| 1 | /* @title: I/O Request Scheduling */ |
| 2 | #pragma once |
| 3 | #include <block/bcache.h> |
| 4 | #include <block/bio.h> |
| 5 | #include <block/block.h> |
| 6 | #include <fs/detect.h> |
| 7 | #include <sch/sched.h> |
| 8 | #include <stdatomic.h> |
| 9 | #include <stdbool.h> |
| 10 | #include <stdint.h> |
| 11 | #include <sync/mutex.h> |
| 12 | #include <thread/thread.h> |
| 13 | |
| 14 | /* |
| 15 | * This is the asynchronous block device |
| 16 | * IO request scheduler, referred to as the |
| 17 | * "bio scheduler", or "bio sched". |
| 18 | * |
| 19 | * ===================== Overview ======================= |
| 20 | * |
| 21 | * The bio scheduler is a generic block IO scheduler that |
| 22 | * block devices are expected to implement device-specific |
| 23 | * policies to allow the scheduler to perform optimizations. |
| 24 | * |
| 25 | * The two main device-specific optimizations are coalescing and reordering. |
| 26 | * |
| 27 | * Coalescing allows two requests to get merged into one, |
| 28 | * and reordering simply changes the order in which |
| 29 | * requests are dispatched, allowing for some block devices |
| 30 | * to operate more efficiently. |
| 31 | * |
| 32 | * When coalescing succeeds, the request with the lower LBA is marked |
| 33 | * as `is_aggregate`, and the one with the higher LBA is marked `skip`. |
| 34 | * |
| 35 | * ===================== Queue Structure ======================= |
| 36 | * |
| 37 | * The scheduler uses a 5-level MLFQ to keep track of requests. |
| 38 | * Each level is a different request priority level. |
| 39 | * |
| 40 | * The head of each MLFQ is always the first request in the queue |
| 41 | * to get dispatched, and the head->next is the next, and so on. |
| 42 | * |
| 43 | * The highest priority requests (BIO_RQ_URGENT) entirely skip |
| 44 | * the queue and are immediately dispatched, giving them no |
| 45 | * time to get coalesced or reordered. |
| 46 | * |
| 47 | * ===================== Request Boosting ======================= |
| 48 | * |
| 49 | * Over time, requests are boosted according to device specific |
| 50 | * policies. All of the device specific policies are contained |
| 51 | * in a per-device `struct bio_scheduler_ops`. |
| 52 | * |
| 53 | * Boosts prevent request starvation, as sometimes, |
| 54 | * requests may not get manually dispatched, and thus |
| 55 | * spend lots of time in the queue without getting executed. |
| 56 | * |
| 57 | * All timestamps for requests are stored in milliseconds, and are taken |
| 58 | * from the current time during the enqueue. |
| 59 | * |
| 60 | * The FIRST boost will occur when a request has spent enough |
| 61 | * time (`ops->max_wait_time[queue_level]`) in the queue. |
| 62 | * |
| 63 | * This is determined via the following for the FIRST boost: |
| 64 | * |
| 65 | * int base_wait = ops->max_wait_time[queue_level]; |
| 66 | * bool should_boost = current_time > (enqueue_time + base_wait); |
| 67 | * |
| 68 | * After the first boost, however, the following |
| 69 | * boosts become accelerated to prevent large batches |
| 70 | * of initially low-priority requests from taking up space |
| 71 | * in higher priority queues. |
| 72 | * |
| 73 | * This adjusted timestamp is computed by simply |
| 74 | * taking the amount of boosts a request has had, |
| 75 | * and performing the following operation: |
| 76 | * |
| 77 | * int adjusted_wait = base_wait >> req->boost_count; |
| 78 | * |
| 79 | * Where the shift can be limited by changing BIO_SCHED_BOOST_SHIFT_LIMIT. |
| 80 | * |
| 81 | * In addition to this, the levels that the request is boosted by also changes. |
| 82 | * The first request will always boost the request by one priority level, but |
| 83 | * following boosts will be different. |
| 84 | * |
| 85 | * The new priority is computed with the following operation: |
| 86 | * |
| 87 | * int get_boost_depth(int boost_count) { |
| 88 | * if (boost_count >= 3) |
| 89 | * return 2; |
| 90 | * else if (boost_count >= 1) |
| 91 | * return 1; |
| 92 | * |
| 93 | * return 0; |
| 94 | * } |
| 95 | * |
| 96 | * int step = get_boost_depth(req->boost_count); |
| 97 | * int new_prio = req->prio + 1 + step; // '1' is the initial boost |
| 98 | * |
| 99 | * Where `new_prio` is limited to the BIO_SCHED_MAX priority level. |
| 100 | * |
| 101 | * Request starvation is automatically checked every `ops->tick_ms` |
| 102 | * milliseconds. |
| 103 | */ |
| 104 | |
| 105 | /* how many queue levels the bio_scheduler has */ |
| 106 | #define BIO_SCHED_LEVELS 5 |
| 107 | #define BIO_SCHED_MAX (BIO_SCHED_LEVELS - 1) |
| 108 | |
| 109 | /* first boosts can only boost priority by this much */ |
