1 // SPDX-License-Identifier: GPL-2.0 or BSD-3-Clause
2 
3 /* Authors: Bernard Metzler <bmt@zurich.ibm.com> */
4 /* Copyright (c) 2008-2019, IBM Corporation */
5 
6 #include <linux/errno.h>
7 #include <linux/types.h>
8 #include <linux/net.h>
9 #include <linux/scatterlist.h>
10 #include <linux/highmem.h>
11 #include <net/tcp.h>
12 
13 #include <rdma/iw_cm.h>
14 #include <rdma/ib_verbs.h>
15 #include <rdma/ib_user_verbs.h>
16 
17 #include "siw.h"
18 #include "siw_verbs.h"
19 #include "siw_mem.h"
20 
21 #define MAX_HDR_INLINE					\
22 	(((uint32_t)(sizeof(struct siw_rreq_pkt) -	\
23 		     sizeof(struct iwarp_send))) & 0xF8)
24 
siw_get_pblpage(struct siw_mem * mem,u64 addr,int * idx)25 static struct page *siw_get_pblpage(struct siw_mem *mem, u64 addr, int *idx)
26 {
27 	struct siw_pbl *pbl = mem->pbl;
28 	u64 offset = addr - mem->va;
29 	dma_addr_t paddr = siw_pbl_get_buffer(pbl, offset, NULL, idx);
30 
31 	if (paddr)
32 		return virt_to_page(paddr);
33 
34 	return NULL;
35 }
36 
37 /*
38  * Copy short payload at provided destination payload address
39  */
siw_try_1seg(struct siw_iwarp_tx * c_tx,void * paddr)40 static int siw_try_1seg(struct siw_iwarp_tx *c_tx, void *paddr)
41 {
42 	struct siw_wqe *wqe = &c_tx->wqe_active;
43 	struct siw_sge *sge = &wqe->sqe.sge[0];
44 	u32 bytes = sge->length;
45 
46 	if (bytes > MAX_HDR_INLINE || wqe->sqe.num_sge != 1)
47 		return MAX_HDR_INLINE + 1;
48 
49 	if (!bytes)
50 		return 0;
51 
52 	if (tx_flags(wqe) & SIW_WQE_INLINE) {
53 		memcpy(paddr, &wqe->sqe.sge[1], bytes);
54 	} else {
55 		struct siw_mem *mem = wqe->mem[0];
56 
57 		if (!mem->mem_obj) {
58 			/* Kernel client using kva */
59 			memcpy(paddr,
60 			       (const void *)(uintptr_t)sge->laddr, bytes);
61 		} else if (c_tx->in_syscall) {
62 			if (copy_from_user(paddr, u64_to_user_ptr(sge->laddr),
63 					   bytes))
64 				return -EFAULT;
65 		} else {
66 			unsigned int off = sge->laddr & ~PAGE_MASK;
67 			struct page *p;
68 			char *buffer;
69 			int pbl_idx = 0;
70 
71 			if (!mem->is_pbl)
72 				p = siw_get_upage(mem->umem, sge->laddr);
73 			else
74 				p = siw_get_pblpage(mem, sge->laddr, &pbl_idx);
75 
76 			if (unlikely(!p))
77 				return -EFAULT;
78 
79 			buffer = kmap(p);
80 
81 			if (likely(PAGE_SIZE - off >= bytes)) {
82 				memcpy(paddr, buffer + off, bytes);
83 			} else {
84 				unsigned long part = bytes - (PAGE_SIZE - off);
85 
86 				memcpy(paddr, buffer + off, part);
87 				kunmap(p);
88 
89 				if (!mem->is_pbl)
90 					p = siw_get_upage(mem->umem,
91 							  sge->laddr + part);
92 				else
93 					p = siw_get_pblpage(mem,
94 							    sge->laddr + part,
95 							    &pbl_idx);
96 				if (unlikely(!p))
97 					return -EFAULT;
98 
99 				buffer = kmap(p);
100 				memcpy(paddr + part, buffer, bytes - part);
101 			}
102 			kunmap(p);
103 		}
104 	}
105 	return (int)bytes;
106 }
107 
108 #define PKT_FRAGMENTED 1
109 #define PKT_COMPLETE 0
110 
111 /*
112  * siw_qp_prepare_tx()
113  *
114  * Prepare tx state for sending out one fpdu. Builds complete pkt
115  * if no user data or only immediate data are present.
116  *
117  * returns PKT_COMPLETE if complete pkt built, PKT_FRAGMENTED otherwise.
118  */
siw_qp_prepare_tx(struct siw_iwarp_tx * c_tx)119 static int siw_qp_prepare_tx(struct siw_iwarp_tx *c_tx)
120 {
121 	struct siw_wqe *wqe = &c_tx->wqe_active;
122 	char *crc = NULL;
123 	int data = 0;
124 
125 	switch (tx_type(wqe)) {
126 	case SIW_OP_READ:
127 	case SIW_OP_READ_LOCAL_INV:
128 		memcpy(&c_tx->pkt.ctrl,
129 		       &iwarp_pktinfo[RDMAP_RDMA_READ_REQ].ctrl,
130 		       sizeof(struct iwarp_ctrl));
131 
132 		c_tx->pkt.rreq.rsvd = 0;
133 		c_tx->pkt.rreq.ddp_qn = htonl(RDMAP_UNTAGGED_QN_RDMA_READ);
134 		c_tx->pkt.rreq.ddp_msn =
135 			htonl(++c_tx->ddp_msn[RDMAP_UNTAGGED_QN_RDMA_READ]);
136 		c_tx->pkt.rreq.ddp_mo = 0;
137 		c_tx->pkt.rreq.sink_stag = htonl(wqe->sqe.sge[0].lkey);
138 		c_tx->pkt.rreq.sink_to =
139 			cpu_to_be64(wqe->sqe.sge[0].laddr);
140 		c_tx->pkt.rreq.source_stag = htonl(wqe->sqe.rkey);
141 		c_tx->pkt.rreq.source_to = cpu_to_be64(wqe->sqe.raddr);
142 		c_tx->pkt.rreq.read_size = htonl(wqe->sqe.sge[0].length);
143 
144 		c_tx->ctrl_len = sizeof(struct iwarp_rdma_rreq);
145 		crc = (char *)&c_tx->pkt.rreq_pkt.crc;
146 		break;
147 
148 	case SIW_OP_SEND:
149 		if (tx_flags(wqe) & SIW_WQE_SOLICITED)
150 			memcpy(&c_tx->pkt.ctrl,
