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 /**
213  * iser_sg_to_page_vec - Translates scatterlist entries to physical addresses
214  * and returns the length of resulting physical address array (may be less than
215  * the original due to possible compaction).
216  *
217  * we build a "page vec" under the assumption that the SG meets the RDMA
218  * alignment requirements. Other then the first and last SG elements, all
219  * the "internal" elements can be compacted into a list whose elements are
220  * dma addresses of physical pages. The code supports also the weird case
221  * where --few fragments of the same page-- are present in the SG as
222  * consecutive elements. Also, it handles one entry SG.
223  */
224 static int iser_sg_to_page_vec(struct iser_data_buf *data,
225 			       struct iser_page_vec *page_vec,
226 			       struct ib_device *ibdev)
227 {
228 	struct scatterlist *sgl = (struct scatterlist *)data->buf;
229 	struct scatterlist *sg;
230 	u64 first_addr, last_addr, page;
231 	int end_aligned;
232 	unsigned int cur_page = 0;
233 	unsigned long total_sz = 0;
234 	int i;
235 
236 	/* compute the offset of first element */
237 	page_vec->offset = (u64) sgl[0].offset & ~MASK_4K;
238 
239 	for_each_sg(sgl, sg, data->dma_nents, i) {
240 		unsigned int dma_len = ib_sg_dma_len(ibdev, sg);
241 
242 		total_sz += dma_len;
243 
244 		first_addr = ib_sg_dma_address(ibdev, sg);
245 		last_addr  = first_addr + dma_len;
246 
247 		end_aligned   = !(last_addr  & ~MASK_4K);
248 
249 		/* continue to collect page fragments till aligned or SG ends */
250 		while (!end_aligned && (i + 1 < data->dma_nents)) {
251 			sg = sg_next(sg);
252 			i++;
253 			dma_len = ib_sg_dma_len(ibdev, sg);
254 			total_sz += dma_len;
255 			last_addr = ib_sg_dma_address(ibdev, sg) + dma_len;
256 			end_aligned = !(last_addr  & ~MASK_4K);
257 		}
258 
259 		/* handle the 1st page in the 1st DMA element */
260 		if (cur_page == 0) {
261 			page = first_addr & MASK_4K;
262 			page_vec->pages[cur_page] = page;
263 			cur_page++;
264 			page += SIZE_4K;
265 		} else
266 			page = first_addr;
267 
268 		for (; page < last_addr; page += SIZE_4K) {
269 			page_vec->pages[cur_page] = page;
270 			cur_page++;
271 		}
272 
273 	}
274 	page_vec->data_size = total_sz;
275 	iser_dbg("page_vec->data_size:%d cur_page %d\n", page_vec->data_size,cur_page);
276 	return cur_page;
277 }
278 
279 #define IS_4K_ALIGNED(addr)	((((unsigned long)addr) & ~MASK_4K) == 0)
280 
281 /**
282  * iser_data_buf_aligned_len - Tries to determine the maximal correctly aligned
283  * for RDMA sub-list of a scatter-gather list of memory buffers, and  returns
284  * the number of entries which are aligned correctly. Supports the case where
285  * consecutive SG elements are actually fragments of the same physcial page.
