Index: stable/10/sys/dev/cxgbe/tom/t4_ddp.c =================================================================== --- stable/10/sys/dev/cxgbe/tom/t4_ddp.c (revision 309555) +++ stable/10/sys/dev/cxgbe/tom/t4_ddp.c (revision 309556) @@ -1,1296 +1,1497 @@ /*- * Copyright (c) 2012 Chelsio Communications, Inc. * All rights reserved. * Written by: Navdeep Parhar * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ #include __FBSDID("$FreeBSD$"); #include "opt_inet.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #define TCPSTATES #include #include #include #include #include #include #include #include #include #ifdef TCP_OFFLOAD #include "common/common.h" #include "common/t4_msg.h" #include "common/t4_regs.h" #include "common/t4_tcb.h" #include "tom/t4_tom.h" VNET_DECLARE(int, tcp_do_autorcvbuf); #define V_tcp_do_autorcvbuf VNET(tcp_do_autorcvbuf) VNET_DECLARE(int, tcp_autorcvbuf_inc); #define V_tcp_autorcvbuf_inc VNET(tcp_autorcvbuf_inc) VNET_DECLARE(int, tcp_autorcvbuf_max); #define V_tcp_autorcvbuf_max VNET(tcp_autorcvbuf_max) static struct mbuf *get_ddp_mbuf(int len); -#define PPOD_SZ(n) ((n) * sizeof(struct pagepod)) -#define PPOD_SIZE (PPOD_SZ(1)) - -/* XXX: must match A_ULP_RX_TDDP_PSZ */ -static int t4_ddp_pgsz[] = {4096, 4096 << 2, 4096 << 4, 4096 << 6}; - #define MAX_DDP_BUFFER_SIZE (M_TCB_RX_DDP_BUF0_LEN) -static int -alloc_ppods(struct tom_data *td, int n, u_int *ppod_addr) -{ - vmem_addr_t v; - int rc; - MPASS(n > 0); - - rc = vmem_alloc(td->ppod_arena, PPOD_SZ(n), M_NOWAIT | M_FIRSTFIT, &v); - *ppod_addr = (u_int)v; - - return (rc); -} - -static void -free_ppods(struct tom_data *td, u_int ppod_addr, int n) +static struct ddp_buffer * +alloc_ddp_buffer(vm_page_t *pages, int npages, int offset, int len) { + struct ddp_buffer *db; - MPASS(n > 0); + db = malloc(sizeof(*db), M_CXGBE, M_NOWAIT | M_ZERO); + if (db == NULL) { + CTR1(KTR_CXGBE, "%s: malloc failed.", __func__); + return (NULL); + } - vmem_free(td->ppod_arena, (vmem_addr_t)ppod_addr, PPOD_SZ(n)); -} + db->npages = npages; + db->pages = pages; + db->offset = offset; + db->len = len; -static inline int -pages_to_nppods(int npages, int ddp_pgsz) -{ - int nsegs = npages * PAGE_SIZE / ddp_pgsz; - - return (howmany(nsegs, PPOD_PAGES)); + return (db); } static void -free_ddp_buffer(struct tom_data *td, struct ddp_buffer *db) +free_ddp_buffer(struct ddp_buffer *db) { if (db == NULL) return; if (db->pages) free(db->pages, M_CXGBE); - if (db->nppods > 0) - free_ppods(td, db->ppod_addr, db->nppods); + if (db->prsv.prsv_nppods > 0) + t4_free_page_pods(&db->prsv); free(db, M_CXGBE); } void release_ddp_resources(struct toepcb *toep) { int i; for (i = 0; i < nitems(toep->db); i++) { if (toep->db[i] != NULL) { - free_ddp_buffer(toep->td, toep->db[i]); + free_ddp_buffer(toep->db[i]); toep->db[i] = NULL; } } } /* XXX: handle_ddp_data code duplication */ void insert_ddp_data(struct toepcb *toep, uint32_t n) { struct inpcb *inp = toep->inp; struct tcpcb *tp = intotcpcb(inp); struct sockbuf *sb = &inp->inp_socket->so_rcv; struct mbuf *m; INP_WLOCK_ASSERT(inp); SOCKBUF_LOCK_ASSERT(sb); m = get_ddp_mbuf(n); tp->rcv_nxt += n; #ifndef USE_DDP_RX_FLOW_CONTROL KASSERT(tp->rcv_wnd >= n, ("%s: negative window size", __func__)); tp->rcv_wnd -= n; #endif KASSERT(toep->sb_cc >= sb->sb_cc, ("%s: