1 // SPDX-License-Identifier: GPL-2.0 or BSD-3-Clause
3 /* Authors: Bernard Metzler <bmt@zurich.ibm.com> */
4 /* Copyright (c) 2008-2019, IBM Corporation */
6 #include <linux/errno.h>
7 #include <linux/types.h>
9 #include <linux/scatterlist.h>
10 #include <linux/highmem.h>
13 #include <rdma/iw_cm.h>
14 #include <rdma/ib_verbs.h>
15 #include <rdma/ib_user_verbs.h>
18 #include "siw_verbs.h"
21 #define MAX_HDR_INLINE \
22 (((uint32_t)(sizeof(struct siw_rreq_pkt) - \
23 sizeof(struct iwarp_send))) & 0xF8)
25 static struct page *siw_get_pblpage(struct siw_mem *mem, u64 addr, int *idx)
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);
32 return virt_to_page(paddr);
38 * Copy short payload at provided destination payload address
40 static int siw_try_1seg(struct siw_iwarp_tx *c_tx, void *paddr)
42 struct siw_wqe *wqe = &c_tx->wqe_active;
43 struct siw_sge *sge = &wqe->sqe.sge[0];
44 u32 bytes = sge->length;
46 if (bytes > MAX_HDR_INLINE || wqe->sqe.num_sge != 1)
47 return MAX_HDR_INLINE + 1;
52 if (tx_flags(wqe) & SIW_WQE_INLINE) {
53 memcpy(paddr, &wqe->sqe.sge[1], bytes);
55 struct siw_mem *mem = wqe->mem[0];
58 /* Kernel client using kva */
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),
66 unsigned int off = sge->laddr & ~PAGE_MASK;
72 p = siw_get_upage(mem->umem, sge->laddr);
74 p = siw_get_pblpage(mem, sge->laddr, &pbl_idx);
79 buffer = kmap_local_page(p);
81 if (likely(PAGE_SIZE - off >= bytes)) {
82 memcpy(paddr, buffer + off, bytes);
84 unsigned long part = bytes - (PAGE_SIZE - off);
86 memcpy(paddr, buffer + off, part);
90 p = siw_get_upage(mem->umem,
93 p = siw_get_pblpage(mem,
99 buffer = kmap_local_page(p);
100 memcpy(paddr + part, buffer, bytes - part);
102 kunmap_local(buffer);
108 #define PKT_FRAGMENTED 1
109 #define PKT_COMPLETE 0
112 * siw_qp_prepare_tx()
114 * Prepare tx state for sending out one fpdu. Builds complete pkt
115 * if no user data or only immediate data are present.
117 * returns PKT_COMPLETE if complete pkt built, PKT_FRAGMENTED otherwise.
119 static int siw_qp_prepare_tx(struct siw_iwarp_tx *c_tx)
121 struct siw_wqe *wqe = &c_tx->wqe_active;
125 switch (tx_type(wqe)) {
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));
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);
144 c_tx->ctrl_len = sizeof(struct iwarp_rdma_rreq);
145 crc = (char *)&c_tx->pkt.rreq_pkt.crc;
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));
154 memcpy(&c_tx->pkt.ctrl, &iwarp_pktinfo[RDMAP_SEND].ctrl,
155 sizeof(struct iwarp_ctrl));
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;
162 c_tx->pkt.send_inv.inval_stag = 0;
164 c_tx->ctrl_len = sizeof(struct iwarp_send);
166 crc = (char *)&c_tx->pkt.send_pkt.crc;
167 data = siw_try_1seg(c_tx, crc);
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));
176 memcpy(&c_tx->pkt.ctrl,
177 &iwarp_pktinfo[RDMAP_SEND_INVAL].ctrl,
178 sizeof(struct iwarp_ctrl));
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;
185 c_tx->pkt.send_inv.inval_stag = cpu_to_be32(wqe->sqe.rkey);
187 c_tx->ctrl_len = sizeof(struct iwarp_send_inv);
189 crc = (char *)&c_tx->pkt.send_pkt.crc;
190 data = siw_try_1seg(c_tx, crc);
194 memcpy(&c_tx->pkt.ctrl, &iwarp_pktinfo[RDMAP_RDMA_WRITE].ctrl,
195 sizeof(struct iwarp_ctrl));
