]>
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 | ||
9bd1f875 | 26 | static void pblk_rb_data_free(struct pblk_rb *rb) |
a4bd217b JG |
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 | ||
9bd1f875 JG |
39 | void pblk_rb_free(struct pblk_rb *rb) |
40 | { | |
41 | pblk_rb_data_free(rb); | |
42 | vfree(rb->entries); | |
43 | } | |
44 | ||
45 | /* | |
46 | * pblk_rb_calculate_size -- calculate the size of the write buffer | |
47 | */ | |
48 | static unsigned int pblk_rb_calculate_size(unsigned int nr_entries) | |
49 | { | |
50 | /* Alloc a write buffer that can at least fit 128 entries */ | |
51 | return (1 << max(get_count_order(nr_entries), 7)); | |
52 | } | |
53 | ||
a4bd217b JG |
54 | /* |
55 | * Initialize ring buffer. The data and metadata buffers must be previously | |
56 | * allocated and their size must be a power of two | |
5fb94e9c | 57 | * (Documentation/core-api/circular-buffers.rst) |
a4bd217b | 58 | */ |
766c8ceb JG |
59 | int pblk_rb_init(struct pblk_rb *rb, unsigned int size, unsigned int threshold, |
60 | unsigned int seg_size) | |
a4bd217b JG |
61 | { |
62 | struct pblk *pblk = container_of(rb, struct pblk, rwb); | |
9bd1f875 | 63 | struct pblk_rb_entry *entries; |
a4bd217b | 64 | unsigned int init_entry = 0; |
a4bd217b | 65 | unsigned int max_order = MAX_ORDER - 1; |
9bd1f875 JG |
66 | unsigned int power_size, power_seg_sz; |
67 | unsigned int alloc_order, order, iter; | |
68 | unsigned int nr_entries; | |
69 | ||
70 | nr_entries = pblk_rb_calculate_size(size); | |
71 | entries = vzalloc(array_size(nr_entries, sizeof(struct pblk_rb_entry))); | |
72 | if (!entries) | |
73 | return -ENOMEM; | |
74 | ||
75 | power_size = get_count_order(size); | |
76 | power_seg_sz = get_count_order(seg_size); | |
a4bd217b JG |
77 | |
78 | down_write(&pblk_rb_lock); | |
9bd1f875 | 79 | rb->entries = entries; |
a4bd217b JG |
80 | rb->seg_size = (1 << power_seg_sz); |
81 | rb->nr_entries = (1 << power_size); | |
82 | rb->mem = rb->subm = rb->sync = rb->l2p_update = 0; | |
766c8ceb | 83 | rb->back_thres = threshold; |
8154d296 | 84 | rb->flush_point = EMPTY_ENTRY; |
a4bd217b JG |
85 | |
86 | spin_lock_init(&rb->w_lock); | |
87 | spin_lock_init(&rb->s_lock); | |
88 | ||
89 | INIT_LIST_HEAD(&rb->pages); | |
90 | ||
9bd1f875 | 91 | alloc_order = power_size; |
a4bd217b JG |
92 | if (alloc_order >= max_order) { |
93 | order = max_order; | |
94 | iter = (1 << (alloc_order - max_order)); | |
95 | } else { | |
96 | order = alloc_order; | |
97 | iter = 1; | |
98 | } | |
99 | ||
100 | do { | |
101 | struct pblk_rb_entry *entry; | |
102 | struct pblk_rb_pages *page_set; | |
103 | void *kaddr; | |
104 | unsigned long set_size; | |
105 | int i; | |
106 | ||
107 | page_set = kmalloc(sizeof(struct pblk_rb_pages), GFP_KERNEL); | |
108 | if (!page_set) { | |
109 | up_write(&pblk_rb_lock); | |
9bd1f875 | 110 | vfree(entries); |
a4bd217b JG |
111 | return -ENOMEM; |
112 | } | |
113 | ||
114 | page_set->order = order; | |
115 | page_set->pages = alloc_pages(GFP_KERNEL, order); | |
116 | if (!page_set->pages) { | |
117 | kfree(page_set); | |
118 | pblk_rb_data_free(rb); | |
119 | up_write(&pblk_rb_lock); | |
9bd1f875 | 120 | vfree(entries); |
a4bd217b JG |
121 | return -ENOMEM; |
122 | } | |
123 | kaddr = page_address(page_set->pages); | |
124 | ||
125 | entry = &rb->entries[init_entry]; | |
126 | entry->data = kaddr; | |
127 | entry->cacheline = pblk_cacheline_to_addr(init_entry++); | |
128 | entry->w_ctx.flags = PBLK_WRITABLE_ENTRY; | |
129 | ||
130 | set_size = (1 << order); | |
131 | for (i = 1; i < set_size; i++) { | |
132 | entry = &rb->entries[init_entry]; | |
133 | entry->cacheline = pblk_cacheline_to_addr(init_entry++); | |