sb %p has more data (%d) than last time (%d).", __func__, sb, sb->sb_cc, toep->sb_cc)); toep->rx_credits += toep->sb_cc - sb->sb_cc; #ifdef USE_DDP_RX_FLOW_CONTROL toep->rx_credits -= n; /* adjust for F_RX_FC_DDP */ #endif sbappendstream_locked(sb, m); toep->sb_cc = sb->sb_cc; } /* SET_TCB_FIELD sent as a ULP command looks like this */ #define LEN__SET_TCB_FIELD_ULP (sizeof(struct ulp_txpkt) + \ sizeof(struct ulptx_idata) + sizeof(struct cpl_set_tcb_field_core)) /* RX_DATA_ACK sent as a ULP command looks like this */ #define LEN__RX_DATA_ACK_ULP (sizeof(struct ulp_txpkt) + \ sizeof(struct ulptx_idata) + sizeof(struct cpl_rx_data_ack_core)) static inline void * mk_set_tcb_field_ulp(struct ulp_txpkt *ulpmc, struct toepcb *toep, uint64_t word, uint64_t mask, uint64_t val) { struct ulptx_idata *ulpsc; struct cpl_set_tcb_field_core *req; ulpmc->cmd_dest = htonl(V_ULPTX_CMD(ULP_TX_PKT) | V_ULP_TXPKT_DEST(0)); ulpmc->len = htobe32(howmany(LEN__SET_TCB_FIELD_ULP, 16)); ulpsc = (struct ulptx_idata *)(ulpmc + 1); ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM)); ulpsc->len = htobe32(sizeof(*req)); req = (struct cpl_set_tcb_field_core *)(ulpsc + 1); OPCODE_TID(req) = htobe32(MK_OPCODE_TID(CPL_SET_TCB_FIELD, toep->tid)); req->reply_ctrl = htobe16(V_NO_REPLY(1) | V_QUEUENO(toep->ofld_rxq->iq.abs_id)); req->word_cookie = htobe16(V_WORD(word) | V_COOKIE(0)); req->mask = htobe64(mask); req->val = htobe64(val); ulpsc = (struct ulptx_idata *)(req + 1); if (LEN__SET_TCB_FIELD_ULP % 16) { ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_NOOP)); ulpsc->len = htobe32(0); return (ulpsc + 1); } return (ulpsc); } static inline void * mk_rx_data_ack_ulp(struct ulp_txpkt *ulpmc, struct toepcb *toep) { struct ulptx_idata *ulpsc; struct cpl_rx_data_ack_core *req; ulpmc->cmd_dest = htonl(V_ULPTX_CMD(ULP_TX_PKT) | V_ULP_TXPKT_DEST(0)); ulpmc->len = htobe32(howmany(LEN__RX_DATA_ACK_ULP, 16)); ulpsc = (struct ulptx_idata *)(ulpmc + 1); ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM)); ulpsc->len = htobe32(sizeof(*req)); req = (struct cpl_rx_data_ack_core *)(ulpsc + 1); OPCODE_TID(req) = htobe32(MK_OPCODE_TID(CPL_RX_DATA_ACK, toep->tid)); req->credit_dack = htobe32(F_RX_MODULATE_RX); ulpsc = (struct ulptx_idata *)(req + 1); if (LEN__RX_DATA_ACK_ULP % 16) { ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_NOOP)); ulpsc->len = htobe32(0); return (ulpsc + 1); } return (ulpsc); } static inline uint64_t select_ddp_flags(struct socket *so, int flags, int db_idx) { uint64_t ddp_flags = V_TF_DDP_INDICATE_OUT(0); int waitall = flags & MSG_WAITALL; int nb = so->so_state & SS_NBIO || flags & (MSG_DONTWAIT | MSG_NBIO); KASSERT(db_idx == 0 || db_idx == 1, ("%s: bad DDP buffer index %d", __func__, db_idx)); if (db_idx == 0) { ddp_flags |= V_TF_DDP_BUF0_VALID(1) | V_TF_DDP_ACTIVE_BUF(0); if (waitall) ddp_flags |= V_TF_DDP_PUSH_DISABLE_0(1); else if (nb) ddp_flags |= V_TF_DDP_BUF0_FLUSH(1); else ddp_flags |= V_TF_DDP_BUF0_FLUSH(0); } else { ddp_flags |= V_TF_DDP_BUF1_VALID(1) | V_TF_DDP_ACTIVE_BUF(1); if (waitall) ddp_flags |= V_TF_DDP_PUSH_DISABLE_1(1); else if (nb) ddp_flags |= V_TF_DDP_BUF1_FLUSH(1); else ddp_flags |= V_TF_DDP_BUF1_FLUSH(0); } return (ddp_flags); } static struct wrqe * mk_update_tcb_for_ddp(struct adapter *sc, struct