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, u_int *ppod_addr)
108 rc = vmem_alloc(td->ppod_arena, PPOD_SZ(n), M_NOWAIT | M_FIRSTFIT, &v);
109 *ppod_addr = (u_int)v;
115 free_ppods(struct tom_data *td, u_int ppod_addr, int n)
120 vmem_free(td->ppod_arena, (vmem_addr_t)ppod_addr, PPOD_SZ(n));
124 pages_to_nppods(int npages, int ddp_pgsz)
126 int nsegs = npages * PAGE_SIZE / ddp_pgsz;
128 return (howmany(nsegs, PPOD_PAGES));
132 free_ddp_buffer(struct tom_data *td, struct ddp_buffer *db)
139 free(db->pages, M_CXGBE);
142 free_ppods(td, db->ppod_addr, db->nppods);
148 release_ddp_resources(struct toepcb *toep)
152 for (i = 0; i < nitems(toep->db); i++) {
153 if (toep->db[i] != NULL) {
154 free_ddp_buffer(toep->td, toep->db[i]);
160 /* XXX: handle_ddp_data code duplication */
162 insert_ddp_data(struct toepcb *toep, uint32_t n)
164 struct inpcb *inp = toep->inp;
165 struct tcpcb *tp = intotcpcb(inp);
166 struct sockbuf *sb = &inp->inp_socket->so_rcv;
169 INP_WLOCK_ASSERT(inp);
170 SOCKBUF_LOCK_ASSERT(sb);
174 #ifndef USE_DDP_RX_FLOW_CONTROL
175 KASSERT(tp->rcv_wnd >= n, ("%s: negative window size", __func__));
179 KASSERT(toep->sb_cc >= sbused(sb),
180 ("%s: sb %p has more data (%d) than last time (%d).",
181 __func__, sb, sbused(sb), toep->sb_cc));
182 toep->rx_credits += toep->sb_cc - sbused(sb);
183 #ifdef USE_DDP_RX_FLOW_CONTROL
184 toep->rx_credits -= n; /* adjust for F_RX_FC_DDP */
186 sbappendstream_locked(sb, m, 0);
187 toep->sb_cc = sbused(sb);
190 /* SET_TCB_FIELD sent as a ULP command looks like this */
191 #define LEN__SET_TCB_FIELD_ULP (sizeof(struct ulp_txpkt) + \
192 sizeof(struct ulptx_idata) + sizeof(struct cpl_set_tcb_field_core))
194 /* RX_DATA_ACK sent as a ULP command looks like this */
195 #define LEN__RX_DATA_ACK_ULP (sizeof(struct ulp_txpkt) + \
196 sizeof(struct ulptx_idata) + sizeof(struct cpl_rx_data_ack_core))
199 mk_set_tcb_field_ulp(struct ulp_txpkt *ulpmc, struct toepcb *toep,
200 uint64_t word, uint64_t mask, uint64_t val)
202 struct ulptx_idata *ulpsc;
203 struct cpl_set_tcb_field_core *req;
205 ulpmc->cmd_dest = htonl(V_ULPTX_CMD(ULP_TX_PKT) | V_ULP_TXPKT_DEST(0));
206 ulpmc->len = htobe32(howmany(LEN__SET_TCB_FIELD_ULP, 16));
208 ulpsc = (struct ulptx_idata *)(ulpmc + 1);
209 ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM));
210 ulpsc->len = htobe32(sizeof(*req));
212 req = (struct cpl_set_tcb_field_core *)(ulpsc + 1);
213 OPCODE_TID(req) = htobe32(MK_OPCODE_TID(CPL_SET_TCB_FIELD, toep->tid));
214 req->reply_ctrl = htobe16(V_NO_REPLY(1) |
215 V_QUEUENO(toep->ofld_rxq->iq.abs_id));
216 req->word_cookie = htobe16(V_WORD(word) | V_COOKIE(0));
217 req->mask = htobe64(mask);
218 req->val = htobe64(val);
220 ulpsc = (struct ulptx_idata *)(req + 1);
221 if (LEN__SET_TCB_FIELD_ULP % 16) {
222 ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_NOOP));
223 ulpsc->len = htobe32(0);
230 mk_rx_data_ack_ulp(struct ulp_txpkt *ulpmc, struct toepcb *toep)
232 struct ulptx_idata *ulpsc;
233 struct cpl_rx_data_ack_core *req;
235 ulpmc->cmd_dest = htonl(V_ULPTX_CMD(ULP_TX_PKT) | V_ULP_TXPKT_DEST(0));
236 ulpmc->len = htobe32(howmany(LEN__RX_DATA_ACK_ULP, 16));
