1 /*
2  * Copyright (c) 2004, 2005, 2006 Voltaire, Inc. All rights reserved.
3  *
4  * This software is available to you under a choice of one of two
5  * licenses.  You may choose to be licensed under the terms of the GNU
6  * General Public License (GPL) Version 2, available from the file
7  * COPYING in the main directory of this source tree, or the
8  * OpenIB.org BSD license below:
9  *
10  *     Redistribution and use in source and binary forms, with or
11  *     without modification, are permitted provided that the following
12  *     conditions are met:
13  *
14  *	- Redistributions of source code must retain the above
15  *	  copyright notice, this list of conditions and the following
16  *	  disclaimer.
17  *
18  *	- Redistributions in binary form must reproduce the above
19  *	  copyright notice, this list of conditions and the following
20  *	  disclaimer in the documentation and/or other materials
21  *	  provided with the distribution.
22  *
23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30  * SOFTWARE.
31  */
32 #include <linux/module.h>
33 #include <linux/kernel.h>
34 #include <linux/slab.h>
35 #include <linux/mm.h>
36 #include <linux/highmem.h>
37 #include <linux/scatterlist.h>
38 
39 #include "iscsi_iser.h"
40 
41 #define ISER_KMALLOC_THRESHOLD 0x20000 /* 128K - kmalloc limit */
42 
43 /**
44  * iser_start_rdma_unaligned_sg
45  */
46 static int iser_start_rdma_unaligned_sg(struct iscsi_iser_task *iser_task,
47 					enum iser_data_dir cmd_dir)
48 {
49 	int dma_nents;
50 	struct ib_device *dev;
51 	char *mem = NULL;
52 	struct iser_data_buf *data = &iser_task->data[cmd_dir];
53 	unsigned long  cmd_data_len = data->data_len;
54 
55 	if (cmd_data_len > ISER_KMALLOC_THRESHOLD)
56 		mem = (void *)__get_free_pages(GFP_ATOMIC,
57 		      ilog2(roundup_pow_of_two(cmd_data_len)) - PAGE_SHIFT);
58 	else
59 		mem = kmalloc(cmd_data_len, GFP_ATOMIC);
60 
61 	if (mem == NULL) {
62 		iser_err("Failed to allocate mem size %d %d for copying sglist\n",
63 			 data->size,(int)cmd_data_len);
64 		return -ENOMEM;
65 	}
66 
67 	if (cmd_dir == ISER_DIR_OUT) {
68 		/* copy the unaligned sg the buffer which is used for RDMA */
69 		struct scatterlist *sgl = (struct scatterlist *)data->buf;
70 		struct scatterlist *sg;
71 		int i;
72 		char *p, *from;
73 
74 		p = mem;
75 		for_each_sg(sgl, sg, data->size, i) {
76 			from = kmap_atomic(sg_page(sg));
77 			memcpy(p,
78 			       from + sg->offset,
79 			       sg->length);
80 			kunmap_atomic(from);
81 			p += sg->length;
82 		}
83 	}
84 
85 	sg_init_one(&iser_task->data_copy[cmd_dir].sg_single, mem, cmd_data_len);
86 	iser_task->data_copy[cmd_dir].buf  =
87 		&iser_task->data_copy[cmd_dir].sg_single;
88 	iser_task->data_copy[cmd_dir].size = 1;
89 
90 	iser_task->data_copy[cmd_dir].copy_buf  = mem;
91 
92 	dev = iser_task->iser_conn->ib_conn->device->ib_device;
93 	dma_nents = ib_dma_map_sg(dev,
94 				  &iser_task->data_copy[cmd_dir].sg_single,
95 				  1,
96 				  (cmd_dir == ISER_DIR_OUT) ?
97 				  DMA_TO_DEVICE : DMA_FROM_DEVICE);
98 	BUG_ON(dma_nents == 0);
99 
100 	iser_task->data_copy[cmd_dir].dma_nents = dma_nents;
101 	return 0;
102 }
103 
104 /**
105  * iser_finalize_rdma_unaligned_sg
106  */
107 void iser_finalize_rdma_unaligned_sg(struct iscsi_iser_task *iser_task,
108 				     enum iser_data_dir         cmd_dir)
109 {
110 	struct ib_device *dev;
111 	struct iser_data_buf *mem_copy;
112 	unsigned long  cmd_data_len;
113 
114 	dev = iser_task->iser_conn->ib_conn->device->ib_device;
115 	mem_copy = &iser_task->data_copy[cmd_dir];
116 
117 	ib_dma_unmap_sg(dev, &mem_copy->sg_single, 1,
118 			(cmd_dir == ISER_DIR_OUT) ?
