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 	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 
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 
85 static ssize_t nvmet_file_submit_bvec(struct nvmet_req *req, loff_t pos,
86 		unsigned long nr_segs, size_t count, int ki_flags)
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 rw;
92 
93 	if (req->cmd->rw.opcode == nvme_cmd_write) {
94 		if (req->cmd->rw.control & cpu_to_le16(NVME_RW_FUA))
95 			ki_flags |= IOCB_DSYNC;
96 		call_iter = req->ns->file->f_op->write_iter;
97 		rw = WRITE;
98 	} else {
99 		call_iter = req->ns->file->f_op->read_iter;
100 		rw = READ;
101 	}
102 
103 	iov_iter_bvec(&iter, rw, req->f.bvec, nr_segs, count);
104 
105 	iocb->ki_pos = pos;
106 	iocb->ki_filp = req->ns->file;
107 	iocb->ki_flags = ki_flags | iocb_flags(req->ns->file);
108 
109 	return call_iter(iocb, &iter);
110 }
111 
112 static void nvmet_file_io_done(struct kiocb *iocb, long ret, long ret2)
113 {
114 	struct nvmet_req *req = container_of(iocb, struct nvmet_req, f.iocb);
115 	u16 status = NVME_SC_SUCCESS;
116 
117 	if (req->f.bvec != req->inline_bvec) {
118 		if (likely(req->f.mpool_alloc == false))
119 			kfree(req->f.bvec);
120 		else
121 			mempool_free(req->f.bvec, req->ns->bvec_pool);
122 	}
123 
124 	if (unlikely(ret != req->data_len))
125 		status = errno_to_nvme_status(req, ret);
126 	nvmet_req_complete(req, status);
127 }
128 
129 static bool nvmet_file_execute_io(struct nvmet_req *req, int ki_flags)
130 {
131 	ssize_t nr_bvec = DIV_ROUND_UP(req->data_len, PAGE_SIZE);
132 	struct sg_page_iter sg_pg_iter;
133 	unsigned long bv_cnt = 0;
134 	bool is_sync = false;
135 	size_t len = 0, total_len = 0;
136 	ssize_t ret = 0;
137 	loff_t pos;
138 
139 
140 	if (req->f.mpool_alloc && nr_bvec > NVMET_MAX_MPOOL_BVEC)
141 		is_sync = true;
142 
143 	pos = le64_to_cpu(req->cmd->rw.slba) << req->ns->blksize_shift;
144 	if (unlikely(pos + req->data_len > req->ns->size)) {
145 		nvmet_req_complete(req, errno_to_nvme_status(req, -ENOSPC));
146 		return true;
147 	}
148 
149 	memset(&req->f.iocb, 0, sizeof(struct kiocb));
150 	for_each_sg_page(req->sg, &sg_pg_iter, req->sg_cnt, 0) {
151 		nvmet_file_init_bvec(&req->f.bvec[bv_cnt], &sg_pg_iter);
152 		len += req->f.bvec[bv_cnt].bv_len;
153 		total_len += req->f.bvec[bv_cnt].bv_len;
154 		bv_cnt++;
155 
156 		WARN_ON_ONCE((nr_bvec - 1) < 0);
157 
158 		if (unlikely(is_sync) &&
159 		    (nr_bvec - 1 == 0 || bv_cnt == NVMET_MAX_MPOOL_BVEC)) {
160 			ret = nvmet_file_submit_bvec(req, pos, bv_cnt, len, 0);
161 			if (ret < 0)
162 				goto complete;
163 
164 			pos += len;
165 			bv_cnt = 0;
166 			len = 0;
167 		}
168 		nr_bvec--;
169 	}
170 
171 	if (WARN_ON_ONCE(total_len != req->data_len)) {
172 		ret = -EIO;
173 		goto complete;
174 	}
175 
176 	if (unlikely(is_sync)) {
177 		ret = total_len;
178 		goto complete;
179 	}
180 
181 	/*
182 	 * A NULL ki_complete ask for synchronous execution, which we want
183 	 * for the IOCB_NOWAIT case.