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);
201 crc = (char *)&c_tx->pkt.write_pkt.crc;
202 data = siw_try_1seg(c_tx, crc);
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));
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);
214 c_tx->ctrl_len = sizeof(struct iwarp_rdma_rresp);
216 crc = (char *)&c_tx->pkt.write_pkt.crc;
217 data = siw_try_1seg(c_tx, crc);
221 siw_dbg_qp(tx_qp(c_tx), "stale wqe type %d\n", tx_type(wqe));
224 if (unlikely(data < 0))
229 if (data <= MAX_HDR_INLINE) {
231 wqe->processed = data;
233 c_tx->pkt.ctrl.mpa_len =
234 htons(c_tx->ctrl_len + data - MPA_HDR_SIZE);
236 /* Add pad, if needed */
237 data += -(int)data & 0x3;
238 /* advance CRC location after payload */
240 c_tx->ctrl_len += data;
242 if (!(c_tx->pkt.ctrl.ddp_rdmap_ctrl & DDP_FLAG_TAGGED))
243 c_tx->pkt.c_untagged.ddp_mo = 0;
245 c_tx->pkt.c_tagged.ddp_to =
246 cpu_to_be64(wqe->sqe.raddr);
251 * Do complete CRC if enabled and short packet
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,
259 crypto_shash_final(c_tx->mpa_crc_hd, (u8 *)crc);
261 c_tx->ctrl_len += MPA_CRC_SIZE;
265 c_tx->ctrl_len += MPA_CRC_SIZE;
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,
279 if (c_tx->zcopy_tx && wqe->bytes >= SENDPAGE_THRESH &&
280 !(tx_flags(wqe) & SIW_WQE_SIGNALLED))
281 c_tx->use_sendpage = 1;
283 c_tx->use_sendpage = 0;
285 return PKT_FRAGMENTED;
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.
293 static int siw_tx_ctrl(struct siw_iwarp_tx *c_tx, struct socket *s,
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 };
301 int rv = kernel_sendmsg(s, &msg, &iov, 1,
302 c_tx->ctrl_len - c_tx->ctrl_sent);
305 c_tx->ctrl_sent += rv;
307 if (c_tx->ctrl_sent == c_tx->ctrl_len)
316 * 0copy TCP transmit interface: Use do_tcp_sendpages.
318 * Using sendpage to push page by page appears to be less efficient
319 * than using sendmsg, even if data are copied.
321 * A general performance limitation might be the extra four bytes
322 * trailer checksum segment to be pushed after user data.
324 static int siw_tcp_sendpages(struct socket *s, struct page **page, int offset,
327 struct sock *sk = s->sk;
328 int i = 0, rv = 0, sent = 0,
329 flags = MSG_MORE | MSG_DONTWAIT | MSG_SENDPAGE_NOTLAST;
332 size_t bytes = min_t(size_t, PAGE_SIZE - offset, size);
334 if (size + offset <= PAGE_SIZE)
335 flags = MSG_MORE | MSG_DONTWAIT;
337 tcp_rate_check_app_limited(sk);
340 rv = do_tcp_sendpages(sk, page[i], offset, bytes, flags);
353 if (rv == -EAGAIN || rv == 0)
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
369 static int siw_0copy_tx(struct socket *s, struct page **page,
370 struct siw_sge *sge, unsigned int offset,
373 int i = 0, sent = 0, rv;
374 int sge_bytes = min(sge->length - offset, size);
376 offset = (sge->laddr + offset) & ~PAGE_MASK;
378 while (sent != size) {
379 rv = siw_tcp_sendpages(s, &page[i], offset, sge_bytes);
382 if (size == sent || sge_bytes > rv)
385 i += PAGE_ALIGN(sge_bytes + offset) >> PAGE_SHIFT;
387 sge_bytes = min(sge->length, size - sent);
388 offset = sge->laddr & ~PAGE_MASK;
397 #define MAX_TRAILER (MPA_CRC_SIZE + 4)
399 static void siw_unmap_pages(struct kvec *iov, unsigned long kmap_mask, int len)
404 * Work backwards through the array to honor the kmap_local_page()
405 * ordering requirements.
