1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * NVMe Over Fabrics Target File I/O commands implementation. 4 * Copyright (c) 2017-2018 Western Digital Corporation or its 5 * affiliates. 6 */ 7 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt 8 #include <linux/uio.h> 9 #include <linux/falloc.h> 10 #include <linux/file.h> 11 #include <linux/fs.h> 12 #include "nvmet.h" 13 14 #define NVMET_MIN_MPOOL_OBJ 16 15 16 void nvmet_file_ns_revalidate(struct nvmet_ns *ns) 17 { 18 ns->size = i_size_read(ns->file->f_mapping->host); 19 } 20 21 void nvmet_file_ns_disable(struct nvmet_ns *ns) 22 { 23 if (ns->file) { 24 if (ns->buffered_io) 25 flush_workqueue(buffered_io_wq); 26 mempool_destroy(ns->bvec_pool); 27 ns->bvec_pool = NULL; 28 fput(ns->file); 29 ns->file = NULL; 30 } 31 } 32 33 int nvmet_file_ns_enable(struct nvmet_ns *ns) 34 { 35 int flags = O_RDWR | O_LARGEFILE; 36 int ret = 0; 37 38 if (!ns->buffered_io) 39 flags |= O_DIRECT; 40 41 ns->file = filp_open(ns->device_path, flags, 0); 42 if (IS_ERR(ns->file)) { 43 ret = PTR_ERR(ns->file); 44 pr_err("failed to open file %s: (%d)\n", 45 ns->device_path, ret); 46 ns->file = NULL; 47 return ret; 48 } 49 50 nvmet_file_ns_revalidate(ns); 51 52 /* 53 * i_blkbits can be greater than the universally accepted upper bound, 54 * so make sure we export a sane namespace lba_shift. 55 */ 56 ns->blksize_shift = min_t(u8, 57 file_inode(ns->file)->i_blkbits, 12); 58 59 ns->bvec_pool = mempool_create(NVMET_MIN_MPOOL_OBJ, mempool_alloc_slab, 60 mempool_free_slab, nvmet_bvec_cache); 61 62 if (!ns->bvec_pool) { 63 ret = -ENOMEM; 64 goto err; 65 } 66 67 return ret; 68 err: 69 fput(ns->file); 70 ns->file = NULL; 71 ns->size = 0; 72 ns->blksize_shift = 0; 73 return ret; 74 } 75 76 static void nvmet_file_init_bvec(struct bio_vec *bv, struct scatterlist *sg) 77 { 78 bv->bv_page = sg_page(sg); 79 bv->bv_offset = sg->offset; 80 bv->bv_len = sg->length; 81 } 82 83 static ssize_t nvmet_file_submit_bvec(struct nvmet_req *req, loff_t pos, 84 unsigned long nr_segs, size_t count, int ki_flags) 85 { 86 struct kiocb *iocb = &req->f.iocb; 87 ssize_t (*call_iter)(struct kiocb *iocb, struct iov_iter *iter); 88 struct iov_iter iter; 89 int rw; 90 91 if (req->cmd->rw.opcode == nvme_cmd_write) { 92 if (req->cmd->rw.control & cpu_to_le16(NVME_RW_FUA)) 93 ki_flags |= IOCB_DSYNC; 94 call_iter = req->ns->file->f_op->write_iter; 95 rw = ITER_SOURCE; 96 } else { 97 call_iter = req->ns->file->f_op->read_iter; 98 rw = ITER_DEST; 99 } 100 101 iov_iter_bvec(&iter, rw, req->f.bvec, nr_segs, count); 102 103 iocb->ki_pos = pos; 104 iocb->ki_filp = req->ns->file; 105 iocb->ki_flags = ki_flags | iocb->ki_filp->f_iocb_flags; 106 107 return call_iter(iocb, &iter); 108 } 109 110 static void nvmet_file_io_done(struct kiocb *iocb, long ret) 111 { 112 struct nvmet_req *req = container_of(iocb, struct nvmet_req, f.iocb); 113 u16 status = NVME_SC_SUCCESS; 114 115 if (req->f.bvec != req->inline_bvec) { 116 if (likely(req->f.mpool_alloc == false)) 117 kfree(req->f.bvec); 118 else 119 mempool_free(req->f.bvec, req->ns->bvec_pool); 120 } 121 122 if (unlikely(ret != req->transfer_len)) 123 status = errno_to_nvme_status(req, ret); 124 nvmet_req_complete(req, status); 125 } 126 127 static bool nvmet_file_execute_io(struct nvmet_req *req, int ki_flags) 128 { 129 ssize_t nr_bvec = req->sg_cnt; 130 unsigned long bv_cnt = 0; 131 bool is_sync = false; 132 size_t len = 0, total_len = 0; 133 ssize_t ret = 0; 134 loff_t pos; 135 