1 // SPDX-License-Identifier: GPL-2.0
3 * NVMe Over Fabrics Target File I/O commands implementation.
4 * Copyright (c) 2017-2018 Western Digital Corporation or its
7 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
9 #include <linux/falloc.h>
10 #include <linux/file.h>
13 #define NVMET_MAX_MPOOL_BVEC 16
14 #define NVMET_MIN_MPOOL_OBJ 16
16 void nvmet_file_ns_disable(struct nvmet_ns *ns)
20 flush_workqueue(buffered_io_wq);
21 mempool_destroy(ns->bvec_pool);
23 kmem_cache_destroy(ns->bvec_cache);
24 ns->bvec_cache = NULL;
30 int nvmet_file_ns_enable(struct nvmet_ns *ns)
32 int flags = O_RDWR | O_LARGEFILE;
39 ns->file = filp_open(ns->device_path, flags, 0);
40 if (IS_ERR(ns->file)) {
41 pr_err("failed to open file %s: (%ld)\n",
42 ns->device_path, PTR_ERR(ns->file));
43 return PTR_ERR(ns->file);
46 ret = vfs_getattr(&ns->file->f_path,
47 &stat, STATX_SIZE, AT_STATX_FORCE_SYNC);
53 * i_blkbits can be greater than the universally accepted upper bound,
54 * so make sure we export a sane namespace lba_shift.
56 ns->blksize_shift = min_t(u8,
57 file_inode(ns->file)->i_blkbits, 12);
59 ns->bvec_cache = kmem_cache_create("nvmet-bvec",
60 NVMET_MAX_MPOOL_BVEC * sizeof(struct bio_vec),
61 0, SLAB_HWCACHE_ALIGN, NULL);
62 if (!ns->bvec_cache) {
67 ns->bvec_pool = mempool_create(NVMET_MIN_MPOOL_OBJ, mempool_alloc_slab,
68 mempool_free_slab, ns->bvec_cache);
78 ns->blksize_shift = 0;
79 nvmet_file_ns_disable(ns);
83 static void nvmet_file_init_bvec(struct bio_vec *bv, struct scatterlist *sg)
85 bv->bv_page = sg_page(sg);
86 bv->bv_offset = sg->offset;
87 bv->bv_len = sg->length;
90 static ssize_t nvmet_file_submit_bvec(struct nvmet_req *req, loff_t pos,
91 unsigned long nr_segs, size_t count, int ki_flags)
93 struct kiocb *iocb = &req->f.iocb;
94 ssize_t (*call_iter)(struct kiocb *iocb, struct iov_iter *iter);
98 if (req->cmd->rw.opcode == nvme_cmd_write) {
99 if (req->cmd->rw.control & cpu_to_le16(NVME_RW_FUA))
100 ki_flags |= IOCB_DSYNC;
101 call_iter = req->ns->file->f_op->write_iter;
104 call_iter = req->ns->file->f_op->read_iter;
108 iov_iter_bvec(&iter, rw, req->f.bvec, nr_segs, count);
111 iocb->ki_filp = req->ns->file;
112 iocb->ki_flags = ki_flags | iocb_flags(req->ns->file);
114 return call_iter(iocb, &iter);
117 static void nvmet_file_io_done(struct kiocb *iocb, long ret, long ret2)
119 struct nvmet_req *req = container_of(iocb, struct nvmet_req, f.iocb);
120 u16 status = NVME_SC_SUCCESS;
122 if (req->f.bvec != req->inline_bvec) {
123 if (likely(req->f.mpool_alloc == false))
126 mempool_free(req->f.bvec, req->ns->bvec_pool);
129 if (unlikely(ret != req->transfer_len))
130 status = errno_to_nvme_status(req, ret);
131 nvmet_req_complete(req, status);
134 static bool nvmet_file_execute_io(struct nvmet_req *req, int ki_flags)
136 ssize_t nr_bvec = req->sg_cnt;
137 unsigned long bv_cnt = 0;