407 for (i = (len-1); i >= 0; i--) {
408 if (kmap_mask & BIT(i)) {
409 unsigned long addr = (unsigned long)iov[i].iov_base;
411 kunmap_local((void *)(addr & PAGE_MASK));
417 * siw_tx_hdt() tries to push a complete packet to TCP where all
418 * packet fragments are referenced by the elements of one iovec.
419 * For the data portion, each involved page must be referenced by
420 * one extra element. All sge's data can be non-aligned to page
421 * boundaries. Two more elements are referencing iWARP header
423 * MAX_ARRAY = 64KB/PAGE_SIZE + 1 + (2 * (SIW_MAX_SGE - 1) + HDR + TRL
425 #define MAX_ARRAY ((0xffff / PAGE_SIZE) + 1 + (2 * (SIW_MAX_SGE - 1) + 2))
428 * Write out iov referencing hdr, data and trailer of current FPDU.
429 * Update transmit state dependent on write return status
431 static int siw_tx_hdt(struct siw_iwarp_tx *c_tx, struct socket *s)
433 struct siw_wqe *wqe = &c_tx->wqe_active;
434 struct siw_sge *sge = &wqe->sqe.sge[c_tx->sge_idx];
435 struct kvec iov[MAX_ARRAY];
436 struct page *page_array[MAX_ARRAY];
437 struct msghdr msg = { .msg_flags = MSG_DONTWAIT | MSG_EOR };
439 int seg = 0, do_crc = c_tx->do_crc, is_kva = 0, rv;
440 unsigned int data_len = c_tx->bytes_unsent, hdr_len = 0, trl_len = 0,
441 sge_off = c_tx->sge_off, sge_idx = c_tx->sge_idx,
442 pbl_idx = c_tx->pbl_idx;
443 unsigned long kmap_mask = 0L;
445 if (c_tx->state == SIW_SEND_HDR) {
446 if (c_tx->use_sendpage) {
447 rv = siw_tx_ctrl(c_tx, s, MSG_DONTWAIT | MSG_MORE);
451 c_tx->state = SIW_SEND_DATA;
454 (char *)&c_tx->pkt.ctrl + c_tx->ctrl_sent;
455 iov[0].iov_len = hdr_len =
456 c_tx->ctrl_len - c_tx->ctrl_sent;
461 wqe->processed += data_len;
463 while (data_len) { /* walk the list of SGE's */
464 unsigned int sge_len = min(sge->length - sge_off, data_len);
465 unsigned int fp_off = (sge->laddr + sge_off) & ~PAGE_MASK;
468 if (!(tx_flags(wqe) & SIW_WQE_INLINE)) {
469 mem = wqe->mem[sge_idx];
470 is_kva = mem->mem_obj == NULL ? 1 : 0;
474 if (is_kva && !c_tx->use_sendpage) {
476 * tx from kernel virtual address: either inline data
477 * or memory region with assigned kernel buffer
480 (void *)(uintptr_t)(sge->laddr + sge_off);
481 iov[seg].iov_len = sge_len;
484 crypto_shash_update(c_tx->mpa_crc_hd,
494 size_t plen = min((int)PAGE_SIZE - fp_off, sge_len);
502 mem, sge->laddr + sge_off,
505 p = siw_get_upage(mem->umem,
506 sge->laddr + sge_off);
508 siw_unmap_pages(iov, kmap_mask, seg);
509 wqe->processed -= c_tx->bytes_unsent;
515 if (!c_tx->use_sendpage) {
516 void *kaddr = kmap_local_page(p);
518 /* Remember for later kunmap() */
519 kmap_mask |= BIT(seg);
520 iov[seg].iov_base = kaddr + fp_off;
521 iov[seg].iov_len = plen;
529 kaddr = kmap_local_page(p);
530 crypto_shash_update(c_tx->mpa_crc_hd,
536 u64 va = sge->laddr + sge_off;
538 page_array[seg] = virt_to_page(va & PAGE_MASK);
