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 "nvmet.h"
12
13 #define NVMET_MAX_MPOOL_BVEC 16
14 #define NVMET_MIN_MPOOL_OBJ 16
15
nvmet_file_ns_disable(struct nvmet_ns * ns)16 void nvmet_file_ns_disable(struct nvmet_ns *ns)
17 {
18 if (ns->file) {
19 if (ns->buffered_io)
20 flush_workqueue(buffered_io_wq);
21 mempool_destroy(ns->bvec_pool);
22 ns->bvec_pool = NULL;
23 kmem_cache_destroy(ns->bvec_cache);
24 ns->bvec_cache = NULL;
25 fput(ns->file);
26 ns->file = NULL;
27 }
28 }
29
nvmet_file_ns_enable(struct nvmet_ns * ns)30 int nvmet_file_ns_enable(struct nvmet_ns *ns)
31 {
32 int flags = O_RDWR | O_LARGEFILE;
33 struct kstat stat;
34 int ret;
35
36 if (!ns->buffered_io)
37 flags |= O_DIRECT;
38
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);
44 }
45
46 ret = vfs_getattr(&ns->file->f_path,
47 &stat, STATX_SIZE, AT_STATX_FORCE_SYNC);
48 if (ret)
49 goto err;
50
51 ns->size = stat.size;
52 ns->blksize_shift = file_inode(ns->file)->i_blkbits;
53
54 ns->bvec_cache = kmem_cache_create("nvmet-bvec",
55 NVMET_MAX_MPOOL_BVEC * sizeof(struct bio_vec),
56 0, SLAB_HWCACHE_ALIGN, NULL);
57 if (!ns->bvec_cache) {
58 ret = -ENOMEM;
59 goto err;
60 }
61
62 ns->bvec_pool = mempool_create(NVMET_MIN_MPOOL_OBJ, mempool_alloc_slab,
63 mempool_free_slab, ns->bvec_cache);
64
65 if (!ns->bvec_pool) {
66 ret = -ENOMEM;
67 goto err;
68 }
69
70 return ret;
71 err:
72 ns->size = 0;
73 ns->blksize_shift = 0;
74 nvmet_file_ns_disable(ns);
75 return ret;
76 }
77
nvmet_file_init_bvec(struct bio_vec * bv,struct sg_page_iter * iter)78 static void nvmet_file_init_bvec(struct bio_vec *bv, struct sg_page_iter *iter)
79 {
80 bv->bv_page = sg_page_iter_page(iter);
81 bv->bv_offset = iter->sg->offset;
82 bv->bv_len = PAGE_SIZE - iter->sg->offset;
83 }
84
nvmet_file_submit_bvec(struct nvmet_req * req,loff_t pos,unsigned long nr_segs,size_t count)85 static ssize_t nvmet_file_submit_bvec(struct nvmet_req *req, loff_t pos,
86 unsigned long nr_segs, size_t count)
87 {
88 struct kiocb *iocb = &req->f.iocb;
89 ssize_t (*call_iter)(struct kiocb *iocb, struct iov_iter *iter);
90 struct iov_iter iter;
91 int ki_flags = 0, rw;
92 ssize_t ret;
93
94 if (req->cmd->rw.opcode == nvme_cmd_write) {
95 if (req->cmd->rw.control & cpu_to_le16(NVME_RW_FUA))
96 ki_flags = IOCB_DSYNC;
97 call_iter = req->ns->file->f_op->write_iter;
98 rw = WRITE;
99 } else {
100 call_iter = req->ns->file->f_op->read_iter;
101 rw = READ;
102 }
103
104 iov_iter_bvec(&iter, ITER_BVEC | rw, req->f.bvec, nr_segs, count);
105
106 iocb->ki_pos = pos;
107 iocb->ki_filp = req->ns->file;
108 iocb->ki_flags = ki_flags | iocb_flags(req->ns->file);
109
110 ret = call_iter(iocb, &iter);
111
112 if (ret != -EIOCBQUEUED && iocb->ki_complete)
113 iocb->ki_complete(iocb, ret, 0);
114
115 return ret;
116 }
117
nvmet_file_io_done(struct kiocb * iocb,long ret,long ret2)118 static void nvmet_file_io_done(struct kiocb *iocb, long ret, long ret2)
119 {
120 struct nvmet_req *req = container_of(iocb, struct nvmet_req, f.iocb);
121
122 if (req->f.bvec != req->inline_bvec) {
123 if (likely(req->f.mpool_alloc == false))
124 kfree(req->f.bvec);
125 else
126 mempool_free(req->f.bvec, req->ns->bvec_pool);
127 }
128
129 nvmet_req_complete(req, ret != req->data_len ?