| 110 | #define BIO_SCHED_STARVATION_BOOST 1 |
| 111 | |
| 112 | /* max of 2^4 threshold reduction */ |
| 113 | #define BIO_SCHED_BOOST_SHIFT_LIMIT 4 |
| 114 | |
| 115 | /* max to scan before bail */ |
| 116 | #define BIO_SCHED_COALESCE_SCAN_LIMIT 8 |
| 117 | |
| 118 | #define BIO_SCHED_MAX_BOOST_SCAN 32 |
| 119 | |
| 120 | /* max coalesces in one enqueue() */ |
| 121 | #define BIO_SCHED_MAX_COALESCES 4 |
| 122 | |
| 123 | struct bio_rqueue { |
| 124 | struct list_head list; |
| 125 | |
| 126 | uint64_t request_count; |
| 127 | |
| 128 | /* coalescing */ |
| 129 | bool dirty; |
| 130 | }; |
| 131 | |
| 132 | struct bio_scheduler { |
| 133 | struct block_device *disk; |
| 134 | struct mutex lock; |
| 135 | uint64_t total_requests; |
| 136 | struct bio_rqueue queues[BIO_SCHED_LEVELS]; |
| 137 | bool defer_pending; |
| 138 | }; |
| 139 | |
| 140 | struct bio_scheduler_ops { |
| 141 | bool (*should_coalesce)(struct block_device *dev, |
| 142 | const struct bio_request *a, |
| 143 | const struct bio_request *b); |
| 144 | |
| 145 | void (*do_coalesce)(struct block_device *dev, struct bio_request *into, |
| 146 | struct bio_request *from); |
| 147 | |
| 148 | void (*reorder)(struct block_device *dev); |
| 149 | |
| 150 | /* maximum request wait time for each queue level before first boost */ |
| 151 | uint32_t max_wait_time[BIO_SCHED_LEVELS]; |
| 152 | |
| 153 | /* maximum requests in all queues before high priority |
| 154 | * requests start getting auto-dispatched */ |
| 155 | uint32_t dispatch_threshold; |
| 156 | |
| 157 | /* maximum number of requests in a target queue to allow |
| 158 | * a boosted request to boost to that level */ |
| 159 | uint64_t boost_occupance_limit[BIO_SCHED_LEVELS]; |
| 160 | |
| 161 | /* amount of time to wait between starvation |
| 162 | * checks to boost starving requests */ |
| 163 | uint64_t tick_ms; |
| 164 | |
| 165 | /* absolute minimum amount of time waited for |
| 166 | * any request to get boosted |
| 167 | * |
| 168 | * used for requests that have been starving |
| 169 | * for a long time to prevent them from |
| 170 | * immediately boosting at the first |
| 171 | * possible opportunity */ |
| 172 | uint64_t min_wait_ms; |
| 173 | }; |
| 174 | |
| 175 | bool noop_should_coalesce(struct block_device *disk, |
| 176 | const struct bio_request *a, |
| 177 | const struct bio_request *b); |
| 178 | |
| 179 | void noop_do_coalesce(struct block_device *disk, struct bio_request *into, |
| 180 | struct bio_request *from); |
| 181 | |
| 182 | void noop_reorder(struct block_device *disk); |
| 183 | |
| 184 | void bio_sched_enqueue(struct block_device *disk, struct bio_request *req); |
| 185 | |
| 186 | void bio_sched_dequeue(struct block_device *disk, struct bio_request *req, |
| 187 | bool already_locked); |
| 188 | |
| 189 | void bio_sched_enqueue_internal(struct bio_scheduler *sched, |
| 190 | struct bio_request *req); |
| 191 | void bio_sched_dequeue_internal(struct bio_scheduler *sched, |
| 192 | struct bio_request *req); |
| 193 | |
| 194 | void bio_sched_dispatch_partial(struct block_device *disk, |
| 195 | enum bio_request_priority prio); |
| 196 | |
| 197 | void bio_sched_dispatch_all(struct block_device *disk); |
| 198 | |
| 199 | void bio_sched_try_early_dispatch(struct bio_scheduler *sched); |
| 200 | |
| 201 | bool bio_sched_try_coalesce(struct bio_scheduler *sched); |
| 202 | bool bio_sched_boost_starved(struct bio_scheduler *sched); |
| 203 | |
| 204 | struct bio_scheduler *bio_sched_create(struct block_device *disk, |
| 205 | struct bio_scheduler_ops *ops); |
| 206 | |
| 207 | static inline void update_request_timestamp(struct bio_request *req) { |
| 208 | req->enqueue_time = time_get_ms(); |
| 209 | } |
| 210 | |
| 211 | static inline bool submit_if_urgent(struct bio_scheduler *sched, |
| 212 | struct bio_request *req) { |
| 213 | if (req->priority == BIO_RQ_URGENT) { |
| 214 | /* VIP request - skip the queue ! */ |
| 215 | sched->disk->submit_bio_async(sched->disk, req); |
| 216 | return true; |
| 217 | } |
| 218 | return false; |
| 219 | } |
| 220 | |
| 221 | static inline bool sched_is_empty(struct bio_scheduler *sched) { |
| 222 | for (uint32_t i = 0; i < BIO_SCHED_LEVELS; i++) |
| 223 | if (!list_empty(head: &sched->queues[i].list)) |
| 224 | return false; |
| 225 | |
| 226 | return true; |
| 227 | } |
| 228 | |
| 229 | static inline bool submit_if_skip_sched(struct bio_scheduler *sched, |
| 230 | struct bio_request *req) { |
| 231 | if (bdev_skip_sched(disk: sched->disk)) { |
| 232 | sched->disk->submit_bio_async(sched->disk, req); |
| 233 | return true; |
| 234 | } |
| 235 | return false; |
| 236 | } |
| 237 | |