2 * Copyright (c) 2012 Chelsio Communications, Inc.
4 * Written by: Navdeep Parhar <np@FreeBSD.org>
6 * Redistribution and use in source and binary forms, with or without
7 * modification, are permitted provided that the following conditions
9 * 1. Redistributions of source code must retain the above copyright
10 * notice, this list of conditions and the following disclaimer.
11 * 2. Redistributions in binary form must reproduce the above copyright
12 * notice, this list of conditions and the following disclaimer in the
13 * documentation and/or other materials provided with the distribution.
15 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
16 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
19 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
28 #include <sys/cdefs.h>
29 __FBSDID("$FreeBSD$");
33 #include <sys/param.h>
34 #include <sys/types.h>
35 #include <sys/systm.h>
36 #include <sys/kernel.h>
38 #include <sys/module.h>
39 #include <sys/protosw.h>
41 #include <sys/domain.h>
42 #include <sys/socket.h>
43 #include <sys/socketvar.h>
45 #include <netinet/in.h>
46 #include <netinet/in_pcb.h>
47 #include <netinet/ip.h>
48 #include <netinet/tcp_var.h>
50 #include <netinet/tcp_fsm.h>
51 #include <netinet/toecore.h>
54 #include <vm/vm_extern.h>
55 #include <vm/vm_param.h>
57 #include <vm/vm_map.h>
58 #include <vm/vm_page.h>
59 #include <vm/vm_object.h>
62 #include "common/common.h"
63 #include "common/t4_msg.h"
64 #include "common/t4_regs.h"
65 #include "common/t4_tcb.h"
66 #include "tom/t4_tom.h"
68 #define PPOD_SZ(n) ((n) * sizeof(struct pagepod))
69 #define PPOD_SIZE (PPOD_SZ(1))
71 /* XXX: must match A_ULP_RX_TDDP_PSZ */
72 static int t4_ddp_pgsz[] = {4096, 4096 << 2, 4096 << 4, 4096 << 6};
76 t4_dump_tcb(struct adapter *sc, int tid)
78 uint32_t tcb_base, off, i, j;
80 /* Dump TCB for the tid */
81 tcb_base = t4_read_reg(sc, A_TP_CMM_TCB_BASE);
82 t4_write_reg(sc, PCIE_MEM_ACCESS_REG(A_PCIE_MEM_ACCESS_OFFSET, 2),
83 tcb_base + tid * TCB_SIZE);
84 t4_read_reg(sc, PCIE_MEM_ACCESS_REG(A_PCIE_MEM_ACCESS_OFFSET, 2));
87 for (i = 0; i < 4; i++) {
89 for (j = 0; j < 8; j++, off += 4)
90 buf[j] = htonl(t4_read_reg(sc, MEMWIN2_BASE + off));
92 printf("%08x %08x %08x %08x %08x %08x %08x %08x\n",
93 buf[0], buf[1], buf[2], buf[3], buf[4], buf[5], buf[6],
99 #define MAX_DDP_BUFFER_SIZE (M_TCB_RX_DDP_BUF0_LEN)
101 alloc_ppods(struct tom_data *td, int n)
107 if (vmem_alloc(td->ppod_arena, n, M_NOWAIT | M_FIRSTFIT, &ppod) != 0)
113 free_ppods(struct tom_data *td, int ppod, int n)
118 vmem_free(td->ppod_arena, ppod, n);
122 pages_to_nppods(int npages, int ddp_pgsz)
124 int nsegs = npages * PAGE_SIZE / ddp_pgsz;
126 return (howmany(nsegs, PPOD_PAGES));
130 free_ddp_buffer(struct tom_data *td, struct ddp_buffer *db)
137 free(db->pages, M_CXGBE);
140 free_ppods(td, G_PPOD_TAG(db->tag), db->nppods);
146 release_ddp_resources(struct toepcb *toep)
150 for (i = 0; i < nitems(toep->db); i++) {
151 if (toep->db[i] != NULL) {
152 free_ddp_buffer(toep->td, toep->db[i]);
158 /* XXX: handle_ddp_data code duplication */
160 insert_ddp_data(struct toepcb *toep, uint32_t n)
162 struct inpcb *inp = toep->inp;
163 struct tcpcb *tp = intotcpcb(inp);
164 struct sockbuf *sb = &inp->inp_socket->so_rcv;
167 INP_WLOCK_ASSERT(inp);
168 SOCKBUF_LOCK_ASSERT(sb);
172 #ifndef USE_DDP_RX_FLOW_CONTROL
173 KASSERT(tp->rcv_wnd >= n, ("%s: negative window size", __func__));
177 KASSERT(toep->sb_cc >= sbused(sb),
178 ("%s: sb %p has more data (%d) than last time (%d).",
179 __func__, sb, sbused(sb), toep->sb_cc));