238 ulpsc = (struct ulptx_idata *)(ulpmc + 1);
239 ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM));
240 ulpsc->len = htobe32(sizeof(*req));
242 req = (struct cpl_rx_data_ack_core *)(ulpsc + 1);
243 OPCODE_TID(req) = htobe32(MK_OPCODE_TID(CPL_RX_DATA_ACK, toep->tid));
244 req->credit_dack = htobe32(F_RX_MODULATE_RX);
246 ulpsc = (struct ulptx_idata *)(req + 1);
247 if (LEN__RX_DATA_ACK_ULP % 16) {
248 ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_NOOP));
249 ulpsc->len = htobe32(0);
255 static inline uint64_t
256 select_ddp_flags(struct socket *so, int flags, int db_idx)
258 uint64_t ddp_flags = V_TF_DDP_INDICATE_OUT(0);
259 int waitall = flags & MSG_WAITALL;
260 int nb = so->so_state & SS_NBIO || flags & (MSG_DONTWAIT | MSG_NBIO);
262 KASSERT(db_idx == 0 || db_idx == 1,
263 ("%s: bad DDP buffer index %d", __func__, db_idx));
266 ddp_flags |= V_TF_DDP_BUF0_VALID(1) | V_TF_DDP_ACTIVE_BUF(0);
268 ddp_flags |= V_TF_DDP_PUSH_DISABLE_0(1);
270 ddp_flags |= V_TF_DDP_BUF0_FLUSH(1);
272 ddp_flags |= V_TF_DDP_BUF0_FLUSH(0);
274 ddp_flags |= V_TF_DDP_BUF1_VALID(1) | V_TF_DDP_ACTIVE_BUF(1);
276 ddp_flags |= V_TF_DDP_PUSH_DISABLE_1(1);
278 ddp_flags |= V_TF_DDP_BUF1_FLUSH(1);
280 ddp_flags |= V_TF_DDP_BUF1_FLUSH(0);
287 mk_update_tcb_for_ddp(struct adapter *sc, struct toepcb *toep, int db_idx,
288 int offset, uint64_t ddp_flags)
290 struct ddp_buffer *db = toep->db[db_idx];
292 struct work_request_hdr *wrh;
293 struct ulp_txpkt *ulpmc;
296 KASSERT(db_idx == 0 || db_idx == 1,
297 ("%s: bad DDP buffer index %d", __func__, db_idx));
300 * We'll send a compound work request that has 3 SET_TCB_FIELDs and an
301 * RX_DATA_ACK (with RX_MODULATE to speed up delivery).
303 * The work request header is 16B and always ends at a 16B boundary.
304 * The ULPTX master commands that follow must all end at 16B boundaries
305 * too so we round up the size to 16.
307 len = sizeof(*wrh) + 3 * roundup2(LEN__SET_TCB_FIELD_ULP, 16) +
308 roundup2(LEN__RX_DATA_ACK_ULP, 16);
310 wr = alloc_wrqe(len, toep->ctrlq);
314 INIT_ULPTX_WRH(wrh, len, 1, 0); /* atomic */
315 ulpmc = (struct ulp_txpkt *)(wrh + 1);
317 /* Write the buffer's tag */
318 ulpmc = mk_set_tcb_field_ulp(ulpmc, toep,
319 W_TCB_RX_DDP_BUF0_TAG + db_idx,
320 V_TCB_RX_DDP_BUF0_TAG(M_TCB_RX_DDP_BUF0_TAG),
321 V_TCB_RX_DDP_BUF0_TAG(db->tag));
323 /* Update the current offset in the DDP buffer and its total length */
325 ulpmc = mk_set_tcb_field_ulp(ulpmc, toep,
326 W_TCB_RX_DDP_BUF0_OFFSET,
327 V_TCB_RX_DDP_BUF0_OFFSET(M_TCB_RX_DDP_BUF0_OFFSET) |
328 V_TCB_RX_DDP_BUF0_LEN(M_TCB_RX_DDP_BUF0_LEN),
329 V_TCB_RX_DDP_BUF0_OFFSET(offset) |
330 V_TCB_RX_DDP_BUF0_LEN(db->len));
332 ulpmc = mk_set_tcb_field_ulp(ulpmc, toep,
333 W_TCB_RX_DDP_BUF1_OFFSET,
334 V_TCB_RX_DDP_BUF1_OFFSET(M_TCB_RX_DDP_BUF1_OFFSET) |
335 V_TCB_RX_DDP_BUF1_LEN((u64)M_TCB_RX_DDP_BUF1_LEN << 32),
336 V_TCB_RX_DDP_BUF1_OFFSET(offset) |
337 V_TCB_RX_DDP_BUF1_LEN((u64)db->len << 32));
339 /* Update DDP flags */
340 ulpmc = mk_set_tcb_field_ulp(ulpmc, toep, W_TCB_RX_DDP_FLAGS,
341 V_TF_DDP_BUF0_FLUSH(1) | V_TF_DDP_BUF1_FLUSH(1) |
342 V_TF_DDP_PUSH_DISABLE_0(1) | V_TF_DDP_PUSH_DISABLE_1(1) |
343 V_TF_DDP_BUF0_VALID(1) | V_TF_DDP_BUF1_VALID(1) |
