1 /*
2  * Copyright (c) 2004, 2005, 2006 Voltaire, Inc. All rights reserved.
3  * Copyright (c) 2013-2014 Mellanox Technologies. All rights reserved.
4  *
5  * This software is available to you under a choice of one of two
6  * licenses.  You may choose to be licensed under the terms of the GNU
7  * General Public License (GPL) Version 2, available from the file
8  * COPYING in the main directory of this source tree, or the
9  * OpenIB.org BSD license below:
10  *
11  *     Redistribution and use in source and binary forms, with or
12  *     without modification, are permitted provided that the following
13  *     conditions are met:
14  *
15  *	- Redistributions of source code must retain the above
16  *	  copyright notice, this list of conditions and the following
17  *	  disclaimer.
18  *
19  *	- Redistributions in binary form must reproduce the above
20  *	  copyright notice, this list of conditions and the following
21  *	  disclaimer in the documentation and/or other materials
22  *	  provided with the distribution.
23  *
24  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
25  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
26  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
27  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
28  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
29  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
30  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
31  * SOFTWARE.
32  */
33 #include <linux/kernel.h>
34 #include <linux/slab.h>
35 #include <linux/mm.h>
36 #include <linux/scatterlist.h>
37 #include <linux/kfifo.h>
38 #include <scsi/scsi_cmnd.h>
39 #include <scsi/scsi_host.h>
40 
41 #include "iscsi_iser.h"
42 
43 /* Register user buffer memory and initialize passive rdma
44  *  dto descriptor. Data size is stored in
45  *  task->data[ISER_DIR_IN].data_len, Protection size
46  *  os stored in task->prot[ISER_DIR_IN].data_len
47  */
48 static int iser_prepare_read_cmd(struct iscsi_task *task)
49 
50 {
51 	struct iscsi_iser_task *iser_task = task->dd_data;
52 	struct iser_device  *device = iser_task->ib_conn->device;
53 	struct iser_regd_buf *regd_buf;
54 	int err;
55 	struct iser_hdr *hdr = &iser_task->desc.iser_header;
56 	struct iser_data_buf *buf_in = &iser_task->data[ISER_DIR_IN];
57 
58 	err = iser_dma_map_task_data(iser_task,
59 				     buf_in,
60 				     ISER_DIR_IN,
61 				     DMA_FROM_DEVICE);
62 	if (err)
63 		return err;
64 
65 	if (scsi_prot_sg_count(iser_task->sc)) {
66 		struct iser_data_buf *pbuf_in = &iser_task->prot[ISER_DIR_IN];
67 
68 		err = iser_dma_map_task_data(iser_task,
69 					     pbuf_in,
70 					     ISER_DIR_IN,
71 					     DMA_FROM_DEVICE);
72 		if (err)
73 			return err;
74 	}
75 
76 	err = device->iser_reg_rdma_mem(iser_task, ISER_DIR_IN);
77 	if (err) {
78 		iser_err("Failed to set up Data-IN RDMA\n");
79 		return err;