134 | entry->data = kaddr + (i * rb->seg_size); | |
135 | entry->w_ctx.flags = PBLK_WRITABLE_ENTRY; | |
136 | bio_list_init(&entry->w_ctx.bios); | |
137 | } | |
138 | ||
139 | list_add_tail(&page_set->list, &rb->pages); | |
140 | iter--; | |
141 | } while (iter > 0); | |
142 | up_write(&pblk_rb_lock); | |
143 | ||
880eda54 | 144 | #ifdef CONFIG_NVM_PBLK_DEBUG |
8154d296 | 145 | atomic_set(&rb->inflight_flush_point, 0); |
a4bd217b JG |
146 | #endif |
147 | ||
148 | /* | |
149 | * Initialize rate-limiter, which controls access to the write buffer | |
6e82f0ba | 150 | * by user and GC I/O |
a4bd217b JG |
151 | */ |
152 | pblk_rl_init(&pblk->rl, rb->nr_entries); | |
153 | ||
154 | return 0; | |
155 | } | |
156 | ||
a4bd217b JG |
157 | static void clean_wctx(struct pblk_w_ctx *w_ctx) |
158 | { | |
159 | int flags; | |
160 | ||
a4bd217b | 161 | flags = READ_ONCE(w_ctx->flags); |
e37d0798 JG |
162 | WARN_ONCE(!(flags & PBLK_SUBMITTED_ENTRY), |
163 | "pblk: overwriting unsubmitted data\n"); | |
a4bd217b JG |
164 | |
165 | /* Release flags on context. Protect from writes and reads */ | |
166 | smp_store_release(&w_ctx->flags, PBLK_WRITABLE_ENTRY); | |
167 | pblk_ppa_set_empty(&w_ctx->ppa); | |
07698466 | 168 | w_ctx->lba = ADDR_EMPTY; |
a4bd217b JG |
169 | } |
170 | ||
171 | #define pblk_rb_ring_count(head, tail, size) CIRC_CNT(head, tail, size) | |
172 | #define pblk_rb_ring_space(rb, head, tail, size) \ | |
173 | (CIRC_SPACE(head, tail, size)) | |
174 | ||
175 | /* | |
176 | * Buffer space is calculated with respect to the back pointer signaling | |
177 | * synchronized entries to the media. | |
178 | */ | |
179 | static unsigned int pblk_rb_space(struct pblk_rb *rb) | |
180 | { | |
181 | unsigned int mem = READ_ONCE(rb->mem); | |
182 | unsigned int sync = READ_ONCE(rb->sync); | |
183 | ||
184 | return pblk_rb_ring_space(rb, mem, sync, rb->nr_entries); | |
185 | } | |
186 | ||
40b8657d JG |
187 | unsigned int pblk_rb_ptr_wrap(struct pblk_rb *rb, unsigned int p, |
188 | unsigned int nr_entries) | |
189 | { | |
190 | return (p + nr_entries) & (rb->nr_entries - 1); | |
191 | } | |
192 | ||
a4bd217b JG |
193 | /* |
194 | * Buffer count is calculated with respect to the submission entry signaling the | |
195 | * entries that are available to send to the media | |
196 | */ | |
197 | unsigned int pblk_rb_read_count(struct pblk_rb *rb) | |
198 | { | |
199 | unsigned int mem = READ_ONCE(rb->mem); | |
200 | unsigned int subm = READ_ONCE(rb->subm); | |
201 | ||
202 | return pblk_rb_ring_count(mem, subm, rb->nr_entries); | |
203 | } | |
204 | ||
ee8d5c1a JG |
205 | unsigned int pblk_rb_sync_count(struct pblk_rb *rb) |
206 | { | |
207 | unsigned int mem = READ_ONCE(rb->mem); | |
208 | unsigned int sync = READ_ONCE(rb->sync); | |
209 | ||
210 | return pblk_rb_ring_count(mem, sync, rb->nr_entries); | |
211 | } | |
212 | ||
a4bd217b JG |
213 | unsigned int pblk_rb_read_commit(struct pblk_rb *rb, unsigned int nr_entries) |
214 | { | |
215 | unsigned int subm; | |
216 | ||
217 | subm = READ_ONCE(rb->subm); | |
218 | /* Commit read means updating submission pointer */ | |
40b8657d | 219 | smp_store_release(&rb->subm, pblk_rb_ptr_wrap(rb, subm, nr_entries)); |
a4bd217b JG |
220 | |
221 | return subm; | |
222 | } | |
223 | ||
05ed3447 | 224 | static int __pblk_rb_update_l2p(struct pblk_rb *rb, unsigned int to_update) |
a4bd217b JG |
225 | { |
226 | struct pblk *pblk = container_of(rb, struct pblk, rwb); | |
227 | struct pblk_line *line; | |
228 | struct pblk_rb_entry *entry; | |
229 | struct pblk_w_ctx *w_ctx; | |
b20ba1bc | 230 | unsigned int user_io = 0, gc_io = 0; |
a4bd217b | 231 | unsigned int i; |
b20ba1bc | 232 | int flags; |
a4bd217b JG |
233 | |