542 (void *)(uintptr_t)va,
551 if (++seg > (int)MAX_ARRAY) {
552 siw_dbg_qp(tx_qp(c_tx), "to many fragments\n");
553 siw_unmap_pages(iov, kmap_mask, seg-1);
554 wqe->processed -= c_tx->bytes_unsent;
560 /* Update SGE variables at end of SGE */
561 if (sge_off == sge->length &&
562 (data_len != 0 || wqe->processed < wqe->bytes)) {
569 if (likely(c_tx->state != SIW_SEND_TRAILER)) {
570 iov[seg].iov_base = &c_tx->trailer.pad[4 - c_tx->pad];
571 iov[seg].iov_len = trl_len = MAX_TRAILER - (4 - c_tx->pad);
573 iov[seg].iov_base = &c_tx->trailer.pad[c_tx->ctrl_sent];
574 iov[seg].iov_len = trl_len = MAX_TRAILER - c_tx->ctrl_sent;
578 *(u32 *)c_tx->trailer.pad = 0;
580 crypto_shash_update(c_tx->mpa_crc_hd,
581 (u8 *)&c_tx->trailer.crc - c_tx->pad,
584 if (!c_tx->mpa_crc_hd)
585 c_tx->trailer.crc = 0;
587 crypto_shash_final(c_tx->mpa_crc_hd, (u8 *)&c_tx->trailer.crc);
589 data_len = c_tx->bytes_unsent;
591 if (c_tx->use_sendpage) {
592 rv = siw_0copy_tx(s, page_array, &wqe->sqe.sge[c_tx->sge_idx],
593 c_tx->sge_off, data_len);
594 if (rv == data_len) {
595 rv = kernel_sendmsg(s, &msg, &iov[seg], 1, trl_len);
602 rv = kernel_sendmsg(s, &msg, iov, seg + 1,
603 hdr_len + data_len + trl_len);
604 siw_unmap_pages(iov, kmap_mask, seg);
606 if (rv < (int)hdr_len) {
607 /* Not even complete hdr pushed or negative rv */
608 wqe->processed -= data_len;
610 c_tx->ctrl_sent += rv;
617 if (rv >= (int)data_len) {
618 /* all user data pushed to TCP or no data to push */
619 if (data_len > 0 && wqe->processed < wqe->bytes) {
620 /* Save the current state for next tx */
621 c_tx->sge_idx = sge_idx;
622 c_tx->sge_off = sge_off;
623 c_tx->pbl_idx = pbl_idx;
627 if (rv == trl_len) /* all pushed */
630 c_tx->state = SIW_SEND_TRAILER;
631 c_tx->ctrl_len = MAX_TRAILER;
632 c_tx->ctrl_sent = rv + 4 - c_tx->pad;
633 c_tx->bytes_unsent = 0;
637 } else if (data_len > 0) {
638 /* Maybe some user data pushed to TCP */
639 c_tx->state = SIW_SEND_DATA;
640 wqe->processed -= data_len - rv;
644 * Some bytes out. Recompute tx state based
645 * on old state and bytes pushed
647 unsigned int sge_unsent;
649 c_tx->bytes_unsent -= rv;
650 sge = &wqe->sqe.sge[c_tx->sge_idx];
651 sge_unsent = sge->length - c_tx->sge_off;
653 while (sge_unsent <= rv) {
658 sge_unsent = sge->length;
670 static void siw_update_tcpseg(struct siw_iwarp_tx *c_tx,
673 struct tcp_sock *tp = tcp_sk(s->sk);
676 if (c_tx->gso_seg_limit == 0)
677 c_tx->tcp_seglen = tp->mss_cache * tp->gso_segs;
681 min_t(u16, c_tx->gso_seg_limit, tp->gso_segs);
683 c_tx->tcp_seglen = tp->mss_cache;
685 /* Loopback may give odd numbers */
686 c_tx->tcp_seglen &= 0xfffffff8;
692 * Prepares transmit context to send out one FPDU if FPDU will contain
693 * user data and user data are not immediate data.
694 * Computes maximum FPDU length to fill up TCP MSS if possible.