119 			DMA_TO_DEVICE : DMA_FROM_DEVICE);
120 
121 	if (cmd_dir == ISER_DIR_IN) {
122 		char *mem;
123 		struct scatterlist *sgl, *sg;
124 		unsigned char *p, *to;
125 		unsigned int sg_size;
126 		int i;
127 
128 		/* copy back read RDMA to unaligned sg */
129 		mem	= mem_copy->copy_buf;
130 
131 		sgl	= (struct scatterlist *)iser_task->data[ISER_DIR_IN].buf;
132 		sg_size = iser_task->data[ISER_DIR_IN].size;
133 
134 		p = mem;
135 		for_each_sg(sgl, sg, sg_size, i) {
136 			to = kmap_atomic(sg_page(sg));
137 			memcpy(to + sg->offset,
138 			       p,
139 			       sg->length);
140 			kunmap_atomic(to);
141 			p += sg->length;
142 		}
143 	}
144 
145 	cmd_data_len = iser_task->data[cmd_dir].data_len;
146 
147 	if (cmd_data_len > ISER_KMALLOC_THRESHOLD)
148 		free_pages((unsigned long)mem_copy->copy_buf,
149 			   ilog2(roundup_pow_of_two(cmd_data_len)) - PAGE_SHIFT);
150 	else
151 		kfree(mem_copy->copy_buf);
152 
153 	mem_copy->copy_buf = NULL;
154 }
155 
156 #define IS_4K_ALIGNED(addr)	((((unsigned long)addr) & ~MASK_4K) == 0)
157 
158 /**
159  * iser_sg_to_page_vec - Translates scatterlist entries to physical addresses
160  * and returns the length of resulting physical address array (may be less than
161  * the original due to possible compaction).
162  *
163  * we build a "page vec" under the assumption that the SG meets the RDMA
164  * alignment requirements. Other then the first and last SG elements, all
165  * the "internal" elements can be compacted into a list whose elements are
166  * dma addresses of physical pages. The code supports also the weird case
167  * where --few fragments of the same page-- are present in the SG as
168  * consecutive elements. Also, it handles one entry SG.
169  */
170 
171 static int iser_sg_to_page_vec(struct iser_data_buf *data,
172 			       struct iser_page_vec *page_vec,
173 			       struct ib_device *ibdev)
174 {
175 	struct scatterlist *sg, *sgl = (struct scatterlist *)data->buf;
176 	u64 start_addr, end_addr, page, chunk_start = 0;
177 	unsigned long total_sz = 0;
178 	unsigned int dma_len;
179 	int i, new_chunk, cur_page, last_ent = data->dma_nents - 1;
180 
181 	/* compute the offset of first element */
182 	page_vec->offset = (u64) sgl[0].offset & ~MASK_4K;
183 
184 	new_chunk = 1;
185 	cur_page  = 0;
186 	for_each_sg(sgl, sg, data->dma_nents, i) {
187 		start_addr = ib_sg_dma_address(ibdev, sg);
188 		if (new_chunk)
189 			chunk_start = start_addr;
190 		dma_len = ib_sg_dma_len(ibdev, sg);
191 		end_addr = start_addr + dma_len;
192 		total_sz += dma_len;
193 
194 		/* collect page fragments until aligned or end of SG list */
195 		if (!IS_4K_ALIGNED(end_addr) && i < last_ent) {
196 			new_chunk = 0;
197 			continue;
198 		}
199 		new_chunk = 1;
200 
201 		/* address of the first page in the contiguous chunk;
202 		   masking relevant for the very first SG entry,
203 		   which might be unaligned */
204 		page = chunk_start & MASK_4K;
205 		do {
206 			page_vec->pages[cur_page++] = page;
207 			page += SIZE_4K;
208 		} while (page < end_addr);
209 	}
210 
211 	page_vec->data_size = total_sz;
212 	iser_dbg("page_vec->data_size:%d cur_page %d\n", page_vec->data_size,cur_page);
213 	return cur_page;
214 }
215 
216 
217 /**
218  * iser_data_buf_aligned_len - Tries to determine the maximal correctly aligned
219  * for RDMA sub-list of a scatter-gather list of memory buffers, and  returns
220  * the number of entries which are aligned correctly. Supports the case where
221  * consecutive SG elements are actually fragments of the same physcial page.