234 | for (i = 0; i < to_update; i++) { | |
05ed3447 | 235 | entry = &rb->entries[rb->l2p_update]; |
a4bd217b JG |
236 | w_ctx = &entry->w_ctx; |
237 | ||
b20ba1bc JG |
238 | flags = READ_ONCE(entry->w_ctx.flags); |
239 | if (flags & PBLK_IOTYPE_USER) | |
240 | user_io++; | |
241 | else if (flags & PBLK_IOTYPE_GC) | |
242 | gc_io++; | |
243 | else | |
244 | WARN(1, "pblk: unknown IO type\n"); | |
245 | ||
a4bd217b JG |
246 | pblk_update_map_dev(pblk, w_ctx->lba, w_ctx->ppa, |
247 | entry->cacheline); | |
248 | ||
cb21665c | 249 | line = pblk_ppa_to_line(pblk, w_ctx->ppa); |
a4bd217b JG |
250 | kref_put(&line->ref, pblk_line_put); |
251 | clean_wctx(w_ctx); | |
40b8657d | 252 | rb->l2p_update = pblk_rb_ptr_wrap(rb, rb->l2p_update, 1); |
a4bd217b JG |
253 | } |
254 | ||
b20ba1bc JG |
255 | pblk_rl_out(&pblk->rl, user_io, gc_io); |
256 | ||
a4bd217b JG |
257 | return 0; |
258 | } | |
259 | ||
260 | /* | |
261 | * When we move the l2p_update pointer, we update the l2p table - lookups will | |
262 | * point to the physical address instead of to the cacheline in the write buffer | |
263 | * from this moment on. | |
264 | */ | |
265 | static int pblk_rb_update_l2p(struct pblk_rb *rb, unsigned int nr_entries, | |
266 | unsigned int mem, unsigned int sync) | |
267 | { | |
268 | unsigned int space, count; | |
269 | int ret = 0; | |
270 | ||
271 | lockdep_assert_held(&rb->w_lock); | |
272 | ||
273 | /* Update l2p only as buffer entries are being overwritten */ | |
274 | space = pblk_rb_ring_space(rb, mem, rb->l2p_update, rb->nr_entries); | |
275 | if (space > nr_entries) | |
276 | goto out; | |
277 | ||
278 | count = nr_entries - space; | |
279 | /* l2p_update used exclusively under rb->w_lock */ | |
05ed3447 | 280 | ret = __pblk_rb_update_l2p(rb, count); |
a4bd217b JG |
281 | |
282 | out: | |
283 | return ret; | |
284 | } | |
285 | ||
286 | /* | |
287 | * Update the l2p entry for all sectors stored on the write buffer. This means | |
288 | * that all future lookups to the l2p table will point to a device address, not | |
289 | * to the cacheline in the write buffer. | |
290 | */ | |
291 | void pblk_rb_sync_l2p(struct pblk_rb *rb) | |
292 | { | |
293 | unsigned int sync; | |
294 | unsigned int to_update; | |
295 | ||
296 | spin_lock(&rb->w_lock); | |
297 | ||
298 | /* Protect from reads and writes */ | |
299 | sync = smp_load_acquire(&rb->sync); | |
300 | ||
301 | to_update = pblk_rb_ring_count(sync, rb->l2p_update, rb->nr_entries); | |
05ed3447 | 302 | __pblk_rb_update_l2p(rb, to_update); |
a4bd217b JG |
303 | |
304 | spin_unlock(&rb->w_lock); | |
305 | } | |
306 | ||
307 | /* | |
308 | * Write @nr_entries to ring buffer from @data buffer if there is enough space. | |
309 | * Typically, 4KB data chunks coming from a bio will be copied to the ring | |
310 | * buffer, thus the write will fail if not all incoming data can be copied. | |
311 | * | |
312 | */ | |
313 | static void __pblk_rb_write_entry(struct pblk_rb *rb, void *data, | |
314 | struct pblk_w_ctx w_ctx, | |
315 | struct pblk_rb_entry *entry) | |
316 | { | |
317 | memcpy(entry->data, data, rb->seg_size); | |
318 | ||
319 | entry->w_ctx.lba = w_ctx.lba; | |
320 | entry->w_ctx.ppa = w_ctx.ppa; | |
321 | } | |
322 | ||
323 | void pblk_rb_write_entry_user(struct pblk_rb *rb, void *data, | |
324 | struct pblk_w_ctx w_ctx, unsigned int ring_pos) | |
325 | { | |
326 | struct pblk *pblk = container_of(rb, struct pblk, rwb); | |
327 | struct pblk_rb_entry *entry; | |
328 | int flags; | |
329 | ||
330 | entry = &rb->entries[ring_pos]; | |
331 | flags = READ_ONCE(entry->w_ctx.flags); | |
880eda54 | 332 | #ifdef CONFIG_NVM_PBLK_DEBUG |
a4bd217b JG |
333 | /* Caller must guarantee that the entry is free */ |