180 toep->rx_credits += toep->sb_cc - sbused(sb);
181 #ifdef USE_DDP_RX_FLOW_CONTROL
182 toep->rx_credits -= n; /* adjust for F_RX_FC_DDP */
184 sbappendstream_locked(sb, m, 0);
185 toep->sb_cc = sbused(sb);
188 /* SET_TCB_FIELD sent as a ULP command looks like this */
189 #define LEN__SET_TCB_FIELD_ULP (sizeof(struct ulp_txpkt) + \
190 sizeof(struct ulptx_idata) + sizeof(struct cpl_set_tcb_field_core))
192 /* RX_DATA_ACK sent as a ULP command looks like this */
193 #define LEN__RX_DATA_ACK_ULP (sizeof(struct ulp_txpkt) + \
194 sizeof(struct ulptx_idata) + sizeof(struct cpl_rx_data_ack_core))
197 mk_set_tcb_field_ulp(struct ulp_txpkt *ulpmc, struct toepcb *toep,
198 uint64_t word, uint64_t mask, uint64_t val)
200 struct ulptx_idata *ulpsc;
201 struct cpl_set_tcb_field_core *req;
203 ulpmc->cmd_dest = htonl(V_ULPTX_CMD(ULP_TX_PKT) | V_ULP_TXPKT_DEST(0));
204 ulpmc->len = htobe32(howmany(LEN__SET_TCB_FIELD_ULP, 16));
206 ulpsc = (struct ulptx_idata *)(ulpmc + 1);
207 ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM));
208 ulpsc->len = htobe32(sizeof(*req));
210 req = (struct cpl_set_tcb_field_core *)(ulpsc + 1);
211 OPCODE_TID(req) = htobe32(MK_OPCODE_TID(CPL_SET_TCB_FIELD, toep->tid));
212 req->reply_ctrl = htobe16(V_NO_REPLY(1) |
213 V_QUEUENO(toep->ofld_rxq->iq.abs_id));
214 req->word_cookie = htobe16(V_WORD(word) | V_COOKIE(0));
215 req->mask = htobe64(mask);
216 req->val = htobe64(val);
218 ulpsc = (struct ulptx_idata *)(req + 1);
219 if (LEN__SET_TCB_FIELD_ULP % 16) {
220 ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_NOOP));
221 ulpsc->len = htobe32(0);
228 mk_rx_data_ack_ulp(struct ulp_txpkt *ulpmc, struct toepcb *toep)
230 struct ulptx_idata *ulpsc;
231 struct cpl_rx_data_ack_core *req;
233 ulpmc->cmd_dest = htonl(V_ULPTX_CMD(ULP_TX_PKT) | V_ULP_TXPKT_DEST(0));
234 ulpmc->len = htobe32(howmany(LEN__RX_DATA_ACK_ULP, 16));
236 ulpsc = (struct ulptx_idata *)(ulpmc + 1);
237 ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM));
238 ulpsc->len = htobe32(sizeof(*req));
240 req = (struct cpl_rx_data_ack_core *)(ulpsc + 1);
241 OPCODE_TID(req) = htobe32(MK_OPCODE_TID(CPL_RX_DATA_ACK, toep->tid));
242 req->credit_dack = htobe32(F_RX_MODULATE_RX);
244 ulpsc = (struct ulptx_idata *)(req + 1);
245 if (LEN__RX_DATA_ACK_ULP % 16) {
246 ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_NOOP));
247 ulpsc->len = htobe32(0);
253 static inline uint64_t
254 select_ddp_flags(struct socket *so, int flags, int db_idx)
256 uint64_t ddp_flags = V_TF_DDP_INDICATE_OUT(0);
257 int waitall = flags & MSG_WAITALL;
258 int nb = so->so_state & SS_NBIO || flags & (MSG_DONTWAIT | MSG_NBIO);
260 KASSERT(db_idx == 0 || db_idx == 1,
261 ("%s: bad DDP buffer index %d", __func__, db_idx));
264 ddp_flags |= V_TF_DDP_BUF0_VALID(1) | V_TF_DDP_ACTIVE_BUF(0);
266 ddp_flags |= V_TF_DDP_PUSH_DISABLE_0(1);
268 ddp_flags |= V_TF_DDP_BUF0_FLUSH(1);
270 ddp_flags |= V_TF_DDP_BUF0_FLUSH(0);
272 ddp_flags |= V_TF_DDP_BUF1_VALID(1) | V_TF_DDP_ACTIVE_BUF(1);
274 ddp_flags |= V_TF_DDP_PUSH_DISABLE_1(1);
276 ddp_flags |= V_TF_DDP_BUF1_FLUSH(1);
278 ddp_flags |= V_TF_DDP_BUF1_FLUSH(0);
285 mk_update_tcb_for_ddp(struct adapter *sc, struct toepcb *toep, int db_idx,
286 int offset, uint64_t ddp_flags)
288 struct ddp_buffer *db = toep->db[db_idx];
290 struct work_request_hdr *wrh;
291 struct ulp_txpkt *ulpmc;
294 KASSERT(db_idx == 0 || db_idx == 1,
295 ("%s: bad DDP buffer index %d", __func__, db_idx));
298 * We'll send a compound work request that has 3 SET_TCB_FIELDs and an
299 * RX_DATA_ACK (with RX_MODULATE to speed up delivery).