344 V_TF_DDP_ACTIVE_BUF(1) | V_TF_DDP_INDICATE_OUT(1), ddp_flags);
346 /* Gratuitous RX_DATA_ACK with RX_MODULATE set to speed up delivery. */
347 ulpmc = mk_rx_data_ack_ulp(ulpmc, toep);
353 discourage_ddp(struct toepcb *toep)
356 if (toep->ddp_score && --toep->ddp_score == 0) {
357 toep->ddp_flags &= ~DDP_OK;
358 toep->ddp_disabled = time_uptime;
359 CTR3(KTR_CXGBE, "%s: tid %u !DDP_OK @ %u",
360 __func__, toep->tid, time_uptime);
365 handle_ddp_data(struct toepcb *toep, __be32 ddp_report, __be32 rcv_nxt, int len)
367 uint32_t report = be32toh(ddp_report);
368 unsigned int db_flag;
369 struct inpcb *inp = toep->inp;
375 db_flag = report & F_DDP_BUF_IDX ? DDP_BUF1_ACTIVE : DDP_BUF0_ACTIVE;
377 if (__predict_false(!(report & F_DDP_INV)))
378 CXGBE_UNIMPLEMENTED("DDP buffer still valid");
381 so = inp_inpcbtosocket(inp);
383 if (__predict_false(inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT))) {
386 * XXX: think a bit more.
387 * tcpcb probably gone, but socket should still be around
388 * because we always wait for DDP completion in soreceive no
389 * matter what. Just wake it up and let it clean up.
392 CTR5(KTR_CXGBE, "%s: tid %u, seq 0x%x, len %d, inp_flags 0x%x",
393 __func__, toep->tid, be32toh(rcv_nxt), len, inp->inp_flags);
399 len += be32toh(rcv_nxt) - tp->rcv_nxt;
401 tp->t_rcvtime = ticks;
402 #ifndef USE_DDP_RX_FLOW_CONTROL
403 KASSERT(tp->rcv_wnd >= len, ("%s: negative window size", __func__));
406 m = get_ddp_mbuf(len);
409 if (report & F_DDP_BUF_COMPLETE)
410 toep->ddp_score = DDP_HIGH_SCORE;
412 discourage_ddp(toep);
414 KASSERT(toep->sb_cc >= sbused(sb),
415 ("%s: sb %p has more data (%d) than last time (%d).",
416 __func__, sb, sbused(sb), toep->sb_cc));
417 toep->rx_credits += toep->sb_cc - sbused(sb);
418 #ifdef USE_DDP_RX_FLOW_CONTROL
419 toep->rx_credits -= len; /* adjust for F_RX_FC_DDP */
421 sbappendstream_locked(sb, m, 0);
422 toep->sb_cc = sbused(sb);
424 KASSERT(toep->ddp_flags & db_flag,
425 ("%s: DDP buffer not active. toep %p, ddp_flags 0x%x, report 0x%x",
426 __func__, toep, toep->ddp_flags, report));
427 toep->ddp_flags &= ~db_flag;
428 sorwakeup_locked(so);
429 SOCKBUF_UNLOCK_ASSERT(sb);
435 #define DDP_ERR (F_DDP_PPOD_MISMATCH | F_DDP_LLIMIT_ERR | F_DDP_ULIMIT_ERR |\
436 F_DDP_PPOD_PARITY_ERR | F_DDP_PADDING_ERR | F_DDP_OFFSET_ERR |\
437 F_DDP_INVALID_TAG | F_DDP_COLOR_ERR | F_DDP_TID_MISMATCH |\
438 F_DDP_INVALID_PPOD | F_DDP_HDRCRC_ERR | F_DDP_DATACRC_ERR)
441 do_rx_data_ddp(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
443 struct adapter *sc = iq->adapter;
444 const struct cpl_rx_data_ddp *cpl = (const void *)(rss + 1);
445 unsigned int tid = GET_TID(cpl);
447 struct toepcb *toep = lookup_tid(sc, tid);
448 struct tom_data *td = toep->td;
450 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
451 KASSERT(toep->tid == tid, ("%s: toep tid/atid mismatch", __func__));
452 KASSERT(!(toep->flags & TPF_SYNQE),
453 ("%s: toep %p claims to be a synq entry", __func__, toep));
455 vld = be32toh(cpl->ddpvld);
456 if (__predict_false(vld & DDP_ERR)) {
457 panic("%s: DDP error 0x%x (tid %d, toep %p)",
458 __func__, vld, tid, toep);