toepcb *toep, int db_idx, int offset, uint64_t ddp_flags) { struct ddp_buffer *db = toep->db[db_idx]; struct wrqe *wr; struct work_request_hdr *wrh; struct ulp_txpkt *ulpmc; int len; KASSERT(db_idx == 0 || db_idx == 1, ("%s: bad DDP buffer index %d", __func__, db_idx)); /* * We'll send a compound work request that has 3 SET_TCB_FIELDs and an * RX_DATA_ACK (with RX_MODULATE to speed up delivery). * * The work request header is 16B and always ends at a 16B boundary. * The ULPTX master commands that follow must all end at 16B boundaries * too so we round up the size to 16. */ len = sizeof(*wrh) + 3 * roundup2(LEN__SET_TCB_FIELD_ULP, 16) + roundup2(LEN__RX_DATA_ACK_ULP, 16); wr = alloc_wrqe(len, toep->ctrlq); if (wr == NULL) return (NULL); wrh = wrtod(wr); INIT_ULPTX_WRH(wrh, len, 1, 0); /* atomic */ ulpmc = (struct ulp_txpkt *)(wrh + 1); /* Write the buffer's tag */ ulpmc = mk_set_tcb_field_ulp(ulpmc, toep, W_TCB_RX_DDP_BUF0_TAG + db_idx, V_TCB_RX_DDP_BUF0_TAG(M_TCB_RX_DDP_BUF0_TAG), - V_TCB_RX_DDP_BUF0_TAG(db->tag)); + V_TCB_RX_DDP_BUF0_TAG(db->prsv.prsv_tag)); /* Update the current offset in the DDP buffer and its total length */ if (db_idx == 0) ulpmc = mk_set_tcb_field_ulp(ulpmc, toep, W_TCB_RX_DDP_BUF0_OFFSET, V_TCB_RX_DDP_BUF0_OFFSET(M_TCB_RX_DDP_BUF0_OFFSET) | V_TCB_RX_DDP_BUF0_LEN(M_TCB_RX_DDP_BUF0_LEN), V_TCB_RX_DDP_BUF0_OFFSET(offset) | V_TCB_RX_DDP_BUF0_LEN(db->len)); else ulpmc = mk_set_tcb_field_ulp(ulpmc, toep, W_TCB_RX_DDP_BUF1_OFFSET, V_TCB_RX_DDP_BUF1_OFFSET(M_TCB_RX_DDP_BUF1_OFFSET) | V_TCB_RX_DDP_BUF1_LEN((u64)M_TCB_RX_DDP_BUF1_LEN << 32), V_TCB_RX_DDP_BUF1_OFFSET(offset) | V_TCB_RX_DDP_BUF1_LEN((u64)db->len << 32)); /* Update DDP flags */ ulpmc = mk_set_tcb_field_ulp(ulpmc, toep, W_TCB_RX_DDP_FLAGS, V_TF_DDP_BUF0_FLUSH(1) | V_TF_DDP_BUF1_FLUSH(1) | V_TF_DDP_PUSH_DISABLE_0(1) | V_TF_DDP_PUSH_DISABLE_1(1) | V_TF_DDP_BUF0_VALID(1) | V_TF_DDP_BUF1_VALID(1) | V_TF_DDP_ACTIVE_BUF(1) | V_TF_DDP_INDICATE_OUT(1), ddp_flags); /* Gratuitous RX_DATA_ACK with RX_MODULATE set to speed up delivery. */ ulpmc = mk_rx_data_ack_ulp(ulpmc, toep); return (wr); } static void discourage_ddp(struct toepcb *toep) { if (toep->ddp_score && --toep->ddp_score == 0) { toep->ddp_flags &= ~DDP_OK; toep->ddp_disabled = time_uptime; CTR3(KTR_CXGBE, "%s: tid %u !DDP_OK @ %u", __func__, toep->tid, time_uptime); } } static int handle_ddp_data(struct toepcb *toep, __be32 ddp_report, __be32 rcv_nxt, int len) { uint32_t report = be32toh(ddp_report); unsigned int db_flag; struct inpcb *inp = toep->inp; struct tcpcb *tp; struct socket *so; struct sockbuf *sb; struct mbuf *m; db_flag = report & F_DDP_BUF_IDX ? DDP_BUF1_ACTIVE : DDP_BUF0_ACTIVE; if (__predict_false(!(report & F_DDP_INV))) CXGBE_UNIMPLEMENTED("DDP buffer still valid"); INP_WLOCK(inp); so = inp_inpcbtosocket(inp); sb = &so->so_rcv; if (__predict_false(inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT))) { /* * XXX: think a bit more. * tcpcb probably gone, but socket should still be around * because we always wait for DDP completion in soreceive no * matter what. Just wake it up and let it clean up. */ CTR5(KTR_CXGBE, "%s: tid %u, seq 0x%x, len %d, inp_flags 0x%x", __func__, toep->tid, be32toh(rcv_nxt), len, inp->inp_flags); SOCKBUF_LOCK(sb); goto wakeup; } tp = intotcpcb(inp); /* * For