184 	 */
185 	if (!(ki_flags & IOCB_NOWAIT))
186 		req->f.iocb.ki_complete = nvmet_file_io_done;
187 
188 	ret = nvmet_file_submit_bvec(req, pos, bv_cnt, total_len, ki_flags);
189 
190 	switch (ret) {
191 	case -EIOCBQUEUED:
192 		return true;
193 	case -EAGAIN:
194 		if (WARN_ON_ONCE(!(ki_flags & IOCB_NOWAIT)))
195 			goto complete;
196 		return false;
197 	case -EOPNOTSUPP:
198 		/*
199 		 * For file systems returning error -EOPNOTSUPP, handle
200 		 * IOCB_NOWAIT error case separately and retry without
201 		 * IOCB_NOWAIT.
202 		 */
203 		if ((ki_flags & IOCB_NOWAIT))
204 			return false;
205 		break;
206 	}
207 
208 complete:
209 	nvmet_file_io_done(&req->f.iocb, ret, 0);
210 	return true;
211 }
212 
213 static void nvmet_file_buffered_io_work(struct work_struct *w)
214 {
215 	struct nvmet_req *req = container_of(w, struct nvmet_req, f.work);
216 
217 	nvmet_file_execute_io(req, 0);
218 }
219 
220 static void nvmet_file_submit_buffered_io(struct nvmet_req *req)
221 {
222 	INIT_WORK(&req->f.work, nvmet_file_buffered_io_work);
223 	queue_work(buffered_io_wq, &req->f.work);
224 }
225 
226 static void nvmet_file_execute_rw(struct nvmet_req *req)
227 {
228 	ssize_t nr_bvec = DIV_ROUND_UP(req->data_len, PAGE_SIZE);
229 
230 	if (!req->sg_cnt || !nr_bvec) {
231 		nvmet_req_complete(req, 0);
232 		return;
233 	}
234 
235 	if (nr_bvec > NVMET_MAX_INLINE_BIOVEC)
236 		req->f.bvec = kmalloc_array(nr_bvec, sizeof(struct bio_vec),
237 				GFP_KERNEL);
238 	else
239 		req->f.bvec = req->inline_bvec;
240 
241 	if (unlikely(!req->f.bvec)) {
242 		/* fallback under memory pressure */
243 		req->f.bvec = mempool_alloc(req->ns->bvec_pool, GFP_KERNEL);
244 		req->f.mpool_alloc = true;
245 	} else
246 		req->f.mpool_alloc = false;
247 
248 	if (req->ns->buffered_io) {
249 		if (likely(!req->f.mpool_alloc) &&
250 				nvmet_file_execute_io(req, IOCB_NOWAIT))
251 			return;
252 		nvmet_file_submit_buffered_io(req);
253 	} else
254 		nvmet_file_execute_io(req, 0);
255 }
256 
257 u16 nvmet_file_flush(struct nvmet_req *req)
258 {
259 	return errno_to_nvme_status(req, vfs_fsync(req->ns->file, 1));
260 }
261 
262 static void nvmet_file_flush_work(struct work_struct *w)
263 {
264 	struct nvmet_req *req = container_of(w, struct nvmet_req, f.work);
265 
266 	nvmet_req_complete(req, nvmet_file_flush(req));
267 }
268 
269 static void nvmet_file_execute_flush(struct nvmet_req *req)
270 {
271 	INIT_WORK(&req->f.work, nvmet_file_flush_work);
272 	schedule_work(&req->f.work);
273 }
274 
275 static void nvmet_file_execute_discard(struct nvmet_req *req)
276 {
277 	int mode = FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE;
278 	struct nvme_dsm_range range;
279 	loff_t offset, len;
280 	u16 status = 0;
281 	int ret;
282 	int i;
283 
284 	for (i = 0; i <= le32_to_cpu(req->cmd->dsm.nr); i++) {
285 		status = nvmet_copy_from_sgl(req, i * sizeof(range), &range,
286 					sizeof(range));
287 		if (status)
288 			break;
289 
290 		offset = le64_to_cpu(range.slba) << req->ns->blksize_shift;
291 		len = le32_to_cpu(range.nlb);
