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  * Decrements the reference count for the
45  * registered buffer & releases it
46  *
47  * returns 0 if released, 1 if deferred
48  */
49 int iser_regd_buff_release(struct iser_regd_buf *regd_buf)
50 {
51 	struct ib_device *dev;
52 
53 	if ((atomic_read(&regd_buf->ref_count) == 0) ||
54 	    atomic_dec_and_test(&regd_buf->ref_count)) {
55 		/* if we used the dma mr, unreg is just NOP */
56 		if (regd_buf->reg.is_fmr)
57 			iser_unreg_mem(&regd_buf->reg);
58 
59 		if (regd_buf->dma_addr) {
60 			dev = regd_buf->device->ib_device;
61 			ib_dma_unmap_single(dev,
62 					 regd_buf->dma_addr,
63 					 regd_buf->data_size,
64 					 regd_buf->direction);
65 		}
66 		/* else this regd buf is associated with task which we */
67 		/* dma_unmap_single/sg later */
68 		return 0;
69 	} else {
70 		iser_dbg("Release deferred, regd.buff: 0x%p\n", regd_buf);
71 		return 1;
72 	}
73 }
74 
75 /**
76  * iser_reg_single - fills registered buffer descriptor with
77  *		     registration information
78  */
79 void iser_reg_single(struct iser_device *device,
80 		     struct iser_regd_buf *regd_buf,
81 		     enum dma_data_direction direction)
82 {
83 	u64 dma_addr;
84 
85 	dma_addr = ib_dma_map_single(device->ib_device,
86 				     regd_buf->virt_addr,
87 				     regd_buf->data_size, direction);
88 	BUG_ON(ib_dma_mapping_error(device->ib_device, dma_addr));
89 
90 	regd_buf->reg.lkey = device->mr->lkey;
91 	regd_buf->reg.len  = regd_buf->data_size;
92 	regd_buf->reg.va   = dma_addr;
93 	regd_buf->reg.is_fmr = 0;
94 
95 	regd_buf->dma_addr  = dma_addr;
96 	regd_buf->direction = direction;
97 }
98 
99 /**
100  * iser_start_rdma_unaligned_sg
101  */
102 static int iser_start_rdma_unaligned_sg(struct iscsi_iser_task *iser_task,
103 					enum iser_data_dir cmd_dir)
104 {
105 	int dma_nents;
106 	struct ib_device *dev;
107 	char *mem = NULL;
108 	struct iser_data_buf *data = &iser_task->data[cmd_dir];
109 	unsigned long  cmd_data_len = data->data_len;
110 
111 	if (cmd_data_len > ISER_KMALLOC_THRESHOLD)
112 		mem = (void *)__get_free_pages(GFP_NOIO,
113 		      ilog2(roundup_pow_of_two(cmd_data_len)) - PAGE_SHIFT);
114 	else
115 		mem = kmalloc(cmd_data_len, GFP_NOIO);
116 
117 	if (mem == NULL) {
118 		iser_err("Failed to allocate mem size %d %d for copying sglist\n",
119 			 data->size,(int)cmd_data_len);
120 		return -ENOMEM;
121 	}
122 
123 	if (cmd_dir == ISER_DIR_OUT) {
124 		/* copy the unaligned sg the buffer which is used for RDMA */
125 		struct scatterlist *sgl = (struct scatterlist *)data->buf;
126 		struct scatterlist *sg;
127 		int i;
128 		char *p, *from;
129 
130 		p = mem;
131 		for_each_sg(sgl, sg, data->size, i) {
132 			from = kmap_atomic(sg_page(sg), KM_USER0);
133 			memcpy(p,
134 			       from + sg->offset,
135 			       sg->length);
136 			kunmap_atomic(from, KM_USER0);
137 			p += sg->length;
138 		}
139 	}
140 
141 	sg_init_one(&iser_task->data_copy[cmd_dir].sg_single, mem, cmd_data_len);
142 	iser_task->data_copy[cmd_dir].buf  =
143 		&iser_task->data_copy[cmd_dir].sg_single;
144 	iser_task->data_copy[cmd_dir].size = 1;
145 
146 	iser_task->data_copy[cmd_dir].copy_buf  = mem;
147 
148 	dev = iser_task->iser_conn->ib_conn->device->ib_device;
149 	dma_nents = ib_dma_map_sg(dev,
150 				  &iser_task->data_copy[cmd_dir].sg_single,
151 				  1,
152 				  (cmd_dir == ISER_DIR_OUT) ?