334 | BUG_ON(!(flags & PBLK_WRITABLE_ENTRY)); | |
335 | #endif | |
336 | ||
337 | __pblk_rb_write_entry(rb, data, w_ctx, entry); | |
338 | ||
339 | pblk_update_map_cache(pblk, w_ctx.lba, entry->cacheline); | |
340 | flags = w_ctx.flags | PBLK_WRITTEN_DATA; | |
341 | ||
342 | /* Release flags on write context. Protect from writes */ | |
343 | smp_store_release(&entry->w_ctx.flags, flags); | |
344 | } | |
345 | ||
346 | void pblk_rb_write_entry_gc(struct pblk_rb *rb, void *data, | |
d340121e JG |
347 | struct pblk_w_ctx w_ctx, struct pblk_line *line, |
348 | u64 paddr, unsigned int ring_pos) | |
a4bd217b JG |
349 | { |
350 | struct pblk *pblk = container_of(rb, struct pblk, rwb); | |
351 | struct pblk_rb_entry *entry; | |
352 | int flags; | |
353 | ||
354 | entry = &rb->entries[ring_pos]; | |
355 | flags = READ_ONCE(entry->w_ctx.flags); | |
880eda54 | 356 | #ifdef CONFIG_NVM_PBLK_DEBUG |
a4bd217b JG |
357 | /* Caller must guarantee that the entry is free */ |
358 | BUG_ON(!(flags & PBLK_WRITABLE_ENTRY)); | |
359 | #endif | |
360 | ||
361 | __pblk_rb_write_entry(rb, data, w_ctx, entry); | |
362 | ||
d340121e | 363 | if (!pblk_update_map_gc(pblk, w_ctx.lba, entry->cacheline, line, paddr)) |
a4bd217b JG |
364 | entry->w_ctx.lba = ADDR_EMPTY; |
365 | ||
366 | flags = w_ctx.flags | PBLK_WRITTEN_DATA; | |
367 | ||
368 | /* Release flags on write context. Protect from writes */ | |
369 | smp_store_release(&entry->w_ctx.flags, flags); | |
370 | } | |
371 | ||
8154d296 | 372 | static int pblk_rb_flush_point_set(struct pblk_rb *rb, struct bio *bio, |
cc9c9a00 | 373 | unsigned int pos) |
a4bd217b JG |
374 | { |
375 | struct pblk_rb_entry *entry; | |
533657c1 | 376 | unsigned int sync, flush_point; |
a4bd217b | 377 | |
3c05ef11 | 378 | pblk_rb_sync_init(rb, NULL); |
533657c1 HH |
379 | sync = READ_ONCE(rb->sync); |
380 | ||
3c05ef11 HH |
381 | if (pos == sync) { |
382 | pblk_rb_sync_end(rb, NULL); | |
533657c1 | 383 | return 0; |
3c05ef11 | 384 | } |
a4bd217b | 385 | |
880eda54 | 386 | #ifdef CONFIG_NVM_PBLK_DEBUG |
8154d296 | 387 | atomic_inc(&rb->inflight_flush_point); |
a4bd217b JG |
388 | #endif |
389 | ||
8154d296 HH |
390 | flush_point = (pos == 0) ? (rb->nr_entries - 1) : (pos - 1); |
391 | entry = &rb->entries[flush_point]; | |
a4bd217b | 392 | |
8154d296 HH |
393 | /* Protect flush points */ |
394 | smp_store_release(&rb->flush_point, flush_point); | |
a4bd217b | 395 | |
b36bbf9d HH |
396 | if (bio) |
397 | bio_list_add(&entry->w_ctx.bios, bio); | |
588726d3 | 398 | |
b36bbf9d | 399 | pblk_rb_sync_end(rb, NULL); |
a4bd217b | 400 | |
b36bbf9d | 401 | return bio ? 1 : 0; |
a4bd217b JG |
402 | } |
403 | ||
404 | static int __pblk_rb_may_write(struct pblk_rb *rb, unsigned int nr_entries, | |
405 | unsigned int *pos) | |
406 | { | |
407 | unsigned int mem; | |
408 | unsigned int sync; | |
766c8ceb | 409 | unsigned int threshold; |
a4bd217b JG |
410 | |
411 | sync = READ_ONCE(rb->sync); | |
412 | mem = READ_ONCE(rb->mem); | |
413 | ||
766c8ceb JG |
414 | threshold = nr_entries + rb->back_thres; |
415 | ||
416 | if (pblk_rb_ring_space(rb, mem, sync, rb->nr_entries) < threshold) | |
a4bd217b JG |
417 | return 0; |
418 | ||
419 | if (pblk_rb_update_l2p(rb, nr_entries, mem, sync)) | |
420 | return 0; | |
421 | ||
422 | *pos = mem; | |
423 | ||
424 | return 1; | |
425 | } | |
426 | ||
427 | static int pblk_rb_may_write(struct pblk_rb *rb, unsigned int nr_entries, | |
428 | unsigned int *pos) | |
429 | { | |
430 | if (!__pblk_rb_may_write(rb, nr_entries, pos)) | |
431 | return 0; | |
432 | ||
433 | /* Protect from read count */ | |
40b8657d | 434 | smp_store_release(&rb->mem, pblk_rb_ptr_wrap(rb, *pos, nr_entries)); |
a4bd217b JG |