301 * The work request header is 16B and always ends at a 16B boundary.
302 * The ULPTX master commands that follow must all end at 16B boundaries
303 * too so we round up the size to 16.
305 len = sizeof(*wrh) + 3 * roundup2(LEN__SET_TCB_FIELD_ULP, 16) +
306 roundup2(LEN__RX_DATA_ACK_ULP, 16);
308 wr = alloc_wrqe(len, toep->ctrlq);
312 INIT_ULPTX_WRH(wrh, len, 1, 0); /* atomic */
313 ulpmc = (struct ulp_txpkt *)(wrh + 1);
315 /* Write the buffer's tag */
316 ulpmc = mk_set_tcb_field_ulp(ulpmc, toep,
317 W_TCB_RX_DDP_BUF0_TAG + db_idx,
318 V_TCB_RX_DDP_BUF0_TAG(M_TCB_RX_DDP_BUF0_TAG),
319 V_TCB_RX_DDP_BUF0_TAG(db->tag));
321 /* Update the current offset in the DDP buffer and its total length */
323 ulpmc = mk_set_tcb_field_ulp(ulpmc, toep,
324 W_TCB_RX_DDP_BUF0_OFFSET,
325 V_TCB_RX_DDP_BUF0_OFFSET(M_TCB_RX_DDP_BUF0_OFFSET) |
326 V_TCB_RX_DDP_BUF0_LEN(M_TCB_RX_DDP_BUF0_LEN),
327 V_TCB_RX_DDP_BUF0_OFFSET(offset) |
328 V_TCB_RX_DDP_BUF0_LEN(db->len));
330 ulpmc = mk_set_tcb_field_ulp(ulpmc, toep,
331 W_TCB_RX_DDP_BUF1_OFFSET,
332 V_TCB_RX_DDP_BUF1_OFFSET(M_TCB_RX_DDP_BUF1_OFFSET) |
333 V_TCB_RX_DDP_BUF1_LEN((u64)M_TCB_RX_DDP_BUF1_LEN << 32),
334 V_TCB_RX_DDP_BUF1_OFFSET(offset) |
335 V_TCB_RX_DDP_BUF1_LEN((u64)db->len << 32));
337 /* Update DDP flags */
338 ulpmc = mk_set_tcb_field_ulp(ulpmc, toep, W_TCB_RX_DDP_FLAGS,
339 V_TF_DDP_BUF0_FLUSH(1) | V_TF_DDP_BUF1_FLUSH(1) |
340 V_TF_DDP_PUSH_DISABLE_0(1) | V_TF_DDP_PUSH_DISABLE_1(1) |
341 V_TF_DDP_BUF0_VALID(1) | V_TF_DDP_BUF1_VALID(1) |
342 V_TF_DDP_ACTIVE_BUF(1) | V_TF_DDP_INDICATE_OUT(1), ddp_flags);
344 /* Gratuitous RX_DATA_ACK with RX_MODULATE set to speed up delivery. */
345 ulpmc = mk_rx_data_ack_ulp(ulpmc, toep);
351 discourage_ddp(struct toepcb *toep)
354 if (toep->ddp_score && --toep->ddp_score == 0) {
355 toep->ddp_flags &= ~DDP_OK;
356 toep->ddp_disabled = time_uptime;
357 CTR3(KTR_CXGBE, "%s: tid %u !DDP_OK @ %u",
358 __func__, toep->tid, time_uptime);
363 handle_ddp_data(struct toepcb *toep, __be32 ddp_report, __be32 rcv_nxt, int len)
365 uint32_t report = be32toh(ddp_report);
366 unsigned int db_flag;
367 struct inpcb *inp = toep->inp;
373 db_flag = report & F_DDP_BUF_IDX ? DDP_BUF1_ACTIVE : DDP_BUF0_ACTIVE;
375 if (__predict_false(!(report & F_DDP_INV)))
376 CXGBE_UNIMPLEMENTED("DDP buffer still valid");
379 so = inp_inpcbtosocket(inp);
381 if (__predict_false(inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT))) {
384 * XXX: think a bit more.
385 * tcpcb probably gone, but socket should still be around
386 * because we always wait for DDP completion in soreceive no
387 * matter what. Just wake it up and let it clean up.