696 * @qp: QP from which to transmit
697 * @wqe: Current WQE causing transmission
699 * TODO: Take into account real available sendspace on socket
700 * to avoid header misalignment due to send pausing within
703 static void siw_prepare_fpdu(struct siw_qp *qp, struct siw_wqe *wqe)
705 struct siw_iwarp_tx *c_tx = &qp->tx_ctx;
709 iwarp_pktinfo[__rdmap_get_opcode(&c_tx->pkt.ctrl)].hdr_len;
713 * Update target buffer offset if any
715 if (!(c_tx->pkt.ctrl.ddp_rdmap_ctrl & DDP_FLAG_TAGGED))
716 /* Untagged message */
717 c_tx->pkt.c_untagged.ddp_mo = cpu_to_be32(wqe->processed);
718 else /* Tagged message */
719 c_tx->pkt.c_tagged.ddp_to =
720 cpu_to_be64(wqe->sqe.raddr + wqe->processed);
722 data_len = wqe->bytes - wqe->processed;
723 if (data_len + c_tx->ctrl_len + MPA_CRC_SIZE > c_tx->tcp_seglen) {
724 /* Trim DDP payload to fit into current TCP segment */
725 data_len = c_tx->tcp_seglen - (c_tx->ctrl_len + MPA_CRC_SIZE);
726 c_tx->pkt.ctrl.ddp_rdmap_ctrl &= ~DDP_FLAG_LAST;
729 c_tx->pkt.ctrl.ddp_rdmap_ctrl |= DDP_FLAG_LAST;
730 c_tx->pad = -data_len & 0x3;
732 c_tx->bytes_unsent = data_len;
734 c_tx->pkt.ctrl.mpa_len =
735 htons(c_tx->ctrl_len + data_len - MPA_HDR_SIZE);
738 * Init MPA CRC computation
740 if (c_tx->mpa_crc_hd) {
741 crypto_shash_init(c_tx->mpa_crc_hd);
742 crypto_shash_update(c_tx->mpa_crc_hd, (u8 *)&c_tx->pkt,
751 * Check permissions for a list of SGE's (SGL).
752 * A successful check will have all memory referenced
753 * for transmission resolved and assigned to the WQE.
755 * @pd: Protection Domain SGL should belong to
756 * @wqe: WQE to be checked
757 * @perms: requested access permissions
761 static int siw_check_sgl_tx(struct ib_pd *pd, struct siw_wqe *wqe,
762 enum ib_access_flags perms)
764 struct siw_sge *sge = &wqe->sqe.sge[0];
765 int i, len, num_sge = wqe->sqe.num_sge;
767 if (unlikely(num_sge > SIW_MAX_SGE))
770 for (i = 0, len = 0; num_sge; num_sge--, i++, sge++) {
772 * rdma verbs: do not check stag for a zero length sge
775 int rv = siw_check_sge(pd, sge, &wqe->mem[i], perms, 0,
778 if (unlikely(rv != E_ACCESS_OK))
787 * siw_qp_sq_proc_tx()
789 * Process one WQE which needs transmission on the wire.
791 static int siw_qp_sq_proc_tx(struct siw_qp *qp, struct siw_wqe *wqe)
793 struct siw_iwarp_tx *c_tx = &qp->tx_ctx;
794 struct socket *s = qp->attrs.sk;
795 int rv = 0, burst_len = qp->tx_ctx.burst;
796 enum rdmap_ecode ecode = RDMAP_ECODE_CATASTROPHIC_STREAM;
798 if (unlikely(wqe->wr_status == SIW_WR_IDLE))
802 burst_len = SQ_USER_MAXBURST;
804 if (wqe->wr_status == SIW_WR_QUEUED) {
805 if (!(wqe->sqe.flags & SIW_WQE_INLINE)) {
806 if (tx_type(wqe) == SIW_OP_READ_RESPONSE)
807 wqe->sqe.num_sge = 1;
809 if (tx_type(wqe) != SIW_OP_READ &&
810 tx_type(wqe) != SIW_OP_READ_LOCAL_INV) {
812 * Reference memory to be tx'd w/o checking
813 * access for LOCAL_READ permission, since
814 * not defined in RDMA core.
