1 /*
2  * Copyright (C) 2016 CNEX Labs
3  * Initial release: Javier Gonzalez <javier@cnexlabs.com>
4  *                  Matias Bjorling <matias@cnexlabs.com>
5  *
6  * This program is free software; you can redistribute it and/or
7  * modify it under the terms of the GNU General Public License version
8  * 2 as published by the Free Software Foundation.
9  *
10  * This program is distributed in the hope that it will be useful, but
11  * WITHOUT ANY WARRANTY; without even the implied warranty of
12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
13  * General Public License for more details.
14  *
15  * pblk-rl.c - pblk's rate limiter for user I/O
16  *
17  */
18 
19 #include "pblk.h"
20 
pblk_rl_kick_u_timer(struct pblk_rl * rl)21 static void pblk_rl_kick_u_timer(struct pblk_rl *rl)
22 {
23 	mod_timer(&rl->u_timer, jiffies + msecs_to_jiffies(5000));
24 }
25 
pblk_rl_is_limit(struct pblk_rl * rl)26 int pblk_rl_is_limit(struct pblk_rl *rl)
27 {
28 	int rb_space;
29 
30 	rb_space = atomic_read(&rl->rb_space);
31 
32 	return (rb_space == 0);
33 }
34 
pblk_rl_user_may_insert(struct pblk_rl * rl,int nr_entries)35 int pblk_rl_user_may_insert(struct pblk_rl *rl, int nr_entries)
36 {
37 	int rb_user_cnt = atomic_read(&rl->rb_user_cnt);
38 	int rb_space = atomic_read(&rl->rb_space);
39 
40 	if (unlikely(rb_space >= 0) && (rb_space - nr_entries < 0))
41 		return NVM_IO_ERR;
42 
43 	if (rb_user_cnt >= rl->rb_user_max)
44 		return NVM_IO_REQUEUE;
45 
46 	return NVM_IO_OK;
47 }
48 
pblk_rl_inserted(struct pblk_rl * rl,int nr_entries)49 void pblk_rl_inserted(struct pblk_rl *rl, int nr_entries)
50 {
51 	int rb_space = atomic_read(&rl->rb_space);
52 
53 	if (unlikely(rb_space >= 0))
54 		atomic_sub(nr_entries, &rl->rb_space);
55 }
56 
pblk_rl_gc_may_insert(struct pblk_rl * rl,int nr_entries)57 int pblk_rl_gc_may_insert(struct pblk_rl *rl, int nr_entries)
58 {
59 	int rb_gc_cnt = atomic_read(&rl->rb_gc_cnt);
60 	int rb_user_active;
61 
62 	/* If there is no user I/O let GC take over space on the write buffer */
63 	rb_user_active = READ_ONCE(rl->rb_user_active);
64 	return (!(rb_gc_cnt >= rl->rb_gc_max && rb_user_active));
65 }
66 
pblk_rl_user_in(struct pblk_rl * rl,int nr_entries)67 void pblk_rl_user_in(struct pblk_rl *rl, int nr_entries)
68 {
69 	atomic_add(nr_entries, &rl->rb_user_cnt);
70 
71 	/* Release user I/O state. Protect from GC */
72 	smp_store_release(&rl->rb_user_active, 1);
73 	pblk_rl_kick_u_timer(rl);
74 }
75 
pblk_rl_werr_line_in(struct pblk_rl * rl)76 void pblk_rl_werr_line_in(struct pblk_rl *rl)
77 {
78 	atomic_inc(&rl->werr_lines);
79 }
80 
pblk_rl_werr_line_out(struct pblk_rl * rl)81 void pblk_rl_werr_line_out(struct pblk_rl *rl)
82 {
83 	atomic_dec(&rl->werr_lines);
84 }
85 
pblk_rl_gc_in(struct pblk_rl * rl,int nr_entries)86 void pblk_rl_gc_in(struct pblk_rl *rl, int nr_entries)
87 {
88 	atomic_add(nr_entries, &rl->rb_gc_cnt);
89 }
90 
pblk_rl_out(struct pblk_rl * rl,int nr_user,int nr_gc)91 void pblk_rl_out(struct pblk_rl *rl, int nr_user, int nr_gc)
92 {
93 	atomic_sub(nr_user, &rl->rb_user_cnt);
94 	atomic_sub(nr_gc, &rl->rb_gc_cnt);
95 }
96 
pblk_rl_nr_free_blks(struct pblk_rl * rl)97 unsigned long pblk_rl_nr_free_blks(struct pblk_rl *rl)
98 {
99 	return atomic_read(&rl->free_blocks);
100 }
101 
pblk_rl_nr_user_free_blks(struct pblk_rl * rl)102 unsigned long pblk_rl_nr_user_free_blks(struct pblk_rl *rl)
103 {
104 	return atomic_read(&rl->free_user_blocks);
105 }
106 
__pblk_rl_update_rates(struct pblk_rl * rl,unsigned long free_blocks)107 static void __pblk_rl_update_rates(struct pblk_rl *rl,
108 				   unsigned long free_blocks)
109 {
110 	struct pblk *pblk = container_of(rl, struct pblk, rl);
111 	int max = rl->rb_budget;
112 	int werr_gc_needed = atomic_read(&rl->werr_lines);
113 
114 	if (free_blocks >= rl->high) {
115 		if (werr_gc_needed) {
116 			/* Allocate a small budget for recovering
117 			 * lines with write errors
118 			 */
119 			rl->rb_gc_max = 1 << rl->rb_windows_pw;
120 			rl->rb_user_max = max - rl->rb_gc_max;
121 			rl->rb_state = PBLK_RL_WERR;
122 		} else {
