]>
Commit | Line | Data |
---|---|---|
a4bd217b JG |
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 | ||
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 | ||
26 | int pblk_rl_user_may_insert(struct pblk_rl *rl, int nr_entries) | |
27 | { | |
28 | int rb_user_cnt = atomic_read(&rl->rb_user_cnt); | |
29 | ||
b20ba1bc | 30 | return (!(rb_user_cnt >= rl->rb_user_max)); |
a4bd217b JG |
31 | } |
32 | ||
33 | int pblk_rl_gc_may_insert(struct pblk_rl *rl, int nr_entries) | |
34 | { | |
35 | int rb_gc_cnt = atomic_read(&rl->rb_gc_cnt); | |
36 | int rb_user_active; | |
37 | ||
38 | /* If there is no user I/O let GC take over space on the write buffer */ | |
39 | rb_user_active = READ_ONCE(rl->rb_user_active); | |
b20ba1bc | 40 | return (!(rb_gc_cnt >= rl->rb_gc_max && rb_user_active)); |
a4bd217b JG |
41 | } |
42 | ||
43 | void pblk_rl_user_in(struct pblk_rl *rl, int nr_entries) | |
44 | { | |
45 | atomic_add(nr_entries, &rl->rb_user_cnt); | |
46 | ||
47 | /* Release user I/O state. Protect from GC */ | |
48 | smp_store_release(&rl->rb_user_active, 1); | |
49 | pblk_rl_kick_u_timer(rl); | |
50 | } | |
51 | ||
52 | void pblk_rl_gc_in(struct pblk_rl *rl, int nr_entries) | |
53 | { | |
54 | atomic_add(nr_entries, &rl->rb_gc_cnt); | |
55 | } | |
56 | ||
57 | void pblk_rl_out(struct pblk_rl *rl, int nr_user, int nr_gc) | |
58 | { | |
59 | atomic_sub(nr_user, &rl->rb_user_cnt); | |
60 | atomic_sub(nr_gc, &rl->rb_gc_cnt); | |
61 | } | |
62 | ||
63 | unsigned long pblk_rl_nr_free_blks(struct pblk_rl *rl) | |
64 | { | |
65 | return atomic_read(&rl->free_blocks); | |
66 | } | |
67 | ||
68 | /* | |
69 | * We check for (i) the number of free blocks in the current LUN and (ii) the | |
70 | * total number of free blocks in the pblk instance. This is to even out the | |
71 | * number of free blocks on each LUN when GC kicks in. | |
72 | * | |
73 | * Only the total number of free blocks is used to configure the rate limiter. | |
74 | */ | |
75 | static int pblk_rl_update_rates(struct pblk_rl *rl, unsigned long max) | |
76 | { | |
77 | unsigned long free_blocks = pblk_rl_nr_free_blks(rl); | |
78 | ||
79 | if (free_blocks >= rl->high) { | |
b20ba1bc JG |
80 | rl->rb_user_max = max; |
81 | rl->rb_gc_max = 0; | |
a4bd217b JG |
82 | rl->rb_state = PBLK_RL_HIGH; |
83 | } else if (free_blocks < rl->high) { | |
84 | int shift = rl->high_pw - rl->rb_windows_pw; | |
85 | int user_windows = free_blocks >> shift; | |
86 | int user_max = user_windows << PBLK_MAX_REQ_ADDRS_PW; | |
a4bd217b JG |
87 | |
88 | rl->rb_user_max = user_max; | |
b20ba1bc JG |
89 | rl->rb_gc_max = max - user_max; |
90 | ||
91 | if (free_blocks <= rl->rsv_blocks) { | |
92 | rl->rb_user_max = 0; | |
93 | rl->rb_gc_max = max; | |
94 | } | |
95 | ||
96 | /* In the worst case, we will need to GC lines in the low list | |
97 | * (high valid sector count). If there are lines to GC on high | |
98 | * or mid lists, these will be prioritized | |
99 | */ | |
100 | rl->rb_state = PBLK_RL_LOW; | |
a4bd217b JG |
101 | } |
102 | ||
103 | return rl->rb_state; | |
104 | } | |
105 | ||
a4bd217b JG |
106 | void pblk_rl_free_lines_inc(struct pblk_rl *rl, struct pblk_line *line) |
107 | { | |
108 | struct pblk *pblk = container_of(rl, struct pblk, rl); | |
a44f53fa | 109 | int blk_in_line = atomic_read(&line->blk_in_line); |