816 rv = siw_check_sgl_tx(qp->pd, wqe, 0);
819 SIW_OP_READ_RESPONSE)
820 ecode = siw_rdmap_error(-rv);
829 wqe->bytes = wqe->sqe.sge[0].length;
830 if (!rdma_is_kernel_res(&qp->base_qp.res)) {
831 if (wqe->bytes > SIW_MAX_INLINE) {
835 wqe->sqe.sge[0].laddr =
836 (u64)(uintptr_t)&wqe->sqe.sge[1];
839 wqe->wr_status = SIW_WR_INPROGRESS;
842 siw_update_tcpseg(c_tx, s);
844 rv = siw_qp_prepare_tx(c_tx);
845 if (rv == PKT_FRAGMENTED) {
846 c_tx->state = SIW_SEND_HDR;
847 siw_prepare_fpdu(qp, wqe);
848 } else if (rv == PKT_COMPLETE) {
849 c_tx->state = SIW_SEND_SHORT_FPDU;
856 siw_dbg_qp(qp, "wr type %d, state %d, data %u, sent %u, id %llx\n",
857 tx_type(wqe), wqe->wr_status, wqe->bytes, wqe->processed,
860 if (--burst_len == 0) {
864 if (c_tx->state == SIW_SEND_SHORT_FPDU) {
865 enum siw_opcode tx_type = tx_type(wqe);
866 unsigned int msg_flags;
868 if (siw_sq_empty(qp) || !siw_tcp_nagle || burst_len == 1)
870 * End current TCP segment, if SQ runs empty,
871 * or siw_tcp_nagle is not set, or we bail out
872 * soon due to no burst credit left.
874 msg_flags = MSG_DONTWAIT;
876 msg_flags = MSG_DONTWAIT | MSG_MORE;
878 rv = siw_tx_ctrl(c_tx, s, msg_flags);
880 if (!rv && tx_type != SIW_OP_READ &&
881 tx_type != SIW_OP_READ_LOCAL_INV)
882 wqe->processed = wqe->bytes;
887 rv = siw_tx_hdt(c_tx, s);
891 * One segment sent. Processing completed if last
892 * segment, Do next segment otherwise.
894 if (unlikely(c_tx->tx_suspend)) {
896 * Verbs, 6.4.: Try stopping sending after a full
897 * DDP segment if the connection goes down
898 * (== peer halfclose)
903 if (c_tx->pkt.ctrl.ddp_rdmap_ctrl & DDP_FLAG_LAST) {
904 siw_dbg_qp(qp, "WQE completed\n");
907 c_tx->state = SIW_SEND_HDR;
909 siw_update_tcpseg(c_tx, s);
911 siw_prepare_fpdu(qp, wqe);
915 qp->tx_ctx.burst = burst_len;
919 if (ecode != RDMAP_ECODE_CATASTROPHIC_STREAM)
920 siw_init_terminate(qp, TERM_ERROR_LAYER_RDMAP,
921 RDMAP_ETYPE_REMOTE_PROTECTION, ecode, 1);
923 siw_init_terminate(qp, TERM_ERROR_LAYER_RDMAP,
924 RDMAP_ETYPE_CATASTROPHIC,
925 RDMAP_ECODE_UNSPECIFIED, 1);
929 static int siw_fastreg_mr(struct ib_pd *pd, struct siw_sqe *sqe)
931 struct ib_mr *base_mr = (struct ib_mr *)(uintptr_t)sqe->base_mr;
932 struct siw_device *sdev = to_siw_dev(pd->device);
936 siw_dbg_pd(pd, "STag 0x%08x\n", sqe->rkey);
938 if (unlikely(!base_mr)) {
939 pr_warn("siw: fastreg: STag 0x%08x unknown\n", sqe->rkey);
943 if (unlikely(base_mr->rkey >> 8 != sqe->rkey >> 8)) {
944 pr_warn("siw: fastreg: STag 0x%08x: bad MR\n", sqe->rkey);
948 mem = siw_mem_id2obj(sdev, sqe->rkey >> 8);
949 if (unlikely(!mem)) {
950 pr_warn("siw: fastreg: STag 0x%08x unknown\n", sqe->rkey);
954 if (unlikely(mem->pd != pd)) {
955 pr_warn("siw: fastreg: PD mismatch\n");
959 if (unlikely(mem->stag_valid)) {
960 pr_warn("siw: fastreg: STag 0x%08x already valid\n", sqe->rkey);
964 /* Refresh STag since user may have changed key part */
965 mem->stag = sqe->rkey;
966 mem->perms = sqe->access;
968 siw_dbg_mem(mem, "STag 0x%08x now valid\n", sqe->rkey);
969 mem->va = base_mr->iova;
976 static int siw_qp_sq_proc_local(struct siw_qp *qp, struct siw_wqe *wqe)
980 switch (tx_type(wqe)) {
982 rv = siw_fastreg_mr(qp->pd, &wqe->sqe);
985 case SIW_OP_INVAL_STAG:
986 rv = siw_invalidate_stag(qp->pd, wqe->sqe.rkey);
996 * siw_qp_sq_process()
998 * Core TX path routine for RDMAP/DDP/MPA using a TCP kernel socket.
