| 1 | #include <block/bio.h> |
| 2 | #include <block/block.h> |
| 3 | #include <block/sched.h> |
| 4 | #include <drivers/nvme.h> |
| 5 | #include <mem/alloc.h> |
| 6 | #include <mem/hhdm.h> |
| 7 | #include <mem/pmm.h> |
| 8 | #include <mem/vmm.h> |
| 9 | #include <stdbool.h> |
| 10 | #include <stddef.h> |
| 11 | #include <stdint.h> |
| 12 | #include <structures/sll.h> |
| 13 | #include <thread/io_wait.h> |
| 14 | #include <thread/workqueue.h> |
| 15 | |
| 16 | #include "internal.h" |
| 17 | #include <thread/thread.h> |
| 18 | |
| 19 | typedef bool (*sync_fn)(struct block_device *, uint64_t, uint8_t *, uint16_t, |
| 20 | struct io_wait_token *iowt); |
| 21 | typedef bool (*async_fn)(struct block_device *, struct nvme_request *); |
| 22 | |
| 23 | static void enqueue_request(struct nvme_device *dev, struct nvme_request *req) { |
| 24 | struct nvme_waiting_requests *q = &dev->waiting_requests; |
| 25 | |
| 26 | enum irql irql = spin_lock_irq_disable(&q->lock); |
| 27 | |
| 28 | list_add_tail(new: &req->list_node, head: &q->list); |
| 29 | |
| 30 | spin_unlock(&q->lock, irql); |
| 31 | } |
| 32 | |
| 33 | static bool nvme_bio_fill_prps(struct nvme_bio_data *data, const void *buffer, |
| 34 | uint64_t size) { |
| 35 | uint64_t offset = (uintptr_t) buffer & (PAGE_SIZE - 1); |
| 36 | uint64_t num_pages = PAGES_NEEDED_FOR(offset + size); |
| 37 | |
| 38 | data->prps = kmalloc(sizeof(uint64_t) * num_pages); |
| 39 | if (!data->prps) |
| 40 | return false; |
| 41 | |
| 42 | /* PRP1 is the only one that can have a non-zero offset, and |
| 43 | * because the transfer begins some way into the first page, it needs |
| 44 | * to use this. Later entries describe full pages so must be page aligned */ |
| 45 | vaddr_t vaddr = (vaddr_t) buffer; |
| 46 | |
| 47 | for (size_t i = 0; i < num_pages; i++) { |
| 48 | data->prps[i] = vmm_get_phys(vaddr, VMM_FLAG_NONE); |
| 49 | nvme_check_dma_addr(phys: data->prps[i], what: "PRP entry" ); |
| 50 | vaddr = PAGE_ALIGN_DOWN(vaddr) + PAGE_SIZE; |
| 51 | } |
| 52 | |
| 53 | data->prp_count = num_pages; |
| 54 | |
| 55 | return true; |
| 56 | } |
| 57 | |
| 58 | static bool nvme_setup_prps(struct nvme_command *cmd, |
| 59 | struct nvme_bio_data *data) { |
| 60 | kassert(data->prp_count > 0); |
| 61 | |
| 62 | cmd->prp1 = data->prps[0]; |
| 63 | nvme_check_dma_addr(phys: cmd->prp1, what: "prp1" ); |
| 64 | |
| 65 | kassert(IS_ALIGNED(cmd->prp1, 4)); /* dword aligned */ |
| 66 | if (data->prp_count == 1) { |
| 67 | cmd->prp2 = 0; |
| 68 | goto free_prps; |
| 69 | } else if (data->prp_count == 2) { |
| 70 | cmd->prp2 = data->prps[1]; |
| 71 | nvme_check_dma_addr(phys: cmd->prp2, what: "prp2" ); |
| 72 | goto free_prps; |
| 73 | } |
| 74 | |
| 75 | /* After 2 pages PRP2 is a PRP List pointer. This means |
| 76 | * we need to provide another page for the list, |
| 77 | * with max_transfer_size always preventing too many */ |
| 78 | uint64_t entries = data->prp_count - 1; |
| 79 | kassert(entries <= NVME_PRPS_PER_PAGE); |
| 80 | |
| 81 | paddr_t list_phys = pmm_alloc_page(); |
| 82 | if (!list_phys) |
| 83 | goto free_prps_fail; |
| 84 | |
| 85 | uint64_t *list = hhdm_paddr_to_ptr(p: list_phys); |
| 86 | for (uint64_t i = 0; i < entries; i++) |
| 87 | list[i] = data->prps[i + 1]; |
| 88 | |
| 89 | data->prp_list_phys = list_phys; |
| 90 | cmd->prp2 = list_phys; |