460 if (toep->ulp_mode == ULP_MODE_ISCSI) {
461 m = m_get(M_NOWAIT, MT_DATA);
463 CXGBE_UNIMPLEMENTED("mbuf alloc failure");
464 memcpy(mtod(m, unsigned char *), cpl,
465 sizeof(struct cpl_rx_data_ddp));
466 if (!t4_cpl_iscsi_callback(td, toep, m, CPL_RX_DATA_DDP))
471 handle_ddp_data(toep, cpl->u.ddp_report, cpl->seq, be16toh(cpl->len));
477 do_rx_ddp_complete(struct sge_iq *iq, const struct rss_header *rss,
480 struct adapter *sc = iq->adapter;
481 const struct cpl_rx_ddp_complete *cpl = (const void *)(rss + 1);
482 unsigned int tid = GET_TID(cpl);
483 struct toepcb *toep = lookup_tid(sc, tid);
485 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
486 KASSERT(toep->tid == tid, ("%s: toep tid/atid mismatch", __func__));
487 KASSERT(!(toep->flags & TPF_SYNQE),
488 ("%s: toep %p claims to be a synq entry", __func__, toep));
490 handle_ddp_data(toep, cpl->ddp_report, cpl->rcv_nxt, 0);
496 enable_ddp(struct adapter *sc, struct toepcb *toep)
499 KASSERT((toep->ddp_flags & (DDP_ON | DDP_OK | DDP_SC_REQ)) == DDP_OK,
500 ("%s: toep %p has bad ddp_flags 0x%x",
501 __func__, toep, toep->ddp_flags));
503 CTR3(KTR_CXGBE, "%s: tid %u (time %u)",
504 __func__, toep->tid, time_uptime);
506 toep->ddp_flags |= DDP_SC_REQ;
507 t4_set_tcb_field(sc, toep, 1, W_TCB_RX_DDP_FLAGS,
508 V_TF_DDP_OFF(1) | V_TF_DDP_INDICATE_OUT(1) |
509 V_TF_DDP_BUF0_INDICATE(1) | V_TF_DDP_BUF1_INDICATE(1) |
510 V_TF_DDP_BUF0_VALID(1) | V_TF_DDP_BUF1_VALID(1),
511 V_TF_DDP_BUF0_INDICATE(1) | V_TF_DDP_BUF1_INDICATE(1));
512 t4_set_tcb_field(sc, toep, 1, W_TCB_T_FLAGS,
513 V_TF_RCV_COALESCE_ENABLE(1), 0);
517 disable_ddp(struct adapter *sc, struct toepcb *toep)
520 KASSERT((toep->ddp_flags & (DDP_ON | DDP_SC_REQ)) == DDP_ON,
521 ("%s: toep %p has bad ddp_flags 0x%x",
522 __func__, toep, toep->ddp_flags));
524 CTR3(KTR_CXGBE, "%s: tid %u (time %u)",
525 __func__, toep->tid, time_uptime);
527 toep->ddp_flags |= DDP_SC_REQ;
528 t4_set_tcb_field(sc, toep, 1, W_TCB_T_FLAGS,
529 V_TF_RCV_COALESCE_ENABLE(1), V_TF_RCV_COALESCE_ENABLE(1));
530 t4_set_tcb_field(sc, toep, 1, W_TCB_RX_DDP_FLAGS, V_TF_DDP_OFF(1),
535 hold_uio(struct uio *uio, vm_page_t **ppages, int *pnpages)
539 vm_offset_t start, end;
543 KASSERT(uio->uio_iovcnt == 1,
544 ("%s: uio_iovcnt %d", __func__, uio->uio_iovcnt));
545 KASSERT(uio->uio_td->td_proc == curproc,
546 ("%s: uio proc (%p) is not curproc (%p)",
547 __func__, uio->uio_td->td_proc, curproc));
549 map = &curproc->p_vmspace->vm_map;
550 iov = &uio->uio_iov[0];
551 start = trunc_page((uintptr_t)iov->iov_base);
552 end = round_page((vm_offset_t)iov->iov_base + iov->iov_len);
553 n = howmany(end - start, PAGE_SIZE);
555 if (end - start > MAX_DDP_BUFFER_SIZE)
558 pp = malloc(n * sizeof(vm_page_t), M_CXGBE, M_NOWAIT);
562 if (vm_fault_quick_hold_pages(map, (vm_offset_t)iov->iov_base,
563 iov->iov_len, VM_PROT_WRITE, pp, n) < 0) {
575 bufcmp(struct ddp_buffer *db, vm_page_t *pages, int npages, int offset, int len)
579 if (db == NULL || db->npages != npages || db->offset != offset ||
583 for (i = 0; i < npages; i++) {
584 if (pages[i]->phys_addr != db->pages[i]->phys_addr)
592 calculate_hcf(int n1, int n2)
613 static struct ddp_buffer *