151 			       &iwarp_pktinfo[RDMAP_SEND_SE].ctrl,
152 			       sizeof(struct iwarp_ctrl));
153 		else
154 			memcpy(&c_tx->pkt.ctrl, &iwarp_pktinfo[RDMAP_SEND].ctrl,
155 			       sizeof(struct iwarp_ctrl));
156 
157 		c_tx->pkt.send.ddp_qn = RDMAP_UNTAGGED_QN_SEND;
158 		c_tx->pkt.send.ddp_msn =
159 			htonl(++c_tx->ddp_msn[RDMAP_UNTAGGED_QN_SEND]);
160 		c_tx->pkt.send.ddp_mo = 0;
161 
162 		c_tx->pkt.send_inv.inval_stag = 0;
163 
164 		c_tx->ctrl_len = sizeof(struct iwarp_send);
165 
166 		crc = (char *)&c_tx->pkt.send_pkt.crc;
167 		data = siw_try_1seg(c_tx, crc);
168 		break;
169 
170 	case SIW_OP_SEND_REMOTE_INV:
171 		if (tx_flags(wqe) & SIW_WQE_SOLICITED)
172 			memcpy(&c_tx->pkt.ctrl,
173 			       &iwarp_pktinfo[RDMAP_SEND_SE_INVAL].ctrl,
174 			       sizeof(struct iwarp_ctrl));
175 		else
176 			memcpy(&c_tx->pkt.ctrl,
177 			       &iwarp_pktinfo[RDMAP_SEND_INVAL].ctrl,
178 			       sizeof(struct iwarp_ctrl));
179 
180 		c_tx->pkt.send.ddp_qn = RDMAP_UNTAGGED_QN_SEND;
181 		c_tx->pkt.send.ddp_msn =
182 			htonl(++c_tx->ddp_msn[RDMAP_UNTAGGED_QN_SEND]);
183 		c_tx->pkt.send.ddp_mo = 0;
184 
185 		c_tx->pkt.send_inv.inval_stag = cpu_to_be32(wqe->sqe.rkey);
186 
187 		c_tx->ctrl_len = sizeof(struct iwarp_send_inv);
188 
189 		crc = (char *)&c_tx->pkt.send_pkt.crc;
190 		data = siw_try_1seg(c_tx, crc);
191 		break;
192 
193 	case SIW_OP_WRITE:
194 		memcpy(&c_tx->pkt.ctrl, &iwarp_pktinfo[RDMAP_RDMA_WRITE].ctrl,
195 		       sizeof(struct iwarp_ctrl));
196 
197 		c_tx->pkt.rwrite.sink_stag = htonl(wqe->sqe.rkey);
198 		c_tx->pkt.rwrite.sink_to = cpu_to_be64(wqe->sqe.raddr);
199 		c_tx->ctrl_len = sizeof(struct iwarp_rdma_write);
200 
201 		crc = (char *)&c_tx->pkt.write_pkt.crc;
202 		data = siw_try_1seg(c_tx, crc);
203 		break;
204 
205 	case SIW_OP_READ_RESPONSE:
206 		memcpy(&c_tx->pkt.ctrl,
207 		       &iwarp_pktinfo[RDMAP_RDMA_READ_RESP].ctrl,
208 		       sizeof(struct iwarp_ctrl));
209 
210 		/* NBO */
211 		c_tx->pkt.rresp.sink_stag = cpu_to_be32(wqe->sqe.rkey);
212 		c_tx->pkt.rresp.sink_to = cpu_to_be64(wqe->sqe.raddr);
213 
214 		c_tx->ctrl_len = sizeof(struct iwarp_rdma_rresp);
215 
216 		crc = (char *)&c_tx->pkt.write_pkt.crc;
217 		data = siw_try_1seg(c_tx, crc);
218 		break;
219 
220 	default:
221 		siw_dbg_qp(tx_qp(c_tx), "stale wqe type %d\n", tx_type(wqe));
222 		return -EOPNOTSUPP;
223 	}
224 	if (unlikely(data < 0))
225 		return data;
226 
227 	c_tx->ctrl_sent = 0;
228 
229 	if (data <= MAX_HDR_INLINE) {
230 		if (data) {
231 			wqe->processed = data;
232 
233 			c_tx->pkt.ctrl.mpa_len =
234 				htons(c_tx->ctrl_len + data - MPA_HDR_SIZE);
235 
236 			/* Add pad, if needed */
237 			data += -(int)data & 0x3;
238 			/* advance CRC location after payload */
239 			crc += data;
240 			c_tx->ctrl_len += data;
241 
242 			if (!(c_tx->pkt.ctrl.ddp_rdmap_ctrl & DDP_FLAG_TAGGED))
243 				c_tx->pkt.c_untagged.ddp_mo = 0;
244 			else
245 				c_tx->pkt.c_tagged.ddp_to =
246 					cpu_to_be64(wqe->sqe.raddr);
247 		}
248 
249 		*(u32 *)crc = 0;
250 		/*
251 		 * Do complete CRC if enabled and short packet
252 		 */
253 		if (c_tx->mpa_crc_hd) {
254 			crypto_shash_init(c_tx->mpa_crc_hd);
255 			if (crypto_shash_update(c_tx->mpa_crc_hd,
256 						(u8 *)&c_tx->pkt,
257 						c_tx->ctrl_len))
258 				return -EINVAL;
259 			crypto_shash_final(c_tx->mpa_crc_hd, (u8 *)crc);
260 		}
261 		c_tx->ctrl_len += MPA_CRC_SIZE;
262 
263 		return PKT_COMPLETE;
264 	}
265 	c_tx->ctrl_len += MPA_CRC_SIZE;
266 	c_tx->sge_idx = 0;
267 	c_tx->sge_off = 0;
268 	c_tx->pbl_idx = 0;
269 
270 	/*
271 	 * Allow direct sending out of user buffer if WR is non signalled
272 	 * and payload is over threshold.
273 	 * Per RDMA verbs, the application should not change the send buffer
274 	 * until the work completed. In iWarp, work completion is only
275 	 * local delivery to TCP. TCP may reuse the buffer for
276 	 * retransmission. Changing unsent data also breaks the CRC,
277 	 * if applied.
278 	 */
279 	if (c_tx->zcopy_tx && wqe->bytes >= SENDPAGE_THRESH &&
280 	    !(tx_flags(wqe) & SIW_WQE_SIGNALLED))
281 		c_tx->use_sendpage = 1;
282 	else
283 		c_tx->use_sendpage = 0;
284 
285 	return PKT_FRAGMENTED;
286 }
287 
288 /*
289  * Send out one complete control type FPDU, or header of FPDU carrying
290  * data. Used for fixed sized packets like Read.Requests or zero length
291  * SENDs, WRITEs, READ.Responses, or header only.