222  */
223 static int iser_data_buf_aligned_len(struct iser_data_buf *data,
224 				      struct ib_device *ibdev)
225 {
226 	struct scatterlist *sgl, *sg, *next_sg = NULL;
227 	u64 start_addr, end_addr;
228 	int i, ret_len, start_check = 0;
229 
230 	if (data->dma_nents == 1)
231 		return 1;
232 
233 	sgl = (struct scatterlist *)data->buf;
234 	start_addr  = ib_sg_dma_address(ibdev, sgl);
235 
236 	for_each_sg(sgl, sg, data->dma_nents, i) {
237 		if (start_check && !IS_4K_ALIGNED(start_addr))
238 			break;
239 
240 		next_sg = sg_next(sg);
241 		if (!next_sg)
242 			break;
243 
244 		end_addr    = start_addr + ib_sg_dma_len(ibdev, sg);
245 		start_addr  = ib_sg_dma_address(ibdev, next_sg);
246 
247 		if (end_addr == start_addr) {
248 			start_check = 0;
249 			continue;
250 		} else
251 			start_check = 1;
252 
253 		if (!IS_4K_ALIGNED(end_addr))
254 			break;
255 	}
256 	ret_len = (next_sg) ? i : i+1;
257 	iser_dbg("Found %d aligned entries out of %d in sg:0x%p\n",
258 		 ret_len, data->dma_nents, data);
259 	return ret_len;
260 }
261 
262 static void iser_data_buf_dump(struct iser_data_buf *data,
263 			       struct ib_device *ibdev)
264 {
265 	struct scatterlist *sgl = (struct scatterlist *)data->buf;
266 	struct scatterlist *sg;
267 	int i;
268 
269 	if (iser_debug_level == 0)
270 		return;
271 
272 	for_each_sg(sgl, sg, data->dma_nents, i)
273 		iser_warn("sg[%d] dma_addr:0x%lX page:0x%p "
274 			 "off:0x%x sz:0x%x dma_len:0x%x\n",
275 			 i, (unsigned long)ib_sg_dma_address(ibdev, sg),
276 			 sg_page(sg), sg->offset,
277 			 sg->length, ib_sg_dma_len(ibdev, sg));
278 }
279 
280 static void iser_dump_page_vec(struct iser_page_vec *page_vec)
281 {
282 	int i;
283 
284 	iser_err("page vec length %d data size %d\n",
285 		 page_vec->length, page_vec->data_size);
286 	for (i = 0; i < page_vec->length; i++)
287 		iser_err("%d %lx\n",i,(unsigned long)page_vec->pages[i]);
288 }
289 
290 static void iser_page_vec_build(struct iser_data_buf *data,
291 				struct iser_page_vec *page_vec,
292 				struct ib_device *ibdev)
293 {
294 	int page_vec_len = 0;
295 
296 	page_vec->length = 0;
297 	page_vec->offset = 0;
298 
299 	iser_dbg("Translating sg sz: %d\n", data->dma_nents);
300 	page_vec_len = iser_sg_to_page_vec(data, page_vec, ibdev);
301 	iser_dbg("sg len %d page_vec_len %d\n", data->dma_nents,page_vec_len);
302 
303 	page_vec->length = page_vec_len;
304 
305 	if (page_vec_len * SIZE_4K < page_vec->data_size) {
306 		iser_err("page_vec too short to hold this SG\n");
307 		iser_data_buf_dump(data, ibdev);
308 		iser_dump_page_vec(page_vec);
309 		BUG();
310 	}
311 }
312 
313 int iser_dma_map_task_data(struct iscsi_iser_task *iser_task,
314 			    struct iser_data_buf *data,
315 			    enum iser_data_dir iser_dir,
316 			    enum dma_data_direction dma_dir)
317 {
318 	struct ib_device *dev;
319 
320 	iser_task->dir[iser_dir] = 1;
321 	dev = iser_task->iser_conn->ib_conn->device->ib_device;
322 
323 	data->dma_nents = ib_dma_map_sg(dev, data->buf, data->size, dma_dir);
324 	if (data->dma_nents == 0) {
325 		iser_err("dma_map_sg failed!!!\n");
326 		return -EINVAL;
327 	}
328 	return 0;
329 }
330 
331 void iser_dma_unmap_task_data(struct iscsi_iser_task *iser_task)
332 {
333 	struct ib_device *dev;