153 				  DMA_TO_DEVICE : DMA_FROM_DEVICE);
154 	BUG_ON(dma_nents == 0);
155 
156 	iser_task->data_copy[cmd_dir].dma_nents = dma_nents;
157 	return 0;
158 }
159 
160 /**
161  * iser_finalize_rdma_unaligned_sg
162  */
163 void iser_finalize_rdma_unaligned_sg(struct iscsi_iser_task *iser_task,
164 				     enum iser_data_dir         cmd_dir)
165 {
166 	struct ib_device *dev;
167 	struct iser_data_buf *mem_copy;
168 	unsigned long  cmd_data_len;
169 
170 	dev = iser_task->iser_conn->ib_conn->device->ib_device;
171 	mem_copy = &iser_task->data_copy[cmd_dir];
172 
173 	ib_dma_unmap_sg(dev, &mem_copy->sg_single, 1,
174 			(cmd_dir == ISER_DIR_OUT) ?
175 			DMA_TO_DEVICE : DMA_FROM_DEVICE);
176 
177 	if (cmd_dir == ISER_DIR_IN) {
178 		char *mem;
179 		struct scatterlist *sgl, *sg;
180 		unsigned char *p, *to;
181 		unsigned int sg_size;
182 		int i;
183 
184 		/* copy back read RDMA to unaligned sg */
185 		mem	= mem_copy->copy_buf;
186 
187 		sgl	= (struct scatterlist *)iser_task->data[ISER_DIR_IN].buf;
188 		sg_size = iser_task->data[ISER_DIR_IN].size;
189 
190 		p = mem;
191 		for_each_sg(sgl, sg, sg_size, i) {
192 			to = kmap_atomic(sg_page(sg), KM_SOFTIRQ0);
193 			memcpy(to + sg->offset,
194 			       p,
195 			       sg->length);
196 			kunmap_atomic(to, KM_SOFTIRQ0);
197 			p += sg->length;
198 		}
199 	}
200 
201 	cmd_data_len = iser_task->data[cmd_dir].data_len;
202 
203 	if (cmd_data_len > ISER_KMALLOC_THRESHOLD)
204 		free_pages((unsigned long)mem_copy->copy_buf,
205 			   ilog2(roundup_pow_of_two(cmd_data_len)) - PAGE_SHIFT);
206 	else
207 		kfree(mem_copy->copy_buf);
208 
209 	mem_copy->copy_buf = NULL;
210 }
211 
212 #define IS_4K_ALIGNED(addr)	((((unsigned long)addr) & ~MASK_4K) == 0)
213 
214 /**
215  * iser_sg_to_page_vec - Translates scatterlist entries to physical addresses
216  * and returns the length of resulting physical address array (may be less than
217  * the original due to possible compaction).
218  *
219  * we build a "page vec" under the assumption that the SG meets the RDMA
220  * alignment requirements. Other then the first and last SG elements, all
221  * the "internal" elements can be compacted into a list whose elements are
222  * dma addresses of physical pages. The code supports also the weird case
223  * where --few fragments of the same page-- are present in the SG as
224  * consecutive elements. Also, it handles one entry SG.
225  */
226 
227 static int iser_sg_to_page_vec(struct iser_data_buf *data,
228 			       struct iser_page_vec *page_vec,
229 			       struct ib_device *ibdev)
230 {
231 	struct scatterlist *sg, *sgl = (struct scatterlist *)data->buf;
232 	u64 start_addr, end_addr, page, chunk_start = 0;
233 	unsigned long total_sz = 0;
234 	unsigned int dma_len;
235 	int i, new_chunk, cur_page, last_ent = data->dma_nents - 1;
236 
237 	/* compute the offset of first element */
238 	page_vec->offset = (u64) sgl[0].offset & ~MASK_4K;
239 
240 	new_chunk = 1;
241 	cur_page  = 0;
242 	for_each_sg(sgl, sg, data->dma_nents, i) {
243 		start_addr = ib_sg_dma_address(ibdev, sg);
244 		if (new_chunk)
245 			chunk_start = start_addr;
246 		dma_len = ib_sg_dma_len(ibdev, sg);
247 		end_addr = start_addr + dma_len;
248 		total_sz += dma_len;
249 
250 		/* collect page fragments until aligned or end of SG list */
251 		if (!IS_4K_ALIGNED(end_addr) && i < last_ent) {
252 			new_chunk = 0;
253 			continue;
254 		}
255 		new_chunk = 1;
256 
257 		/* address of the first page in the contiguous chunk;
258 		   masking relevant for the very first SG entry,
259 		   which might be unaligned */
260 		page = chunk_start & MASK_4K;
261 		do {
262 			page_vec->pages[cur_page++] = page;
263 			page += SIZE_4K;
264 		} while (page < end_addr);
265 	}
266 
267 	page_vec->data_size = total_sz;
268 	iser_dbg("page_vec->data_size:%d cur_page %d\n", page_vec->data_size,cur_page);
269 	return cur_page;
270 }
271 
272 
273 /**
274  * iser_data_buf_aligned_len - Tries to determine the maximal correctly aligned
275  * for RDMA sub-list of a scatter-gather list of memory buffers, and  returns
276  * the number of entries which are aligned correctly. Supports the case where
277  * consecutive SG elements are actually fragments of the same physcial page.