390 CTR5(KTR_CXGBE, "%s: tid %u, seq 0x%x, len %d, inp_flags 0x%x",
391 __func__, toep->tid, be32toh(rcv_nxt), len, inp->inp_flags);
397 len += be32toh(rcv_nxt) - tp->rcv_nxt;
399 tp->t_rcvtime = ticks;
400 #ifndef USE_DDP_RX_FLOW_CONTROL
401 KASSERT(tp->rcv_wnd >= len, ("%s: negative window size", __func__));
404 m = get_ddp_mbuf(len);
407 if (report & F_DDP_BUF_COMPLETE)
408 toep->ddp_score = DDP_HIGH_SCORE;
410 discourage_ddp(toep);
412 KASSERT(toep->sb_cc >= sbused(sb),
413 ("%s: sb %p has more data (%d) than last time (%d).",
414 __func__, sb, sbused(sb), toep->sb_cc));
415 toep->rx_credits += toep->sb_cc - sbused(sb);
416 #ifdef USE_DDP_RX_FLOW_CONTROL
417 toep->rx_credits -= len; /* adjust for F_RX_FC_DDP */
419 sbappendstream_locked(sb, m, 0);
420 toep->sb_cc = sbused(sb);
422 KASSERT(toep->ddp_flags & db_flag,
423 ("%s: DDP buffer not active. toep %p, ddp_flags 0x%x, report 0x%x",
424 __func__, toep, toep->ddp_flags, report));
425 toep->ddp_flags &= ~db_flag;
426 sorwakeup_locked(so);
427 SOCKBUF_UNLOCK_ASSERT(sb);
433 #define DDP_ERR (F_DDP_PPOD_MISMATCH | F_DDP_LLIMIT_ERR | F_DDP_ULIMIT_ERR |\
434 F_DDP_PPOD_PARITY_ERR | F_DDP_PADDING_ERR | F_DDP_OFFSET_ERR |\
435 F_DDP_INVALID_TAG | F_DDP_COLOR_ERR | F_DDP_TID_MISMATCH |\
436 F_DDP_INVALID_PPOD | F_DDP_HDRCRC_ERR | F_DDP_DATACRC_ERR)
439 do_rx_data_ddp(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
441 struct adapter *sc = iq->adapter;
442 const struct cpl_rx_data_ddp *cpl = (const void *)(rss + 1);
443 unsigned int tid = GET_TID(cpl);
445 struct toepcb *toep = lookup_tid(sc, tid);
446 struct tom_data *td = toep->td;
448 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
449 KASSERT(toep->tid == tid, ("%s: toep tid/atid mismatch", __func__));
450 KASSERT(!(toep->flags & TPF_SYNQE),
451 ("%s: toep %p claims to be a synq entry", __func__, toep));
453 vld = be32toh(cpl->ddpvld);
454 if (__predict_false(vld & DDP_ERR)) {
455 panic("%s: DDP error 0x%x (tid %d, toep %p)",
456 __func__, vld, tid, toep);
458 if (toep->ulp_mode == ULP_MODE_ISCSI) {
459 m = m_get(M_NOWAIT, MT_DATA);
461 CXGBE_UNIMPLEMENTED("mbuf alloc failure");
462 memcpy(mtod(m, unsigned char *), cpl,
463 sizeof(struct cpl_rx_data_ddp));
464 if (!t4_cpl_iscsi_callback(td, toep, m, CPL_RX_DATA_DDP))
469 handle_ddp_data(toep, cpl->u.ddp_report, cpl->seq, be16toh(cpl->len));
475 do_rx_ddp_complete(struct sge_iq *iq, const struct rss_header *rss,
478 struct adapter *sc = iq->adapter;
479 const struct cpl_rx_ddp_complete *cpl = (const void *)(rss + 1);
480 unsigned int tid = GET_TID(cpl);
481 struct toepcb *toep = lookup_tid(sc, tid);
483 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
484 KASSERT(toep->tid == tid, ("%s: toep tid/atid mismatch", __func__));
485 KASSERT(!(toep->flags & TPF_SYNQE),
486 ("%s: toep %p claims to be a synq entry", __func__, toep));
488 handle_ddp_data(toep, cpl->ddp_report, cpl->rcv_nxt, 0);
494 enable_ddp(struct adapter *sc, struct toepcb *toep)
497 KASSERT((toep->ddp_flags & (DDP_ON | DDP_OK | DDP_SC_REQ)) == DDP_OK,
498 ("%s: toep %p has bad ddp_flags 0x%x",
499 __func__, toep, toep->ddp_flags));
501 CTR3(KTR_CXGBE, "%s: tid %u (time %u)",
502 __func__, toep->tid, time_uptime);
504 toep->ddp_flags |= DDP_SC_REQ;
505 t4_set_tcb_field(sc, toep, 1, W_TCB_RX_DDP_FLAGS,
506 V_TF_DDP_OFF(1) | V_TF_DDP_INDICATE_OUT(1) |
507 V_TF_DDP_BUF0_INDICATE(1) | V_TF_DDP_BUF1_INDICATE(1) |
508 V_TF_DDP_BUF0_VALID(1) | V_TF_DDP_BUF1_VALID(1),
509 V_TF_DDP_BUF0_INDICATE(1) | V_TF_DDP_BUF1_INDICATE(1));
510 t4_set_tcb_field(sc, toep, 1, W_TCB_T_FLAGS,
511 V_TF_RCV_COALESCE_ENABLE(1), 0);
515 disable_ddp(struct adapter *sc, struct toepcb *toep)