292  */
siw_tx_ctrl(struct siw_iwarp_tx * c_tx,struct socket * s,int flags)293 static int siw_tx_ctrl(struct siw_iwarp_tx *c_tx, struct socket *s,
294 			      int flags)
295 {
296 	struct msghdr msg = { .msg_flags = flags };
297 	struct kvec iov = { .iov_base =
298 				    (char *)&c_tx->pkt.ctrl + c_tx->ctrl_sent,
299 			    .iov_len = c_tx->ctrl_len - c_tx->ctrl_sent };
300 
301 	int rv = kernel_sendmsg(s, &msg, &iov, 1,
302 				c_tx->ctrl_len - c_tx->ctrl_sent);
303 
304 	if (rv >= 0) {
305 		c_tx->ctrl_sent += rv;
306 
307 		if (c_tx->ctrl_sent == c_tx->ctrl_len)
308 			rv = 0;
309 		else
310 			rv = -EAGAIN;
311 	}
312 	return rv;
313 }
314 
315 /*
316  * 0copy TCP transmit interface: Use do_tcp_sendpages.
317  *
318  * Using sendpage to push page by page appears to be less efficient
319  * than using sendmsg, even if data are copied.
320  *
321  * A general performance limitation might be the extra four bytes
322  * trailer checksum segment to be pushed after user data.
323  */
siw_tcp_sendpages(struct socket * s,struct page ** page,int offset,size_t size)324 static int siw_tcp_sendpages(struct socket *s, struct page **page, int offset,
325 			     size_t size)
326 {
327 	struct sock *sk = s->sk;
328 	int i = 0, rv = 0, sent = 0,
329 	    flags = MSG_MORE | MSG_DONTWAIT | MSG_SENDPAGE_NOTLAST;
330 
331 	while (size) {
332 		size_t bytes = min_t(size_t, PAGE_SIZE - offset, size);
333 
334 		if (size + offset <= PAGE_SIZE)
335 			flags = MSG_MORE | MSG_DONTWAIT;
336 
337 		tcp_rate_check_app_limited(sk);
338 try_page_again:
339 		lock_sock(sk);
340 		rv = do_tcp_sendpages(sk, page[i], offset, bytes, flags);
341 		release_sock(sk);
342 
343 		if (rv > 0) {
344 			size -= rv;
345 			sent += rv;
346 			if (rv != bytes) {
347 				offset += rv;
348 				bytes -= rv;
349 				goto try_page_again;
350 			}
351 			offset = 0;
352 		} else {
353 			if (rv == -EAGAIN || rv == 0)
354 				break;
355 			return rv;
356 		}
357 		i++;
358 	}
359 	return sent;
360 }
361 
362 /*
363  * siw_0copy_tx()
364  *
365  * Pushes list of pages to TCP socket. If pages from multiple
366  * SGE's, all referenced pages of each SGE are pushed in one
367  * shot.
368  */
siw_0copy_tx(struct socket * s,struct page ** page,struct siw_sge * sge,unsigned int offset,unsigned int size)369 static int siw_0copy_tx(struct socket *s, struct page **page,
370 			struct siw_sge *sge, unsigned int offset,
371 			unsigned int size)
372 {
373 	int i = 0, sent = 0, rv;
374 	int sge_bytes = min(sge->length - offset, size);
375 
376 	offset = (sge->laddr + offset) & ~PAGE_MASK;
377 
378 	while (sent != size) {
379 		rv = siw_tcp_sendpages(s, &page[i], offset, sge_bytes);
380 		if (rv >= 0) {
381 			sent += rv;
382 			if (size == sent || sge_bytes > rv)
383 				break;
384 
385 			i += PAGE_ALIGN(sge_bytes + offset) >> PAGE_SHIFT;
386 			sge++;
387 			sge_bytes = min(sge->length, size - sent);
388 			offset = sge->laddr & ~PAGE_MASK;
389 		} else {
390 			sent = rv;
391 			break;
392 		}
393 	}
394 	return sent;
395 }
396 
397 #define MAX_TRAILER (MPA_CRC_SIZE + 4)
398 
siw_unmap_pages(struct page ** pp,unsigned long kmap_mask)399 static void siw_unmap_pages(struct page **pp, unsigned long kmap_mask)
400 {
401 	while (kmap_mask) {
402 		if (kmap_mask & BIT(0))
403 			kunmap(*pp);
404 		pp++;
405 		kmap_mask >>= 1;
406 	}
407 }
408 
409 /*
410  * siw_tx_hdt() tries to push a complete packet to TCP where all
411  * packet fragments are referenced by the elements of one iovec.
412  * For the data portion, each involved page must be referenced by
413  * one extra element. All sge's data can be non-aligned to page
414  * boundaries. Two more elements are referencing iWARP header
415  * and trailer:
416  * MAX_ARRAY = 64KB/PAGE_SIZE + 1 + (2 * (SIW_MAX_SGE - 1) + HDR + TRL
417  */
418 #define MAX_ARRAY ((0xffff / PAGE_SIZE) + 1 + (2 * (SIW_MAX_SGE - 1) + 2))
419 
420 /*
421  * Write out iov referencing hdr, data and trailer of current FPDU.