614 alloc_ddp_buffer(struct tom_data *td, vm_page_t *pages, int npages, int offset,
617 int i, hcf, seglen, idx, ppod, nppods;
618 struct ddp_buffer *db;
621 * The DDP page size is unrelated to the VM page size. We combine
622 * contiguous physical pages into larger segments to get the best DDP
623 * page size possible. This is the largest of the four sizes in
624 * A_ULP_RX_TDDP_PSZ that evenly divides the HCF of the segment sizes in
628 for (i = 0; i < npages; i++) {
630 while (i < npages - 1 &&
631 pages[i]->phys_addr + PAGE_SIZE == pages[i + 1]->phys_addr) {
636 hcf = calculate_hcf(hcf, seglen);
637 if (hcf < t4_ddp_pgsz[1]) {
639 goto have_pgsz; /* give up, short circuit */
643 if (hcf % t4_ddp_pgsz[0] != 0) {
644 /* hmmm. This could only happen when PAGE_SIZE < 4K */
645 KASSERT(PAGE_SIZE < 4096,
646 ("%s: PAGE_SIZE %d, hcf %d", __func__, PAGE_SIZE, hcf));
647 CTR3(KTR_CXGBE, "%s: PAGE_SIZE %d, hcf %d",
648 __func__, PAGE_SIZE, hcf);
652 for (idx = nitems(t4_ddp_pgsz) - 1; idx > 0; idx--) {
653 if (hcf % t4_ddp_pgsz[idx] == 0)
657 MPASS(idx <= M_PPOD_PGSZ);
659 db = malloc(sizeof(*db), M_CXGBE, M_NOWAIT);
661 CTR1(KTR_CXGBE, "%s: malloc failed.", __func__);
665 nppods = pages_to_nppods(npages, t4_ddp_pgsz[idx]);
666 if (alloc_ppods(td, nppods, &db->ppod_addr) != 0) {
668 CTR4(KTR_CXGBE, "%s: no pods, nppods %d, resid %d, pgsz %d",
669 __func__, nppods, len, t4_ddp_pgsz[idx]);
672 ppod = (db->ppod_addr - td->ppod_start) / PPOD_SIZE;
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;
702 cmd = htobe32(V_ULPTX_CMD(ULP_TX_MEM_WRITE));
704 cmd |= htobe32(F_ULP_MEMIO_ORDER);
706 cmd |= htobe32(F_T5_ULP_MEMIO_IMM);
707 ddp_pgsz = t4_ddp_pgsz[G_PPOD_PGSZ(db->tag)];
708 ppod_addr = db->ppod_addr;
709 for (i = 0; i < db->nppods; ppod_addr += chunk) {
711 /* How many page pods are we writing in this cycle */
712 n = min(db->nppods - i, NUM_ULP_TX_SC_IMM_PPODS);
714 len = roundup2(sizeof(*ulpmc) + sizeof(*ulpsc) + chunk, 16);
716 wr = alloc_wrqe(len, toep->ctrlq);
718 return (ENOMEM); /* ok to just bail out */
721 INIT_ULPTX_WR(ulpmc, len, 0, 0);
723 ulpmc->dlen = htobe32(V_ULP_MEMIO_DATA_LEN(chunk / 32));
724 ulpmc->len16 = htobe32(howmany(len - sizeof(ulpmc->wr), 16));
725 ulpmc->lock_addr = htobe32(V_ULP_MEMIO_ADDR(ppod_addr >> 5));
727 ulpsc = (struct ulptx_idata *)(ulpmc + 1);
728 ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM));
729 ulpsc->len = htobe32(chunk);
731 ppod = (struct pagepod *)(ulpsc + 1);
732 for (j = 0; j < n; i++, j++, ppod++) {
733 ppod->vld_tid_pgsz_tag_color = htobe64(F_PPOD_VALID |
734 V_PPOD_TID(toep->tid) | db->tag);
735 ppod->len_offset = htobe64(V_PPOD_LEN(db->len) |
736 V_PPOD_OFST(db->offset));
738 idx = i * PPOD_PAGES * (ddp_pgsz / PAGE_SIZE);
739 for (k = 0; k < nitems(ppod->addr); k++) {
740 if (idx < db->npages) {
742 htobe64(db->pages[idx]->phys_addr);
743 idx += ddp_pgsz / PAGE_SIZE;
748 "%s: tid %d ppod[%d]->addr[%d] = %p",
749 __func__, toep->tid, i, k,
750 htobe64(ppod->addr[k]));
763 * Reuse, or allocate (and program the page pods for) a new DDP buffer. The
764 * "pages" array is handed over to this function and should not be used in any
765 * way by the caller after that.