int i; 136 struct scatterlist *sg; 137 138 if (req->f.mpool_alloc && nr_bvec > NVMET_MAX_MPOOL_BVEC) 139 is_sync = true; 140 141 pos = le64_to_cpu(req->cmd->rw.slba) << req->ns->blksize_shift; 142 if (unlikely(pos + req->transfer_len > req->ns->size)) { 143 nvmet_req_complete(req, errno_to_nvme_status(req, -ENOSPC)); 144 return true; 145 } 146 147 memset(&req->f.iocb, 0, sizeof(struct kiocb)); 148 for_each_sg(req->sg, sg, req->sg_cnt, i) { 149 nvmet_file_init_bvec(&req->f.bvec[bv_cnt], sg); 150 len += req->f.bvec[bv_cnt].bv_len; 151 total_len += req->f.bvec[bv_cnt].bv_len; 152 bv_cnt++; 153 154 WARN_ON_ONCE((nr_bvec - 1) < 0); 155 156 if (unlikely(is_sync) && 157 (nr_bvec - 1 == 0 || bv_cnt == NVMET_MAX_MPOOL_BVEC)) { 158 ret = nvmet_file_submit_bvec(req, pos, bv_cnt, len, 0); 159 if (ret < 0) 160 goto complete; 161 162 pos += len; 163 bv_cnt = 0; 164 len = 0; 165 } 166 nr_bvec--; 167 } 168 169 if (WARN_ON_ONCE(total_len != req->transfer_len)) { 170 ret = -EIO; 171 goto complete; 172 } 173 174 if (unlikely(is_sync)) { 175 ret = total_len; 176 goto complete; 177 } 178 179 /* 180 * A NULL ki_complete ask for synchronous execution, which we want 181 * for the IOCB_NOWAIT case. 182 */ 183 if (!(ki_flags & IOCB_NOWAIT)) 184 req->f.iocb.ki_complete = nvmet_file_io_done; 185 186 ret = nvmet_file_submit_bvec(req, pos, bv_cnt, total_len, ki_flags); 187 188 switch (ret) { 189 case -EIOCBQUEUED: 190 return true; 191 case -EAGAIN: 192 if (WARN_ON_ONCE(!(ki_flags & IOCB_NOWAIT))) 193 goto complete; 194 return false; 195 case -EOPNOTSUPP: 196 /* 197 * For file systems returning error -EOPNOTSUPP, handle 198 * IOCB_NOWAIT error case separately and retry without 199 * IOCB_NOWAIT. 200 */ 201 if ((ki_flags & IOCB_NOWAIT)) 202 return false; 203 break; 204 } 205 206 complete: 207 nvmet_file_io_done(&req->f.iocb, ret); 208 return true; 209 } 210 211 static void nvmet_file_buffered_io_work(struct work_struct *w) 212 { 213 struct nvmet_req *req = container_of(w, struct nvmet_req, f.work); 214 215 nvmet_file_execute_io(req, 0); 216 } 217 218 static void nvmet_file_submit_buffered_io(struct nvmet_req *req) 219 { 220 INIT_WORK(&req->f.work, nvmet_file_buffered_io_work); 221 queue_work(buffered_io_wq, &req->f.work); 222 } 223 224 static void nvmet_file_execute_rw(struct nvmet_req *req) 225 { 226 ssize_t nr_bvec = req->sg_cnt; 227 228 if (!nvmet_check_transfer_len(req, nvmet_rw_data_len(req))) 229 return; 230 231 if (!req->sg_cnt || !nr_bvec) { 232 nvmet_req_complete(req, 0); 233 return; 234 } 235 236 if (nr_bvec > NVMET_MAX_INLINE_BIOVEC) 237 req->f.bvec = kmalloc_array(nr_bvec, sizeof(struct bio_vec), 238 GFP_KERNEL); 239 else 240 req->f.bvec = req->inline_bvec; 241 242 if (unlikely(!req->f.bvec)) { 243 /* fallback under memory pressure */ 244 req->f.bvec = mempool_alloc(req->ns->bvec_pool, GFP_KERNEL); 245 req->f.mpool_alloc = true; 246 } else 247 req->f.mpool_alloc = false; 248 249 if (req->ns->buffered_io) { 250 if (likely(!req->f.mpool_alloc) && 251 (req->ns->file->f_mode & FMODE_NOWAIT) && 252 nvmet_file_execute_io(req, IOCB_NOWAIT)) 253 return; 254 nvmet_file_submit_buffered_io(req); 255 } else 256 nvmet_file_execute_io(req, 0); 257 } 258 259 u16 nvmet_file_flush(struct nvmet_req *req) 260 { 261 return errno_to_nvme_status(req, vfs_fsync(req->ns->file, 1)); 262 } 263 264 static void nvmet_file_flush_work(struct work_struct *w) 265 { 266 struct nvmet_req *req = container_of(w, struct nvmet_req, f.work); 267 268 