422  * Update transmit state dependent on write return status
423  */
siw_tx_hdt(struct siw_iwarp_tx * c_tx,struct socket * s)424 static int siw_tx_hdt(struct siw_iwarp_tx *c_tx, struct socket *s)
425 {
426 	struct siw_wqe *wqe = &c_tx->wqe_active;
427 	struct siw_sge *sge = &wqe->sqe.sge[c_tx->sge_idx];
428 	struct kvec iov[MAX_ARRAY];
429 	struct page *page_array[MAX_ARRAY];
430 	struct msghdr msg = { .msg_flags = MSG_DONTWAIT | MSG_EOR };
431 
432 	int seg = 0, do_crc = c_tx->do_crc, is_kva = 0, rv;
433 	unsigned int data_len = c_tx->bytes_unsent, hdr_len = 0, trl_len = 0,
434 		     sge_off = c_tx->sge_off, sge_idx = c_tx->sge_idx,
435 		     pbl_idx = c_tx->pbl_idx;
436 	unsigned long kmap_mask = 0L;
437 
438 	if (c_tx->state == SIW_SEND_HDR) {
439 		if (c_tx->use_sendpage) {
440 			rv = siw_tx_ctrl(c_tx, s, MSG_DONTWAIT | MSG_MORE);
441 			if (rv)
442 				goto done;
443 
444 			c_tx->state = SIW_SEND_DATA;
445 		} else {
446 			iov[0].iov_base =
447 				(char *)&c_tx->pkt.ctrl + c_tx->ctrl_sent;
448 			iov[0].iov_len = hdr_len =
449 				c_tx->ctrl_len - c_tx->ctrl_sent;
450 			seg = 1;
451 		}
452 	}
453 
454 	wqe->processed += data_len;
455 
456 	while (data_len) { /* walk the list of SGE's */
457 		unsigned int sge_len = min(sge->length - sge_off, data_len);
458 		unsigned int fp_off = (sge->laddr + sge_off) & ~PAGE_MASK;
459 		struct siw_mem *mem;
460 
461 		if (!(tx_flags(wqe) & SIW_WQE_INLINE)) {
462 			mem = wqe->mem[sge_idx];
463 			is_kva = mem->mem_obj == NULL ? 1 : 0;
464 		} else {
465 			is_kva = 1;
466 		}
467 		if (is_kva && !c_tx->use_sendpage) {
468 			/*
469 			 * tx from kernel virtual address: either inline data
470 			 * or memory region with assigned kernel buffer
471 			 */
472 			iov[seg].iov_base =
473 				(void *)(uintptr_t)(sge->laddr + sge_off);
474 			iov[seg].iov_len = sge_len;
475 
476 			if (do_crc)
477 				crypto_shash_update(c_tx->mpa_crc_hd,
478 						    iov[seg].iov_base,
479 						    sge_len);
480 			sge_off += sge_len;
481 			data_len -= sge_len;
482 			seg++;
483 			goto sge_done;
484 		}
485 
486 		while (sge_len) {
487 			size_t plen = min((int)PAGE_SIZE - fp_off, sge_len);
488 
489 			if (!is_kva) {
490 				struct page *p;
491 
492 				if (mem->is_pbl)
493 					p = siw_get_pblpage(
494 						mem, sge->laddr + sge_off,
495 						&pbl_idx);
496 				else
497 					p = siw_get_upage(mem->umem,
498 							  sge->laddr + sge_off);
499 				if (unlikely(!p)) {
500 					siw_unmap_pages(page_array, kmap_mask);
501 					wqe->processed -= c_tx->bytes_unsent;
502 					rv = -EFAULT;
503 					goto done_crc;
504 				}
505 				page_array[seg] = p;
506 
507 				if (!c_tx->use_sendpage) {
508 					iov[seg].iov_base = kmap(p) + fp_off;
509 					iov[seg].iov_len = plen;
510 
511 					/* Remember for later kunmap() */
512 					kmap_mask |= BIT(seg);
513 
514 					if (do_crc)
515 						crypto_shash_update(
516 							c_tx->mpa_crc_hd,
517 							iov[seg].iov_base,
518 							plen);
519 				} else if (do_crc) {
520 					crypto_shash_update(c_tx->mpa_crc_hd,
521 							    kmap(p) + fp_off,
522 							    plen);
523 					kunmap(p);
524 				}
525 			} else {
526 				u64 va = sge->laddr + sge_off;
527 
528 				page_array[seg] = virt_to_page(va & PAGE_MASK);
529 				if (do_crc)
530 					crypto_shash_update(
531 						c_tx->mpa_crc_hd,
532 						(void *)(uintptr_t)va,
533 						plen);
534 			}
535 
536 			sge_len -= plen;
537 			sge_off += plen;
538 			data_len -= plen;
539 			fp_off = 0;
540 
541 			if (++seg > (int)MAX_ARRAY) {
542 				siw_dbg_qp(tx_qp(c_tx), "to many fragments\n");
543 				siw_unmap_pages(page_array, kmap_mask);
544 				wqe->processed -= c_tx->bytes_unsent;
545 				rv = -EMSGSIZE;
546 				goto done_crc;
547 			}
548 		}
549 sge_done:
550 		/* Update SGE variables at end of SGE */
551 		if (sge_off == sge->length &&
552 		    (data_len != 0 || wqe->processed < wqe->bytes)) {
553 			sge_idx++;
554 			sge++;
555 			sge_off = 0;
556 		}
557 	}
558 	/* trailer */
559 	if (likely(c_tx->state != SIW_SEND_TRAILER)) {
560 		iov[seg].iov_base = &c_tx->trailer.pad[4 - c_tx->pad];
561 		iov[seg].iov_len = trl_len = MAX_TRAILER - (4 - c_tx->pad);
562 	} else {
563 		iov[seg].iov_base = &c_tx->trailer.pad[c_tx->ctrl_sent];
564 		iov[seg].iov_len = trl_len = MAX_TRAILER - c_tx->ctrl_sent;
565 	}
566 
567 	if (c_tx->pad) {
568 		*(u32 *)c_tx->trailer.pad = 0;
569 		if (do_crc)
570 			crypto_shash_update(c_tx->mpa_crc_hd,
571 				(u8 *)&c_tx->trailer.crc - c_tx->pad,
572 				c_tx->pad);
573 	}
574 	if (!c_tx->mpa_crc_hd)
575 		c_tx->trailer.crc = 0;
576 	else if (do_crc)
577 		crypto_shash_final(c_tx->mpa_crc_hd, (u8 *)&c_tx->trailer.crc);
578 
579 	data_len = c_tx->bytes_unsent;
580 
581 	if (c_tx->use_sendpage) {
582 		rv = siw_0copy_tx(s, page_array, &wqe->sqe.sge[c_tx->sge_idx],
583 				  c_tx->sge_off, data_len);
584 		if (rv == data_len) {
585 			rv = kernel_sendmsg(s, &msg, &iov[seg], 1, trl_len);
586 			if (rv > 0)
587 				rv += data_len;
588 			else
589 				rv = data_len;
590 		}
591 	} else {
592 		rv = kernel_sendmsg(s, &msg, iov, seg + 1,
593 				    hdr_len + data_len + trl_len);
594 		siw_unmap_pages(page_array, kmap_mask);
595 	}
596 	if (rv < (int)hdr_len) {
597 		/* Not even complete hdr pushed or negative rv */
598 		wqe->processed -= data_len;