768 select_ddp_buffer(struct adapter *sc, struct toepcb *toep, vm_page_t *pages,
769 int npages, int db_off, int db_len)
771 struct ddp_buffer *db;
772 struct tom_data *td = sc->tom_softc;
773 int i, empty_slot = -1;
776 for (i = 0; i < nitems(toep->db); i++) {
777 if (bufcmp(toep->db[i], pages, npages, db_off, db_len) == 0) {
778 free(pages, M_CXGBE);
779 return (i); /* pages still held */
780 } else if (toep->db[i] == NULL && empty_slot < 0)
784 /* Allocate new buffer, write its page pods. */
785 db = alloc_ddp_buffer(td, pages, npages, db_off, db_len);
787 vm_page_unhold_pages(pages, npages);
788 free(pages, M_CXGBE);
791 if (write_page_pods(sc, toep, db) != 0) {
792 vm_page_unhold_pages(pages, npages);
793 free_ddp_buffer(td, db);
799 i = arc4random() % nitems(toep->db);
800 free_ddp_buffer(td, toep->db[i]);
804 CTR5(KTR_CXGBE, "%s: tid %d, DDP buffer[%d] = %p (tag 0x%x)",
805 __func__, toep->tid, i, db, db->tag);
811 wire_ddp_buffer(struct ddp_buffer *db)
816 for (i = 0; i < db->npages; i++) {
826 unwire_ddp_buffer(struct ddp_buffer *db)
831 for (i = 0; i < db->npages; i++) {
834 vm_page_unwire(p, PQ_INACTIVE);
840 handle_ddp(struct socket *so, struct uio *uio, int flags, int error)
842 struct sockbuf *sb = &so->so_rcv;
843 struct tcpcb *tp = so_sototcpcb(so);
844 struct toepcb *toep = tp->t_toe;
845 struct adapter *sc = td_adapter(toep->td);
847 int npages, db_idx, rc, buf_flag;
848 struct ddp_buffer *db;
852 SOCKBUF_LOCK_ASSERT(sb);
855 if (sb->sb_cc + sc->tt.ddp_thres > uio->uio_resid) {
856 CTR4(KTR_CXGBE, "%s: sb_cc %d, threshold %d, resid %d",
857 __func__, sb->sb_cc, sc->tt.ddp_thres, uio->uio_resid);
861 /* XXX: too eager to disable DDP, could handle NBIO better than this. */
862 if (sbused(sb) >= uio->uio_resid || uio->uio_resid < sc->tt.ddp_thres ||
863 uio->uio_resid > MAX_DDP_BUFFER_SIZE || uio->uio_iovcnt > 1 ||
864 so->so_state & SS_NBIO || flags & (MSG_DONTWAIT | MSG_NBIO) ||
865 error || so->so_error || sb->sb_state & SBS_CANTRCVMORE)
869 * Fault in and then hold the pages of the uio buffers. We'll wire them
870 * a bit later if everything else works out.
873 if (hold_uio(uio, &pages, &npages) != 0) {
878 if (__predict_false(so->so_error || sb->sb_state & SBS_CANTRCVMORE)) {
879 vm_page_unhold_pages(pages, npages);
880 free(pages, M_CXGBE);
885 * Figure out which one of the two DDP buffers to use this time.