80 	}
81 	regd_buf = &iser_task->rdma_regd[ISER_DIR_IN];
82 
83 	hdr->flags    |= ISER_RSV;
84 	hdr->read_stag = cpu_to_be32(regd_buf->reg.rkey);
85 	hdr->read_va   = cpu_to_be64(regd_buf->reg.va);
86 
87 	iser_dbg("Cmd itt:%d READ tags RKEY:%#.4X VA:%#llX\n",
88 		 task->itt, regd_buf->reg.rkey,
89 		 (unsigned long long)regd_buf->reg.va);
90 
91 	return 0;
92 }
93 
94 /* Register user buffer memory and initialize passive rdma
95  *  dto descriptor. Data size is stored in
96  *  task->data[ISER_DIR_OUT].data_len, Protection size
97  *  is stored at task->prot[ISER_DIR_OUT].data_len
98  */
99 static int
100 iser_prepare_write_cmd(struct iscsi_task *task,
101 		       unsigned int imm_sz,
102 		       unsigned int unsol_sz,
103 		       unsigned int edtl)
104 {
105 	struct iscsi_iser_task *iser_task = task->dd_data;
106 	struct iser_device  *device = iser_task->ib_conn->device;
107 	struct iser_regd_buf *regd_buf;
108 	int err;
109 	struct iser_hdr *hdr = &iser_task->desc.iser_header;
110 	struct iser_data_buf *buf_out = &iser_task->data[ISER_DIR_OUT];
111 	struct ib_sge *tx_dsg = &iser_task->desc.tx_sg[1];
112 
113 	err = iser_dma_map_task_data(iser_task,
114 				     buf_out,
115 				     ISER_DIR_OUT,
116 				     DMA_TO_DEVICE);
117 	if (err)
118 		return err;
119 
120 	if (scsi_prot_sg_count(iser_task->sc)) {
121 		struct iser_data_buf *pbuf_out = &iser_task->prot[ISER_DIR_OUT];
122 
123 		err = iser_dma_map_task_data(iser_task,
124 					     pbuf_out,
125 					     ISER_DIR_OUT,
126 					     DMA_TO_DEVICE);
127 		if (err)
128 			return err;
129 	}
130 
131 	err = device->iser_reg_rdma_mem(iser_task, ISER_DIR_OUT);
132 	if (err != 0) {
133 		iser_err("Failed to register write cmd RDMA mem\n");
134 		return err;
135 	}
136 
137 	regd_buf = &iser_task->rdma_regd[ISER_DIR_OUT];
138 
139 	if (unsol_sz < edtl) {
140 		hdr->flags     |= ISER_WSV;
141 		hdr->write_stag = cpu_to_be32(regd_buf->reg.rkey);
142 		hdr->write_va   = cpu_to_be64(regd_buf->reg.va + unsol_sz);
143 
144 		iser_dbg("Cmd itt:%d, WRITE tags, RKEY:%#.4X "
145 			 "VA:%#llX + unsol:%d\n",
146 			 task->itt, regd_buf->reg.rkey,
147 			 (unsigned long long)regd_buf->reg.va, unsol_sz);
148 	}
149 
150 	if (imm_sz > 0) {
151 		iser_dbg("Cmd itt:%d, WRITE, adding imm.data sz: %d\n",
152 			 task->itt, imm_sz);
153 		tx_dsg->addr   = regd_buf->reg.va;
154 		tx_dsg->length = imm_sz;
155 		tx_dsg->lkey   = regd_buf->reg.lkey;
156 		iser_task->desc.num_sge = 2;
157 	}
158 
159 	return 0;
160 }
161 
162 /* creates a new tx descriptor and adds header regd buffer */
163 static void iser_create_send_desc(struct iser_conn	*ib_conn,
164 				  struct iser_tx_desc	*tx_desc)
165 {
166 	struct iser_device *device = ib_conn->device;
167 
168 	ib_dma_sync_single_for_cpu(device->ib_device,
169 		tx_desc->dma_addr, ISER_HEADERS_LEN, DMA_TO_DEVICE);
170 