599 		if (rv >= 0) {
600 			c_tx->ctrl_sent += rv;
601 			rv = -EAGAIN;
602 		}
603 		goto done_crc;
604 	}
605 	rv -= hdr_len;
606 
607 	if (rv >= (int)data_len) {
608 		/* all user data pushed to TCP or no data to push */
609 		if (data_len > 0 && wqe->processed < wqe->bytes) {
610 			/* Save the current state for next tx */
611 			c_tx->sge_idx = sge_idx;
612 			c_tx->sge_off = sge_off;
613 			c_tx->pbl_idx = pbl_idx;
614 		}
615 		rv -= data_len;
616 
617 		if (rv == trl_len) /* all pushed */
618 			rv = 0;
619 		else {
620 			c_tx->state = SIW_SEND_TRAILER;
621 			c_tx->ctrl_len = MAX_TRAILER;
622 			c_tx->ctrl_sent = rv + 4 - c_tx->pad;
623 			c_tx->bytes_unsent = 0;
624 			rv = -EAGAIN;
625 		}
626 
627 	} else if (data_len > 0) {
628 		/* Maybe some user data pushed to TCP */
629 		c_tx->state = SIW_SEND_DATA;
630 		wqe->processed -= data_len - rv;
631 
632 		if (rv) {
633 			/*
634 			 * Some bytes out. Recompute tx state based
635 			 * on old state and bytes pushed
636 			 */
637 			unsigned int sge_unsent;
638 
639 			c_tx->bytes_unsent -= rv;
640 			sge = &wqe->sqe.sge[c_tx->sge_idx];
641 			sge_unsent = sge->length - c_tx->sge_off;
642 
643 			while (sge_unsent <= rv) {
644 				rv -= sge_unsent;
645 				c_tx->sge_idx++;
646 				c_tx->sge_off = 0;
647 				sge++;
648 				sge_unsent = sge->length;
649 			}
650 			c_tx->sge_off += rv;
651 		}
652 		rv = -EAGAIN;
653 	}
654 done_crc:
655 	c_tx->do_crc = 0;
656 done:
657 	return rv;
658 }
659 
siw_update_tcpseg(struct siw_iwarp_tx * c_tx,struct socket * s)660 static void siw_update_tcpseg(struct siw_iwarp_tx *c_tx,
661 				     struct socket *s)
662 {
663 	struct tcp_sock *tp = tcp_sk(s->sk);
664 
665 	if (tp->gso_segs) {
666 		if (c_tx->gso_seg_limit == 0)
667 			c_tx->tcp_seglen = tp->mss_cache * tp->gso_segs;
668 		else
669 			c_tx->tcp_seglen =
670 				tp->mss_cache *
671 				min_t(u16, c_tx->gso_seg_limit, tp->gso_segs);
672 	} else {
673 		c_tx->tcp_seglen = tp->mss_cache;
674 	}
675 	/* Loopback may give odd numbers */
676 	c_tx->tcp_seglen &= 0xfffffff8;
677 }
678 
679 /*
680  * siw_prepare_fpdu()
681  *
682  * Prepares transmit context to send out one FPDU if FPDU will contain
683  * user data and user data are not immediate data.
684  * Computes maximum FPDU length to fill up TCP MSS if possible.
685  *
686  * @qp:		QP from which to transmit
687  * @wqe:	Current WQE causing transmission
688  *
689  * TODO: Take into account real available sendspace on socket
690  *       to avoid header misalignment due to send pausing within
691  *       fpdu transmission
692  */
siw_prepare_fpdu(struct siw_qp * qp,struct siw_wqe * wqe)693 static void siw_prepare_fpdu(struct siw_qp *qp, struct siw_wqe *wqe)
694 {
695 	struct siw_iwarp_tx *c_tx = &qp->tx_ctx;
696 	int data_len;
697 
698 	c_tx->ctrl_len =
699 		iwarp_pktinfo[__rdmap_get_opcode(&c_tx->pkt.ctrl)].hdr_len;
700 	c_tx->ctrl_sent = 0;
701 
702 	/*
703 	 * Update target buffer offset if any
704 	 */
705 	if (!(c_tx->pkt.ctrl.ddp_rdmap_ctrl & DDP_FLAG_TAGGED))
706 		/* Untagged message */
707 		c_tx->pkt.c_untagged.ddp_mo = cpu_to_be32(wqe->processed);
708 	else /* Tagged message */
709 		c_tx->pkt.c_tagged.ddp_to =
710 			cpu_to_be64(wqe->sqe.raddr + wqe->processed);
711 
712 	data_len = wqe->bytes - wqe->processed;
713 	if (data_len + c_tx->ctrl_len + MPA_CRC_SIZE > c_tx->tcp_seglen) {
714 		/* Trim DDP payload to fit into current TCP segment */
715 		data_len = c_tx->tcp_seglen - (c_tx->ctrl_len + MPA_CRC_SIZE);
716 		c_tx->pkt.ctrl.ddp_rdmap_ctrl &= ~DDP_FLAG_LAST;
717 		c_tx->pad = 0;
718 	} else {
719 		c_tx->pkt.ctrl.ddp_rdmap_ctrl |= DDP_FLAG_LAST;
720 		c_tx->pad = -data_len & 0x3;
721 	}
722 	c_tx->bytes_unsent = data_len;
723 
724 	c_tx->pkt.ctrl.mpa_len =
725 		htons(c_tx->ctrl_len + data_len - MPA_HDR_SIZE);
726 
727 	/*
728 	 * Init MPA CRC computation
729 	 */
730 	if (c_tx->mpa_crc_hd) {
731 		crypto_shash_init(c_tx->mpa_crc_hd);
732 		crypto_shash_update(c_tx->mpa_crc_hd, (u8 *)&c_tx->pkt,
733 				    c_tx->ctrl_len);
734 		c_tx->do_crc = 1;
735 	}
736 }
737 
738 /*
739  * siw_check_sgl_tx()
740  *
741  * Check permissions for a list of SGE's (SGL).
742  * A successful check will have all memory referenced
743  * for transmission resolved and assigned to the WQE.
744  *
745  * @pd:		Protection Domain SGL should belong to
746  * @wqe:	WQE to be checked
747  * @perms:	requested access permissions
748  *
749  */
750 
siw_check_sgl_tx(struct ib_pd * pd,struct siw_wqe * wqe,enum ib_access_flags perms)751 static int siw_check_sgl_tx(struct ib_pd *pd, struct siw_wqe *wqe,
752 			    enum ib_access_flags perms)
753 {
754 	struct siw_sge *sge = &wqe->sqe.sge[0];
755 	int i, len, num_sge = wqe->sqe.num_sge;
756 
757 	if (unlikely(num_sge > SIW_MAX_SGE))
758 		return -EINVAL;
759 
760 	for (i = 0, len = 0; num_sge; num_sge--, i++, sge++) {
761 		/*
762 		 * rdma verbs: do not check stag for a zero length sge
763 		 */
764 		if (sge->length) {
765 			int rv = siw_check_sge(pd, sge, &wqe->mem[i], perms, 0,
766 					       sge->length);
767 
768 			if (unlikely(rv != E_ACCESS_OK))
769 				return rv;
770 		}
771 		len += sge->length;
772 	}
773 	return len;
774 }
775 
776 /*
777  * siw_qp_sq_proc_tx()
778  *
779  * Process one WQE which needs transmission on the wire.