138 bool is_sync = false;
139 size_t len = 0, total_len = 0;
143 struct scatterlist *sg;
145 if (req->f.mpool_alloc && nr_bvec > NVMET_MAX_MPOOL_BVEC)
148 pos = le64_to_cpu(req->cmd->rw.slba) << req->ns->blksize_shift;
149 if (unlikely(pos + req->transfer_len > req->ns->size)) {
150 nvmet_req_complete(req, errno_to_nvme_status(req, -ENOSPC));
154 memset(&req->f.iocb, 0, sizeof(struct kiocb));
155 for_each_sg(req->sg, sg, req->sg_cnt, i) {
156 nvmet_file_init_bvec(&req->f.bvec[bv_cnt], sg);
157 len += req->f.bvec[bv_cnt].bv_len;
158 total_len += req->f.bvec[bv_cnt].bv_len;
161 WARN_ON_ONCE((nr_bvec - 1) < 0);
163 if (unlikely(is_sync) &&
164 (nr_bvec - 1 == 0 || bv_cnt == NVMET_MAX_MPOOL_BVEC)) {
165 ret = nvmet_file_submit_bvec(req, pos, bv_cnt, len, 0);
176 if (WARN_ON_ONCE(total_len != req->transfer_len)) {
181 if (unlikely(is_sync)) {
187 * A NULL ki_complete ask for synchronous execution, which we want
188 * for the IOCB_NOWAIT case.
190 if (!(ki_flags & IOCB_NOWAIT))
191 req->f.iocb.ki_complete = nvmet_file_io_done;
193 ret = nvmet_file_submit_bvec(req, pos, bv_cnt, total_len, ki_flags);
199 if (WARN_ON_ONCE(!(ki_flags & IOCB_NOWAIT)))
204 * For file systems returning error -EOPNOTSUPP, handle
205 * IOCB_NOWAIT error case separately and retry without
208 if ((ki_flags & IOCB_NOWAIT))
214 nvmet_file_io_done(&req->f.iocb, ret, 0);
218 static void nvmet_file_buffered_io_work(struct work_struct *w)
220 struct nvmet_req *req = container_of(w, struct nvmet_req, f.work);
222 nvmet_file_execute_io(req, 0);
225 static void nvmet_file_submit_buffered_io(struct nvmet_req *req)
227 INIT_WORK(&req->f.work, nvmet_file_buffered_io_work);
228 queue_work(buffered_io_wq, &req->f.work);
231 static void nvmet_file_execute_rw(struct nvmet_req *req)
233 ssize_t nr_bvec = req->sg_cnt;
235 if (!nvmet_check_data_len(req, nvmet_rw_len(req)))
238 if (!req->sg_cnt || !nr_bvec) {
239 nvmet_req_complete(req, 0);
243 if (nr_bvec > NVMET_MAX_INLINE_BIOVEC)
244 req->f.bvec = kmalloc_array(nr_bvec, sizeof(struct bio_vec),
247 req->f.bvec = req->inline_bvec;
249 if (unlikely(!req->f.bvec)) {
250 /* fallback under memory pressure */
251 req->f.bvec = mempool_alloc(req->ns->bvec_pool, GFP_KERNEL);
252 req->f.mpool_alloc = true;
254 req->f.mpool_alloc = false;
256 if (req->ns->buffered_io) {
257 if (likely(!req->f.mpool_alloc) &&
258 nvmet_file_execute_io(req, IOCB_NOWAIT))
260 nvmet_file_submit_buffered_io(req);
262 nvmet_file_execute_io(req, 0);
265 u16 nvmet_file_flush(struct nvmet_req *req)
267 return errno_to_nvme_status(req, vfs_fsync(req->ns->file, 1));
270 static void nvmet_file_flush_work(struct work_struct *w)
272 struct nvmet_req *req = container_of(w, struct nvmet_req, f.work);
274 nvmet_req_complete(req, nvmet_file_flush(req));
277 static void nvmet_file_execute_flush(struct nvmet_req *req)
279 if (!nvmet_check_data_len(req, 0))