171 	memset(&tx_desc->iser_header, 0, sizeof(struct iser_hdr));
172 	tx_desc->iser_header.flags = ISER_VER;
173 
174 	tx_desc->num_sge = 1;
175 
176 	if (tx_desc->tx_sg[0].lkey != device->mr->lkey) {
177 		tx_desc->tx_sg[0].lkey = device->mr->lkey;
178 		iser_dbg("sdesc %p lkey mismatch, fixing\n", tx_desc);
179 	}
180 }
181 
182 static void iser_free_login_buf(struct iser_conn *ib_conn)
183 {
184 	if (!ib_conn->login_buf)
185 		return;
186 
187 	if (ib_conn->login_req_dma)
188 		ib_dma_unmap_single(ib_conn->device->ib_device,
189 				    ib_conn->login_req_dma,
190 				    ISCSI_DEF_MAX_RECV_SEG_LEN, DMA_TO_DEVICE);
191 
192 	if (ib_conn->login_resp_dma)
193 		ib_dma_unmap_single(ib_conn->device->ib_device,
194 				    ib_conn->login_resp_dma,
195 				    ISER_RX_LOGIN_SIZE, DMA_FROM_DEVICE);
196 
197 	kfree(ib_conn->login_buf);
198 
199 	/* make sure we never redo any unmapping */
200 	ib_conn->login_req_dma = 0;
201 	ib_conn->login_resp_dma = 0;
202 	ib_conn->login_buf = NULL;
203 }
204 
205 static int iser_alloc_login_buf(struct iser_conn *ib_conn)
206 {
207 	struct iser_device	*device;
208 	int			req_err, resp_err;
209 
210 	BUG_ON(ib_conn->device == NULL);
211 
212 	device = ib_conn->device;
213 
214 	ib_conn->login_buf = kmalloc(ISCSI_DEF_MAX_RECV_SEG_LEN +
215 				     ISER_RX_LOGIN_SIZE, GFP_KERNEL);
216 	if (!ib_conn->login_buf)
217 		goto out_err;
218 
219 	ib_conn->login_req_buf  = ib_conn->login_buf;
220 	ib_conn->login_resp_buf = ib_conn->login_buf +
221 						ISCSI_DEF_MAX_RECV_SEG_LEN;
222 
223 	ib_conn->login_req_dma = ib_dma_map_single(ib_conn->device->ib_device,
224 				(void *)ib_conn->login_req_buf,
225 				ISCSI_DEF_MAX_RECV_SEG_LEN, DMA_TO_DEVICE);
226 
227 	ib_conn->login_resp_dma = ib_dma_map_single(ib_conn->device->ib_device,
228 				(void *)ib_conn->login_resp_buf,
229 				ISER_RX_LOGIN_SIZE, DMA_FROM_DEVICE);
230 
231 	req_err  = ib_dma_mapping_error(device->ib_device,
232 					ib_conn->login_req_dma);
233 	resp_err = ib_dma_mapping_error(device->ib_device,
234 					ib_conn->login_resp_dma);
235 
236 	if (req_err || resp_err) {
237 		if (req_err)
238 			ib_conn->login_req_dma = 0;
239 		if (resp_err)
240 			ib_conn->login_resp_dma = 0;
241 		goto free_login_buf;
242 	}
243 	return 0;
244 
245 free_login_buf:
246 	iser_free_login_buf(ib_conn);
247 
248 out_err:
249 	iser_err("unable to alloc or map login buf\n");
250 	return -ENOMEM;
251 }
252 
253 int iser_alloc_rx_descriptors(struct iser_conn *ib_conn, struct iscsi_session *session)
254 {
255 	int i, j;
256 	u64 dma_addr;
257 	struct iser_rx_desc *rx_desc;
258 	struct ib_sge       *rx_sg;
259 	struct iser_device  *device = ib_conn->device;
260 
261 	ib_conn->qp_max_recv_dtos = session->cmds_max;
262 	ib_conn->qp_max_recv_dtos_mask = session->cmds_max - 1; /* cmds_max is 2^N */
263 	ib_conn->min_posted_rx = ib_conn->qp_max_recv_dtos >> 2;
264 