999 * Sends RDMAP payload for the current SQ WR @wqe of @qp in one or more
1000 * MPA FPDUs, each containing a DDP segment.
1002 * SQ processing may occur in user context as a result of posting
1003 * new WQE's or from siw_sq_work_handler() context. Processing in
1004 * user context is limited to non-kernel verbs users.
1006 * SQ processing may get paused anytime, possibly in the middle of a WR
1007 * or FPDU, if insufficient send space is available. SQ processing
1008 * gets resumed from siw_sq_work_handler(), if send space becomes
1011 * Must be called with the QP state read-locked.
1014 * An outbound RREQ can be satisfied by the corresponding RRESP
1015 * _before_ it gets assigned to the ORQ. This happens regularly
1016 * in RDMA READ via loopback case. Since both outbound RREQ and
1017 * inbound RRESP can be handled by the same CPU, locking the ORQ
1018 * is dead-lock prone and thus not an option. With that, the
1019 * RREQ gets assigned to the ORQ _before_ being sent - see
1020 * siw_activate_tx() - and pulled back in case of send failure.
1022 int siw_qp_sq_process(struct siw_qp *qp)
1024 struct siw_wqe *wqe = tx_wqe(qp);
1025 enum siw_opcode tx_type;
1026 unsigned long flags;
1029 siw_dbg_qp(qp, "enter for type %d\n", tx_type(wqe));
1033 * Stop QP processing if SQ state changed
1035 if (unlikely(qp->tx_ctx.tx_suspend)) {
1036 siw_dbg_qp(qp, "tx suspended\n");
1039 tx_type = tx_type(wqe);
1041 if (tx_type <= SIW_OP_READ_RESPONSE)
1042 rv = siw_qp_sq_proc_tx(qp, wqe);
1044 rv = siw_qp_sq_proc_local(qp, wqe);
1048 * WQE processing done
1052 case SIW_OP_SEND_REMOTE_INV:
1054 siw_wqe_put_mem(wqe, tx_type);
1057 case SIW_OP_INVAL_STAG:
1059 if (tx_flags(wqe) & SIW_WQE_SIGNALLED)
1060 siw_sqe_complete(qp, &wqe->sqe, wqe->bytes,
1065 case SIW_OP_READ_LOCAL_INV:
1067 * already enqueued to ORQ queue
1071 case SIW_OP_READ_RESPONSE:
1072 siw_wqe_put_mem(wqe, tx_type);
1076 WARN(1, "undefined WQE type %d\n", tx_type);
1081 spin_lock_irqsave(&qp->sq_lock, flags);
1082 wqe->wr_status = SIW_WR_IDLE;
1083 rv = siw_activate_tx(qp);
1084 spin_unlock_irqrestore(&qp->sq_lock, flags);
1091 } else if (rv == -EAGAIN) {
1092 siw_dbg_qp(qp, "sq paused: hd/tr %d of %d, data %d\n",
1093 qp->tx_ctx.ctrl_sent, qp->tx_ctx.ctrl_len,
1094 qp->tx_ctx.bytes_unsent);
1097 } else if (rv == -EINPROGRESS) {
1098 rv = siw_sq_start(qp);
1102 * WQE processing failed.
1104 * o It turns any WQE into a signalled WQE.
1105 * o Local catastrophic error must be surfaced
1106 * o QP must be moved into Terminate state: done by code
1107 * doing socket state change processing
1109 * o TODO: Termination message must be sent.
1110 * o TODO: Implement more precise work completion errors,
1111 * see enum ib_wc_status in ib_verbs.h
1113 siw_dbg_qp(qp, "wqe type %d processing failed: %d\n",
1116 spin_lock_irqsave(&qp->sq_lock, flags);
1118 * RREQ may have already been completed by inbound RRESP!