nvmet_req_complete(req, nvmet_file_flush(req)); 269 } 270 271 static void nvmet_file_execute_flush(struct nvmet_req *req) 272 { 273 if (!nvmet_check_transfer_len(req, 0)) 274 return; 275 INIT_WORK(&req->f.work, nvmet_file_flush_work); 276 queue_work(nvmet_wq, &req->f.work); 277 } 278 279 static void nvmet_file_execute_discard(struct nvmet_req *req) 280 { 281 int mode = FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE; 282 struct nvme_dsm_range range; 283 loff_t offset, len; 284 u16 status = 0; 285 int ret; 286 int i; 287 288 for (i = 0; i <= le32_to_cpu(req->cmd->dsm.nr); i++) { 289 status = nvmet_copy_from_sgl(req, i * sizeof(range), &range, 290 sizeof(range)); 291 if (status) 292 break; 293 294 offset = le64_to_cpu(range.slba) << req->ns->blksize_shift; 295 len = le32_to_cpu(range.nlb); 296 len <<= req->ns->blksize_shift; 297 if (offset + len > req->ns->size) { 298 req->error_slba = le64_to_cpu(range.slba); 299 status = errno_to_nvme_status(req, -ENOSPC); 300 break; 301 } 302 303 ret = vfs_fallocate(req->ns->file, mode, offset, len); 304 if (ret && ret != -EOPNOTSUPP) { 305 req->error_slba = le64_to_cpu(range.slba); 306 status = errno_to_nvme_status(req, ret); 307 break; 308 } 309 } 310 311 nvmet_req_complete(req, status); 312 } 313 314 static void nvmet_file_dsm_work(struct work_struct *w) 315 { 316 struct nvmet_req *req = container_of(w, struct nvmet_req, f.work); 317 318 switch (le32_to_cpu(req->cmd->dsm.attributes)) { 319 case NVME_DSMGMT_AD: 320 nvmet_file_execute_discard(req); 321 return; 322 case NVME_DSMGMT_IDR: 323 case NVME_DSMGMT_IDW: 324 default: 325 /* Not supported yet */ 326 nvmet_req_complete(req, 0); 327 return; 328 } 329 } 330 331 static void nvmet_file_execute_dsm(struct nvmet_req *req) 332 { 333 if (!nvmet_check_data_len_lte(req, nvmet_dsm_len(req))) 334 return; 335 INIT_WORK(&req->f.work, nvmet_file_dsm_work); 336 queue_work(nvmet_wq, &req->f.work); 337 } 338 339 static void nvmet_file_write_zeroes_work(struct work_struct *w) 340 { 341 struct nvmet_req *req = container_of(w, struct nvmet_req, f.work); 342 struct nvme_write_zeroes_cmd *write_zeroes = &req->cmd->write_zeroes; 343 int mode = FALLOC_FL_ZERO_RANGE | FALLOC_FL_KEEP_SIZE; 344 loff_t offset; 345 loff_t len; 346 int ret; 347 348 offset = le64_to_cpu(write_zeroes->slba) << req->ns->blksize_shift; 349 len = (((sector_t)le16_to_cpu(write_zeroes->length) + 1) << 350 req->ns->blksize_shift); 351 352 if (unlikely(offset + len > req->ns->size)) { 353 nvmet_req_complete(req, errno_to_nvme_status(req, -ENOSPC)); 354 return; 355 } 356 357 ret = vfs_fallocate(req->ns->file, mode, offset, len); 358 nvmet_req_complete(req, ret < 0 ? errno_to_nvme_status(req, ret) : 0); 359 } 360 361 static void nvmet_file_execute_write_zeroes(struct nvmet_req *req) 362 { 363 if (!nvmet_check_transfer_len(req, 0)) 364 return; 365 INIT_WORK(&req->f.work, nvmet_file_write_zeroes_work); 366 queue_work(nvmet_wq, &req->f.work); 367 } 368 369 u16 nvmet_file_parse_io_cmd(struct nvmet_req *req) 370 { 371 switch (req->cmd->common.opcode) { 372 case nvme_cmd_read: 373 case nvme_cmd_write: 374 req->execute = nvmet_file_execute_rw; 375 return 0; 376 case nvme_cmd_flush: 377 req->execute = nvmet_file_execute_flush; 378 return 0; 379 case nvme_cmd_dsm: 380 req->execute = nvmet_file_execute_dsm; 381 return 0; 382 case nvme_cmd_write_zeroes: 383 req->execute = nvmet_file_execute_write_zeroes; 384 return 0; 385 default: 386 return nvmet_report_invalid_opcode(req); 387 } 388 } 389