||
588726d3 | 475 | if (!pblk_rb_may_write_flush(rb, nr_entries, pos, bio, &io_ret)) { |
a4bd217b JG |
476 | spin_unlock(&rb->w_lock); |
477 | return NVM_IO_REQUEUE; | |
478 | } | |
479 | ||
480 | pblk_rl_user_in(&pblk->rl, nr_entries); | |
481 | spin_unlock(&rb->w_lock); | |
482 | ||
588726d3 | 483 | return io_ret; |
a4bd217b JG |
484 | } |
485 | ||
486 | /* | |
487 | * Look at pblk_rb_may_write_user comment | |
488 | */ | |
489 | int pblk_rb_may_write_gc(struct pblk_rb *rb, unsigned int nr_entries, | |
490 | unsigned int *pos) | |
491 | { | |
492 | struct pblk *pblk = container_of(rb, struct pblk, rwb); | |
493 | ||
494 | spin_lock(&rb->w_lock); | |
495 | if (!pblk_rl_gc_may_insert(&pblk->rl, nr_entries)) { | |
496 | spin_unlock(&rb->w_lock); | |
497 | return 0; | |
498 | } | |
499 | ||
500 | if (!pblk_rb_may_write(rb, nr_entries, pos)) { | |
501 | spin_unlock(&rb->w_lock); | |
502 | return 0; | |
503 | } | |
504 | ||
505 | pblk_rl_gc_in(&pblk->rl, nr_entries); | |
506 | spin_unlock(&rb->w_lock); | |
507 | ||
508 | return 1; | |
509 | } | |
510 | ||
a4bd217b JG |
511 | /* |
512 | * Read available entries on rb and add them to the given bio. To avoid a memory | |
513 | * copy, a page reference to the write buffer is used to be added to the bio. | |
514 | * | |
515 | * This function is used by the write thread to form the write bio that will | |
516 | * persist data on the write buffer to the media. | |
517 | */ | |
d624f371 | 518 | unsigned int pblk_rb_read_to_bio(struct pblk_rb *rb, struct nvm_rq *rqd, |
875d94f3 JG |
519 | unsigned int pos, unsigned int nr_entries, |
520 | unsigned int count) | |
a4bd217b JG |
521 | { |
522 | struct pblk *pblk = container_of(rb, struct pblk, rwb); | |
d624f371 JG |
523 | struct request_queue *q = pblk->dev->q; |
524 | struct pblk_c_ctx *c_ctx = nvm_rq_to_pdu(rqd); | |
875d94f3 | 525 | struct bio *bio = rqd->bio; |
a4bd217b JG |
526 | struct pblk_rb_entry *entry; |
527 | struct page *page; | |
d624f371 | 528 | unsigned int pad = 0, to_read = nr_entries; |
a4bd217b JG |
529 | unsigned int i; |
530 | int flags; | |
a4bd217b JG |
531 | |
532 | if (count < nr_entries) { | |
533 | pad = nr_entries - count; | |
534 | to_read = count; | |
535 | } | |
536 | ||
537 | c_ctx->sentry = pos; | |
538 | c_ctx->nr_valid = to_read; | |
539 | c_ctx->nr_padded = pad; | |
540 | ||
541 | for (i = 0; i < to_read; i++) { | |
542 | entry = &rb->entries[pos]; | |
543 | ||
544 | /* A write has been allowed into the buffer, but data is still | |
545 | * being copied to it. It is ok to busy wait. | |
546 | */ | |
547 | try: | |
548 | flags = READ_ONCE(entry->w_ctx.flags); | |
10888129 JG |
549 | if (!(flags & PBLK_WRITTEN_DATA)) { |
550 | io_schedule(); | |
a4bd217b | 551 | goto try; |
10888129 | 552 | } |
a4bd217b | 553 | |
a4bd217b JG |
554 | page = virt_to_page(entry->data); |
555 | if (!page) { | |
4e495a46 | 556 | pblk_err(pblk, "could not allocate write bio page\n"); |
a4bd217b JG |
557 | flags &= ~PBLK_WRITTEN_DATA; |
558 | flags |= PBLK_SUBMITTED_ENTRY; | |
559 | /* Release flags on context. Protect from writes */ | |
560 | smp_store_release(&entry->w_ctx.flags, flags); | |