887 db_idx = select_ddp_buffer(sc, toep, pages, npages,
888 (uintptr_t)uio->uio_iov->iov_base & PAGE_MASK, uio->uio_resid);
889 pages = NULL; /* handed off to select_ddp_buffer */
892 db = toep->db[db_idx];
893 buf_flag = db_idx == 0 ? DDP_BUF0_ACTIVE : DDP_BUF1_ACTIVE;
896 * Build the compound work request that tells the chip where to DMA the
899 ddp_flags = select_ddp_flags(so, flags, db_idx);
900 wr = mk_update_tcb_for_ddp(sc, toep, db_idx, sbused(sb), ddp_flags);
903 * Just unhold the pages. The DDP buffer's software state is
904 * left as-is in the toep. The page pods were written
905 * successfully and we may have an opportunity to use it in the
908 vm_page_unhold_pages(db->pages, db->npages);
912 /* Wire (and then unhold) the pages, and give the chip the go-ahead. */
915 sb->sb_flags &= ~SB_DDP_INDICATE;
916 toep->ddp_flags |= buf_flag;
919 * Wait for the DDP operation to complete and then unwire the pages.
920 * The return code from the sbwait will be the final return code of this
921 * function. But we do need to wait for DDP no matter what.
924 while (toep->ddp_flags & buf_flag) {
925 /* XXXGL: shouldn't here be sbwait() call? */
926 sb->sb_flags |= SB_WAIT;
927 msleep(&sb->sb_acc, &sb->sb_mtx, PSOCK , "sbwait", 0);
929 unwire_ddp_buffer(db);
932 disable_ddp(sc, toep);
933 discourage_ddp(toep);
934 sb->sb_flags &= ~SB_DDP_INDICATE;
939 t4_init_ddp(struct adapter *sc, struct tom_data *td)
942 td->ppod_start = sc->vres.ddp.start;
943 td->ppod_arena = vmem_create("DDP page pods", sc->vres.ddp.start,
944 sc->vres.ddp.size, 1, 32, M_FIRSTFIT | M_NOWAIT);
946 t4_register_cpl_handler(sc, CPL_RX_DATA_DDP, do_rx_data_ddp);
947 t4_register_cpl_handler(sc, CPL_RX_DDP_COMPLETE, do_rx_ddp_complete);
951 t4_uninit_ddp(struct adapter *sc __unused, struct tom_data *td)
954 if (td->ppod_arena != NULL) {
955 vmem_destroy(td->ppod_arena);
956 td->ppod_arena = NULL;
960 #define VNET_SO_ASSERT(so) \
961 VNET_ASSERT(curvnet != NULL, \
962 ("%s:%d curvnet is NULL, so=%p", __func__, __LINE__, (so)));
963 #define SBLOCKWAIT(f) (((f) & MSG_DONTWAIT) ? 0 : SBL_WAIT)
965 soreceive_rcvoob(struct socket *so, struct uio *uio, int flags)
968 CXGBE_UNIMPLEMENTED(__func__);
971 static char ddp_magic_str[] = "nothing to see here";
974 get_ddp_mbuf(int len)
978 m = m_get(M_NOWAIT, MT_DATA);
980 CXGBE_UNIMPLEMENTED("mbuf alloc failure");
982 m->m_data = &ddp_magic_str[0];
988 is_ddp_mbuf(struct mbuf *m)
991 return (m->m_data == &ddp_magic_str[0]);
995 * Copy an mbuf chain into a uio limited by len if set.
998 m_mbuftouio_ddp(struct uio *uio, struct mbuf *m, int len)
1000 int error, length, total;
1004 total = min(uio->uio_resid, len);
1006 total = uio->uio_resid;
1008 /* Fill the uio with data from the mbufs. */
1009 for (; m != NULL; m = m->m_next) {
1010 length = min(m->m_len, total - progress);
1012 if (is_ddp_mbuf(m)) {
1013 enum uio_seg segflag = uio->uio_segflg;
1015 uio->uio_segflg = UIO_NOCOPY;
1016 error = uiomove(mtod(m, void *), length, uio);
1017 uio->uio_segflg = segflag;
1019 error = uiomove(mtod(m, void *), length, uio);
1030 * Based on soreceive_stream() in uipc_socket.c
1033 t4_soreceive_ddp(struct socket *so, struct sockaddr **psa, struct uio *uio,
1034 struct mbuf **mp0, struct mbuf **controlp, int *flagsp)
1036 int len = 0, error = 0, flags, oresid, ddp_handled = 0;
1038 struct mbuf *m, *n = NULL;
1040 /* We only do stream sockets. */
1041 if (so->so_type != SOCK_STREAM)
1045 if (controlp != NULL)