278  */
279 static int iser_data_buf_aligned_len(struct iser_data_buf *data,
280 				      struct ib_device *ibdev)
281 {
282 	struct scatterlist *sgl, *sg, *next_sg = NULL;
283 	u64 start_addr, end_addr;
284 	int i, ret_len, start_check = 0;
285 
286 	if (data->dma_nents == 1)
287 		return 1;
288 
289 	sgl = (struct scatterlist *)data->buf;
290 	start_addr  = ib_sg_dma_address(ibdev, sgl);
291 
292 	for_each_sg(sgl, sg, data->dma_nents, i) {
293 		if (start_check && !IS_4K_ALIGNED(start_addr))
294 			break;
295 
296 		next_sg = sg_next(sg);
297 		if (!next_sg)
298 			break;
299 
300 		end_addr    = start_addr + ib_sg_dma_len(ibdev, sg);
301 		start_addr  = ib_sg_dma_address(ibdev, next_sg);
302 
303 		if (end_addr == start_addr) {
304 			start_check = 0;
305 			continue;
306 		} else
307 			start_check = 1;
308 
309 		if (!IS_4K_ALIGNED(end_addr))
310 			break;
311 	}
312 	ret_len = (next_sg) ? i : i+1;
313 	iser_dbg("Found %d aligned entries out of %d in sg:0x%p\n",
314 		 ret_len, data->dma_nents, data);
315 	return ret_len;
316 }
317 
318 static void iser_data_buf_dump(struct iser_data_buf *data,
319 			       struct ib_device *ibdev)
320 {
321 	struct scatterlist *sgl = (struct scatterlist *)data->buf;
322 	struct scatterlist *sg;
323 	int i;
324 
325 	if (iser_debug_level == 0)
326 		return;
327 
328 	for_each_sg(sgl, sg, data->dma_nents, i)
329 		iser_warn("sg[%d] dma_addr:0x%lX page:0x%p "
330 			 "off:0x%x sz:0x%x dma_len:0x%x\n",
331 			 i, (unsigned long)ib_sg_dma_address(ibdev, sg),
332 			 sg_page(sg), sg->offset,
333 			 sg->length, ib_sg_dma_len(ibdev, sg));
334 }
335 
336 static void iser_dump_page_vec(struct iser_page_vec *page_vec)
337 {
338 	int i;
339 
340 	iser_err("page vec length %d data size %d\n",
341 		 page_vec->length, page_vec->data_size);
342 	for (i = 0; i < page_vec->length; i++)
343 		iser_err("%d %lx\n",i,(unsigned long)page_vec->pages[i]);
344 }
345 
346 static void iser_page_vec_build(struct iser_data_buf *data,
347 				struct iser_page_vec *page_vec,
348 				struct ib_device *ibdev)
349 {
350 	int page_vec_len = 0;
351 
352 	page_vec->length = 0;
353 	page_vec->offset = 0;
354 
355 	iser_dbg("Translating sg sz: %d\n", data->dma_nents);
356 	page_vec_len = iser_sg_to_page_vec(data, page_vec, ibdev);
357 	iser_dbg("sg len %d page_vec_len %d\n", data->dma_nents,page_vec_len);
358 
359 	page_vec->length = page_vec_len;
360 
361 	if (page_vec_len * SIZE_4K < page_vec->data_size) {
362 		iser_err("page_vec too short to hold this SG\n");
363 		iser_data_buf_dump(data, ibdev);
364 		iser_dump_page_vec(page_vec);
365 		BUG();
366 	}
367 }
368 
369 int iser_dma_map_task_data(struct iscsi_iser_task *iser_task,
370 			    struct iser_data_buf *data,
371 			    enum iser_data_dir iser_dir,
372 			    enum dma_data_direction dma_dir)
373 {
374 	struct ib_device *dev;
375 
376 	iser_task->dir[iser_dir] = 1;
377 	dev = iser_task->iser_conn->ib_conn->device->ib_device;
378 
379 	data->dma_nents = ib_dma_map_sg(dev, data->buf, data->size, dma_dir);
380 	if (data->dma_nents == 0) {
381 		iser_err("dma_map_sg failed!!!\n");
382 		return -EINVAL;
383 	}
384 	return 0;
385 }
386 
387 void iser_dma_unmap_task_data(struct iscsi_iser_task *iser_task)
388 {
389 	struct ib_device *dev;
390 	struct iser_data_buf *data;
391 
392 	dev = iser_task->iser_conn->ib_conn->device->ib_device;
393 