1120 if ((tx_type == SIW_OP_READ ||
1121 tx_type == SIW_OP_READ_LOCAL_INV) && qp->attrs.orq_size) {
1122 /* Cleanup pending entry in ORQ */
1124 qp->orq[qp->orq_put % qp->attrs.orq_size].flags = 0;
1126 spin_unlock_irqrestore(&qp->sq_lock, flags);
1128 * immediately suspends further TX processing
1130 if (!qp->tx_ctx.tx_suspend)
1131 siw_qp_cm_drop(qp, 0);
1135 case SIW_OP_SEND_REMOTE_INV:
1136 case SIW_OP_SEND_WITH_IMM:
1139 case SIW_OP_READ_LOCAL_INV:
1140 siw_wqe_put_mem(wqe, tx_type);
1143 case SIW_OP_INVAL_STAG:
1145 siw_sqe_complete(qp, &wqe->sqe, wqe->bytes,
1146 SIW_WC_LOC_QP_OP_ERR);
1148 siw_qp_event(qp, IB_EVENT_QP_FATAL);
1152 case SIW_OP_READ_RESPONSE:
1153 siw_dbg_qp(qp, "proc. read.response failed: %d\n", rv);
1155 siw_qp_event(qp, IB_EVENT_QP_REQ_ERR);
1157 siw_wqe_put_mem(wqe, SIW_OP_READ_RESPONSE);
1162 WARN(1, "undefined WQE type %d\n", tx_type);
1165 wqe->wr_status = SIW_WR_IDLE;
1171 static void siw_sq_resume(struct siw_qp *qp)
1173 if (down_read_trylock(&qp->state_lock)) {
1174 if (likely(qp->attrs.state == SIW_QP_STATE_RTS &&
1175 !qp->tx_ctx.tx_suspend)) {
1176 int rv = siw_qp_sq_process(qp);
1178 up_read(&qp->state_lock);
1180 if (unlikely(rv < 0)) {
1181 siw_dbg_qp(qp, "SQ task failed: err %d\n", rv);
1183 if (!qp->tx_ctx.tx_suspend)
1184 siw_qp_cm_drop(qp, 0);
1187 up_read(&qp->state_lock);
1190 siw_dbg_qp(qp, "Resume SQ while QP locked\n");
1196 struct llist_head active;
1197 wait_queue_head_t waiting;
1200 static DEFINE_PER_CPU(struct tx_task_t, siw_tx_task_g);
1202 void siw_stop_tx_thread(int nr_cpu)
1204 kthread_stop(siw_tx_thread[nr_cpu]);
1205 wake_up(&per_cpu(siw_tx_task_g, nr_cpu).waiting);
1208 int siw_run_sq(void *data)
1210 const int nr_cpu = (unsigned int)(long)data;
1211 struct llist_node *active;
1213 struct tx_task_t *tx_task = &per_cpu(siw_tx_task_g, nr_cpu);
1215 init_llist_head(&tx_task->active);
1216 init_waitqueue_head(&tx_task->waiting);
1219 struct llist_node *fifo_list = NULL;
1221 wait_event_interruptible(tx_task->waiting,
1222 !llist_empty(&tx_task->active) ||
1223 kthread_should_stop());
1225 if (kthread_should_stop())
1228 active = llist_del_all(&tx_task->active);
1230 * llist_del_all returns a list with newest entry first.
1231 * Re-order list for fairness among QP's.
1234 struct llist_node *tmp = active;
1236 active = llist_next(active);
1237 tmp->next = fifo_list;
1241 qp = container_of(fifo_list, struct siw_qp, tx_list);
1242 fifo_list = llist_next(fifo_list);
1243 qp->tx_list.next = NULL;
1248 active = llist_del_all(&tx_task->active);
1250 llist_for_each_entry(qp, active, tx_list) {
1251 qp->tx_list.next = NULL;
1258 int siw_sq_start(struct siw_qp *qp)
1260 if (tx_wqe(qp)->wr_status == SIW_WR_IDLE)
1263 if (unlikely(!cpu_online(qp->tx_cpu))) {
1264 siw_put_tx_cpu(qp->tx_cpu);
1265 qp->tx_cpu = siw_get_tx_cpu(qp->sdev);
1266 if (qp->tx_cpu < 0) {
1267 pr_warn("siw: no tx cpu available\n");
1274 llist_add(&qp->tx_list, &per_cpu(siw_tx_task_g, qp->tx_cpu).active);
1276 wake_up(&per_cpu(siw_tx_task_g, qp->tx_cpu).waiting);