RX_DDP_COMPLETE, len will be zero and rcv_nxt is the * sequence number of the next byte to receive. The length of * the data received for this message must be computed by * comparing the new and old values of rcv_nxt. * * For RX_DATA_DDP, len might be non-zero, but it is only the * length of the most recent DMA. It does not include the * total length of the data received since the previous update * for this DDP buffer. rcv_nxt is the sequence number of the * first received byte from the most recent DMA. */ len += be32toh(rcv_nxt) - tp->rcv_nxt; tp->rcv_nxt += len; tp->t_rcvtime = ticks; #ifndef USE_DDP_RX_FLOW_CONTROL KASSERT(tp->rcv_wnd >= len, ("%s: negative window size", __func__)); tp->rcv_wnd -= len; #endif m = get_ddp_mbuf(len); SOCKBUF_LOCK(sb); if (report & F_DDP_BUF_COMPLETE) toep->ddp_score = DDP_HIGH_SCORE; else discourage_ddp(toep); /* receive buffer autosize */ if (sb->sb_flags & SB_AUTOSIZE && V_tcp_do_autorcvbuf && sb->sb_hiwat < V_tcp_autorcvbuf_max && len > (sbspace(sb) / 8 * 7)) { unsigned int hiwat = sb->sb_hiwat; unsigned int newsize = min(hiwat + V_tcp_autorcvbuf_inc, V_tcp_autorcvbuf_max); if (!sbreserve_locked(sb, newsize, so, NULL)) sb->sb_flags &= ~SB_AUTOSIZE; else toep->rx_credits += newsize - hiwat; } KASSERT(toep->sb_cc >= sb->sb_cc, ("%s: sb %p has more data (%d) than last time (%d).", __func__, sb, sb->sb_cc, toep->sb_cc)); toep->rx_credits += toep->sb_cc - sb->sb_cc; #ifdef USE_DDP_RX_FLOW_CONTROL toep->rx_credits -= len; /* adjust for F_RX_FC_DDP */ #endif sbappendstream_locked(sb, m); toep->sb_cc = sb->sb_cc; wakeup: KASSERT(toep->ddp_flags & db_flag, ("%s: DDP buffer not active. toep %p, ddp_flags 0x%x, report 0x%x", __func__, toep, toep->ddp_flags, report)); toep->ddp_flags &= ~db_flag; sorwakeup_locked(so); SOCKBUF_UNLOCK_ASSERT(sb); INP_WUNLOCK(inp); return (0); } void handle_ddp_close(struct toepcb *toep, struct tcpcb *tp, struct sockbuf *sb, __be32 rcv_nxt) { struct mbuf *m; int len; SOCKBUF_LOCK_ASSERT(sb); INP_WLOCK_ASSERT(toep->inp); len = be32toh(rcv_nxt) - tp->rcv_nxt; /* Signal handle_ddp() to break out of its sleep loop. */ toep->ddp_flags &= ~(DDP_BUF0_ACTIVE | DDP_BUF1_ACTIVE); if (len == 0) return; tp->rcv_nxt += len; KASSERT(toep->sb_cc >= sb->sb_cc, ("%s: sb %p has more data (%d) than last time (%d).", __func__, sb, sb->sb_cc, toep->sb_cc)); toep->rx_credits += toep->sb_cc - sb->sb_cc; #ifdef USE_DDP_RX_FLOW_CONTROL toep->rx_credits -= len; /* adjust for F_RX_FC_DDP */ #endif m = get_ddp_mbuf(len); sbappendstream_locked(sb, m); toep->sb_cc = sb->sb_cc; } #define DDP_ERR (F_DDP_PPOD_MISMATCH | F_DDP_LLIMIT_ERR | F_DDP_ULIMIT_ERR |\ F_DDP_PPOD_PARITY_ERR | F_DDP_PADDING_ERR | F_DDP_OFFSET_ERR |\ F_DDP_INVALID_TAG | F_DDP_COLOR_ERR | F_DDP_TID_MISMATCH |\ F_DDP_INVALID_PPOD | F_DDP_HDRCRC_ERR | F_DDP_DATACRC_ERR) extern cpl_handler_t t4_cpl_handler[]; static int do_rx_data_ddp(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m) { struct adapter *sc = iq->adapter; const struct cpl_rx_data_ddp *cpl = (const void *)(rss + 1); unsigned int tid = GET_TID(cpl); uint32_t vld; struct toepcb *toep = lookup_tid(sc, tid); KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__)); KASSERT(toep->tid == tid, ("%s: toep tid/atid mismatch", __func__)); KASSERT(!