1048 flags = *flagsp &~ MSG_EOR;
1051 if (flags & MSG_OOB)
1052 return (soreceive_rcvoob(so, uio, flags));
1058 /* Prevent other readers from entering the socket. */
1059 error = sblock(sb, SBLOCKWAIT(flags));
1064 /* Easy one, no space to copyout anything. */
1065 if (uio->uio_resid == 0) {
1069 oresid = uio->uio_resid;
1071 /* We will never ever get anything unless we are or were connected. */
1072 if (!(so->so_state & (SS_ISCONNECTED|SS_ISDISCONNECTED))) {
1078 SOCKBUF_LOCK_ASSERT(&so->so_rcv);
1080 if (sb->sb_flags & SB_DDP_INDICATE && !ddp_handled) {
1082 /* uio should be just as it was at entry */
1083 KASSERT(oresid == uio->uio_resid,
1084 ("%s: oresid = %d, uio_resid = %zd, sbused = %d",
1085 __func__, oresid, uio->uio_resid, sbused(sb)));
1087 error = handle_ddp(so, uio, flags, 0);
1093 /* Abort if socket has reported problems. */
1097 if (oresid > uio->uio_resid)
1099 error = so->so_error;
1100 if (!(flags & MSG_PEEK))
1105 /* Door is closed. Deliver what is left, if any. */
1106 if (sb->sb_state & SBS_CANTRCVMORE) {
1113 /* Socket buffer is empty and we shall not block. */
1114 if (sbused(sb) == 0 &&
1115 ((so->so_state & SS_NBIO) || (flags & (MSG_DONTWAIT|MSG_NBIO)))) {
1120 /* Socket buffer got some data that we shall deliver now. */
1121 if (sbused(sb) && !(flags & MSG_WAITALL) &&
1122 ((so->so_state & SS_NBIO) ||
1123 (flags & (MSG_DONTWAIT|MSG_NBIO)) ||
1124 sbused(sb) >= sb->sb_lowat ||
1125 sbused(sb) >= uio->uio_resid ||
1126 sbused(sb) >= sb->sb_hiwat) ) {
1130 /* On MSG_WAITALL we must wait until all data or error arrives. */
1131 if ((flags & MSG_WAITALL) &&
1132 (sbused(sb) >= uio->uio_resid || sbused(sb) >= sb->sb_lowat))
1136 * Wait and block until (more) data comes in.
1137 * NB: Drops the sockbuf lock during wait.
1141 if (sb->sb_flags & SB_DDP_INDICATE && !ddp_handled) {
1142 (void) handle_ddp(so, uio, flags, 1);
1150 SOCKBUF_LOCK_ASSERT(&so->so_rcv);
1151 KASSERT(sbused(sb) > 0, ("%s: sockbuf empty", __func__));
1152 KASSERT(sb->sb_mb != NULL, ("%s: sb_mb == NULL", __func__));
1154 if (sb->sb_flags & SB_DDP_INDICATE && !ddp_handled)
1159 uio->uio_td->td_ru.ru_msgrcv++;
1161 /* Fill uio until full or current end of socket buffer is reached. */
1162 len = min(uio->uio_resid, sbused(sb));
1164 /* Dequeue as many mbufs as possible. */
1165 if (!(flags & MSG_PEEK) && len >= sb->sb_mb->m_len) {
1166 for (*mp0 = m = sb->sb_mb;
1167 m != NULL && m->m_len <= len;
1170 uio->uio_resid -= m->m_len;
1175 if (sb->sb_mb == NULL)
1179 /* Copy the remainder. */
1181 KASSERT(sb->sb_mb != NULL,
1182 ("%s: len > 0 && sb->sb_mb empty", __func__));
1184 m = m_copym(sb->sb_mb, 0, len, M_NOWAIT);
1186 len = 0; /* Don't flush data from sockbuf. */
1188 uio->uio_resid -= m->m_len;
1199 /* NB: Must unlock socket buffer as uiomove may sleep. */
1201 error = m_mbuftouio_ddp(uio, sb->sb_mb, len);
1206 SBLASTRECORDCHK(sb);
1210 * Remove the delivered data from the socket buffer unless we
1211 * were only peeking.
1213 if (!(flags & MSG_PEEK)) {
1215 sbdrop_locked(sb, len);
1217 /* Notify protocol that we drained some data. */
1218 if ((so->so_proto->pr_flags & PR_WANTRCVD) &&
1219 (((flags & MSG_WAITALL) && uio->uio_resid > 0) ||
1220 !(flags & MSG_SOCALLBCK))) {
1223 (*so->so_proto->pr_usrreqs->pru_rcvd)(so, flags);
1229 * For MSG_WAITALL we may have to loop again and wait for
1230 * more data to come in.
1232 if ((flags & MSG_WAITALL) && uio->uio_resid > 0)
1235 SOCKBUF_LOCK_ASSERT(sb);
1236 SBLASTRECORDCHK(sb);