518 KASSERT((toep->ddp_flags & (DDP_ON | DDP_SC_REQ)) == DDP_ON,
519 ("%s: toep %p has bad ddp_flags 0x%x",
520 __func__, toep, toep->ddp_flags));
522 CTR3(KTR_CXGBE, "%s: tid %u (time %u)",
523 __func__, toep->tid, time_uptime);
525 toep->ddp_flags |= DDP_SC_REQ;
526 t4_set_tcb_field(sc, toep, 1, W_TCB_T_FLAGS,
527 V_TF_RCV_COALESCE_ENABLE(1), V_TF_RCV_COALESCE_ENABLE(1));
528 t4_set_tcb_field(sc, toep, 1, W_TCB_RX_DDP_FLAGS, V_TF_DDP_OFF(1),
533 hold_uio(struct uio *uio, vm_page_t **ppages, int *pnpages)
537 vm_offset_t start, end;
541 KASSERT(uio->uio_iovcnt == 1,
542 ("%s: uio_iovcnt %d", __func__, uio->uio_iovcnt));
543 KASSERT(uio->uio_td->td_proc == curproc,
544 ("%s: uio proc (%p) is not curproc (%p)",
545 __func__, uio->uio_td->td_proc, curproc));
547 map = &curproc->p_vmspace->vm_map;
548 iov = &uio->uio_iov[0];
549 start = trunc_page((uintptr_t)iov->iov_base);
550 end = round_page((vm_offset_t)iov->iov_base + iov->iov_len);
551 n = howmany(end - start, PAGE_SIZE);
553 if (end - start > MAX_DDP_BUFFER_SIZE)
556 pp = malloc(n * sizeof(vm_page_t), M_CXGBE, M_NOWAIT);
560 if (vm_fault_quick_hold_pages(map, (vm_offset_t)iov->iov_base,
561 iov->iov_len, VM_PROT_WRITE, pp, n) < 0) {
573 bufcmp(struct ddp_buffer *db, vm_page_t *pages, int npages, int offset, int len)
577 if (db == NULL || db->npages != npages || db->offset != offset ||
581 for (i = 0; i < npages; i++) {
582 if (pages[i]->phys_addr != db->pages[i]->phys_addr)
590 calculate_hcf(int n1, int n2)
611 static struct ddp_buffer *
612 alloc_ddp_buffer(struct tom_data *td, vm_page_t *pages, int npages, int offset,
615 int i, hcf, seglen, idx, ppod, nppods;
616 struct ddp_buffer *db;
619 * The DDP page size is unrelated to the VM page size. We combine
620 * contiguous physical pages into larger segments to get the best DDP
621 * page size possible. This is the largest of the four sizes in
622 * A_ULP_RX_TDDP_PSZ that evenly divides the HCF of the segment sizes in
626 for (i = 0; i < npages; i++) {
628 while (i < npages - 1 &&
629 pages[i]->phys_addr + PAGE_SIZE == pages[i + 1]->phys_addr) {
634 hcf = calculate_hcf(hcf, seglen);
635 if (hcf < t4_ddp_pgsz[1]) {
637 goto have_pgsz; /* give up, short circuit */
641 if (hcf % t4_ddp_pgsz[0] != 0) {
642 /* hmmm. This could only happen when PAGE_SIZE < 4K */
643 KASSERT(PAGE_SIZE < 4096,
644 ("%s: PAGE_SIZE %d, hcf %d", __func__, PAGE_SIZE, hcf));
645 CTR3(KTR_CXGBE, "%s: PAGE_SIZE %d, hcf %d",
646 __func__, PAGE_SIZE, hcf);
650 for (idx = nitems(t4_ddp_pgsz) - 1; idx > 0; idx--) {
651 if (hcf % t4_ddp_pgsz[idx] == 0)
656 db = malloc(sizeof(*db), M_CXGBE, M_NOWAIT);
658 CTR1(KTR_CXGBE, "%s: malloc failed.", __func__);
662 nppods = pages_to_nppods(npages, t4_ddp_pgsz[idx]);
663 ppod = alloc_ppods(td, nppods);
666 CTR4(KTR_CXGBE, "%s: no pods, nppods %d, resid %d, pgsz %d",
667 __func__, nppods, len, t4_ddp_pgsz[idx]);
671 KASSERT(idx <= M_PPOD_PGSZ && ppod <= M_PPOD_TAG,
672 ("%s: DDP pgsz_idx = %d, ppod = %d", __func__, idx, ppod));
674 db->tag = V_PPOD_PGSZ(idx) | V_PPOD_TAG(ppod);
681 CTR6(KTR_CXGBE, "New DDP buffer. "
682 "ddp_pgsz %d, ppod 0x%x, npages %d, nppods %d, offset %d, len %d",
683 t4_ddp_pgsz[idx], ppod, db->npages, db->nppods, db->offset,
689 #define NUM_ULP_TX_SC_IMM_PPODS (256 / PPOD_SIZE)
692 write_page_pods(struct adapter *sc, struct toepcb *toep, struct ddp_buffer *db)
695 struct ulp_mem_io *ulpmc;
696 struct ulptx_idata *ulpsc;
697 struct pagepod *ppod;
698 int i, j, k, n, chunk, len, ddp_pgsz, idx, ppod_addr;
701 cmd = htobe32(V_ULPTX_CMD(ULP_TX_MEM_WRITE));
703 cmd |= htobe32(F_ULP_MEMIO_ORDER);
705 cmd |= htobe32(F_T5_ULP_MEMIO_IMM);
706 ddp_pgsz = t4_ddp_pgsz[G_PPOD_PGSZ(db->tag)];