d624f371 | 561 | return NVM_IO_ERR; |
a4bd217b JG |
562 | } |
563 | ||
d624f371 JG |
564 | if (bio_add_pc_page(q, bio, page, rb->seg_size, 0) != |
565 | rb->seg_size) { | |
4e495a46 | 566 | pblk_err(pblk, "could not add page to write bio\n"); |
a4bd217b JG |
567 | flags &= ~PBLK_WRITTEN_DATA; |
568 | flags |= PBLK_SUBMITTED_ENTRY; | |
569 | /* Release flags on context. Protect from writes */ | |
570 | smp_store_release(&entry->w_ctx.flags, flags); | |
d624f371 | 571 | return NVM_IO_ERR; |
a4bd217b JG |
572 | } |
573 | ||
a4bd217b JG |
574 | flags &= ~PBLK_WRITTEN_DATA; |
575 | flags |= PBLK_SUBMITTED_ENTRY; | |
576 | ||
577 | /* Release flags on context. Protect from writes */ | |
578 | smp_store_release(&entry->w_ctx.flags, flags); | |
579 | ||
40b8657d | 580 | pos = pblk_rb_ptr_wrap(rb, pos, 1); |
a4bd217b JG |
581 | } |
582 | ||
d624f371 JG |
583 | if (pad) { |
584 | if (pblk_bio_add_pages(pblk, bio, GFP_KERNEL, pad)) { | |
4e495a46 | 585 | pblk_err(pblk, "could not pad page in write bio\n"); |
d624f371 JG |
586 | return NVM_IO_ERR; |
587 | } | |
d624f371 | 588 | |
5d149bfa HH |
589 | if (pad < pblk->min_write_pgs) |
590 | atomic64_inc(&pblk->pad_dist[pad - 1]); | |
591 | else | |
4e495a46 | 592 | pblk_warn(pblk, "padding more than min. sectors\n"); |
5d149bfa HH |
593 | |
594 | atomic64_add(pad, &pblk->pad_wa); | |
595 | } | |
76758390 | 596 | |
880eda54 | 597 | #ifdef CONFIG_NVM_PBLK_DEBUG |
5d149bfa | 598 | atomic_long_add(pad, &pblk->padded_writes); |
a4bd217b | 599 | #endif |
d624f371 JG |
600 | |
601 | return NVM_IO_OK; | |
a4bd217b JG |
602 | } |
603 | ||
604 | /* | |
605 | * Copy to bio only if the lba matches the one on the given cache entry. | |
606 | * Otherwise, it means that the entry has been overwritten, and the bio should | |
607 | * be directed to disk. | |
608 | */ | |
609 | int pblk_rb_copy_to_bio(struct pblk_rb *rb, struct bio *bio, sector_t lba, | |
75cb8e93 | 610 | struct ppa_addr ppa, int bio_iter, bool advanced_bio) |
a4bd217b | 611 | { |
07698466 | 612 | struct pblk *pblk = container_of(rb, struct pblk, rwb); |
a4bd217b JG |
613 | struct pblk_rb_entry *entry; |
614 | struct pblk_w_ctx *w_ctx; | |
07698466 JG |
615 | struct ppa_addr l2p_ppa; |
616 | u64 pos = pblk_addr_to_cacheline(ppa); | |
a4bd217b JG |
617 | void *data; |
618 | int flags; | |
619 | int ret = 1; | |
620 | ||
a4bd217b | 621 | |
880eda54 | 622 | #ifdef CONFIG_NVM_PBLK_DEBUG |
a4bd217b JG |
623 | /* Caller must ensure that the access will not cause an overflow */ |
624 | BUG_ON(pos >= rb->nr_entries); | |
625 | #endif | |
626 | entry = &rb->entries[pos]; | |
627 | w_ctx = &entry->w_ctx; | |
628 | flags = READ_ONCE(w_ctx->flags); | |
629 | ||
07698466 JG |
630 | spin_lock(&rb->w_lock); |
631 | spin_lock(&pblk->trans_lock); | |
632 | l2p_ppa = pblk_trans_map_get(pblk, lba); | |
633 | spin_unlock(&pblk->trans_lock); | |
634 | ||
a4bd217b | 635 | /* Check if the entry has been overwritten or is scheduled to be */ |
07698466 JG |
636 | if (!pblk_ppa_comp(l2p_ppa, ppa) || w_ctx->lba != lba || |
637 | flags & PBLK_WRITABLE_ENTRY) { | |
a4bd217b JG |
638 | ret = 0; |
639 | goto out; | |
640 | } | |
641 | ||