(toep->flags & TPF_SYNQE), ("%s: toep %p claims to be a synq entry", __func__, toep)); vld = be32toh(cpl->ddpvld); if (__predict_false(vld & DDP_ERR)) { panic("%s: DDP error 0x%x (tid %d, toep %p)", __func__, vld, tid, toep); } if (toep->ulp_mode == ULP_MODE_ISCSI) { t4_cpl_handler[CPL_RX_ISCSI_DDP](iq, rss, m); return (0); } handle_ddp_data(toep, cpl->u.ddp_report, cpl->seq, be16toh(cpl->len)); return (0); } static int do_rx_ddp_complete(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m) { struct adapter *sc = iq->adapter; const struct cpl_rx_ddp_complete *cpl = (const void *)(rss + 1); unsigned int tid = GET_TID(cpl); struct toepcb *toep = lookup_tid(sc, tid); KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__)); KASSERT(toep->tid == tid, ("%s: toep tid/atid mismatch", __func__)); KASSERT(!(toep->flags & TPF_SYNQE), ("%s: toep %p claims to be a synq entry", __func__, toep)); handle_ddp_data(toep, cpl->ddp_report, cpl->rcv_nxt, 0); return (0); } void enable_ddp(struct adapter *sc, struct toepcb *toep) { KASSERT((toep->ddp_flags & (DDP_ON | DDP_OK | DDP_SC_REQ)) == DDP_OK, ("%s: toep %p has bad ddp_flags 0x%x", __func__, toep, toep->ddp_flags)); CTR3(KTR_CXGBE, "%s: tid %u (time %u)", __func__, toep->tid, time_uptime); toep->ddp_flags |= DDP_SC_REQ; t4_set_tcb_field(sc, toep->ctrlq, toep->tid, W_TCB_RX_DDP_FLAGS, V_TF_DDP_OFF(1) | V_TF_DDP_INDICATE_OUT(1) | V_TF_DDP_BUF0_INDICATE(1) | V_TF_DDP_BUF1_INDICATE(1) | V_TF_DDP_BUF0_VALID(1) | V_TF_DDP_BUF1_VALID(1), V_TF_DDP_BUF0_INDICATE(1) | V_TF_DDP_BUF1_INDICATE(1), 0, 0, toep->ofld_rxq->iq.abs_id); t4_set_tcb_field(sc, toep->ctrlq, toep->tid, W_TCB_T_FLAGS, V_TF_RCV_COALESCE_ENABLE(1), 0, 0, 0, toep->ofld_rxq->iq.abs_id); } static inline void disable_ddp(struct adapter *sc, struct toepcb *toep) { KASSERT((toep->ddp_flags & (DDP_ON | DDP_SC_REQ)) == DDP_ON, ("%s: toep %p has bad ddp_flags 0x%x", __func__, toep, toep->ddp_flags)); CTR3(KTR_CXGBE, "%s: tid %u (time %u)", __func__, toep->tid, time_uptime); toep->ddp_flags |= DDP_SC_REQ; t4_set_tcb_field(sc, toep->ctrlq, toep->tid, W_TCB_T_FLAGS, V_TF_RCV_COALESCE_ENABLE(1), V_TF_RCV_COALESCE_ENABLE(1), 0, 0, toep->ofld_rxq->iq.abs_id); t4_set_tcb_field(sc, toep->ctrlq, toep->tid, W_TCB_RX_DDP_FLAGS, V_TF_DDP_OFF(1), V_TF_DDP_OFF(1), 0, 0, toep->ofld_rxq->iq.abs_id); } static int hold_uio(struct uio *uio, vm_page_t **ppages, int *pnpages) { struct vm_map *map; struct iovec *iov; vm_offset_t start, end; vm_page_t *pp; int n; KASSERT(uio->uio_iovcnt == 1, ("%s: uio_iovcnt %d", __func__, uio->uio_iovcnt)); KASSERT(uio->uio_td->td_proc == curproc, ("%s: uio proc (%p) is not curproc (%p)", __func__, uio->uio_td->td_proc, curproc)); map = &curproc->p_vmspace->vm_map; iov = &uio->uio_iov[0]; start = trunc_page((uintptr_t)iov->iov_base); end = round_page((vm_offset_t)iov->iov_base + iov->iov_len); n = howmany(end - start, PAGE_SIZE); if (end - start > MAX_DDP_BUFFER_SIZE) return (E2BIG); pp = malloc(n * sizeof(vm_page_t), M_CXGBE, M_NOWAIT); if (pp == NULL) return (ENOMEM); if (vm_fault_quick_hold_pages(map, (vm_offset_t)iov->iov_base, iov->iov_len, VM_PROT_WRITE, pp, n) < 0) { free(pp, M_CXGBE); return (EFAULT); } *ppages = pp; *pnpages = n; return (0); } static int bufcmp(struct ddp_buffer *db, vm_page_t *pages, int npages, int