265 	if (device->iser_alloc_rdma_reg_res(ib_conn, session->scsi_cmds_max))
266 		goto create_rdma_reg_res_failed;
267 
268 	if (iser_alloc_login_buf(ib_conn))
269 		goto alloc_login_buf_fail;
270 
271 	ib_conn->rx_descs = kmalloc(session->cmds_max *
272 				sizeof(struct iser_rx_desc), GFP_KERNEL);
273 	if (!ib_conn->rx_descs)
274 		goto rx_desc_alloc_fail;
275 
276 	rx_desc = ib_conn->rx_descs;
277 
278 	for (i = 0; i < ib_conn->qp_max_recv_dtos; i++, rx_desc++)  {
279 		dma_addr = ib_dma_map_single(device->ib_device, (void *)rx_desc,
280 					ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
281 		if (ib_dma_mapping_error(device->ib_device, dma_addr))
282 			goto rx_desc_dma_map_failed;
283 
284 		rx_desc->dma_addr = dma_addr;
285 
286 		rx_sg = &rx_desc->rx_sg;
287 		rx_sg->addr   = rx_desc->dma_addr;
288 		rx_sg->length = ISER_RX_PAYLOAD_SIZE;
289 		rx_sg->lkey   = device->mr->lkey;
290 	}
291 
292 	ib_conn->rx_desc_head = 0;
293 	return 0;
294 
295 rx_desc_dma_map_failed:
296 	rx_desc = ib_conn->rx_descs;
297 	for (j = 0; j < i; j++, rx_desc++)
298 		ib_dma_unmap_single(device->ib_device, rx_desc->dma_addr,
299 				    ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
300 	kfree(ib_conn->rx_descs);
301 	ib_conn->rx_descs = NULL;
302 rx_desc_alloc_fail:
303 	iser_free_login_buf(ib_conn);
304 alloc_login_buf_fail:
305 	device->iser_free_rdma_reg_res(ib_conn);
306 create_rdma_reg_res_failed:
307 	iser_err("failed allocating rx descriptors / data buffers\n");
308 	return -ENOMEM;
309 }
310 
311 void iser_free_rx_descriptors(struct iser_conn *ib_conn)
312 {
313 	int i;
314 	struct iser_rx_desc *rx_desc;
315 	struct iser_device *device = ib_conn->device;
316 
317 	if (!ib_conn->rx_descs)
318 		goto free_login_buf;
319 
320 	if (device->iser_free_rdma_reg_res)
321 		device->iser_free_rdma_reg_res(ib_conn);
322 
323 	rx_desc = ib_conn->rx_descs;
324 	for (i = 0; i < ib_conn->qp_max_recv_dtos; i++, rx_desc++)
325 		ib_dma_unmap_single(device->ib_device, rx_desc->dma_addr,
326 				    ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
327 	kfree(ib_conn->rx_descs);
328 	/* make sure we never redo any unmapping */
329 	ib_conn->rx_descs = NULL;
330 
331 free_login_buf:
332 	iser_free_login_buf(ib_conn);
333 }
334 
335 static int iser_post_rx_bufs(struct iscsi_conn *conn, struct iscsi_hdr *req)
336 {
337 	struct iser_conn *ib_conn = conn->dd_data;
338 	struct iscsi_session *session = conn->session;
339 
340 	iser_dbg("req op %x flags %x\n", req->opcode, req->flags);
341 	/* check if this is the last login - going to full feature phase */
342 	if ((req->flags & ISCSI_FULL_FEATURE_PHASE) != ISCSI_FULL_FEATURE_PHASE)
343 		return 0;
344 
345 	/*
346 	 * Check that there is one posted recv buffer (for the last login
347 	 * response) and no posted send buffers left - they must have been
348 	 * consumed during previous login phases.
