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 
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 
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 	/*
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_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) {
63 		ret = -ENOMEM;
64 		goto err;
65 	}
66 
67 	ns->bvec_pool = mempool_create(NVMET_MIN_MPOOL_OBJ, mempool_alloc_slab,
68 			mempool_free_slab, ns->bvec_cache);
69 
70 	if (!ns->bvec_pool) {
71 		ret = -ENOMEM;
72 		goto err;
73 	}
74 
75 	return ret;
76 err:
77 	ns->size = 0;
78 	ns->blksize_shift = 0;
79 	nvmet_file_ns_disable(ns);
80 	return ret;
81 }
82 
83 static void nvmet_file_init_bvec(struct bio_vec *bv, struct scatterlist *sg)
84 {
85 	bv->bv_page = sg_page(sg);
86 	bv->bv_offset = sg->offset;
87 	bv->bv_len = sg->length;
88 }
89 
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)
92 {
93 	struct kiocb *iocb = &req->f.iocb;
94 	ssize_t (*call_iter)(struct kiocb *iocb, struct iov_iter *iter);
95 	struct iov_iter iter;
96 	int rw;
97 
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;
102 		rw = WRITE;
103 	} else {
104 		call_iter = req->ns->file->f_op->read_iter;
105 		rw = READ;
106 	}
107 
108 	iov_iter_bvec(&iter, rw, req->f.bvec, nr_segs, count);
109 
110 	iocb->ki_pos = pos;
111 	iocb->ki_filp = req->ns->file;
112 	iocb->ki_flags = ki_flags | iocb_flags(req->ns->file);
113 
114 	return call_iter(iocb, &iter);
115 }
116 
117 static void nvmet_file_io_done(struct kiocb *iocb, long ret, long ret2)
118 {
119 	struct nvmet_req *req = container_of(iocb, struct nvmet_req, f.iocb);
120 	u16 status = NVME_SC_SUCCESS;
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 	if (unlikely(ret != req->transfer_len))
130 		status = errno_to_nvme_status(req, ret);
131 	nvmet_req_complete(req, status);
132 }
133 
134 static bool nvmet_file_execute_io(struct nvmet_req *req, int ki_flags)
135 {
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;
140 	ssize_t ret = 0;
141 	loff_t pos;
142 	int i;
143 	struct scatterlist *sg;
144 
145 	if (req->f.mpool_alloc && nr_bvec > NVMET_MAX_MPOOL_BVEC)
146 		is_sync = true;
147 
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));
151 		return true;
152 	}
153 
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;
159 		bv_cnt++;
160 
161 		WARN_ON_ONCE((nr_bvec - 1) < 0);
162 
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);
166 			if (ret < 0)
167 				goto complete;
168 
169 			pos += len;
170 			bv_cnt = 0;
171 			len = 0;
172 		}
173 		nr_bvec--;
174 	}
175 
176 	if (WARN_ON_ONCE(total_len != req->transfer_len)) {
177 		ret = -EIO;
178 		goto complete;
179 	}
180 
181 	if (unlikely(is_sync)) {
182 		ret = total_len;
183 		goto complete;
184 	}
185 
186 	/*
187 	 * A NULL ki_complete ask for synchronous execution, which we want
188 	 * for the IOCB_NOWAIT case.
189 	 */
190 	if (!(ki_flags & IOCB_NOWAIT))
191 		req->f.iocb.ki_complete = nvmet_file_io_done;
192 
193 	ret = nvmet_file_submit_bvec(req, pos, bv_cnt, total_len, ki_flags);
194 
195 	switch (ret) {
196 	case -EIOCBQUEUED:
197 		return true;
198 	case -EAGAIN:
199 		if (WARN_ON_ONCE(!(ki_flags & IOCB_NOWAIT)))
200 			goto complete;
201 		return false;
202 	case -EOPNOTSUPP:
203 		/*
204 		 * For file systems returning error -EOPNOTSUPP, handle
205 		 * IOCB_NOWAIT error case separately and retry without
206 		 * IOCB_NOWAIT.