123 			rl->rb_user_max = max;
124 			rl->rb_gc_max = 0;
125 			rl->rb_state = PBLK_RL_OFF;
126 		}
127 	} else if (free_blocks < rl->high) {
128 		int shift = rl->high_pw - rl->rb_windows_pw;
129 		int user_windows = free_blocks >> shift;
130 		int user_max = user_windows << PBLK_MAX_REQ_ADDRS_PW;
131 
132 		rl->rb_user_max = user_max;
133 		rl->rb_gc_max = max - user_max;
134 
135 		if (free_blocks <= rl->rsv_blocks) {
136 			rl->rb_user_max = 0;
137 			rl->rb_gc_max = max;
138 		}
139 
140 		/* In the worst case, we will need to GC lines in the low list
141 		 * (high valid sector count). If there are lines to GC on high
142 		 * or mid lists, these will be prioritized
143 		 */
144 		rl->rb_state = PBLK_RL_LOW;
145 	}
146 
147 	if (rl->rb_state != PBLK_RL_OFF)
148 		pblk_gc_should_start(pblk);
149 	else
150 		pblk_gc_should_stop(pblk);
151 }
152 
pblk_rl_update_rates(struct pblk_rl * rl)153 void pblk_rl_update_rates(struct pblk_rl *rl)
154 {
155 	__pblk_rl_update_rates(rl, pblk_rl_nr_user_free_blks(rl));
156 }
157 
pblk_rl_free_lines_inc(struct pblk_rl * rl,struct pblk_line * line)158 void pblk_rl_free_lines_inc(struct pblk_rl *rl, struct pblk_line *line)
159 {
160 	int blk_in_line = atomic_read(&line->blk_in_line);
161 	int free_blocks;
162 
163 	atomic_add(blk_in_line, &rl->free_blocks);
164 	free_blocks = atomic_add_return(blk_in_line, &rl->free_user_blocks);
165 
166 	__pblk_rl_update_rates(rl, free_blocks);
167 }
168 
pblk_rl_free_lines_dec(struct pblk_rl * rl,struct pblk_line * line,bool used)169 void pblk_rl_free_lines_dec(struct pblk_rl *rl, struct pblk_line *line,
170 			    bool used)
171 {
172 	int blk_in_line = atomic_read(&line->blk_in_line);
173 	int free_blocks;
174 
175 	atomic_sub(blk_in_line, &rl->free_blocks);
176 
177 	if (used)
178 		free_blocks = atomic_sub_return(blk_in_line,
179 							&rl->free_user_blocks);
180 	else
181 		free_blocks = atomic_read(&rl->free_user_blocks);
182 
183 	__pblk_rl_update_rates(rl, free_blocks);
184 }
185 
pblk_rl_high_thrs(struct pblk_rl * rl)186 int pblk_rl_high_thrs(struct pblk_rl *rl)
187 {
188 	return rl->high;
189 }
190 
pblk_rl_max_io(struct pblk_rl * rl)191 int pblk_rl_max_io(struct pblk_rl *rl)
192 {
193 	return rl->rb_max_io;
194 }
195 
pblk_rl_u_timer(struct timer_list * t)196 static void pblk_rl_u_timer(struct timer_list *t)
197 {
198 	struct pblk_rl *rl = from_timer(rl, t, u_timer);
199 
200 	/* Release user I/O state. Protect from GC */
201 	smp_store_release(&rl->rb_user_active, 0);
202 }
203 
pblk_rl_free(struct pblk_rl * rl)204 void pblk_rl_free(struct pblk_rl *rl)
205 {
206 	del_timer(&rl->u_timer);
207 }
208 
pblk_rl_init(struct pblk_rl * rl,int budget)209 void pblk_rl_init(struct pblk_rl *rl, int budget)
210 {
211 	struct pblk *pblk = container_of(rl, struct pblk, rl);
212 	struct nvm_tgt_dev *dev = pblk->dev;
213 	struct nvm_geo *geo = &dev->geo;
214 	struct pblk_line_mgmt *l_mg = &pblk->l_mg;
215 	struct pblk_line_meta *lm = &pblk->lm;
216 	int min_blocks = lm->blk_per_line * PBLK_GC_RSV_LINE;
217 	int sec_meta, blk_meta;
218 
219 	unsigned int rb_windows;
220 
221 	/* Consider sectors used for metadata */
222 	sec_meta = (lm->smeta_sec + lm->emeta_sec[0]) * l_mg->nr_free_lines;
223 	blk_meta = DIV_ROUND_UP(sec_meta, geo->clba);
224 
225 	rl->high = pblk->op_blks - blk_meta - lm->blk_per_line;
226 	rl->high_pw = get_count_order(rl->high);
227 
228 	rl->rsv_blocks = min_blocks;
229 
230 	/* This will always be a power-of-2 */
231 	rb_windows = budget / PBLK_MAX_REQ_ADDRS;
232 	rl->rb_windows_pw = get_count_order(rb_windows);
233 
234 	/* To start with, all buffer is available to user I/O writers */
235 	rl->rb_budget = budget;
236 	rl->rb_user_max = budget;
237 	rl->rb_max_io = budget >> 1;
238 	rl->rb_gc_max = 0;
239 	rl->rb_state = PBLK_RL_HIGH;
240 
241 	atomic_set(&rl->rb_user_cnt, 0);
242 	atomic_set(&rl->rb_gc_cnt, 0);
243 	atomic_set(&rl->rb_space, -1);
244 	atomic_set(&rl->werr_lines, 0);
245 
246 	timer_setup(&rl->u_timer, pblk_rl_u_timer, 0);
247 
248 	rl->rb_user_active = 0;
249 	rl->rb_gc_active = 0;
250 }
251