281 INIT_WORK(&req->f.work, nvmet_file_flush_work);
282 schedule_work(&req->f.work);
285 static void nvmet_file_execute_discard(struct nvmet_req *req)
287 int mode = FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE;
288 struct nvme_dsm_range range;
294 for (i = 0; i <= le32_to_cpu(req->cmd->dsm.nr); i++) {
295 status = nvmet_copy_from_sgl(req, i * sizeof(range), &range,
300 offset = le64_to_cpu(range.slba) << req->ns->blksize_shift;
301 len = le32_to_cpu(range.nlb);
302 len <<= req->ns->blksize_shift;
303 if (offset + len > req->ns->size) {
304 req->error_slba = le64_to_cpu(range.slba);
305 status = errno_to_nvme_status(req, -ENOSPC);
309 ret = vfs_fallocate(req->ns->file, mode, offset, len);
310 if (ret && ret != -EOPNOTSUPP) {
311 req->error_slba = le64_to_cpu(range.slba);
312 status = errno_to_nvme_status(req, ret);
317 nvmet_req_complete(req, status);
320 static void nvmet_file_dsm_work(struct work_struct *w)
322 struct nvmet_req *req = container_of(w, struct nvmet_req, f.work);
324 switch (le32_to_cpu(req->cmd->dsm.attributes)) {
326 nvmet_file_execute_discard(req);
328 case NVME_DSMGMT_IDR:
329 case NVME_DSMGMT_IDW:
331 /* Not supported yet */
332 nvmet_req_complete(req, 0);
337 static void nvmet_file_execute_dsm(struct nvmet_req *req)
339 if (!nvmet_check_data_len_lte(req, nvmet_dsm_len(req)))
341 INIT_WORK(&req->f.work, nvmet_file_dsm_work);
342 schedule_work(&req->f.work);
345 static void nvmet_file_write_zeroes_work(struct work_struct *w)
347 struct nvmet_req *req = container_of(w, struct nvmet_req, f.work);
348 struct nvme_write_zeroes_cmd *write_zeroes = &req->cmd->write_zeroes;
349 int mode = FALLOC_FL_ZERO_RANGE | FALLOC_FL_KEEP_SIZE;
354 offset = le64_to_cpu(write_zeroes->slba) << req->ns->blksize_shift;
355 len = (((sector_t)le16_to_cpu(write_zeroes->length) + 1) <<
356 req->ns->blksize_shift);
358 if (unlikely(offset + len > req->ns->size)) {
359 nvmet_req_complete(req, errno_to_nvme_status(req, -ENOSPC));
363 ret = vfs_fallocate(req->ns->file, mode, offset, len);
364 nvmet_req_complete(req, ret < 0 ? errno_to_nvme_status(req, ret) : 0);
367 static void nvmet_file_execute_write_zeroes(struct nvmet_req *req)
369 if (!nvmet_check_data_len(req, 0))
371 INIT_WORK(&req->f.work, nvmet_file_write_zeroes_work);
372 schedule_work(&req->f.work);
375 u16 nvmet_file_parse_io_cmd(struct nvmet_req *req)
377 struct nvme_command *cmd = req->cmd;
379 switch (cmd->common.opcode) {
382 req->execute = nvmet_file_execute_rw;
385 req->execute = nvmet_file_execute_flush;
388 req->execute = nvmet_file_execute_dsm;
390 case nvme_cmd_write_zeroes:
391 req->execute = nvmet_file_execute_write_zeroes;
394 pr_err("unhandled cmd for file ns %d on qid %d\n",
395 cmd->common.opcode, req->sq->qid);
396 req->error_loc = offsetof(struct nvme_common_command, opcode);
397 return NVME_SC_INVALID_OPCODE | NVME_SC_DNR;