286  */
287 static unsigned int iser_data_buf_aligned_len(struct iser_data_buf *data,
288 					      struct ib_device *ibdev)
289 {
290 	struct scatterlist *sgl, *sg;
291 	u64 end_addr, next_addr;
292 	int i, cnt;
293 	unsigned int ret_len = 0;
294 
295 	sgl = (struct scatterlist *)data->buf;
296 
297 	cnt = 0;
298 	for_each_sg(sgl, sg, data->dma_nents, i) {
299 		/* iser_dbg("Checking sg iobuf [%d]: phys=0x%08lX "
300 		   "offset: %ld sz: %ld\n", i,
301 		   (unsigned long)sg_phys(sg),
302 		   (unsigned long)sg->offset,
303 		   (unsigned long)sg->length); */
304 		end_addr = ib_sg_dma_address(ibdev, sg) +
305 			   ib_sg_dma_len(ibdev, sg);
306 		/* iser_dbg("Checking sg iobuf end address "
307 		       "0x%08lX\n", end_addr); */
308 		if (i + 1 < data->dma_nents) {
309 			next_addr = ib_sg_dma_address(ibdev, sg_next(sg));
310 			/* are i, i+1 fragments of the same page? */
311 			if (end_addr == next_addr) {
312 				cnt++;
313 				continue;
314 			} else if (!IS_4K_ALIGNED(end_addr)) {
315 				ret_len = cnt + 1;
316 				break;
317 			}
318 		}
319 		cnt++;
320 	}
321 	if (i == data->dma_nents)
322 		ret_len = cnt;	/* loop ended */
323 	iser_dbg("Found %d aligned entries out of %d in sg:0x%p\n",
324 		 ret_len, data->dma_nents, data);
325 	return ret_len;
326 }
327 
328 static void iser_data_buf_dump(struct iser_data_buf *data,
329 			       struct ib_device *ibdev)
330 {
331 	struct scatterlist *sgl = (struct scatterlist *)data->buf;
332 	struct scatterlist *sg;
333 	int i;
334 
335 	if (iser_debug_level == 0)
336 		return;
337 
338 	for_each_sg(sgl, sg, data->dma_nents, i)
339 		iser_warn("sg[%d] dma_addr:0x%lX page:0x%p "
340 			 "off:0x%x sz:0x%x dma_len:0x%x\n",
341 			 i, (unsigned long)ib_sg_dma_address(ibdev, sg),
342 			 sg_page(sg), sg->offset,
343 			 sg->length, ib_sg_dma_len(ibdev, sg));
344 }
345 
346 static void iser_dump_page_vec(struct iser_page_vec *page_vec)
347 {
348 	int i;
349 
350 	iser_err("page vec length %d data size %d\n",
351 		 page_vec->length, page_vec->data_size);
352 	for (i = 0; i < page_vec->length; i++)
353 		iser_err("%d %lx\n",i,(unsigned long)page_vec->pages[i]);
354 }
355 
356 static void iser_page_vec_build(struct iser_data_buf *data,
357 				struct iser_page_vec *page_vec,
358 				struct ib_device *ibdev)
359 {
360 	int page_vec_len = 0;
361 
362 	page_vec->length = 0;
363 	page_vec->offset = 0;
364 
365 	iser_dbg("Translating sg sz: %d\n", data->dma_nents);
366 	page_vec_len = iser_sg_to_page_vec(data, page_vec, ibdev);
367 	iser_dbg("sg len %d page_vec_len %d\n", data->dma_nents,page_vec_len);
368 
369 	page_vec->length = page_vec_len;
370 
371 	if (page_vec_len * SIZE_4K < page_vec->data_size) {
372 		iser_err("page_vec too short to hold this SG\n");
373 		iser_data_buf_dump(data, ibdev);
374 		iser_dump_page_vec(page_vec);
375 		BUG();
376 	}
377 }
378 
379 int iser_dma_map_task_data(struct iscsi_iser_task *iser_task,
380 			    struct iser_data_buf *data,
381 			    enum iser_data_dir iser_dir,
382 			    enum dma_data_direction dma_dir)
383 {
384 	struct ib_device *dev;
385 
386 	iser_task->dir[iser_dir] = 1;
387 	dev = iser_task->iser_conn->ib_conn->device->ib_device;
388 
389 	data->dma_nents = ib_dma_map_sg(dev, data->buf, data->size, dma_dir);
390 	if (data->dma_nents == 0) {
391 		iser_err("dma_map_sg failed!!!\n");
392 		return -EINVAL;
393 	}
394 	return 0;
395 }
396 
397 void iser_dma_unmap_task_data(struct iscsi_iser_task *iser_task)
398 {
399 	struct ib_device *dev;