| 91 | nvme_check_dma_addr(phys: cmd->prp2, what: "PRP list pointer" ); |
| 92 | |
| 93 | /* List page has everything needed by the controller */ |
| 94 | free_prps: |
| 95 | kfree(data->prps); |
| 96 | data->prps = NULL; |
| 97 | return true; |
| 98 | |
| 99 | free_prps_fail: |
| 100 | kfree(data->prps); |
| 101 | data->prps = NULL; |
| 102 | return false; |
| 103 | } |
| 104 | |
| 105 | static bool rw_send_command(struct block_device *disk, struct nvme_request *req, |
| 106 | uint8_t opc) { |
| 107 | struct nvme_device *nvme = disk->driver_data; |
| 108 | uint16_t qid = THIS_QID(nvme); |
| 109 | uint64_t lba = req->lba; |
| 110 | uint64_t count = req->sector_count; |
| 111 | void *buffer = req->buffer; |
| 112 | |
| 113 | struct nvme_queue *q = nvme->io_queues[qid]; |
| 114 | |
| 115 | if (atomic_load(&q->outstanding) >= q->sq_depth) { |
| 116 | enqueue_request(dev: nvme, req); |
| 117 | |
| 118 | /* No room */ |
| 119 | return true; |
| 120 | } |
| 121 | |
| 122 | struct nvme_bio_data *data = |
| 123 | kmalloc(sizeof(struct nvme_bio_data), ALLOC_FLAGS_ZERO); |
| 124 | if (!data) |
| 125 | return false; |
| 126 | |
| 127 | if (!nvme_bio_fill_prps(data, buffer, size: count * disk->sector_size)) { |
| 128 | kfree(data); |
| 129 | return false; |
| 130 | } |
| 131 | |
| 132 | req->bio_data = data; |
| 133 | |
| 134 | struct nvme_command cmd = {0}; |
| 135 | cmd.opc = opc; |
| 136 | cmd.nsid = 1; |
| 137 | cmd.cdw10 = lba & 0xFFFFFFFFULL; |
| 138 | cmd.cdw11 = lba >> 32ULL; |
| 139 | cmd.cdw12 = count - 1; |
| 140 | |
| 141 | if (!nvme_setup_prps(cmd: &cmd, data)) { |
| 142 | req->bio_data = NULL; |
| 143 | kfree(data); |
| 144 | return false; |
| 145 | } |
| 146 | |
| 147 | req->lba = lba; |
| 148 | req->buffer = buffer; |
| 149 | req->sector_count = count; |
| 150 | req->done = false; |
| 151 | req->status = -1; |
| 152 | |
| 153 | nvme_submit_io_cmd(nvme, cmd: &cmd, qid, req); |
| 154 | |
| 155 | return true; |
| 156 | } |
| 157 | |
| 158 | static bool rw_sync(struct block_device *disk, uint64_t lba, uint8_t *buffer, |
| 159 | uint16_t count, async_fn function, |
| 160 | struct io_wait_token *iowt) { |
| 161 | struct nvme_request req = {0}; |
| 162 | req.lba = lba; |
| 163 | req.buffer = buffer; |
| 164 | req.sector_count = count; |
| 165 | req.remaining_parts = 1; |
| 166 | INIT_LIST_HEAD(list: &req.list_node); |
| 167 | |
| 168 | struct thread *curr = thread_get_current(); |
| 169 | |
| 170 | enum irql irql = irql_raise(new_level: IRQL_DISPATCH_LEVEL); |
| 171 | req.waiter = curr; |
| 172 | |
| 173 | if (io_wait_token_active(t: iowt)) |
| 174 | io_wait_end(t: iowt, act: IO_WAIT_END_NO_OP); |
| 175 | |
| 176 | io_wait_begin(out: iowt, io_object: disk->driver_data); |
| 177 | |
| 178 | function(disk, &req); |
| 179 | irql_lower(old_level: irql); |
| 180 | |
| 181 | /* Go run something else now */ |
| 182 | thread_yield_until_wake_match(); |
| 183 | |
| 184 | return !req.status; |
| 185 | } |
| 186 | |
| 187 | static bool rw_wrapper(struct block_device *disk, uint64_t lba, uint8_t *buf, |
| 188 | uint64_t cnt, sync_fn function) { |
| 189 | struct nvme_device *nvme = (struct nvme_device *) disk->driver_data; |
| 190 | uint16_t max_sectors = nvme->max_transfer_size / disk->sector_size; |
| 191 | struct io_wait_token iowt = IO_WAIT_TOKEN_EMPTY; |
| 192 | |
| 193 | while (cnt > 0) { |
| 194 | uint16_t chunk = (cnt > max_sectors) ? max_sectors : (uint16_t) cnt; |