707 ppod_addr = sc->vres.ddp.start + G_PPOD_TAG(db->tag) * PPOD_SIZE;
708 for (i = 0; i < db->nppods; ppod_addr += chunk) {
710 /* How many page pods are we writing in this cycle */
711 n = min(db->nppods - i, NUM_ULP_TX_SC_IMM_PPODS);
713 len = roundup2(sizeof(*ulpmc) + sizeof(*ulpsc) + chunk, 16);
715 wr = alloc_wrqe(len, toep->ctrlq);
717 return (ENOMEM); /* ok to just bail out */
720 INIT_ULPTX_WR(ulpmc, len, 0, 0);
722 ulpmc->dlen = htobe32(V_ULP_MEMIO_DATA_LEN(chunk / 32));
723 ulpmc->len16 = htobe32(howmany(len - sizeof(ulpmc->wr), 16));
724 ulpmc->lock_addr = htobe32(V_ULP_MEMIO_ADDR(ppod_addr >> 5));
726 ulpsc = (struct ulptx_idata *)(ulpmc + 1);
727 ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM));
728 ulpsc->len = htobe32(chunk);
730 ppod = (struct pagepod *)(ulpsc + 1);
731 for (j = 0; j < n; i++, j++, ppod++) {
732 ppod->vld_tid_pgsz_tag_color = htobe64(F_PPOD_VALID |
733 V_PPOD_TID(toep->tid) | db->tag);
734 ppod->len_offset = htobe64(V_PPOD_LEN(db->len) |
735 V_PPOD_OFST(db->offset));
737 idx = i * PPOD_PAGES * (ddp_pgsz / PAGE_SIZE);
738 for (k = 0; k < nitems(ppod->addr); k++) {
739 if (idx < db->npages) {
741 htobe64(db->pages[idx]->phys_addr);
742 idx += ddp_pgsz / PAGE_SIZE;
747 "%s: tid %d ppod[%d]->addr[%d] = %p",
748 __func__, toep->tid, i, k,
749 htobe64(ppod->addr[k]));
762 * Reuse, or allocate (and program the page pods for) a new DDP buffer. The
763 * "pages" array is handed over to this function and should not be used in any
764 * way by the caller after that.
767 select_ddp_buffer(struct adapter *sc, struct toepcb *toep, vm_page_t *pages,
768 int npages, int db_off, int db_len)
770 struct ddp_buffer *db;
771 struct tom_data *td = sc->tom_softc;
772 int i, empty_slot = -1;
775 for (i = 0; i < nitems(toep->db); i++) {
776 if (bufcmp(toep->db[i], pages, npages, db_off, db_len) == 0) {
777 free(pages, M_CXGBE);
778 return (i); /* pages still held */
779 } else if (toep->db[i] == NULL && empty_slot < 0)
783 /* Allocate new buffer, write its page pods. */
784 db = alloc_ddp_buffer(td, pages, npages, db_off, db_len);
786 vm_page_unhold_pages(pages, npages);
787 free(pages, M_CXGBE);
790 if (write_page_pods(sc, toep, db) != 0) {
791 vm_page_unhold_pages(pages, npages);
792 free_ddp_buffer(td, db);
798 i = arc4random() % nitems(toep->db);
799 free_ddp_buffer(td, toep->db[i]);
803 CTR5(KTR_CXGBE, "%s: tid %d, DDP buffer[%d] = %p (tag 0x%x)",
804 __func__, toep->tid, i, db, db->tag);
810 wire_ddp_buffer(struct ddp_buffer *db)
815 for (i = 0; i < db->npages; i++) {
825 unwire_ddp_buffer(struct ddp_buffer *db)
830 for (i = 0; i < db->npages; i++) {
833 vm_page_unwire(p, PQ_INACTIVE);
839 handle_ddp(struct socket *so, struct uio *uio, int flags, int error)
841 struct sockbuf *sb = &so->so_rcv;
842 struct tcpcb *tp = so_sototcpcb(so);
843 struct toepcb *toep = tp->t_toe;
844 struct adapter *sc = td_adapter(toep->td);
846 int npages, db_idx, rc, buf_flag;
847 struct ddp_buffer *db;
851 SOCKBUF_LOCK_ASSERT(sb);
854 if (sb->sb_cc + sc->tt.ddp_thres > uio->uio_resid) {
855 CTR4(KTR_CXGBE, "%s: sb_cc %d, threshold %d, resid %d",
856 __func__, sb->sb_cc, sc->tt.ddp_thres, uio->uio_resid);
860 /* XXX: too eager to disable DDP, could handle NBIO better than this. */
861 if (sbused(sb) >= uio->uio_resid || uio->uio_resid < sc->tt.ddp_thres ||
862 uio->uio_resid > MAX_DDP_BUFFER_SIZE || uio->uio_iovcnt > 1 ||
863 so->so_state & SS_NBIO || flags & (MSG_DONTWAIT | MSG_NBIO) ||
864 error || so->so_error || sb->sb_state & SBS_CANTRCVMORE)
868 * Fault in and then hold the pages of the uio buffers. We'll wire them
869 * a bit later if everything else works out.