400 	struct iser_data_buf *data;
401 
402 	dev = iser_task->iser_conn->ib_conn->device->ib_device;
403 
404 	if (iser_task->dir[ISER_DIR_IN]) {
405 		data = &iser_task->data[ISER_DIR_IN];
406 		ib_dma_unmap_sg(dev, data->buf, data->size, DMA_FROM_DEVICE);
407 	}
408 
409 	if (iser_task->dir[ISER_DIR_OUT]) {
410 		data = &iser_task->data[ISER_DIR_OUT];
411 		ib_dma_unmap_sg(dev, data->buf, data->size, DMA_TO_DEVICE);
412 	}
413 }
414 
415 /**
416  * iser_reg_rdma_mem - Registers memory intended for RDMA,
417  * obtaining rkey and va
418  *
419  * returns 0 on success, errno code on failure
420  */
421 int iser_reg_rdma_mem(struct iscsi_iser_task *iser_task,
422 		      enum   iser_data_dir        cmd_dir)
423 {
424 	struct iscsi_conn    *iscsi_conn = iser_task->iser_conn->iscsi_conn;
425 	struct iser_conn     *ib_conn = iser_task->iser_conn->ib_conn;
426 	struct iser_device   *device = ib_conn->device;
427 	struct ib_device     *ibdev = device->ib_device;
428 	struct iser_data_buf *mem = &iser_task->data[cmd_dir];
429 	struct iser_regd_buf *regd_buf;
430 	int aligned_len;
431 	int err;
432 	int i;
433 	struct scatterlist *sg;
434 
435 	regd_buf = &iser_task->rdma_regd[cmd_dir];
436 
437 	aligned_len = iser_data_buf_aligned_len(mem, ibdev);
438 	if (aligned_len != mem->dma_nents) {
439 		iscsi_conn->fmr_unalign_cnt++;
440 		iser_warn("rdma alignment violation %d/%d aligned\n",
441 			 aligned_len, mem->size);
442 		iser_data_buf_dump(mem, ibdev);
443 
444 		/* unmap the command data before accessing it */
445 		iser_dma_unmap_task_data(iser_task);
446 
447 		/* allocate copy buf, if we are writing, copy the */
448 		/* unaligned scatterlist, dma map the copy        */
449 		if (iser_start_rdma_unaligned_sg(iser_task, cmd_dir) != 0)
450 				return -ENOMEM;
451 		mem = &iser_task->data_copy[cmd_dir];
452 	}
453 
454 	/* if there a single dma entry, FMR is not needed */
455 	if (mem->dma_nents == 1) {
456 		sg = (struct scatterlist *)mem->buf;
457 
458 		regd_buf->reg.lkey = device->mr->lkey;
459 		regd_buf->reg.rkey = device->mr->rkey;
460 		regd_buf->reg.len  = ib_sg_dma_len(ibdev, &sg[0]);
461 		regd_buf->reg.va   = ib_sg_dma_address(ibdev, &sg[0]);
462 		regd_buf->reg.is_fmr = 0;
463 
464 		iser_dbg("PHYSICAL Mem.register: lkey: 0x%08X rkey: 0x%08X  "
465 			 "va: 0x%08lX sz: %ld]\n",
466 			 (unsigned int)regd_buf->reg.lkey,
467 			 (unsigned int)regd_buf->reg.rkey,
468 			 (unsigned long)regd_buf->reg.va,
469 			 (unsigned long)regd_buf->reg.len);
470 	} else { /* use FMR for multiple dma entries */
471 		iser_page_vec_build(mem, ib_conn->page_vec, ibdev);
472 		err = iser_reg_page_vec(ib_conn, ib_conn->page_vec, &regd_buf->reg);
473 		if (err) {
474 			iser_data_buf_dump(mem, ibdev);
475 			iser_err("mem->dma_nents = %d (dlength = 0x%x)\n",
476 				 mem->dma_nents,
477 				 ntoh24(iser_task->desc.iscsi_header.dlength));
478 			iser_err("page_vec: data_size = 0x%x, length = %d, offset = 0x%x\n",
479 				 ib_conn->page_vec->data_size, ib_conn->page_vec->length,
480 				 ib_conn->page_vec->offset);
481 			for (i=0 ; i<ib_conn->page_vec->length ; i++)
482 				iser_err("page_vec[%d] = 0x%llx\n", i,
483 					 (unsigned long long) ib_conn->page_vec->pages[i]);
484 			return err;
485 		}
486 	}
487 
488 	/* take a reference on this regd buf such that it will not be released *
489 	 * (eg in send dto completion) before we get the scsi response         */
490 	atomic_inc(&regd_buf->ref_count);
491 	return 0;
492 }
493