435 | return 1; |
436 | } | |
437 | ||
588726d3 JG |
438 | void pblk_rb_flush(struct pblk_rb *rb) |
439 | { | |
440 | struct pblk *pblk = container_of(rb, struct pblk, rwb); | |
441 | unsigned int mem = READ_ONCE(rb->mem); | |
442 | ||
8154d296 | 443 | if (pblk_rb_flush_point_set(rb, NULL, mem)) |
588726d3 JG |
444 | return; |
445 | ||
cc9c9a00 | 446 | pblk_write_kick(pblk); |
588726d3 JG |
447 | } |
448 | ||
a4bd217b JG |
449 | static int pblk_rb_may_write_flush(struct pblk_rb *rb, unsigned int nr_entries, |
450 | unsigned int *pos, struct bio *bio, | |
451 | int *io_ret) | |
452 | { | |
453 | unsigned int mem; | |
454 | ||
455 | if (!__pblk_rb_may_write(rb, nr_entries, pos)) | |
456 | return 0; | |
457 | ||
40b8657d | 458 | mem = pblk_rb_ptr_wrap(rb, *pos, nr_entries); |
a4bd217b JG |
459 | *io_ret = NVM_IO_DONE; |
460 | ||
461 | if (bio->bi_opf & REQ_PREFLUSH) { | |
462 | struct pblk *pblk = container_of(rb, struct pblk, rwb); | |
463 | ||
5d149bfa | 464 | atomic64_inc(&pblk->nr_flush); |
8154d296 | 465 | if (pblk_rb_flush_point_set(&pblk->rwb, bio, mem)) |
a4bd217b JG |
466 | *io_ret = NVM_IO_OK; |
467 | } | |
468 | ||
469 | /* Protect from read count */ | |
470 | smp_store_release(&rb->mem, mem); | |
6ca2f71f | 471 | |
a4bd217b JG |
472 | return 1; |
473 | } | |
474 | ||
475 | /* | |
476 | * Atomically check that (i) there is space on the write buffer for the | |
477 | * incoming I/O, and (ii) the current I/O type has enough budget in the write | |
478 | * buffer (rate-limiter). | |
479 | */ | |
480 | int pblk_rb_may_write_user(struct pblk_rb *rb, struct bio *bio, | |
481 | unsigned int nr_entries, unsigned int *pos) | |
482 | { | |
483 | struct pblk *pblk = container_of(rb, struct pblk, rwb); | |
588726d3 | 484 | int io_ret; |
a4bd217b JG |
485 | |
486 | spin_lock(&rb->w_lock); | |
588726d3 JG |
487 | io_ret = pblk_rl_user_may_insert(&pblk->rl, nr_entries); |
488 | if (io_ret) { | |
a4bd217b | 489 | spin_unlock(&rb->w_lock); |
588726d3 | 490 | return io_ret; |
a4bd217b JG |
491 | } |
492 | ||
588726d3 | 493 | if (!pblk_rb_may_write_flush(rb, nr_entries, pos, bio, &io_ret)) { |
a4bd217b JG |
494 | spin_unlock(&rb->w_lock); |
495 | return NVM_IO_REQUEUE; | |
496 | } | |
497 | ||
498 | pblk_rl_user_in(&pblk->rl, nr_entries); | |
499 | spin_unlock(&rb->w_lock); | |
500 | ||
588726d3 | 501 | return io_ret; |
a4bd217b JG |
502 | } |
503 | ||
504 | /* | |
505 | * Look at pblk_rb_may_write_user comment | |
506 | */ | |
507 | int pblk_rb_may_write_gc(struct pblk_rb *rb, unsigned int nr_entries, | |
508 | unsigned int *pos) | |
509 | { | |
510 | struct pblk *pblk = container_of(rb, struct pblk, rwb); | |
511 | ||
512 | spin_lock(&rb->w_lock); | |
513 | if (!pblk_rl_gc_may_insert(&pblk->rl, nr_entries)) { | |
514 | spin_unlock(&rb->w_lock); | |
515 | return 0; | |
516 | } | |
517 | ||
518 | if (!pblk_rb_may_write(rb, nr_entries, pos)) { | |
519 | spin_unlock(&rb->w_lock); | |
520 | return 0; | |
521 | } | |
522 | ||
523 | pblk_rl_gc_in(&pblk->rl, nr_entries); | |
524 | spin_unlock(&rb->w_lock); | |
525 | ||
526 | return 1; | |
527 | } | |
528 | ||
a4bd217b JG |
529 | /* |
530 | * Read available entries on rb and add them to the given bio. To avoid a memory | |
531 | * copy, a page reference to the write buffer is used to be added to the bio. | |
532 | * | |
533 | * This function is used by the write thread to form the write bio that will | |
534 | * persist data on the write buffer to the media. | |
535 | */ | |
d624f371 | 536 | unsigned int pblk_rb_read_to_bio(struct pblk_rb *rb, struct nvm_rq *rqd, |
875d94f3 JG |
537 | unsigned int pos, unsigned int nr_entries, |
538 | unsigned int count) | |
a4bd217b JG |
539 | { |