349 	 */
350 	WARN_ON(ib_conn->post_recv_buf_count != 1);
351 	WARN_ON(atomic_read(&ib_conn->post_send_buf_count) != 0);
352 
353 	if (session->discovery_sess) {
354 		iser_info("Discovery session, re-using login RX buffer\n");
355 		return 0;
356 	} else
357 		iser_info("Normal session, posting batch of RX %d buffers\n",
358 			  ib_conn->min_posted_rx);
359 
360 	/* Initial post receive buffers */
361 	if (iser_post_recvm(ib_conn, ib_conn->min_posted_rx))
362 		return -ENOMEM;
363 
364 	return 0;
365 }
366 
367 /**
368  * iser_send_command - send command PDU
369  */
370 int iser_send_command(struct iscsi_conn *conn,
371 		      struct iscsi_task *task)
372 {
373 	struct iser_conn *ib_conn = conn->dd_data;
374 	struct iscsi_iser_task *iser_task = task->dd_data;
375 	unsigned long edtl;
376 	int err;
377 	struct iser_data_buf *data_buf, *prot_buf;
378 	struct iscsi_scsi_req *hdr = (struct iscsi_scsi_req *)task->hdr;
379 	struct scsi_cmnd *sc  =  task->sc;
380 	struct iser_tx_desc *tx_desc = &iser_task->desc;
381 
382 	edtl = ntohl(hdr->data_length);
383 
384 	/* build the tx desc regd header and add it to the tx desc dto */
385 	tx_desc->type = ISCSI_TX_SCSI_COMMAND;
386 	iser_create_send_desc(ib_conn, tx_desc);
387 
388 	if (hdr->flags & ISCSI_FLAG_CMD_READ) {
389 		data_buf = &iser_task->data[ISER_DIR_IN];
390 		prot_buf = &iser_task->prot[ISER_DIR_IN];
391 	} else {
392 		data_buf = &iser_task->data[ISER_DIR_OUT];
393 		prot_buf = &iser_task->prot[ISER_DIR_OUT];
394 	}
395 
396 	if (scsi_sg_count(sc)) { /* using a scatter list */
397 		data_buf->buf  = scsi_sglist(sc);
398 		data_buf->size = scsi_sg_count(sc);
399 	}
400 	data_buf->data_len = scsi_bufflen(sc);
401 
402 	if (scsi_prot_sg_count(sc)) {
403 		prot_buf->buf  = scsi_prot_sglist(sc);
404 		prot_buf->size = scsi_prot_sg_count(sc);
405 		prot_buf->data_len = data_buf->data_len >>
406 				     ilog2(sc->device->sector_size) * 8;
407 	}
408 
409 	if (hdr->flags & ISCSI_FLAG_CMD_READ) {
410 		err = iser_prepare_read_cmd(task);
411 		if (err)
412 			goto send_command_error;
413 	}
414 	if (hdr->flags & ISCSI_FLAG_CMD_WRITE) {
415 		err = iser_prepare_write_cmd(task,
416 					     task->imm_count,
417 				             task->imm_count +
418 					     task->unsol_r2t.data_length,
419 					     edtl);
420 		if (err)
421 			goto send_command_error;
422 	}
423 
424 	iser_task->status = ISER_TASK_STATUS_STARTED;
425 
426 	err = iser_post_send(ib_conn, tx_desc);
427 	if (!err)
428 		return 0;
429 
430 send_command_error:
431 	iser_err("conn %p failed task->itt %d err %d\n",conn, task->itt, err);
432 	return err;
433 }
434 
435 /**
436  * iser_send_data_out - send data out PDU
437  */
438 int iser_send_data_out(struct iscsi_conn *conn,
439 		       struct iscsi_task *task,
440 		       struct iscsi_data *hdr)
441 {
442 	struct iser_conn *ib_conn = conn->dd_data;
443 	struct iscsi_iser_task *iser_task = task->dd_data;
444 	struct iser_tx_desc *tx_desc = NULL;
445 	struct iser_regd_buf *regd_buf;