334 	struct iser_data_buf *data;
335 
336 	dev = iser_task->iser_conn->ib_conn->device->ib_device;
337 
338 	if (iser_task->dir[ISER_DIR_IN]) {
339 		data = &iser_task->data[ISER_DIR_IN];
340 		ib_dma_unmap_sg(dev, data->buf, data->size, DMA_FROM_DEVICE);
341 	}
342 
343 	if (iser_task->dir[ISER_DIR_OUT]) {
344 		data = &iser_task->data[ISER_DIR_OUT];
345 		ib_dma_unmap_sg(dev, data->buf, data->size, DMA_TO_DEVICE);
346 	}
347 }
348 
349 /**
350  * iser_reg_rdma_mem - Registers memory intended for RDMA,
351  * obtaining rkey and va
352  *
353  * returns 0 on success, errno code on failure
354  */
355 int iser_reg_rdma_mem(struct iscsi_iser_task *iser_task,
356 		      enum   iser_data_dir        cmd_dir)
357 {
358 	struct iscsi_conn    *iscsi_conn = iser_task->iser_conn->iscsi_conn;
359 	struct iser_conn     *ib_conn = iser_task->iser_conn->ib_conn;
360 	struct iser_device   *device = ib_conn->device;
361 	struct ib_device     *ibdev = device->ib_device;
362 	struct iser_data_buf *mem = &iser_task->data[cmd_dir];
363 	struct iser_regd_buf *regd_buf;
364 	int aligned_len;
365 	int err;
366 	int i;
367 	struct scatterlist *sg;
368 
369 	regd_buf = &iser_task->rdma_regd[cmd_dir];
370 
371 	aligned_len = iser_data_buf_aligned_len(mem, ibdev);
372 	if (aligned_len != mem->dma_nents) {
373 		iscsi_conn->fmr_unalign_cnt++;
374 		iser_warn("rdma alignment violation %d/%d aligned\n",
375 			 aligned_len, mem->size);
376 		iser_data_buf_dump(mem, ibdev);
377 
378 		/* unmap the command data before accessing it */
379 		iser_dma_unmap_task_data(iser_task);
380 
381 		/* allocate copy buf, if we are writing, copy the */
382 		/* unaligned scatterlist, dma map the copy        */
383 		if (iser_start_rdma_unaligned_sg(iser_task, cmd_dir) != 0)
384 				return -ENOMEM;
385 		mem = &iser_task->data_copy[cmd_dir];
386 	}
387 
388 	/* if there a single dma entry, FMR is not needed */
389 	if (mem->dma_nents == 1) {
390 		sg = (struct scatterlist *)mem->buf;
391 
392 		regd_buf->reg.lkey = device->mr->lkey;
393 		regd_buf->reg.rkey = device->mr->rkey;
394 		regd_buf->reg.len  = ib_sg_dma_len(ibdev, &sg[0]);
395 		regd_buf->reg.va   = ib_sg_dma_address(ibdev, &sg[0]);
396 		regd_buf->reg.is_fmr = 0;
397 
398 		iser_dbg("PHYSICAL Mem.register: lkey: 0x%08X rkey: 0x%08X  "
399 			 "va: 0x%08lX sz: %ld]\n",
400 			 (unsigned int)regd_buf->reg.lkey,
401 			 (unsigned int)regd_buf->reg.rkey,
402 			 (unsigned long)regd_buf->reg.va,
403 			 (unsigned long)regd_buf->reg.len);
404 	} else { /* use FMR for multiple dma entries */
405 		iser_page_vec_build(mem, ib_conn->page_vec, ibdev);
406 		err = iser_reg_page_vec(ib_conn, ib_conn->page_vec, &regd_buf->reg);
407 		if (err) {
408 			iser_data_buf_dump(mem, ibdev);
409 			iser_err("mem->dma_nents = %d (dlength = 0x%x)\n",
410 				 mem->dma_nents,
411 				 ntoh24(iser_task->desc.iscsi_header.dlength));
412 			iser_err("page_vec: data_size = 0x%x, length = %d, offset = 0x%x\n",
413 				 ib_conn->page_vec->data_size, ib_conn->page_vec->length,
414 				 ib_conn->page_vec->offset);
415 			for (i=0 ; i<ib_conn->page_vec->length ; i++)
416 				iser_err("page_vec[%d] = 0x%llx\n", i,
417 					 (unsigned long long) ib_conn->page_vec->pages[i]);
418 			return err;
419 		}
420 	}
421 	return 0;
422 }
423