780  */
siw_qp_sq_proc_tx(struct siw_qp * qp,struct siw_wqe * wqe)781 static int siw_qp_sq_proc_tx(struct siw_qp *qp, struct siw_wqe *wqe)
782 {
783 	struct siw_iwarp_tx *c_tx = &qp->tx_ctx;
784 	struct socket *s = qp->attrs.sk;
785 	int rv = 0, burst_len = qp->tx_ctx.burst;
786 	enum rdmap_ecode ecode = RDMAP_ECODE_CATASTROPHIC_STREAM;
787 
788 	if (unlikely(wqe->wr_status == SIW_WR_IDLE))
789 		return 0;
790 
791 	if (!burst_len)
792 		burst_len = SQ_USER_MAXBURST;
793 
794 	if (wqe->wr_status == SIW_WR_QUEUED) {
795 		if (!(wqe->sqe.flags & SIW_WQE_INLINE)) {
796 			if (tx_type(wqe) == SIW_OP_READ_RESPONSE)
797 				wqe->sqe.num_sge = 1;
798 
799 			if (tx_type(wqe) != SIW_OP_READ &&
800 			    tx_type(wqe) != SIW_OP_READ_LOCAL_INV) {
801 				/*
802 				 * Reference memory to be tx'd w/o checking
803 				 * access for LOCAL_READ permission, since
804 				 * not defined in RDMA core.
805 				 */
806 				rv = siw_check_sgl_tx(qp->pd, wqe, 0);
807 				if (rv < 0) {
808 					if (tx_type(wqe) ==
809 					    SIW_OP_READ_RESPONSE)
810 						ecode = siw_rdmap_error(-rv);
811 					rv = -EINVAL;
812 					goto tx_error;
813 				}
814 				wqe->bytes = rv;
815 			} else {
816 				wqe->bytes = 0;
817 			}
818 		} else {
819 			wqe->bytes = wqe->sqe.sge[0].length;
820 			if (!rdma_is_kernel_res(&qp->base_qp.res)) {
821 				if (wqe->bytes > SIW_MAX_INLINE) {
822 					rv = -EINVAL;
823 					goto tx_error;
824 				}
825 				wqe->sqe.sge[0].laddr =
826 					(u64)(uintptr_t)&wqe->sqe.sge[1];
827 			}
828 		}
829 		wqe->wr_status = SIW_WR_INPROGRESS;
830 		wqe->processed = 0;
831 
832 		siw_update_tcpseg(c_tx, s);
833 
834 		rv = siw_qp_prepare_tx(c_tx);
835 		if (rv == PKT_FRAGMENTED) {
836 			c_tx->state = SIW_SEND_HDR;
837 			siw_prepare_fpdu(qp, wqe);
838 		} else if (rv == PKT_COMPLETE) {
839 			c_tx->state = SIW_SEND_SHORT_FPDU;
840 		} else {
841 			goto tx_error;
842 		}
843 	}
844 
845 next_segment:
846 	siw_dbg_qp(qp, "wr type %d, state %d, data %u, sent %u, id %llx\n",
847 		   tx_type(wqe), wqe->wr_status, wqe->bytes, wqe->processed,
848 		   wqe->sqe.id);
849 
850 	if (--burst_len == 0) {
851 		rv = -EINPROGRESS;
852 		goto tx_done;
853 	}
854 	if (c_tx->state == SIW_SEND_SHORT_FPDU) {
855 		enum siw_opcode tx_type = tx_type(wqe);
856 		unsigned int msg_flags;
857 
858 		if (siw_sq_empty(qp) || !siw_tcp_nagle || burst_len == 1)
859 			/*
860 			 * End current TCP segment, if SQ runs empty,
861 			 * or siw_tcp_nagle is not set, or we bail out
862 			 * soon due to no burst credit left.
863 			 */
864 			msg_flags = MSG_DONTWAIT;
865 		else
866 			msg_flags = MSG_DONTWAIT | MSG_MORE;
867 
868 		rv = siw_tx_ctrl(c_tx, s, msg_flags);
869 
870 		if (!rv && tx_type != SIW_OP_READ &&
871 		    tx_type != SIW_OP_READ_LOCAL_INV)
872 			wqe->processed = wqe->bytes;
873 
874 		goto tx_done;
875 
876 	} else {
877 		rv = siw_tx_hdt(c_tx, s);
878 	}
879 	if (!rv) {
880 		/*
881 		 * One segment sent. Processing completed if last
882 		 * segment, Do next segment otherwise.
883 		 */
884 		if (unlikely(c_tx->tx_suspend)) {
885 			/*
886 			 * Verbs, 6.4.: Try stopping sending after a full
887 			 * DDP segment if the connection goes down
888 			 * (== peer halfclose)
889 			 */
890 			rv = -ECONNABORTED;
891 			goto tx_done;
892 		}
893 		if (c_tx->pkt.ctrl.ddp_rdmap_ctrl & DDP_FLAG_LAST) {
894 			siw_dbg_qp(qp, "WQE completed\n");
895 			goto tx_done;
896 		}
897 		c_tx->state = SIW_SEND_HDR;
898 
899 		siw_update_tcpseg(c_tx, s);
900 
901 		siw_prepare_fpdu(qp, wqe);
902 		goto next_segment;
903 	}
904 tx_done:
905 	qp->tx_ctx.burst = burst_len;
906 	return rv;
907 
908 tx_error:
909 	if (ecode != RDMAP_ECODE_CATASTROPHIC_STREAM)
910 		siw_init_terminate(qp, TERM_ERROR_LAYER_RDMAP,
911 				   RDMAP_ETYPE_REMOTE_PROTECTION, ecode, 1);
912 	else
913 		siw_init_terminate(qp, TERM_ERROR_LAYER_RDMAP,
914 				   RDMAP_ETYPE_CATASTROPHIC,
915 				   RDMAP_ECODE_UNSPECIFIED, 1);
916 	return rv;
917 }
918 
siw_fastreg_mr(struct ib_pd * pd,struct siw_sqe * sqe)919 static int siw_fastreg_mr(struct ib_pd *pd, struct siw_sqe *sqe)
920 {
921 	struct ib_mr *base_mr = (struct ib_mr *)(uintptr_t)sqe->base_mr;
922 	struct siw_device *sdev = to_siw_dev(pd->device);
923 	struct siw_mem *mem;
924 	int rv = 0;
925 
926 	siw_dbg_pd(pd, "STag 0x%08x\n", sqe->rkey);
927 
928 	if (unlikely(!base_mr)) {
929 		pr_warn("siw: fastreg: STag 0x%08x unknown\n", sqe->rkey);
930 		return -EINVAL;
931 	}
932 
933 	if (unlikely(base_mr->rkey >> 8 != sqe->rkey  >> 8)) {
934 		pr_warn("siw: fastreg: STag 0x%08x: bad MR\n", sqe->rkey);
935 		return -EINVAL;
936 	}
937 
938 	mem = siw_mem_id2obj(sdev, sqe->rkey  >> 8);
939 	if (unlikely(!mem)) {
940 		pr_warn("siw: fastreg: STag 0x%08x unknown\n", sqe->rkey);
941 		return -EINVAL;
942 	}
943 
944 	if (unlikely(mem->pd != pd)) {
945 		pr_warn("siw: fastreg: PD mismatch\n");
946 		rv = -EINVAL;
947 		goto out;
948 	}
949 	if (unlikely(mem->stag_valid)) {
950 		pr_warn("siw: fastreg: STag 0x%08x already valid\n", sqe->rkey);
951 		rv = -EINVAL;
952 		goto out;
953 	}
954 	/* Refresh STag since user may have changed key part */
955 	mem->stag = sqe->rkey;
956 	mem->perms = sqe->access;
957 
958 	siw_dbg_mem(mem, "STag 0x%08x now valid\n", sqe->rkey);
959 	mem->va = base_mr->iova;
960 	mem->stag_valid = 1;
961 out:
962 	siw_mem_put(mem);
963 	return rv;
964 }
965 
siw_qp_sq_proc_local(struct siw_qp * qp,struct siw_wqe * wqe)966 static int siw_qp_sq_proc_local(struct siw_qp *qp, struct siw_wqe *wqe)
967 {
968 	int rv;
969 
970 	switch (tx_type(wqe)) {
971 	case SIW_OP_REG_MR:
972 		rv = siw_fastreg_mr(qp->pd, &wqe->sqe);
973 		break;
974 
975 	case SIW_OP_INVAL_STAG:
976 		rv = siw_invalidate_stag(qp->pd, wqe->sqe.rkey);
977 		break;
978 
979 	default:
980 		rv = -EINVAL;
981 	}
982 	return rv;
983 }
984 
985 /*
986  * siw_qp_sq_process()
987  *
988  * Core TX path routine for RDMAP/DDP/MPA using a TCP kernel socket.