446 	unsigned long buf_offset;
447 	unsigned long data_seg_len;
448 	uint32_t itt;
449 	int err = 0;
450 	struct ib_sge *tx_dsg;
451 
452 	itt = (__force uint32_t)hdr->itt;
453 	data_seg_len = ntoh24(hdr->dlength);
454 	buf_offset   = ntohl(hdr->offset);
455 
456 	iser_dbg("%s itt %d dseg_len %d offset %d\n",
457 		 __func__,(int)itt,(int)data_seg_len,(int)buf_offset);
458 
459 	tx_desc = kmem_cache_zalloc(ig.desc_cache, GFP_ATOMIC);
460 	if (tx_desc == NULL) {
461 		iser_err("Failed to alloc desc for post dataout\n");
462 		return -ENOMEM;
463 	}
464 
465 	tx_desc->type = ISCSI_TX_DATAOUT;
466 	tx_desc->iser_header.flags = ISER_VER;
467 	memcpy(&tx_desc->iscsi_header, hdr, sizeof(struct iscsi_hdr));
468 
469 	/* build the tx desc */
470 	iser_initialize_task_headers(task, tx_desc);
471 
472 	regd_buf = &iser_task->rdma_regd[ISER_DIR_OUT];
473 	tx_dsg = &tx_desc->tx_sg[1];
474 	tx_dsg->addr    = regd_buf->reg.va + buf_offset;
475 	tx_dsg->length  = data_seg_len;
476 	tx_dsg->lkey    = regd_buf->reg.lkey;
477 	tx_desc->num_sge = 2;
478 
479 	if (buf_offset + data_seg_len > iser_task->data[ISER_DIR_OUT].data_len) {
480 		iser_err("Offset:%ld & DSL:%ld in Data-Out "
481 			 "inconsistent with total len:%ld, itt:%d\n",
482 			 buf_offset, data_seg_len,
483 			 iser_task->data[ISER_DIR_OUT].data_len, itt);
484 		err = -EINVAL;
485 		goto send_data_out_error;
486 	}
487 	iser_dbg("data-out itt: %d, offset: %ld, sz: %ld\n",
488 		 itt, buf_offset, data_seg_len);
489 
490 
491 	err = iser_post_send(ib_conn, tx_desc);
492 	if (!err)
493 		return 0;
494 
495 send_data_out_error:
496 	kmem_cache_free(ig.desc_cache, tx_desc);
497 	iser_err("conn %p failed err %d\n",conn, err);
498 	return err;
499 }
500 
501 int iser_send_control(struct iscsi_conn *conn,
502 		      struct iscsi_task *task)
503 {
504 	struct iser_conn *ib_conn = conn->dd_data;
505 	struct iscsi_iser_task *iser_task = task->dd_data;
506 	struct iser_tx_desc *mdesc = &iser_task->desc;
507 	unsigned long data_seg_len;
508 	int err = 0;
509 	struct iser_device *device;
510 
511 	/* build the tx desc regd header and add it to the tx desc dto */
512 	mdesc->type = ISCSI_TX_CONTROL;
513 	iser_create_send_desc(ib_conn, mdesc);
514 
515 	device = ib_conn->device;
516 
517 	data_seg_len = ntoh24(task->hdr->dlength);
518 
519 	if (data_seg_len > 0) {
520 		struct ib_sge *tx_dsg = &mdesc->tx_sg[1];
521 		if (task != conn->login_task) {
522 			iser_err("data present on non login task!!!\n");
523 			goto send_control_error;
524 		}
525 
526 		ib_dma_sync_single_for_cpu(device->ib_device,
527 			ib_conn->login_req_dma, task->data_count,
528 			DMA_TO_DEVICE);
529 
530 		memcpy(ib_conn->login_req_buf, task->data, task->data_count);
531 
532 		ib_dma_sync_single_for_device(device->ib_device,
533 			ib_conn->login_req_dma, task->data_count,
534 			DMA_TO_DEVICE);
535 
536 		tx_dsg->addr    = ib_conn->login_req_dma;
537 		tx_dsg->length  = task->data_count;
538 		tx_dsg->lkey    = device->mr->lkey;