207 		 */
208 		if ((ki_flags & IOCB_NOWAIT))
209 			return false;
210 		break;
211 	}
212 
213 complete:
214 	nvmet_file_io_done(&req->f.iocb, ret, 0);
215 	return true;
216 }
217 
218 static void nvmet_file_buffered_io_work(struct work_struct *w)
219 {
220 	struct nvmet_req *req = container_of(w, struct nvmet_req, f.work);
221 
222 	nvmet_file_execute_io(req, 0);
223 }
224 
225 static void nvmet_file_submit_buffered_io(struct nvmet_req *req)
226 {
227 	INIT_WORK(&req->f.work, nvmet_file_buffered_io_work);
228 	queue_work(buffered_io_wq, &req->f.work);
229 }
230 
231 static void nvmet_file_execute_rw(struct nvmet_req *req)
232 {
233 	ssize_t nr_bvec = req->sg_cnt;
234 
235 	if (!nvmet_check_data_len(req, nvmet_rw_len(req)))
236 		return;
237 
238 	if (!req->sg_cnt || !nr_bvec) {
239 		nvmet_req_complete(req, 0);
240 		return;
241 	}
242 
243 	if (nr_bvec > NVMET_MAX_INLINE_BIOVEC)
244 		req->f.bvec = kmalloc_array(nr_bvec, sizeof(struct bio_vec),
245 				GFP_KERNEL);
246 	else
247 		req->f.bvec = req->inline_bvec;
248 
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;
253 	} else
254 		req->f.mpool_alloc = false;
255 
256 	if (req->ns->buffered_io) {
257 		if (likely(!req->f.mpool_alloc) &&
258 				nvmet_file_execute_io(req, IOCB_NOWAIT))
259 			return;
260 		nvmet_file_submit_buffered_io(req);
261 	} else
262 		nvmet_file_execute_io(req, 0);
263 }
264 
265 u16 nvmet_file_flush(struct nvmet_req *req)
266 {
267 	return errno_to_nvme_status(req, vfs_fsync(req->ns->file, 1));
268 }
269 
270 static void nvmet_file_flush_work(struct work_struct *w)
271 {
272 	struct nvmet_req *req = container_of(w, struct nvmet_req, f.work);
273 
274 	nvmet_req_complete(req, nvmet_file_flush(req));
275 }
276 
277 static void nvmet_file_execute_flush(struct nvmet_req *req)
278 {
279 	if (!nvmet_check_data_len(req, 0))
280 		return;
281 	INIT_WORK(&req->f.work, nvmet_file_flush_work);
282 	schedule_work(&req->f.work);
283 }
284 
285 static void nvmet_file_execute_discard(struct nvmet_req *req)
286 {
287 	int mode = FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE;
288 	struct nvme_dsm_range range;
289 	loff_t offset, len;
290 	u16 status = 0;
291 	int ret;
292 	int i;
293 
294 	for (i = 0; i <= le32_to_cpu(req->cmd->dsm.nr); i++) {
295 		status = nvmet_copy_from_sgl(req, i * sizeof(range), &range,
296 					sizeof(range));
297 		if (status)
298 			break;
299 
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);
306 			break;
307 		}
308 
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);
313 			break;
314 		}
315 	}
316 
317 	nvmet_req_complete(req, status);
318 }
319 
320 static void nvmet_file_dsm_work(struct work_struct *w)
321 {
322 	struct nvmet_req *req = container_of(w, struct nvmet_req, f.work);
323 
324 	switch (le32_to_cpu(req->cmd->dsm.attributes)) {
325 	case NVME_DSMGMT_AD:
326 		nvmet_file_execute_discard(req);
327 		return;
328 	case NVME_DSMGMT_IDR:
329 	case NVME_DSMGMT_IDW:
330 	default:
331 		/* Not supported yet */
332 		nvmet_req_complete(req, 0);
333 		return;
334 	}
335 }
336 
337 static void nvmet_file_execute_dsm(struct nvmet_req *req)
338 {
339 	if (!nvmet_check_data_len(req, nvmet_dsm_len(req)))
340 		return;
341 	INIT_WORK(&req->f.work, nvmet_file_dsm_work);
342 	schedule_work(&req->f.work);
343 }
344 
345 static void nvmet_file_write_zeroes_work(struct work_struct *w)
346 {
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;
350 	loff_t offset;
351 	loff_t len;
352 	int ret;
353 
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);
357 
358 	if (unlikely(offset + len > req->ns->size)) {
359 		nvmet_req_complete(req, errno_to_nvme_status(req, -ENOSPC));
360 		return;
361 	}
362 
363 	ret = vfs_fallocate(req->ns->file, mode, offset, len);
364 	nvmet_req_complete(req, ret < 0 ? errno_to_nvme_status(req, ret) : 0);
365 }
366 
367 static void nvmet_file_execute_write_zeroes(struct nvmet_req *req)
368 {
369 	if (!nvmet_check_data_len(req, 0))
370 		return;
371 	INIT_WORK(&req->f.work, nvmet_file_write_zeroes_work);
372 	schedule_work(&req->f.work);
373 }
374 
375 u16 nvmet_file_parse_io_cmd(struct nvmet_req *req)
376 {
377 	struct nvme_command *cmd = req->cmd;
378 
379 	switch (cmd->common.opcode) {
380 	case nvme_cmd_read:
381 	case nvme_cmd_write:
382 		req->execute = nvmet_file_execute_rw;
383 		return 0;
384 	case nvme_cmd_flush:
385 		req->execute = nvmet_file_execute_flush;
386 		return 0;
387 	case nvme_cmd_dsm:
388 		req->execute = nvmet_file_execute_dsm;
389 		return 0;
390 	case nvme_cmd_write_zeroes:
391 		req->execute = nvmet_file_execute_write_zeroes;
392 		return 0;
393 	default:
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;
398 	}
399 }
400