642 | /* Only advance the bio if it hasn't been advanced already. If advanced, | |
643 | * this bio is at least a partial bio (i.e., it has partially been | |
644 | * filled with data from the cache). If part of the data resides on the | |
645 | * media, we will read later on | |
646 | */ | |
75cb8e93 | 647 | if (unlikely(!advanced_bio)) |
a4bd217b JG |
648 | bio_advance(bio, bio_iter * PBLK_EXPOSED_PAGE_SIZE); |
649 | ||
650 | data = bio_data(bio); | |
651 | memcpy(data, entry->data, rb->seg_size); | |
652 | ||
653 | out: | |
654 | spin_unlock(&rb->w_lock); | |
655 | return ret; | |
656 | } | |
657 | ||
658 | struct pblk_w_ctx *pblk_rb_w_ctx(struct pblk_rb *rb, unsigned int pos) | |
659 | { | |
40b8657d | 660 | unsigned int entry = pblk_rb_ptr_wrap(rb, pos, 0); |
a4bd217b JG |
661 | |
662 | return &rb->entries[entry].w_ctx; | |
663 | } | |
664 | ||
665 | unsigned int pblk_rb_sync_init(struct pblk_rb *rb, unsigned long *flags) | |
666 | __acquires(&rb->s_lock) | |
667 | { | |
668 | if (flags) | |
669 | spin_lock_irqsave(&rb->s_lock, *flags); | |
670 | else | |
671 | spin_lock_irq(&rb->s_lock); | |
672 | ||
673 | return rb->sync; | |
674 | } | |
675 | ||
676 | void pblk_rb_sync_end(struct pblk_rb *rb, unsigned long *flags) | |
677 | __releases(&rb->s_lock) | |
678 | { | |
679 | lockdep_assert_held(&rb->s_lock); | |
680 | ||
681 | if (flags) | |
682 | spin_unlock_irqrestore(&rb->s_lock, *flags); | |
683 | else | |
684 | spin_unlock_irq(&rb->s_lock); | |
685 | } | |
686 | ||
687 | unsigned int pblk_rb_sync_advance(struct pblk_rb *rb, unsigned int nr_entries) | |
688 | { | |
533657c1 | 689 | unsigned int sync, flush_point; |
a4bd217b JG |
690 | lockdep_assert_held(&rb->s_lock); |
691 | ||
692 | sync = READ_ONCE(rb->sync); | |
533657c1 HH |
693 | flush_point = READ_ONCE(rb->flush_point); |
694 | ||
695 | if (flush_point != EMPTY_ENTRY) { | |
696 | unsigned int secs_to_flush; | |
a4bd217b | 697 | |
533657c1 HH |
698 | secs_to_flush = pblk_rb_ring_count(flush_point, sync, |
699 | rb->nr_entries); | |
700 | if (secs_to_flush < nr_entries) { | |
701 | /* Protect flush points */ | |
702 | smp_store_release(&rb->flush_point, EMPTY_ENTRY); | |
703 | } | |
704 | } | |
705 | ||
40b8657d | 706 | sync = pblk_rb_ptr_wrap(rb, sync, nr_entries); |
a4bd217b JG |
707 | |
708 | /* Protect from counts */ | |
709 | smp_store_release(&rb->sync, sync); | |
710 | ||
711 | return sync; | |
712 | } | |
713 | ||
533657c1 | 714 | /* Calculate how many sectors to submit up to the current flush point. */ |
8154d296 | 715 | unsigned int pblk_rb_flush_point_count(struct pblk_rb *rb) |
a4bd217b | 716 | { |
533657c1 HH |
717 | unsigned int subm, sync, flush_point; |
718 | unsigned int submitted, to_flush; | |
a4bd217b | 719 | |
8154d296 HH |
720 | /* Protect flush points */ |
721 | flush_point = smp_load_acquire(&rb->flush_point); | |
722 | if (flush_point == EMPTY_ENTRY) | |
a4bd217b JG |
723 | return 0; |
724 | ||
533657c1 HH |
725 | /* Protect syncs */ |
726 | sync = smp_load_acquire(&rb->sync); | |
727 | ||
a4bd217b | 728 | subm = READ_ONCE(rb->subm); |
533657c1 | 729 | submitted = pblk_rb_ring_count(subm, sync, rb->nr_entries); |