539 		mdesc->num_sge = 2;
540 	}
541 
542 	if (task == conn->login_task) {
543 		iser_dbg("op %x dsl %lx, posting login rx buffer\n",
544 			 task->hdr->opcode, data_seg_len);
545 		err = iser_post_recvl(ib_conn);
546 		if (err)
547 			goto send_control_error;
548 		err = iser_post_rx_bufs(conn, task->hdr);
549 		if (err)
550 			goto send_control_error;
551 	}
552 
553 	err = iser_post_send(ib_conn, mdesc);
554 	if (!err)
555 		return 0;
556 
557 send_control_error:
558 	iser_err("conn %p failed err %d\n",conn, err);
559 	return err;
560 }
561 
562 /**
563  * iser_rcv_dto_completion - recv DTO completion
564  */
565 void iser_rcv_completion(struct iser_rx_desc *rx_desc,
566 			 unsigned long rx_xfer_len,
567 			 struct iser_conn *ib_conn)
568 {
569 	struct iscsi_hdr *hdr;
570 	u64 rx_dma;
571 	int rx_buflen, outstanding, count, err;
572 
573 	/* differentiate between login to all other PDUs */
574 	if ((char *)rx_desc == ib_conn->login_resp_buf) {
575 		rx_dma = ib_conn->login_resp_dma;
576 		rx_buflen = ISER_RX_LOGIN_SIZE;
577 	} else {
578 		rx_dma = rx_desc->dma_addr;
579 		rx_buflen = ISER_RX_PAYLOAD_SIZE;
580 	}
581 
582 	ib_dma_sync_single_for_cpu(ib_conn->device->ib_device, rx_dma,
583 			rx_buflen, DMA_FROM_DEVICE);
584 
585 	hdr = &rx_desc->iscsi_header;
586 
587 	iser_dbg("op 0x%x itt 0x%x dlen %d\n", hdr->opcode,
588 			hdr->itt, (int)(rx_xfer_len - ISER_HEADERS_LEN));
589 
590 	iscsi_iser_recv(ib_conn->iscsi_conn, hdr, rx_desc->data,
591 			rx_xfer_len - ISER_HEADERS_LEN);
592 
593 	ib_dma_sync_single_for_device(ib_conn->device->ib_device, rx_dma,
594 				      rx_buflen, DMA_FROM_DEVICE);
595 
596 	/* decrementing conn->post_recv_buf_count only --after-- freeing the   *
597 	 * task eliminates the need to worry on tasks which are completed in   *
598 	 * parallel to the execution of iser_conn_term. So the code that waits *
599 	 * for the posted rx bufs refcount to become zero handles everything   */
600 	ib_conn->post_recv_buf_count--;
601 
602 	if (rx_dma == ib_conn->login_resp_dma)
603 		return;
604 
605 	outstanding = ib_conn->post_recv_buf_count;
606 	if (outstanding + ib_conn->min_posted_rx <= ib_conn->qp_max_recv_dtos) {
607 		count = min(ib_conn->qp_max_recv_dtos - outstanding,
608 						ib_conn->min_posted_rx);
609 		err = iser_post_recvm(ib_conn, count);
610 		if (err)
611 			iser_err("posting %d rx bufs err %d\n", count, err);
612 	}
613 }
614 
615 void iser_snd_completion(struct iser_tx_desc *tx_desc,
616 			struct iser_conn *ib_conn)
617 {
618 	struct iscsi_task *task;
619 	struct iser_device *device = ib_conn->device;
620 
621 	if (tx_desc->type == ISCSI_TX_DATAOUT) {
622 		ib_dma_unmap_single(device->ib_device, tx_desc->dma_addr,
623 					ISER_HEADERS_LEN, DMA_TO_DEVICE);
624 		kmem_cache_free(ig.desc_cache, tx_desc);
625 		tx_desc = NULL;
626 	}
627 
628 	atomic_dec(&ib_conn->post_send_buf_count);
629 
630 	if (tx_desc && tx_desc->type == ISCSI_TX_CONTROL) {
631 		/* this arithmetic is legal by libiscsi dd_data allocation */