offset, int len) { int i; if (db == NULL || db->npages != npages || db->offset != offset || db->len != len) return (1); for (i = 0; i < npages; i++) { if (pages[i]->phys_addr != db->pages[i]->phys_addr) return (1); } return (0); } static int calculate_hcf(int n1, int n2) { int a, b, t; if (n1 <= n2) { a = n1; b = n2; } else { a = n2; b = n1; } while (a != 0) { t = a; a = b % a; b = t; } return (b); } -static struct ddp_buffer * -alloc_ddp_buffer(struct tom_data *td, vm_page_t *pages, int npages, int offset, - int len) +static inline int +pages_to_nppods(int npages, int ddp_page_shift) { - int i, hcf, seglen, idx, ppod, nppods; - struct ddp_buffer *db; + MPASS(ddp_page_shift >= PAGE_SHIFT); + + return (howmany(npages >> (ddp_page_shift - PAGE_SHIFT), PPOD_PAGES)); +} + +static int +alloc_page_pods(struct ppod_region *pr, u_int nppods, u_int pgsz_idx, + struct ppod_reservation *prsv) +{ + vmem_addr_t addr; /* relative to start of region */ + + if (vmem_alloc(pr->pr_arena, PPOD_SZ(nppods), M_NOWAIT | M_FIRSTFIT, + &addr) != 0) + return (ENOMEM); + + CTR5(KTR_CXGBE, "%-17s arena %p, addr 0x%08x, nppods %d, pgsz %d", + __func__, pr->pr_arena, (uint32_t)addr & pr->pr_tag_mask, + nppods, 1 << pr->pr_page_shift[pgsz_idx]); + /* + * The hardware tagmask includes an extra invalid bit but the arena was + * seeded with valid values only. An allocation out of this arena will + * fit inside the tagmask but won't have the invalid bit set. + */ + MPASS((addr & pr->pr_tag_mask) == addr); + MPASS((addr & pr->pr_invalid_bit) == 0); + + prsv->prsv_pr = pr; + prsv->prsv_tag = V_PPOD_PGSZ(pgsz_idx) | addr; + prsv->prsv_nppods = nppods; + + return (0); +} + +int +t4_alloc_page_pods_for_db(struct ppod_region *pr, struct ddp_buffer *db) +{ + int i, hcf, seglen, idx, nppods; + struct ppod_reservation *prsv = &db->prsv; + + KASSERT(prsv->prsv_nppods == 0, + ("%s: page pods already allocated", __func__)); + + /* * The DDP page size is unrelated to the VM page size. We combine * contiguous physical pages into larger segments to get the best DDP * page size possible. This is the largest of the four sizes in * A_ULP_RX_TDDP_PSZ that evenly divides the HCF of the segment sizes in * the page list. */ hcf = 0; - for (i = 0; i < npages; i++) { + for (i = 0; i < db->npages; i++) { seglen = PAGE_SIZE; - while (i < npages - 1 && - pages[i]->phys_addr + PAGE_SIZE == pages[i + 1]->phys_addr) { + while (i < db->npages - 1 && + db->pages[i]->phys_addr + PAGE_SIZE == + db->pages[i + 1]->phys_addr) { seglen += PAGE_SIZE; i++; } hcf = calculate_hcf(hcf, seglen); - if (hcf < t4_ddp_pgsz[1]) { + if (hcf < (1 << pr->pr_page_shift[1])) { idx = 0; goto have_pgsz; /* give up, short circuit */ } } - if (hcf % t4_ddp_pgsz[0] != 0) { - /* hmmm. This could only happen when PAGE_SIZE < 4K */ - KASSERT(PAGE_SIZE < 4096, - ("%s: PAGE_SIZE %d, hcf %d", __func__, PAGE_SIZE, hcf)); - CTR3(KTR_CXGBE, "%s: PAGE_SIZE %d, hcf %d", - __func__, PAGE_SIZE, hcf); - return (NULL); - } - - for (idx = nitems(t4_ddp_pgsz) - 1; idx > 0; idx--) { - if (hcf % t4_ddp_pgsz[idx] == 0) +#define PR_PAGE_MASK(x) ((1 << pr->pr_page_shift[(x)]) - 1) + MPASS((hcf & PR_PAGE_MASK(0)) == 0); /* PAGE_SIZE is >= 4K everywhere */ + for (idx = nitems(pr->pr_page_shift) - 1; idx > 0; idx--) { + if ((hcf & PR_PAGE_MASK(idx)) == 