130 NVME_SC_INTERNAL | NVME_SC_DNR : 0);
131 }
132
nvmet_file_execute_rw(struct nvmet_req * req)133 static void nvmet_file_execute_rw(struct nvmet_req *req)
134 {
135 ssize_t nr_bvec = DIV_ROUND_UP(req->data_len, PAGE_SIZE);
136 struct sg_page_iter sg_pg_iter;
137 unsigned long bv_cnt = 0;
138 bool is_sync = false;
139 size_t len = 0, total_len = 0;
140 ssize_t ret = 0;
141 loff_t pos;
142
143 if (!req->sg_cnt || !nr_bvec) {
144 nvmet_req_complete(req, 0);
145 return;
146 }
147
148 pos = le64_to_cpu(req->cmd->rw.slba) << req->ns->blksize_shift;
149 if (unlikely(pos + req->data_len > req->ns->size)) {
150 nvmet_req_complete(req, NVME_SC_LBA_RANGE | NVME_SC_DNR);
151 return;
152 }
153
154 if (nr_bvec > NVMET_MAX_INLINE_BIOVEC)
155 req->f.bvec = kmalloc_array(nr_bvec, sizeof(struct bio_vec),
156 GFP_KERNEL);
157 else
158 req->f.bvec = req->inline_bvec;
159
160 req->f.mpool_alloc = false;
161 if (unlikely(!req->f.bvec)) {
162 /* fallback under memory pressure */
163 req->f.bvec = mempool_alloc(req->ns->bvec_pool, GFP_KERNEL);
164 req->f.mpool_alloc = true;
165 if (nr_bvec > NVMET_MAX_MPOOL_BVEC)
166 is_sync = true;
167 }
168
169 memset(&req->f.iocb, 0, sizeof(struct kiocb));
170 for_each_sg_page(req->sg, &sg_pg_iter, req->sg_cnt, 0) {
171 nvmet_file_init_bvec(&req->f.bvec[bv_cnt], &sg_pg_iter);
172 len += req->f.bvec[bv_cnt].bv_len;
173 total_len += req->f.bvec[bv_cnt].bv_len;
174 bv_cnt++;
175
176 WARN_ON_ONCE((nr_bvec - 1) < 0);
177
178 if (unlikely(is_sync) &&
179 (nr_bvec - 1 == 0 || bv_cnt == NVMET_MAX_MPOOL_BVEC)) {
180 ret = nvmet_file_submit_bvec(req, pos, bv_cnt, len);
181 if (ret < 0)
182 goto out;
183 pos += len;
184 bv_cnt = 0;
185 len = 0;
186 }
187 nr_bvec--;
188 }
189
190 if (WARN_ON_ONCE(total_len != req->data_len))
191 ret = -EIO;
192 out:
193 if (unlikely(is_sync || ret)) {
194 nvmet_file_io_done(&req->f.iocb, ret < 0 ? ret : total_len, 0);
195 return;
196 }
197 req->f.iocb.ki_complete = nvmet_file_io_done;
198 nvmet_file_submit_bvec(req, pos, bv_cnt, total_len);
199 }
200
nvmet_file_buffered_io_work(struct work_struct * w)201 static void nvmet_file_buffered_io_work(struct work_struct *w)
202 {
203 struct nvmet_req *req = container_of(w, struct nvmet_req, f.work);
204
205 nvmet_file_execute_rw(req);
206 }
207
nvmet_file_execute_rw_buffered_io(struct nvmet_req * req)208 static void nvmet_file_execute_rw_buffered_io(struct nvmet_req *req)
209 {
210 INIT_WORK(&req->f.work, nvmet_file_buffered_io_work);
211 queue_work(buffered_io_wq, &req->f.work);
212 }
213
nvmet_file_flush(struct nvmet_req * req)214 u16 nvmet_file_flush(struct nvmet_req *req)
215 {
216 if (vfs_fsync(req->ns->file, 1) < 0)
217 return NVME_SC_INTERNAL | NVME_SC_DNR;
218 return 0;
219 }
220
nvmet_file_flush_work(struct work_struct * w)221 static void nvmet_file_flush_work(struct work_struct *w)
222 {
223 struct nvmet_req *req = container_of(w, struct nvmet_req, f.work);
224
225 nvmet_req_complete(req, nvmet_file_flush(req));
226 }
227
nvmet_file_execute_flush(struct nvmet_req * req)228 static void nvmet_file_execute_flush(struct nvmet_req *req)
229 {
230 INIT_WORK(&req->f.work, nvmet_file_flush_work);
231 schedule_work(&req->f.work);
232 }
233
nvmet_file_execute_discard(struct nvmet_req * req)234 static void nvmet_file_execute_discard(struct nvmet_req *req)
235 {
236 int mode = FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE;