989  * Sends RDMAP payload for the current SQ WR @wqe of @qp in one or more
990  * MPA FPDUs, each containing a DDP segment.
991  *
992  * SQ processing may occur in user context as a result of posting
993  * new WQE's or from siw_sq_work_handler() context. Processing in
994  * user context is limited to non-kernel verbs users.
995  *
996  * SQ processing may get paused anytime, possibly in the middle of a WR
997  * or FPDU, if insufficient send space is available. SQ processing
998  * gets resumed from siw_sq_work_handler(), if send space becomes
999  * available again.
1000  *
1001  * Must be called with the QP state read-locked.
1002  *
1003  * Note:
1004  * An outbound RREQ can be satisfied by the corresponding RRESP
1005  * _before_ it gets assigned to the ORQ. This happens regularly
1006  * in RDMA READ via loopback case. Since both outbound RREQ and
1007  * inbound RRESP can be handled by the same CPU, locking the ORQ
1008  * is dead-lock prone and thus not an option. With that, the
1009  * RREQ gets assigned to the ORQ _before_ being sent - see
1010  * siw_activate_tx() - and pulled back in case of send failure.
1011  */
siw_qp_sq_process(struct siw_qp * qp)1012 int siw_qp_sq_process(struct siw_qp *qp)
1013 {
1014 	struct siw_wqe *wqe = tx_wqe(qp);
1015 	enum siw_opcode tx_type;
1016 	unsigned long flags;
1017 	int rv = 0;
1018 
1019 	siw_dbg_qp(qp, "enter for type %d\n", tx_type(wqe));
1020 
1021 next_wqe:
1022 	/*
1023 	 * Stop QP processing if SQ state changed
1024 	 */
1025 	if (unlikely(qp->tx_ctx.tx_suspend)) {
1026 		siw_dbg_qp(qp, "tx suspended\n");
1027 		goto done;
1028 	}
1029 	tx_type = tx_type(wqe);
1030 
1031 	if (tx_type <= SIW_OP_READ_RESPONSE)
1032 		rv = siw_qp_sq_proc_tx(qp, wqe);
1033 	else
1034 		rv = siw_qp_sq_proc_local(qp, wqe);
1035 
1036 	if (!rv) {
1037 		/*
1038 		 * WQE processing done
1039 		 */
1040 		switch (tx_type) {
1041 		case SIW_OP_SEND:
1042 		case SIW_OP_SEND_REMOTE_INV:
1043 		case SIW_OP_WRITE:
1044 			siw_wqe_put_mem(wqe, tx_type);
1045 			fallthrough;
1046 
1047 		case SIW_OP_INVAL_STAG:
1048 		case SIW_OP_REG_MR:
1049 			if (tx_flags(wqe) & SIW_WQE_SIGNALLED)
1050 				siw_sqe_complete(qp, &wqe->sqe, wqe->bytes,
1051 						 SIW_WC_SUCCESS);
1052 			break;
1053 
1054 		case SIW_OP_READ:
1055 		case SIW_OP_READ_LOCAL_INV:
1056 			/*
1057 			 * already enqueued to ORQ queue
1058 			 */
1059 			break;
1060 
1061 		case SIW_OP_READ_RESPONSE:
1062 			siw_wqe_put_mem(wqe, tx_type);
1063 			break;
1064 
1065 		default:
1066 			WARN(1, "undefined WQE type %d\n", tx_type);
1067 			rv = -EINVAL;
1068 			goto done;
1069 		}
1070 
1071 		spin_lock_irqsave(&qp->sq_lock, flags);
1072 		wqe->wr_status = SIW_WR_IDLE;
1073 		rv = siw_activate_tx(qp);
1074 		spin_unlock_irqrestore(&qp->sq_lock, flags);
1075 
1076 		if (rv <= 0)
1077 			goto done;
1078 
1079 		goto next_wqe;
1080 
1081 	} else if (rv == -EAGAIN) {
1082 		siw_dbg_qp(qp, "sq paused: hd/tr %d of %d, data %d\n",
1083 			   qp->tx_ctx.ctrl_sent, qp->tx_ctx.ctrl_len,
1084 			   qp->tx_ctx.bytes_unsent);
1085 		rv = 0;
1086 		goto done;
1087 	} else if (rv == -EINPROGRESS) {
1088 		rv = siw_sq_start(qp);
1089 		goto done;
1090 	} else {
1091 		/*
1092 		 * WQE processing failed.
1093 		 * Verbs 8.3.2:
1094 		 * o It turns any WQE into a signalled WQE.
1095 		 * o Local catastrophic error must be surfaced
1096 		 * o QP must be moved into Terminate state: done by code
1097 		 *   doing socket state change processing
1098 		 *
1099 		 * o TODO: Termination message must be sent.
1100 		 * o TODO: Implement more precise work completion errors,
1101 		 *         see enum ib_wc_status in ib_verbs.h
1102 		 */
1103 		siw_dbg_qp(qp, "wqe type %d processing failed: %d\n",
1104 			   tx_type(wqe), rv);
1105 
1106 		spin_lock_irqsave(&qp->sq_lock, flags);
1107 		/*
1108 		 * RREQ may have already been completed by inbound RRESP!