540 | struct pblk *pblk = container_of(rb, struct pblk, rwb); | |
d624f371 JG |
541 | struct request_queue *q = pblk->dev->q; |
542 | struct pblk_c_ctx *c_ctx = nvm_rq_to_pdu(rqd); | |
875d94f3 | 543 | struct bio *bio = rqd->bio; |
a4bd217b JG |
544 | struct pblk_rb_entry *entry; |
545 | struct page *page; | |
d624f371 | 546 | unsigned int pad = 0, to_read = nr_entries; |
a4bd217b JG |
547 | unsigned int i; |
548 | int flags; | |
a4bd217b JG |
549 | |
550 | if (count < nr_entries) { | |
551 | pad = nr_entries - count; | |
552 | to_read = count; | |
553 | } | |
554 | ||
55d8ec35 IK |
555 | /* Add space for packed metadata if in use*/ |
556 | pad += (pblk->min_write_pgs - pblk->min_write_pgs_data); | |
557 | ||
a4bd217b JG |
558 | c_ctx->sentry = pos; |
559 | c_ctx->nr_valid = to_read; | |
560 | c_ctx->nr_padded = pad; | |
561 | ||
562 | for (i = 0; i < to_read; i++) { | |
563 | entry = &rb->entries[pos]; | |
564 | ||
565 | /* A write has been allowed into the buffer, but data is still | |
566 | * being copied to it. It is ok to busy wait. | |
567 | */ | |
568 | try: | |
569 | flags = READ_ONCE(entry->w_ctx.flags); | |
10888129 JG |
570 | if (!(flags & PBLK_WRITTEN_DATA)) { |
571 | io_schedule(); | |
a4bd217b | 572 | goto try; |
10888129 | 573 | } |
a4bd217b | 574 | |
a4bd217b JG |
575 | page = virt_to_page(entry->data); |
576 | if (!page) { | |
4e495a46 | 577 | pblk_err(pblk, "could not allocate write bio page\n"); |
a4bd217b JG |
578 | flags &= ~PBLK_WRITTEN_DATA; |
579 | flags |= PBLK_SUBMITTED_ENTRY; | |
580 | /* Release flags on context. Protect from writes */ | |
581 | smp_store_release(&entry->w_ctx.flags, flags); | |
d624f371 | 582 | return NVM_IO_ERR; |
a4bd217b JG |
583 | } |
584 | ||
d624f371 JG |
585 | if (bio_add_pc_page(q, bio, page, rb->seg_size, 0) != |
586 | rb->seg_size) { | |
4e495a46 | 587 | pblk_err(pblk, "could not add page to write bio\n"); |
a4bd217b JG |
588 | flags &= ~PBLK_WRITTEN_DATA; |
589 | flags |= PBLK_SUBMITTED_ENTRY; | |
590 | /* Release flags on context. Protect from writes */ | |
591 | smp_store_release(&entry->w_ctx.flags, flags); | |
d624f371 | 592 | return NVM_IO_ERR; |
a4bd217b JG |
593 | } |
594 | ||
a4bd217b JG |
595 | flags &= ~PBLK_WRITTEN_DATA; |
596 | flags |= PBLK_SUBMITTED_ENTRY; | |
597 | ||
598 | /* Release flags on context. Protect from writes */ | |
599 | smp_store_release(&entry->w_ctx.flags, flags); | |
600 | ||
40b8657d | 601 | pos = pblk_rb_ptr_wrap(rb, pos, 1); |
a4bd217b JG |
602 | } |
603 | ||
d624f371 JG |
604 | if (pad) { |
605 | if (pblk_bio_add_pages(pblk, bio, GFP_KERNEL, pad)) { | |
4e495a46 | 606 | pblk_err(pblk, "could not pad page in write bio\n"); |
d624f371 JG |
607 | return NVM_IO_ERR; |
608 | } | |
d624f371 | 609 | |
5d149bfa HH |
610 | if (pad < pblk->min_write_pgs) |
611 | atomic64_inc(&pblk->pad_dist[pad - 1]); | |
612 | else | |
4e495a46 | 613 | pblk_warn(pblk, "padding more than min. sectors\n"); |
5d149bfa HH |
614 | |
615 | atomic64_add(pad, &pblk->pad_wa); | |
616 | } | |
76758390 | 617 | |
880eda54 | 618 | #ifdef CONFIG_NVM_PBLK_DEBUG |
5d149bfa | 619 | atomic_long_add(pad, &pblk->padded_writes); |
a4bd217b | 620 | #endif |
d624f371 JG |
621 | |
622 | return NVM_IO_OK; | |
a4bd217b JG |
623 | } |
624 | ||
625 | /* | |
626 | * Copy to bio only if the lba matches the one on the given cache entry. | |
627 | * Otherwise, it means that the entry has been overwritten, and the bio should | |
628 | * be directed to disk. | |
629 | */ | |
630 | int pblk_rb_copy_to_bio(struct pblk_rb *rb, struct bio *bio, sector_t lba, | |
75cb8e93 | 631 | struct ppa_addr ppa, int bio_iter, bool advanced_bio) |
a4bd217b | 632 | { |
07698466 | 633 | struct pblk *pblk = container_of(rb, struct pblk, rwb); |