237 struct nvme_dsm_range range;
238 loff_t offset, len;
239 u16 ret;
240 int i;
241
242 for (i = 0; i <= le32_to_cpu(req->cmd->dsm.nr); i++) {
243 ret = nvmet_copy_from_sgl(req, i * sizeof(range), &range,
244 sizeof(range));
245 if (ret)
246 break;
247
248 offset = le64_to_cpu(range.slba) << req->ns->blksize_shift;
249 len = le32_to_cpu(range.nlb) << req->ns->blksize_shift;
250 if (offset + len > req->ns->size) {
251 ret = NVME_SC_LBA_RANGE | NVME_SC_DNR;
252 break;
253 }
254
255 if (vfs_fallocate(req->ns->file, mode, offset, len)) {
256 ret = NVME_SC_INTERNAL | NVME_SC_DNR;
257 break;
258 }
259 }
260
261 nvmet_req_complete(req, ret);
262 }
263
nvmet_file_dsm_work(struct work_struct * w)264 static void nvmet_file_dsm_work(struct work_struct *w)
265 {
266 struct nvmet_req *req = container_of(w, struct nvmet_req, f.work);
267
268 switch (le32_to_cpu(req->cmd->dsm.attributes)) {
269 case NVME_DSMGMT_AD:
270 nvmet_file_execute_discard(req);
271 return;
272 case NVME_DSMGMT_IDR:
273 case NVME_DSMGMT_IDW:
274 default:
275 /* Not supported yet */
276 nvmet_req_complete(req, 0);
277 return;
278 }
279 }
280
nvmet_file_execute_dsm(struct nvmet_req * req)281 static void nvmet_file_execute_dsm(struct nvmet_req *req)
282 {
283 INIT_WORK(&req->f.work, nvmet_file_dsm_work);
284 schedule_work(&req->f.work);
285 }
286
nvmet_file_write_zeroes_work(struct work_struct * w)287 static void nvmet_file_write_zeroes_work(struct work_struct *w)
288 {
289 struct nvmet_req *req = container_of(w, struct nvmet_req, f.work);
290 struct nvme_write_zeroes_cmd *write_zeroes = &req->cmd->write_zeroes;
291 int mode = FALLOC_FL_ZERO_RANGE | FALLOC_FL_KEEP_SIZE;
292 loff_t offset;
293 loff_t len;
294 int ret;
295
296 offset = le64_to_cpu(write_zeroes->slba) << req->ns->blksize_shift;
297 len = (((sector_t)le16_to_cpu(write_zeroes->length) + 1) <<
298 req->ns->blksize_shift);
299
300 if (unlikely(offset + len > req->ns->size)) {
301 nvmet_req_complete(req, NVME_SC_LBA_RANGE | NVME_SC_DNR);
302 return;
303 }
304
305 ret = vfs_fallocate(req->ns->file, mode, offset, len);
306 nvmet_req_complete(req, ret < 0 ? NVME_SC_INTERNAL | NVME_SC_DNR : 0);
307 }
308
nvmet_file_execute_write_zeroes(struct nvmet_req * req)309 static void nvmet_file_execute_write_zeroes(struct nvmet_req *req)
310 {
311 INIT_WORK(&req->f.work, nvmet_file_write_zeroes_work);
312 schedule_work(&req->f.work);
313 }
314
nvmet_file_parse_io_cmd(struct nvmet_req * req)315 u16 nvmet_file_parse_io_cmd(struct nvmet_req *req)
316 {
317 struct nvme_command *cmd = req->cmd;
318
319 switch (cmd->common.opcode) {
320 case nvme_cmd_read:
321 case nvme_cmd_write:
322 if (req->ns->buffered_io)
323 req->execute = nvmet_file_execute_rw_buffered_io;
324 else
325 req->execute = nvmet_file_execute_rw;
326 req->data_len = nvmet_rw_len(req);
327 return 0;
328 case nvme_cmd_flush:
329 req->execute = nvmet_file_execute_flush;
330 req->data_len = 0;
331 return 0;
332 case nvme_cmd_dsm:
333 req->execute = nvmet_file_execute_dsm;
334 req->data_len = (le32_to_cpu(cmd->dsm.nr) + 1) *
335 sizeof(struct nvme_dsm_range);
336 return 0;
337 case nvme_cmd_write_zeroes:
338 req->execute = nvmet_file_execute_write_zeroes;
339 req->data_len = 0;
340 return 0;
341 default:
342 pr_err("unhandled cmd for file ns %d on qid %d\n",
343 cmd->common.opcode, req->sq->qid);
344 return NVME_SC_INVALID_OPCODE | NVME_SC_DNR;
345 }
346 }
347