872 if (hold_uio(uio, &pages, &npages) != 0) {
877 if (__predict_false(so->so_error || sb->sb_state & SBS_CANTRCVMORE)) {
878 vm_page_unhold_pages(pages, npages);
879 free(pages, M_CXGBE);
884 * Figure out which one of the two DDP buffers to use this time.
886 db_idx = select_ddp_buffer(sc, toep, pages, npages,
887 (uintptr_t)uio->uio_iov->iov_base & PAGE_MASK, uio->uio_resid);
888 pages = NULL; /* handed off to select_ddp_buffer */
891 db = toep->db[db_idx];
892 buf_flag = db_idx == 0 ? DDP_BUF0_ACTIVE : DDP_BUF1_ACTIVE;
895 * Build the compound work request that tells the chip where to DMA the
898 ddp_flags = select_ddp_flags(so, flags, db_idx);
899 wr = mk_update_tcb_for_ddp(sc, toep, db_idx, sbused(sb), ddp_flags);
902 * Just unhold the pages. The DDP buffer's software state is
903 * left as-is in the toep. The page pods were written
904 * successfully and we may have an opportunity to use it in the
907 vm_page_unhold_pages(db->pages, db->npages);
911 /* Wire (and then unhold) the pages, and give the chip the go-ahead. */
914 sb->sb_flags &= ~SB_DDP_INDICATE;
915 toep->ddp_flags |= buf_flag;
918 * Wait for the DDP operation to complete and then unwire the pages.
919 * The return code from the sbwait will be the final return code of this
920 * function. But we do need to wait for DDP no matter what.
923 while (toep->ddp_flags & buf_flag) {
924 /* XXXGL: shouldn't here be sbwait() call? */
925 sb->sb_flags |= SB_WAIT;
926 msleep(&sb->sb_acc, &sb->sb_mtx, PSOCK , "sbwait", 0);
928 unwire_ddp_buffer(db);
931 disable_ddp(sc, toep);
932 discourage_ddp(toep);
933 sb->sb_flags &= ~SB_DDP_INDICATE;
938 t4_init_ddp(struct adapter *sc, struct tom_data *td)
940 int nppods = sc->vres.ddp.size / PPOD_SIZE;
942 td->ppod_arena = vmem_create("DDP page pods", 0, nppods, 1, 32,
943 M_FIRSTFIT | M_NOWAIT);
945 t4_register_cpl_handler(sc, CPL_RX_DATA_DDP, do_rx_data_ddp);
946 t4_register_cpl_handler(sc, CPL_RX_DDP_COMPLETE, do_rx_ddp_complete);
950 t4_uninit_ddp(struct adapter *sc __unused, struct tom_data *td)
953 if (td->ppod_arena != NULL) {
954 vmem_destroy(td->ppod_arena);
955 td->ppod_arena = NULL;
959 #define VNET_SO_ASSERT(so) \
960 VNET_ASSERT(curvnet != NULL, \
961 ("%s:%d curvnet is NULL, so=%p", __func__, __LINE__, (so)));
962 #define SBLOCKWAIT(f) (((f) & MSG_DONTWAIT) ? 0 : SBL_WAIT)
964 soreceive_rcvoob(struct socket *so, struct uio *uio, int flags)
967 CXGBE_UNIMPLEMENTED(__func__);
970 static char ddp_magic_str[] = "nothing to see here";
973 get_ddp_mbuf(int len)
977 m = m_get(M_NOWAIT, MT_DATA);
979 CXGBE_UNIMPLEMENTED("mbuf alloc failure");
981 m->m_data = &ddp_magic_str[0];
987 is_ddp_mbuf(struct mbuf *m)
990 return (m->m_data == &ddp_magic_str[0]);
994 * Copy an mbuf chain into a uio limited by len if set.