a4bd217b JG |
634 | struct pblk_rb_entry *entry; |
635 | struct pblk_w_ctx *w_ctx; | |
07698466 JG |
636 | struct ppa_addr l2p_ppa; |
637 | u64 pos = pblk_addr_to_cacheline(ppa); | |
a4bd217b JG |
638 | void *data; |
639 | int flags; | |
640 | int ret = 1; | |
641 | ||
a4bd217b | 642 | |
880eda54 | 643 | #ifdef CONFIG_NVM_PBLK_DEBUG |
a4bd217b JG |
644 | /* Caller must ensure that the access will not cause an overflow */ |
645 | BUG_ON(pos >= rb->nr_entries); | |
646 | #endif | |
647 | entry = &rb->entries[pos]; | |
648 | w_ctx = &entry->w_ctx; | |
649 | flags = READ_ONCE(w_ctx->flags); | |
650 | ||
07698466 JG |
651 | spin_lock(&rb->w_lock); |
652 | spin_lock(&pblk->trans_lock); | |
653 | l2p_ppa = pblk_trans_map_get(pblk, lba); | |
654 | spin_unlock(&pblk->trans_lock); | |
655 | ||
a4bd217b | 656 | /* Check if the entry has been overwritten or is scheduled to be */ |
07698466 JG |
657 | if (!pblk_ppa_comp(l2p_ppa, ppa) || w_ctx->lba != lba || |
658 | flags & PBLK_WRITABLE_ENTRY) { | |
a4bd217b JG |
659 | ret = 0; |
660 | goto out; | |
661 | } | |
662 | ||
663 | /* Only advance the bio if it hasn't been advanced already. If advanced, | |
664 | * this bio is at least a partial bio (i.e., it has partially been | |
665 | * filled with data from the cache). If part of the data resides on the | |
666 | * media, we will read later on | |
667 | */ | |
75cb8e93 | 668 | if (unlikely(!advanced_bio)) |
a4bd217b JG |
669 | bio_advance(bio, bio_iter * PBLK_EXPOSED_PAGE_SIZE); |
670 | ||
671 | data = bio_data(bio); | |
672 | memcpy(data, entry->data, rb->seg_size); | |
673 | ||
674 | out: | |
675 | spin_unlock(&rb->w_lock); | |
676 | return ret; | |
677 | } | |
678 | ||
679 | struct pblk_w_ctx *pblk_rb_w_ctx(struct pblk_rb *rb, unsigned int pos) | |
680 | { | |
40b8657d | 681 | unsigned int entry = pblk_rb_ptr_wrap(rb, pos, 0); |
a4bd217b JG |
682 | |
683 | return &rb->entries[entry].w_ctx; | |
684 | } | |
685 | ||
686 | unsigned int pblk_rb_sync_init(struct pblk_rb *rb, unsigned long *flags) | |
687 | __acquires(&rb->s_lock) | |
688 | { | |
689 | if (flags) | |
690 | spin_lock_irqsave(&rb->s_lock, *flags); | |
691 | else | |
692 | spin_lock_irq(&rb->s_lock); | |
693 | ||
694 | return rb->sync; | |
695 | } | |
696 | ||
697 | void pblk_rb_sync_end(struct pblk_rb *rb, unsigned long *flags) | |
698 | __releases(&rb->s_lock) | |
699 | { | |
700 | lockdep_assert_held(&rb->s_lock); | |
701 | ||
702 | if (flags) | |
703 | spin_unlock_irqrestore(&rb->s_lock, *flags); | |
704 | else | |
705 | spin_unlock_irq(&rb->s_lock); | |
706 | } | |
707 | ||
708 | unsigned int pblk_rb_sync_advance(struct pblk_rb *rb, unsigned int nr_entries) | |
709 | { | |
533657c1 | 710 | unsigned int sync, flush_point; |
a4bd217b JG |
711 | lockdep_assert_held(&rb->s_lock); |
712 | ||
713 | sync = READ_ONCE(rb->sync); | |
533657c1 HH |
714 | flush_point = READ_ONCE(rb->flush_point); |
715 | ||
716 | if (flush_point != EMPTY_ENTRY) { | |
717 | unsigned int secs_to_flush; | |
a4bd217b | 718 | |
533657c1 HH |
719 | secs_to_flush = pblk_rb_ring_count(flush_point, sync, |
720 | rb->nr_entries); | |
721 | if (secs_to_flush < nr_entries) { | |
722 | /* Protect flush points */ | |
723 | smp_store_release(&rb->flush_point, EMPTY_ENTRY); | |
724 | } | |
725 | } | |
726 | ||
40b8657d | 727 | sync = pblk_rb_ptr_wrap(rb, sync, nr_entries); |
a4bd217b JG |
728 | |
729 | /* Protect from counts */ | |
730 | smp_store_release(&rb->sync, sync); | |
731 | ||
732 | return sync; | |
733 | } | |
734 | ||
533657c1 | 735 | /* Calculate how many sectors to submit up to the current flush point. */ |
8154d296 | 736 | unsigned int pblk_rb_flush_point_count(struct pblk_rb *rb) |