| 195 | if (!function(disk, lba, buf, chunk, &iowt)) { |
| 196 | io_wait_end(t: &iowt, act: IO_WAIT_END_YIELD); |
| 197 | return false; |
| 198 | } |
| 199 | |
| 200 | lba += chunk; |
| 201 | buf += chunk * disk->sector_size; |
| 202 | cnt -= chunk; |
| 203 | } |
| 204 | |
| 205 | /* Reset priority after boost */ |
| 206 | io_wait_end(t: &iowt, act: IO_WAIT_END_YIELD); |
| 207 | return true; |
| 208 | } |
| 209 | |
| 210 | static bool rw_async_wrapper(struct block_device *disk, |
| 211 | struct nvme_request *req, async_fn function) { |
| 212 | struct nvme_device *nvme = (struct nvme_device *) disk->driver_data; |
| 213 | uint16_t max_sectors = nvme->max_transfer_size / disk->sector_size; |
| 214 | uint16_t chunk; |
| 215 | |
| 216 | uint64_t cnt = req->sector_count; |
| 217 | |
| 218 | int part_count = 0; |
| 219 | uint64_t tmp_cnt = cnt; |
| 220 | while (tmp_cnt > 0) { |
| 221 | chunk = (tmp_cnt > max_sectors) ? max_sectors : (uint16_t) tmp_cnt; |
| 222 | tmp_cnt -= chunk; |
| 223 | part_count++; |
| 224 | } |
| 225 | |
| 226 | req->remaining_parts = part_count; |
| 227 | |
| 228 | while (cnt > 0) { |
| 229 | chunk = (cnt > max_sectors) ? max_sectors : (uint16_t) cnt; |
| 230 | cnt -= chunk; |
| 231 | |
| 232 | if (!function(disk, req)) |
| 233 | return false; |
| 234 | } |
| 235 | |
| 236 | return true; |
| 237 | } |
| 238 | |
| 239 | bool nvme_read_sector(struct block_device *disk, uint64_t lba, uint8_t *buffer, |
| 240 | uint16_t count, struct io_wait_token *i) { |
| 241 | return rw_sync(disk, lba, buffer, count, function: nvme_read_sector_async, iowt: i); |
| 242 | } |
| 243 | |
| 244 | bool nvme_write_sector(struct block_device *disk, uint64_t lba, uint8_t *buffer, |
| 245 | uint16_t count, struct io_wait_token *i) { |
| 246 | return rw_sync(disk, lba, buffer, count, function: nvme_write_sector_async, iowt: i); |
| 247 | } |
| 248 | |
| 249 | bool nvme_read_sector_wrapper(struct block_device *disk, uint64_t lba, |
| 250 | uint8_t *buf, uint64_t cnt) { |
| 251 | return rw_wrapper(disk, lba, buf, cnt, function: nvme_read_sector); |
| 252 | } |
| 253 | |
| 254 | bool nvme_write_sector_wrapper(struct block_device *disk, uint64_t lba, |
| 255 | const uint8_t *buf, uint64_t cnt) { |
| 256 | return rw_wrapper(disk, lba, buf: (uint8_t *) buf, cnt, function: nvme_write_sector); |
| 257 | } |
| 258 | |
| 259 | bool nvme_read_sector_async(struct block_device *disk, |
| 260 | struct nvme_request *req) { |
| 261 | return rw_send_command(disk, req, NVME_OP_IO_READ); |
| 262 | } |
| 263 | |
| 264 | bool nvme_write_sector_async(struct block_device *disk, |
| 265 | struct nvme_request *req) { |
| 266 | return rw_send_command(disk, req, NVME_OP_IO_WRITE); |
| 267 | } |
| 268 | |
| 269 | bool nvme_write_sector_async_wrapper(struct block_device *disk, |
| 270 | struct nvme_request *req) { |
| 271 | return rw_async_wrapper(disk, req, function: nvme_write_sector_async); |
| 272 | } |
| 273 | |
| 274 | bool nvme_read_sector_async_wrapper(struct block_device *disk, |
| 275 | struct nvme_request *req) { |
| 276 | return rw_async_wrapper(disk, req, function: nvme_read_sector_async); |
| 277 | } |
| 278 | |
| 279 | bool nvme_send_nvme_req(struct block_device *d, struct nvme_request *r) { |
| 280 | return rw_send_command(disk: d, req: r, opc: r->write ? NVME_OP_IO_WRITE : NVME_OP_IO_READ); |
| 281 | } |
| 282 | |