997 m_mbuftouio_ddp(struct uio *uio, struct mbuf *m, int len)
999 int error, length, total;
1003 total = min(uio->uio_resid, len);
1005 total = uio->uio_resid;
1007 /* Fill the uio with data from the mbufs. */
1008 for (; m != NULL; m = m->m_next) {
1009 length = min(m->m_len, total - progress);
1011 if (is_ddp_mbuf(m)) {
1012 enum uio_seg segflag = uio->uio_segflg;
1014 uio->uio_segflg = UIO_NOCOPY;
1015 error = uiomove(mtod(m, void *), length, uio);
1016 uio->uio_segflg = segflag;
1018 error = uiomove(mtod(m, void *), length, uio);
1029 * Based on soreceive_stream() in uipc_socket.c
1032 t4_soreceive_ddp(struct socket *so, struct sockaddr **psa, struct uio *uio,
1033 struct mbuf **mp0, struct mbuf **controlp, int *flagsp)
1035 int len = 0, error = 0, flags, oresid, ddp_handled = 0;
1037 struct mbuf *m, *n = NULL;
1039 /* We only do stream sockets. */
1040 if (so->so_type != SOCK_STREAM)
1044 if (controlp != NULL)
1047 flags = *flagsp &~ MSG_EOR;
1050 if (flags & MSG_OOB)
1051 return (soreceive_rcvoob(so, uio, flags));
1057 /* Prevent other readers from entering the socket. */
1058 error = sblock(sb, SBLOCKWAIT(flags));
1063 /* Easy one, no space to copyout anything. */
1064 if (uio->uio_resid == 0) {
1068 oresid = uio->uio_resid;
1070 /* We will never ever get anything unless we are or were connected. */
1071 if (!(so->so_state & (SS_ISCONNECTED|SS_ISDISCONNECTED))) {
1077 SOCKBUF_LOCK_ASSERT(&so->so_rcv);
1079 if (sb->sb_flags & SB_DDP_INDICATE && !ddp_handled) {
1081 /* uio should be just as it was at entry */
1082 KASSERT(oresid == uio->uio_resid,
1083 ("%s: oresid = %d, uio_resid = %zd, sbused = %d",
1084 __func__, oresid, uio->uio_resid, sbused(sb)));
1086 error = handle_ddp(so, uio, flags, 0);
1092 /* Abort if socket has reported problems. */
1096 if (oresid > uio->uio_resid)
1098 error = so->so_error;
1099 if (!(flags & MSG_PEEK))
1104 /* Door is closed. Deliver what is left, if any. */
1105 if (sb->sb_state & SBS_CANTRCVMORE) {
1112 /* Socket buffer is empty and we shall not block. */
1113 if (sbused(sb) == 0 &&
1114 ((so->so_state & SS_NBIO) || (flags & (MSG_DONTWAIT|MSG_NBIO)))) {
1119 /* Socket buffer got some data that we shall deliver now. */
1120 if (sbused(sb) && !(flags & MSG_WAITALL) &&
1121 ((so->so_state & SS_NBIO) ||
1122 (flags & (MSG_DONTWAIT|MSG_NBIO)) ||
1123 sbused(sb) >= sb->sb_lowat ||
1124 sbused(sb) >= uio->uio_resid ||
1125 sbused(sb) >= sb->sb_hiwat) ) {
1129 /* On MSG_WAITALL we must wait until all data or error arrives. */
1130 if ((flags & MSG_WAITALL) &&
1131 (sbused(sb) >= uio->uio_resid || sbused(sb) >= sb->sb_lowat))
1135 * Wait and block until (more) data comes in.
1136 * NB: Drops the sockbuf lock during wait.
1140 if (sb->sb_flags & SB_DDP_INDICATE && !ddp_handled) {
1141 (void) handle_ddp(so, uio, flags, 1);
1149 SOCKBUF_LOCK_ASSERT(&so->so_rcv);
1150 KASSERT(sbused(sb) > 0, ("%s: sockbuf empty", __func__));
1151 KASSERT(sb->sb_mb != NULL, ("%s: sb_mb == NULL", __func__));
1153 if (sb->sb_flags & SB_DDP_INDICATE && !ddp_handled)
1158 uio->uio_td->td_ru.ru_msgrcv++;
1160 /* Fill uio until full or current end of socket buffer is reached. */
1161 len = min(uio->uio_resid, sbused(sb));
1163 /* Dequeue as many mbufs as possible. */
1164 if (!(flags & MSG_PEEK) && len >= sb->sb_mb->m_len) {
1165 for (*mp0 = m = sb->sb_mb;
1166 m != NULL && m->m_len <= len;
1169 uio->uio_resid -= m->m_len;
1174 if (sb->sb_mb == NULL)
1178 /* Copy the remainder. */
1180 KASSERT(sb->sb_mb != NULL,
1181 ("%s: len > 0 && sb->sb_mb empty", __func__));
1183 m = m_copym(sb->sb_mb, 0, len, M_NOWAIT);
1185 len = 0; /* Don't flush data from sockbuf. */
1187 uio->uio_resid -= m->m_len;
1198 /* NB: Must unlock socket buffer as uiomove may sleep. */
1200 error = m_mbuftouio_ddp(uio, sb->sb_mb, len);
1205 SBLASTRECORDCHK(sb);
1209 * Remove the delivered data from the socket buffer unless we
1210 * were only peeking.
1212 if (!(flags & MSG_PEEK)) {
1214 sbdrop_locked(sb, len);
1216 /* Notify protocol that we drained some data. */
1217 if ((so->so_proto->pr_flags & PR_WANTRCVD) &&
1218 (((flags & MSG_WAITALL) && uio->uio_resid > 0) ||
1219 !(flags & MSG_SOCALLBCK))) {
1222 (*so->so_proto->pr_usrreqs->pru_rcvd)(so, flags);
1228 * For MSG_WAITALL we may have to loop again and wait for
1229 * more data to come in.
1231 if ((flags & MSG_WAITALL) && uio->uio_resid > 0)
1234 SOCKBUF_LOCK_ASSERT(sb);
1235 SBLASTRECORDCHK(sb);