394 	if (iser_task->dir[ISER_DIR_IN]) {
395 		data = &iser_task->data[ISER_DIR_IN];
396 		ib_dma_unmap_sg(dev, data->buf, data->size, DMA_FROM_DEVICE);
397 	}
398 
399 	if (iser_task->dir[ISER_DIR_OUT]) {
400 		data = &iser_task->data[ISER_DIR_OUT];
401 		ib_dma_unmap_sg(dev, data->buf, data->size, DMA_TO_DEVICE);
402 	}
403 }
404 
405 /**
406  * iser_reg_rdma_mem - Registers memory intended for RDMA,
407  * obtaining rkey and va
408  *
409  * returns 0 on success, errno code on failure
410  */
411 int iser_reg_rdma_mem(struct iscsi_iser_task *iser_task,
412 		      enum   iser_data_dir        cmd_dir)
413 {
414 	struct iscsi_conn    *iscsi_conn = iser_task->iser_conn->iscsi_conn;
415 	struct iser_conn     *ib_conn = iser_task->iser_conn->ib_conn;
416 	struct iser_device   *device = ib_conn->device;
417 	struct ib_device     *ibdev = device->ib_device;
418 	struct iser_data_buf *mem = &iser_task->data[cmd_dir];
419 	struct iser_regd_buf *regd_buf;
420 	int aligned_len;
421 	int err;
422 	int i;
423 	struct scatterlist *sg;
424 
425 	regd_buf = &iser_task->rdma_regd[cmd_dir];
426 
427 	aligned_len = iser_data_buf_aligned_len(mem, ibdev);
428 	if (aligned_len != mem->dma_nents) {
429 		iscsi_conn->fmr_unalign_cnt++;
430 		iser_warn("rdma alignment violation %d/%d aligned\n",
431 			 aligned_len, mem->size);
432 		iser_data_buf_dump(mem, ibdev);
433 
434 		/* unmap the command data before accessing it */
435 		iser_dma_unmap_task_data(iser_task);
436 
437 		/* allocate copy buf, if we are writing, copy the */
438 		/* unaligned scatterlist, dma map the copy        */
439 		if (iser_start_rdma_unaligned_sg(iser_task, cmd_dir) != 0)
440 				return -ENOMEM;
441 		mem = &iser_task->data_copy[cmd_dir];
442 	}
443 
444 	/* if there a single dma entry, FMR is not needed */
445 	if (mem->dma_nents == 1) {
446 		sg = (struct scatterlist *)mem->buf;
447 
448 		regd_buf->reg.lkey = device->mr->lkey;
449 		regd_buf->reg.rkey = device->mr->rkey;
450 		regd_buf->reg.len  = ib_sg_dma_len(ibdev, &sg[0]);
451 		regd_buf->reg.va   = ib_sg_dma_address(ibdev, &sg[0]);
452 		regd_buf->reg.is_fmr = 0;
453 
454 		iser_dbg("PHYSICAL Mem.register: lkey: 0x%08X rkey: 0x%08X  "
455 			 "va: 0x%08lX sz: %ld]\n",
456 			 (unsigned int)regd_buf->reg.lkey,
457 			 (unsigned int)regd_buf->reg.rkey,
458 			 (unsigned long)regd_buf->reg.va,
459 			 (unsigned long)regd_buf->reg.len);
460 	} else { /* use FMR for multiple dma entries */
461 		iser_page_vec_build(mem, ib_conn->page_vec, ibdev);
462 		err = iser_reg_page_vec(ib_conn, ib_conn->page_vec, &regd_buf->reg);
463 		if (err) {
464 			iser_data_buf_dump(mem, ibdev);
465 			iser_err("mem->dma_nents = %d (dlength = 0x%x)\n",
466 				 mem->dma_nents,
467 				 ntoh24(iser_task->desc.iscsi_header.dlength));
468 			iser_err("page_vec: data_size = 0x%x, length = %d, offset = 0x%x\n",
469 				 ib_conn->page_vec->data_size, ib_conn->page_vec->length,
470 				 ib_conn->page_vec->offset);
471 			for (i=0 ; i<ib_conn->page_vec->length ; i++)
472 				iser_err("page_vec[%d] = 0x%llx\n", i,
473 					 (unsigned long long) ib_conn->page_vec->pages[i]);
474 			return err;
475 		}
476 	}
477 
478 	/* take a reference on this regd buf such that it will not be released *
479 	 * (eg in send dto completion) before we get the scsi response         */
480 	atomic_inc(&regd_buf->ref_count);
481 	return 0;
482 }
483