292 		len <<= req->ns->blksize_shift;
293 		if (offset + len > req->ns->size) {
294 			req->error_slba = le64_to_cpu(range.slba);
295 			status = errno_to_nvme_status(req, -ENOSPC);
296 			break;
297 		}
298 
299 		ret = vfs_fallocate(req->ns->file, mode, offset, len);
300 		if (ret && ret != -EOPNOTSUPP) {
301 			req->error_slba = le64_to_cpu(range.slba);
302 			status = errno_to_nvme_status(req, ret);
303 			break;
304 		}
305 	}
306 
307 	nvmet_req_complete(req, status);
308 }
309 
310 static void nvmet_file_dsm_work(struct work_struct *w)
311 {
312 	struct nvmet_req *req = container_of(w, struct nvmet_req, f.work);
313 
314 	switch (le32_to_cpu(req->cmd->dsm.attributes)) {
315 	case NVME_DSMGMT_AD:
316 		nvmet_file_execute_discard(req);
317 		return;
318 	case NVME_DSMGMT_IDR:
319 	case NVME_DSMGMT_IDW:
320 	default:
321 		/* Not supported yet */
322 		nvmet_req_complete(req, 0);
323 		return;
324 	}
325 }
326 
327 static void nvmet_file_execute_dsm(struct nvmet_req *req)
328 {
329 	INIT_WORK(&req->f.work, nvmet_file_dsm_work);
330 	schedule_work(&req->f.work);
331 }
332 
333 static void nvmet_file_write_zeroes_work(struct work_struct *w)
334 {
335 	struct nvmet_req *req = container_of(w, struct nvmet_req, f.work);
336 	struct nvme_write_zeroes_cmd *write_zeroes = &req->cmd->write_zeroes;
337 	int mode = FALLOC_FL_ZERO_RANGE | FALLOC_FL_KEEP_SIZE;
338 	loff_t offset;
339 	loff_t len;
340 	int ret;
341 
342 	offset = le64_to_cpu(write_zeroes->slba) << req->ns->blksize_shift;
343 	len = (((sector_t)le16_to_cpu(write_zeroes->length) + 1) <<
344 			req->ns->blksize_shift);
345 
346 	if (unlikely(offset + len > req->ns->size)) {
347 		nvmet_req_complete(req, errno_to_nvme_status(req, -ENOSPC));
348 		return;
349 	}
350 
351 	ret = vfs_fallocate(req->ns->file, mode, offset, len);
352 	nvmet_req_complete(req, ret < 0 ? errno_to_nvme_status(req, ret) : 0);
353 }
354 
355 static void nvmet_file_execute_write_zeroes(struct nvmet_req *req)
356 {
357 	INIT_WORK(&req->f.work, nvmet_file_write_zeroes_work);
358 	schedule_work(&req->f.work);
359 }
360 
361 u16 nvmet_file_parse_io_cmd(struct nvmet_req *req)
362 {
363 	struct nvme_command *cmd = req->cmd;
364 
365 	switch (cmd->common.opcode) {
366 	case nvme_cmd_read:
367 	case nvme_cmd_write:
368 		req->execute = nvmet_file_execute_rw;
369 		req->data_len = nvmet_rw_len(req);
370 		return 0;
371 	case nvme_cmd_flush:
372 		req->execute = nvmet_file_execute_flush;
373 		req->data_len = 0;
374 		return 0;
375 	case nvme_cmd_dsm:
376 		req->execute = nvmet_file_execute_dsm;
377 		req->data_len = (le32_to_cpu(cmd->dsm.nr) + 1) *
378 			sizeof(struct nvme_dsm_range);
379 		return 0;
380 	case nvme_cmd_write_zeroes:
381 		req->execute = nvmet_file_execute_write_zeroes;
382 		req->data_len = 0;
383 		return 0;
384 	default:
385 		pr_err("unhandled cmd for file ns %d on qid %d\n",
386 				cmd->common.opcode, req->sq->qid);
387 		req->error_loc = offsetof(struct nvme_common_command, opcode);
388 		return NVME_SC_INVALID_OPCODE | NVME_SC_DNR;
389 	}
390 }
391