1109 		 */
1110 		if ((tx_type == SIW_OP_READ ||
1111 		     tx_type == SIW_OP_READ_LOCAL_INV) && qp->attrs.orq_size) {
1112 			/* Cleanup pending entry in ORQ */
1113 			qp->orq_put--;
1114 			qp->orq[qp->orq_put % qp->attrs.orq_size].flags = 0;
1115 		}
1116 		spin_unlock_irqrestore(&qp->sq_lock, flags);
1117 		/*
1118 		 * immediately suspends further TX processing
1119 		 */
1120 		if (!qp->tx_ctx.tx_suspend)
1121 			siw_qp_cm_drop(qp, 0);
1122 
1123 		switch (tx_type) {
1124 		case SIW_OP_SEND:
1125 		case SIW_OP_SEND_REMOTE_INV:
1126 		case SIW_OP_SEND_WITH_IMM:
1127 		case SIW_OP_WRITE:
1128 		case SIW_OP_READ:
1129 		case SIW_OP_READ_LOCAL_INV:
1130 			siw_wqe_put_mem(wqe, tx_type);
1131 			fallthrough;
1132 
1133 		case SIW_OP_INVAL_STAG:
1134 		case SIW_OP_REG_MR:
1135 			siw_sqe_complete(qp, &wqe->sqe, wqe->bytes,
1136 					 SIW_WC_LOC_QP_OP_ERR);
1137 
1138 			siw_qp_event(qp, IB_EVENT_QP_FATAL);
1139 
1140 			break;
1141 
1142 		case SIW_OP_READ_RESPONSE:
1143 			siw_dbg_qp(qp, "proc. read.response failed: %d\n", rv);
1144 
1145 			siw_qp_event(qp, IB_EVENT_QP_REQ_ERR);
1146 
1147 			siw_wqe_put_mem(wqe, SIW_OP_READ_RESPONSE);
1148 
1149 			break;
1150 
1151 		default:
1152 			WARN(1, "undefined WQE type %d\n", tx_type);
1153 			rv = -EINVAL;
1154 		}
1155 		wqe->wr_status = SIW_WR_IDLE;
1156 	}
1157 done:
1158 	return rv;
1159 }
1160 
siw_sq_resume(struct siw_qp * qp)1161 static void siw_sq_resume(struct siw_qp *qp)
1162 {
1163 	if (down_read_trylock(&qp->state_lock)) {
1164 		if (likely(qp->attrs.state == SIW_QP_STATE_RTS &&
1165 			   !qp->tx_ctx.tx_suspend)) {
1166 			int rv = siw_qp_sq_process(qp);
1167 
1168 			up_read(&qp->state_lock);
1169 
1170 			if (unlikely(rv < 0)) {
1171 				siw_dbg_qp(qp, "SQ task failed: err %d\n", rv);
1172 
1173 				if (!qp->tx_ctx.tx_suspend)
1174 					siw_qp_cm_drop(qp, 0);
1175 			}
1176 		} else {
1177 			up_read(&qp->state_lock);
1178 		}
1179 	} else {
1180 		siw_dbg_qp(qp, "Resume SQ while QP locked\n");
1181 	}
1182 	siw_qp_put(qp);
1183 }
1184 
1185 struct tx_task_t {
1186 	struct llist_head active;
1187 	wait_queue_head_t waiting;
1188 };
1189 
1190 static DEFINE_PER_CPU(struct tx_task_t, siw_tx_task_g);
1191 
siw_stop_tx_thread(int nr_cpu)1192 void siw_stop_tx_thread(int nr_cpu)
1193 {
1194 	kthread_stop(siw_tx_thread[nr_cpu]);
1195 	wake_up(&per_cpu(siw_tx_task_g, nr_cpu).waiting);
1196 }
1197 
siw_run_sq(void * data)1198 int siw_run_sq(void *data)
1199 {
1200 	const int nr_cpu = (unsigned int)(long)data;
1201 	struct llist_node *active;
1202 	struct siw_qp *qp;
1203 	struct tx_task_t *tx_task = &per_cpu(siw_tx_task_g, nr_cpu);
1204 
1205 	init_llist_head(&tx_task->active);
1206 	init_waitqueue_head(&tx_task->waiting);
1207 
1208 	while (1) {
1209 		struct llist_node *fifo_list = NULL;
1210 
1211 		wait_event_interruptible(tx_task->waiting,
1212 					 !llist_empty(&tx_task->active) ||
1213 						 kthread_should_stop());
1214 
1215 		if (kthread_should_stop())
1216 			break;
1217 
1218 		active = llist_del_all(&tx_task->active);
1219 		/*
1220 		 * llist_del_all returns a list with newest entry first.
1221 		 * Re-order list for fairness among QP's.
1222 		 */
1223 		while (active) {
1224 			struct llist_node *tmp = active;
1225 
1226 			active = llist_next(active);
1227 			tmp->next = fifo_list;
1228 			fifo_list = tmp;
1229 		}
1230 		while (fifo_list) {
1231 			qp = container_of(fifo_list, struct siw_qp, tx_list);
1232 			fifo_list = llist_next(fifo_list);
1233 			qp->tx_list.next = NULL;
1234 
1235 			siw_sq_resume(qp);
1236 		}
1237 	}
1238 	active = llist_del_all(&tx_task->active);
1239 	if (active) {
1240 		llist_for_each_entry(qp, active, tx_list) {
1241 			qp->tx_list.next = NULL;
1242 			siw_sq_resume(qp);
1243 		}
1244 	}
1245 	return 0;
1246 }
1247 
siw_sq_start(struct siw_qp * qp)1248 int siw_sq_start(struct siw_qp *qp)
1249 {
1250 	if (tx_wqe(qp)->wr_status == SIW_WR_IDLE)
1251 		return 0;
1252 
1253 	if (unlikely(!cpu_online(qp->tx_cpu))) {
1254 		siw_put_tx_cpu(qp->tx_cpu);
1255 		qp->tx_cpu = siw_get_tx_cpu(qp->sdev);
1256 		if (qp->tx_cpu < 0) {
1257 			pr_warn("siw: no tx cpu available\n");
1258 
1259 			return -EIO;
1260 		}
1261 	}
1262 	siw_qp_get(qp);
1263 
1264 	llist_add(&qp->tx_list, &per_cpu(siw_tx_task_g, qp->tx_cpu).active);
1265 
1266 	wake_up(&per_cpu(siw_tx_task_g, qp->tx_cpu).waiting);
1267 
1268 	return 0;
1269 }
1270