632 		task = (void *) ((long)(void *)tx_desc -
633 				  sizeof(struct iscsi_task));
634 		if (task->hdr->itt == RESERVED_ITT)
635 			iscsi_put_task(task);
636 	}
637 }
638 
639 void iser_task_rdma_init(struct iscsi_iser_task *iser_task)
640 
641 {
642 	iser_task->status = ISER_TASK_STATUS_INIT;
643 
644 	iser_task->dir[ISER_DIR_IN] = 0;
645 	iser_task->dir[ISER_DIR_OUT] = 0;
646 
647 	iser_task->data[ISER_DIR_IN].data_len  = 0;
648 	iser_task->data[ISER_DIR_OUT].data_len = 0;
649 
650 	iser_task->prot[ISER_DIR_IN].data_len  = 0;
651 	iser_task->prot[ISER_DIR_OUT].data_len = 0;
652 
653 	memset(&iser_task->rdma_regd[ISER_DIR_IN], 0,
654 	       sizeof(struct iser_regd_buf));
655 	memset(&iser_task->rdma_regd[ISER_DIR_OUT], 0,
656 	       sizeof(struct iser_regd_buf));
657 }
658 
659 void iser_task_rdma_finalize(struct iscsi_iser_task *iser_task)
660 {
661 	struct iser_device *device = iser_task->ib_conn->device;
662 	int is_rdma_data_aligned = 1;
663 	int is_rdma_prot_aligned = 1;
664 	int prot_count = scsi_prot_sg_count(iser_task->sc);
665 
666 	/* if we were reading, copy back to unaligned sglist,
667 	 * anyway dma_unmap and free the copy
668 	 */
669 	if (iser_task->data_copy[ISER_DIR_IN].copy_buf != NULL) {
670 		is_rdma_data_aligned = 0;
671 		iser_finalize_rdma_unaligned_sg(iser_task,
672 						&iser_task->data[ISER_DIR_IN],
673 						&iser_task->data_copy[ISER_DIR_IN],
674 						ISER_DIR_IN);
675 	}
676 
677 	if (iser_task->data_copy[ISER_DIR_OUT].copy_buf != NULL) {
678 		is_rdma_data_aligned = 0;
679 		iser_finalize_rdma_unaligned_sg(iser_task,
680 						&iser_task->data[ISER_DIR_OUT],
681 						&iser_task->data_copy[ISER_DIR_OUT],
682 						ISER_DIR_OUT);
683 	}
684 
685 	if (iser_task->prot_copy[ISER_DIR_IN].copy_buf != NULL) {
686 		is_rdma_prot_aligned = 0;
687 		iser_finalize_rdma_unaligned_sg(iser_task,
688 						&iser_task->prot[ISER_DIR_IN],
689 						&iser_task->prot_copy[ISER_DIR_IN],
690 						ISER_DIR_IN);
691 	}
692 
693 	if (iser_task->prot_copy[ISER_DIR_OUT].copy_buf != NULL) {
694 		is_rdma_prot_aligned = 0;
695 		iser_finalize_rdma_unaligned_sg(iser_task,
696 						&iser_task->prot[ISER_DIR_OUT],
697 						&iser_task->prot_copy[ISER_DIR_OUT],
698 						ISER_DIR_OUT);
699 	}
700 
701 	if (iser_task->dir[ISER_DIR_IN]) {
702 		device->iser_unreg_rdma_mem(iser_task, ISER_DIR_IN);
703 		if (is_rdma_data_aligned)
704 			iser_dma_unmap_task_data(iser_task,
705 						 &iser_task->data[ISER_DIR_IN]);
706 		if (prot_count && is_rdma_prot_aligned)
707 			iser_dma_unmap_task_data(iser_task,
708 						 &iser_task->prot[ISER_DIR_IN]);
709 	}
710 
711 	if (iser_task->dir[ISER_DIR_OUT]) {
712 		device->iser_unreg_rdma_mem(iser_task, ISER_DIR_OUT);
713 		if (is_rdma_data_aligned)
714 			iser_dma_unmap_task_data(iser_task,
715 						 &iser_task->data[ISER_DIR_OUT]);
716 		if (prot_count && is_rdma_prot_aligned)
717 			iser_dma_unmap_task_data(iser_task,
718 						 &iser_task->prot[ISER_DIR_OUT]);
719 	}
720 }
721