a4bd217b JG |
730 | |
731 | /* The sync point itself counts as a sector to sync */ | |
533657c1 | 732 | to_flush = pblk_rb_ring_count(flush_point, sync, rb->nr_entries) + 1; |
a4bd217b | 733 | |
533657c1 | 734 | return (submitted < to_flush) ? (to_flush - submitted) : 0; |
a4bd217b JG |
735 | } |
736 | ||
a4bd217b JG |
737 | int pblk_rb_tear_down_check(struct pblk_rb *rb) |
738 | { | |
739 | struct pblk_rb_entry *entry; | |
740 | int i; | |
741 | int ret = 0; | |
742 | ||
743 | spin_lock(&rb->w_lock); | |
744 | spin_lock_irq(&rb->s_lock); | |
745 | ||
746 | if ((rb->mem == rb->subm) && (rb->subm == rb->sync) && | |
747 | (rb->sync == rb->l2p_update) && | |
8154d296 | 748 | (rb->flush_point == EMPTY_ENTRY)) { |
a4bd217b JG |
749 | goto out; |
750 | } | |
751 | ||
752 | if (!rb->entries) { | |
753 | ret = 1; | |
754 | goto out; | |
755 | } | |
756 | ||
757 | for (i = 0; i < rb->nr_entries; i++) { | |
758 | entry = &rb->entries[i]; | |
759 | ||
760 | if (!entry->data) { | |
761 | ret = 1; | |
762 | goto out; | |
763 | } | |
764 | } | |
765 | ||
766 | out: | |
767 | spin_unlock(&rb->w_lock); | |
768 | spin_unlock_irq(&rb->s_lock); | |
769 | ||
770 | return ret; | |
771 | } | |
772 | ||
773 | unsigned int pblk_rb_wrap_pos(struct pblk_rb *rb, unsigned int pos) | |
774 | { | |
775 | return (pos & (rb->nr_entries - 1)); | |
776 | } | |
777 | ||
778 | int pblk_rb_pos_oob(struct pblk_rb *rb, u64 pos) | |
779 | { | |
780 | return (pos >= rb->nr_entries); | |
781 | } | |
782 | ||
783 | ssize_t pblk_rb_sysfs(struct pblk_rb *rb, char *buf) | |
784 | { | |
785 | struct pblk *pblk = container_of(rb, struct pblk, rwb); | |
786 | struct pblk_c_ctx *c; | |
787 | ssize_t offset; | |
788 | int queued_entries = 0; | |
789 | ||
790 | spin_lock_irq(&rb->s_lock); | |
791 | list_for_each_entry(c, &pblk->compl_list, list) | |
792 | queued_entries++; | |
793 | spin_unlock_irq(&rb->s_lock); | |
794 | ||
8154d296 | 795 | if (rb->flush_point != EMPTY_ENTRY) |
a4bd217b JG |
796 | offset = scnprintf(buf, PAGE_SIZE, |
797 | "%u\t%u\t%u\t%u\t%u\t%u\t%u - %u/%u/%u - %d\n", | |
798 | rb->nr_entries, | |
799 | rb->mem, | |
800 | rb->subm, | |
801 | rb->sync, | |
802 | rb->l2p_update, | |
880eda54 | 803 | #ifdef CONFIG_NVM_PBLK_DEBUG |
8154d296 | 804 | atomic_read(&rb->inflight_flush_point), |
a4bd217b JG |
805 | #else |
806 | 0, | |
807 | #endif | |
8154d296 | 808 | rb->flush_point, |
a4bd217b JG |
809 | pblk_rb_read_count(rb), |
810 | pblk_rb_space(rb), | |
8154d296 | 811 | pblk_rb_flush_point_count(rb), |
a4bd217b JG |
812 | queued_entries); |
813 | else | |
814 | offset = scnprintf(buf, PAGE_SIZE, | |
815 | "%u\t%u\t%u\t%u\t%u\t%u\tNULL - %u/%u/%u - %d\n", | |
816 | rb->nr_entries, | |
817 | rb->mem, | |
818 | rb->subm, | |
819 | rb->sync, | |
820 | rb->l2p_update, | |
880eda54 | 821 | #ifdef CONFIG_NVM_PBLK_DEBUG |
8154d296 | 822 | atomic_read(&rb->inflight_flush_point), |
a4bd217b JG |
823 | #else |
824 | 0, | |
825 | #endif | |
826 | pblk_rb_read_count(rb), | |
827 | pblk_rb_space(rb), | |
8154d296 | 828 | pblk_rb_flush_point_count(rb), |
a4bd217b JG |
829 | queued_entries); |
830 | ||
831 | return offset; | |
832 | } |