a4bd217b JG |
110 | int ret; |
111 | ||
a44f53fa | 112 | atomic_add(blk_in_line, &rl->free_blocks); |
a4bd217b JG |
113 | /* Rates will not change that often - no need to lock update */ |
114 | ret = pblk_rl_update_rates(rl, rl->rb_budget); | |
115 | ||
116 | if (ret == (PBLK_RL_MID | PBLK_RL_LOW)) | |
117 | pblk_gc_should_start(pblk); | |
118 | else | |
119 | pblk_gc_should_stop(pblk); | |
120 | } | |
121 | ||
122 | void pblk_rl_free_lines_dec(struct pblk_rl *rl, struct pblk_line *line) | |
123 | { | |
a44f53fa | 124 | int blk_in_line = atomic_read(&line->blk_in_line); |
a4bd217b | 125 | |
a44f53fa | 126 | atomic_sub(blk_in_line, &rl->free_blocks); |
b20ba1bc JG |
127 | } |
128 | ||
129 | void pblk_gc_should_kick(struct pblk *pblk) | |
130 | { | |
131 | struct pblk_rl *rl = &pblk->rl; | |
132 | int ret; | |
a4bd217b JG |
133 | |
134 | /* Rates will not change that often - no need to lock update */ | |
135 | ret = pblk_rl_update_rates(rl, rl->rb_budget); | |
136 | if (ret == (PBLK_RL_MID | PBLK_RL_LOW)) | |
137 | pblk_gc_should_start(pblk); | |
138 | else | |
139 | pblk_gc_should_stop(pblk); | |
140 | } | |
141 | ||
b20ba1bc | 142 | int pblk_rl_high_thrs(struct pblk_rl *rl) |
a4bd217b JG |
143 | { |
144 | return rl->high; | |
145 | } | |
146 | ||
b20ba1bc JG |
147 | int pblk_rl_low_thrs(struct pblk_rl *rl) |
148 | { | |
149 | return rl->low; | |
150 | } | |
151 | ||
a4bd217b JG |
152 | int pblk_rl_sysfs_rate_show(struct pblk_rl *rl) |
153 | { | |
154 | return rl->rb_user_max; | |
155 | } | |
156 | ||
157 | static void pblk_rl_u_timer(unsigned long data) | |
158 | { | |
159 | struct pblk_rl *rl = (struct pblk_rl *)data; | |
160 | ||
161 | /* Release user I/O state. Protect from GC */ | |
162 | smp_store_release(&rl->rb_user_active, 0); | |
163 | } | |
164 | ||
165 | void pblk_rl_free(struct pblk_rl *rl) | |
166 | { | |
167 | del_timer(&rl->u_timer); | |
168 | } | |
169 | ||
170 | void pblk_rl_init(struct pblk_rl *rl, int budget) | |
171 | { | |
b20ba1bc JG |
172 | struct pblk *pblk = container_of(rl, struct pblk, rl); |
173 | struct pblk_line_meta *lm = &pblk->lm; | |
174 | int min_blocks = lm->blk_per_line * PBLK_GC_RSV_LINE; | |
a4bd217b JG |
175 | unsigned int rb_windows; |
176 | ||
177 | rl->high = rl->total_blocks / PBLK_USER_HIGH_THRS; | |
a4bd217b JG |
178 | rl->high_pw = get_count_order(rl->high); |
179 | ||
b20ba1bc JG |
180 | rl->low = rl->total_blocks / PBLK_USER_LOW_THRS; |
181 | if (rl->low < min_blocks) | |
182 | rl->low = min_blocks; | |
183 | ||
184 | rl->rsv_blocks = min_blocks; | |
185 | ||
a4bd217b JG |
186 | /* This will always be a power-of-2 */ |
187 | rb_windows = budget / PBLK_MAX_REQ_ADDRS; | |
b20ba1bc | 188 | rl->rb_windows_pw = get_count_order(rb_windows); |
a4bd217b JG |
189 | |
190 | /* To start with, all buffer is available to user I/O writers */ | |
191 | rl->rb_budget = budget; | |
192 | rl->rb_user_max = budget; | |
193 | atomic_set(&rl->rb_user_cnt, 0); | |
194 | rl->rb_gc_max = 0; | |
195 | rl->rb_state = PBLK_RL_HIGH; | |
196 | atomic_set(&rl->rb_gc_cnt, 0); | |
197 | ||
198 | setup_timer(&rl->u_timer, pblk_rl_u_timer, (unsigned long)rl); | |
b20ba1bc | 199 | |
a4bd217b | 200 | rl->rb_user_active = 0; |
b20ba1bc | 201 | rl->rb_gc_active = 0; |
a4bd217b | 202 | } |