a4bd217b | 737 | { |
533657c1 HH |
738 | unsigned int subm, sync, flush_point; |
739 | unsigned int submitted, to_flush; | |
a4bd217b | 740 | |
8154d296 HH |
741 | /* Protect flush points */ |
742 | flush_point = smp_load_acquire(&rb->flush_point); | |
743 | if (flush_point == EMPTY_ENTRY) | |
a4bd217b JG |
744 | return 0; |
745 | ||
533657c1 HH |
746 | /* Protect syncs */ |
747 | sync = smp_load_acquire(&rb->sync); | |
748 | ||
a4bd217b | 749 | subm = READ_ONCE(rb->subm); |
533657c1 | 750 | submitted = pblk_rb_ring_count(subm, sync, rb->nr_entries); |
a4bd217b JG |
751 | |
752 | /* The sync point itself counts as a sector to sync */ | |
533657c1 | 753 | to_flush = pblk_rb_ring_count(flush_point, sync, rb->nr_entries) + 1; |
a4bd217b | 754 | |
533657c1 | 755 | return (submitted < to_flush) ? (to_flush - submitted) : 0; |
a4bd217b JG |
756 | } |
757 | ||
a4bd217b JG |
758 | int pblk_rb_tear_down_check(struct pblk_rb *rb) |
759 | { | |
760 | struct pblk_rb_entry *entry; | |
761 | int i; | |
762 | int ret = 0; | |
763 | ||
764 | spin_lock(&rb->w_lock); | |
765 | spin_lock_irq(&rb->s_lock); | |
766 | ||
767 | if ((rb->mem == rb->subm) && (rb->subm == rb->sync) && | |
768 | (rb->sync == rb->l2p_update) && | |
8154d296 | 769 | (rb->flush_point == EMPTY_ENTRY)) { |
a4bd217b JG |
770 | goto out; |
771 | } | |
772 | ||
773 | if (!rb->entries) { | |
774 | ret = 1; | |
775 | goto out; | |
776 | } | |
777 | ||
778 | for (i = 0; i < rb->nr_entries; i++) { | |
779 | entry = &rb->entries[i]; | |
780 | ||
781 | if (!entry->data) { | |
782 | ret = 1; | |
783 | goto out; | |
784 | } | |
785 | } | |
786 | ||
787 | out: | |
788 | spin_unlock(&rb->w_lock); | |
789 | spin_unlock_irq(&rb->s_lock); | |
790 | ||
791 | return ret; | |
792 | } | |
793 | ||
794 | unsigned int pblk_rb_wrap_pos(struct pblk_rb *rb, unsigned int pos) | |
795 | { | |
796 | return (pos & (rb->nr_entries - 1)); | |
797 | } | |
798 | ||
799 | int pblk_rb_pos_oob(struct pblk_rb *rb, u64 pos) | |
800 | { | |
801 | return (pos >= rb->nr_entries); | |
802 | } | |
803 | ||
804 | ssize_t pblk_rb_sysfs(struct pblk_rb *rb, char *buf) | |
805 | { | |
806 | struct pblk *pblk = container_of(rb, struct pblk, rwb); | |
807 | struct pblk_c_ctx *c; | |
808 | ssize_t offset; | |
809 | int queued_entries = 0; | |
810 | ||
811 | spin_lock_irq(&rb->s_lock); | |
812 | list_for_each_entry(c, &pblk->compl_list, list) | |
813 | queued_entries++; | |
814 | spin_unlock_irq(&rb->s_lock); | |
815 | ||
8154d296 | 816 | if (rb->flush_point != EMPTY_ENTRY) |
a4bd217b JG |
817 | offset = scnprintf(buf, PAGE_SIZE, |
818 | "%u\t%u\t%u\t%u\t%u\t%u\t%u - %u/%u/%u - %d\n", | |
819 | rb->nr_entries, | |
820 | rb->mem, | |
821 | rb->subm, | |
822 | rb->sync, | |
823 | rb->l2p_update, | |
880eda54 | 824 | #ifdef CONFIG_NVM_PBLK_DEBUG |
8154d296 | 825 | atomic_read(&rb->inflight_flush_point), |
a4bd217b JG |
826 | #else |
827 | 0, | |
828 | #endif | |
8154d296 | 829 | rb->flush_point, |
a4bd217b JG |
830 | pblk_rb_read_count(rb), |
831 | pblk_rb_space(rb), | |
8154d296 | 832 | pblk_rb_flush_point_count(rb), |
a4bd217b JG |
833 | queued_entries); |
834 | else | |
835 | offset = scnprintf(buf, PAGE_SIZE, | |
836 | "%u\t%u\t%u\t%u\t%u\t%u\tNULL - %u/%u/%u - %d\n", | |
837 | rb->nr_entries, | |
838 | rb->mem, | |
839 | rb->subm, | |
840 | rb->sync, | |
841 | rb->l2p_update, | |
880eda54 | 842 | #ifdef CONFIG_NVM_PBLK_DEBUG |
8154d296 | 843 | atomic_read(&rb->inflight_flush_point), |
a4bd217b JG |
844 | #else |
845 | 0, | |
846 | #endif | |
847 | pblk_rb_read_count(rb), | |
848 | pblk_rb_space(rb), | |
8154d296 | 849 | pblk_rb_flush_point_count(rb), |
a4bd217b JG |
850 | queued_entries); |
851 | ||
852 | return offset; | |
853 | } |