0) break; } +#undef PR_PAGE_MASK + have_pgsz: MPASS(idx <= M_PPOD_PGSZ); - db = malloc(sizeof(*db), M_CXGBE, M_NOWAIT); - if (db == NULL) { - CTR1(KTR_CXGBE, "%s: malloc failed.", __func__); - return (NULL); + nppods = pages_to_nppods(db->npages, pr->pr_page_shift[idx]); + if (alloc_page_pods(pr, nppods, idx, prsv) != 0) + return (0); + MPASS(prsv->prsv_nppods > 0); + + return (1); +} + +int +t4_alloc_page_pods_for_buf(struct ppod_region *pr, vm_offset_t buf, int len, + struct ppod_reservation *prsv) +{ + int hcf, seglen, idx, npages, nppods; + uintptr_t start_pva, end_pva, pva, p1; + + MPASS(buf > 0); + MPASS(len > 0); + + /* + * The DDP page size is unrelated to the VM page size. We combine + * contiguous physical pages into larger segments to get the best DDP + * page size possible. This is the largest of the four sizes in + * A_ULP_RX_ISCSI_PSZ that evenly divides the HCF of the segment sizes + * in the page list. + */ + hcf = 0; + start_pva = trunc_page(buf); + end_pva = trunc_page(buf + len - 1); + pva = start_pva; + while (pva <= end_pva) { + seglen = PAGE_SIZE; + p1 = pmap_kextract(pva); + pva += PAGE_SIZE; + while (pva <= end_pva && p1 + seglen == pmap_kextract(pva)) { + seglen += PAGE_SIZE; + pva += PAGE_SIZE; + } + + hcf = calculate_hcf(hcf, seglen); + if (hcf < (1 << pr->pr_page_shift[1])) { + idx = 0; + goto have_pgsz; /* give up, short circuit */ + } } - nppods = pages_to_nppods(npages, t4_ddp_pgsz[idx]); - if (alloc_ppods(td, nppods, &db->ppod_addr) != 0) { - free(db, M_CXGBE); - CTR4(KTR_CXGBE, "%s: no pods, nppods %d, resid %d, pgsz %d", - __func__, nppods, len, t4_ddp_pgsz[idx]); - return (NULL); +#define PR_PAGE_MASK(x) ((1 << pr->pr_page_shift[(x)]) - 1) + MPASS((hcf & PR_PAGE_MASK(0)) == 0); /* PAGE_SIZE is >= 4K everywhere */ + for (idx = nitems(pr->pr_page_shift) - 1; idx > 0; idx--) { + if ((hcf & PR_PAGE_MASK(idx)) == 0) + break; } - ppod = (db->ppod_addr - td->ppod_start) / PPOD_SIZE; +#undef PR_PAGE_MASK - db->tag = V_PPOD_PGSZ(idx) | V_PPOD_TAG(ppod); - db->nppods = nppods; - db->npages = npages; - db->pages = pages; - db->offset = offset; - db->len = len; +have_pgsz: + MPASS(idx <= M_PPOD_PGSZ); - CTR6(KTR_CXGBE, "New DDP buffer. " - "ddp_pgsz %d, ppod 0x%x, npages %d, nppods %d, offset %d, len %d", - t4_ddp_pgsz[idx], ppod, db->npages, db->nppods, db->offset, - db->len); + npages = 1; + npages += (end_pva - start_pva) >> pr->pr_page_shift[idx]; + nppods = howmany(npages, PPOD_PAGES); + if (alloc_page_pods(pr, nppods, idx, prsv) != 0) + return (ENOMEM); + MPASS(prsv->prsv_nppods > 0); - return (db); + return (0); } +void +t4_free_page_pods(struct ppod_reservation *prsv) +{ + struct ppod_region *pr = prsv->prsv_pr; + vmem_addr_t addr; + + MPASS(prsv != NULL); + MPASS(prsv->prsv_nppods != 0); + + addr = prsv->prsv_tag & pr->pr_tag_mask; + MPASS((addr & pr->pr_invalid_bit) == 0); + + CTR4(KTR_CXGBE, "%-17s arena %p, addr 0x%08x, nppods %d", __func__, + pr->pr_arena, addr, prsv->prsv_nppods); + + vmem_free(pr->pr_arena, addr, PPOD_SZ(prsv->prsv_nppods)); + prsv->prsv_nppods = 0; +} + #define NUM_ULP_TX_SC_IMM_PPODS (256 / PPOD_SIZE) -static int -write_page_pods(struct adapter *sc, struct toepcb *toep, struct ddp_buffer *db) +int +t4_write_page_pods_for_db(struct adapter *sc, struct sge_wrq *wrq, int tid, + struct ddp_buffer *db) { struct wrqe *wr; struct ulp_mem_io *ulpmc; struct ulptx_idata *ulpsc; struct pagepod *ppod; int i, j, k, n, chunk, len, ddp_pgsz, idx; u_int ppod_addr; uint32_t cmd; + struct ppod_reservation *prsv = &db->prsv; + struct ppod_region *pr = prsv->prsv_pr; + MPASS(prsv->prsv_nppods > 0); + cmd = htobe32(V_ULPTX_CMD(ULP_TX_MEM_WRITE)); if (is_t4(sc)) cmd |= htobe32(F_ULP_MEMIO_ORDER); else cmd |= htobe32(F_T5_ULP_MEMIO_IMM); - ddp_pgsz = t4_ddp_pgsz[G_PPOD_PGSZ(db->tag)]; - ppod_addr = db->ppod_addr; - for (i = 0; i < db->nppods; ppod_addr += chunk) { + ddp_pgsz = 1 << pr->pr_page_shift[G_PPOD_PGSZ(prsv->prsv_tag)]; + ppod_addr = pr->pr_start + (prsv->prsv_tag & pr->pr_tag_mask); + for (i = 0; i < prsv->prsv_nppods; ppod_addr += chunk) { /* How many page pods are we writing in this cycle */ - n = min(db->nppods - i, NUM_ULP_TX_SC_IMM_PPODS); + n = min(prsv->prsv_nppods - i, NUM_ULP_TX_SC_IMM_PPODS); chunk = PPOD_SZ(n); len = roundup2(sizeof(*ulpmc) + sizeof(*ulpsc) + chunk, 16); - wr = alloc_wrqe(len, toep->ctrlq); + wr = alloc_wrqe(len, wrq); if (wr == NULL) return (ENOMEM); /* ok to just bail out */ ulpmc = wrtod(wr); INIT_ULPTX_WR(ulpmc, len, 0, 0); ulpmc->cmd = cmd; ulpmc->dlen = htobe32(V_ULP_MEMIO_DATA_LEN(chunk / 32)); ulpmc->len16 = htobe32(howmany(len - sizeof(ulpmc->wr), 16)); ulpmc->lock_addr = htobe32(V_ULP_MEMIO_ADDR(ppod_addr >> 5)); ulpsc = (struct ulptx_idata *)(ulpmc + 1); ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM)); ulpsc->len = htobe32(chunk); ppod = (struct pagepod *)(ulpsc + 1); for (j = 0; j < n; i++, j++, ppod++) { ppod->vld_tid_pgsz_tag_color = htobe64(F_PPOD_VALID | - V_PPOD_TID(toep->tid) | db->tag); + V_PPOD_TID(tid) | prsv->prsv_tag); ppod->len_offset = htobe64(V_PPOD_LEN(db->len) | V_PPOD_OFST(db->offset)); ppod->rsvd = 0; idx = i * PPOD_PAGES * (ddp_pgsz / PAGE_SIZE); for (k = 0; k < nitems(ppod->addr); k++) { if (idx < db->npages) { ppod->addr[k] = htobe64(db->pages[idx]->phys_addr); idx += ddp_pgsz / PAGE_SIZE; } else ppod->addr[k] = 0; #if 0 CTR5(KTR_CXGBE, "%s: tid %d ppod[%d]->addr[%d] = %p", __func__, toep->tid, i, k, htobe64(ppod->addr[k])); #endif } } t4_wrq_tx(sc, wr); } return (0); } +int +t4_write_page_pods_for_buf(struct adapter *sc, struct sge_wrq *wrq, int tid, + struct ppod_reservation *prsv, vm_offset_t buf, int buflen) +{ + struct wrqe *wr; + struct ulp_mem_io *ulpmc; + struct ulptx_idata *ulpsc; + struct pagepod *ppod; + int i, j, k, n, chunk, len, ddp_pgsz; + u_int ppod_addr, offset; + uint32_t cmd; + struct ppod_region *pr = prsv->prsv_pr; + uintptr_t end_pva, pva, pa; + + cmd = htobe32(V_ULPTX_CMD(ULP_TX_MEM_WRITE)); + if (is_t4(sc)) + cmd |= htobe32(F_ULP_MEMIO_ORDER); + else + cmd |= htobe32(F_T5_ULP_MEMIO_IMM); + ddp_pgsz = 1 << pr->pr_page_shift[G_PPOD_PGSZ(prsv->prsv_tag)]; + offset = buf & PAGE_MASK; + ppod_addr = pr->pr_start + (prsv->prsv_tag & pr->pr_tag_mask); + pva = trunc_page(buf); + end_pva = trunc_page(buf + buflen - 1); + for (i = 0; i < prsv->prsv_nppods; ppod_addr += chunk) { + + /* How many page pods are we writing in this cycle */ + n = min(prsv->prsv_nppods - i, NUM_ULP_TX_SC_IMM_PPODS); + MPASS(n > 0); + chunk = PPOD_SZ(n); + len = roundup2(sizeof(*ulpmc) + sizeof(*ulpsc) + chunk, 16); + + wr = alloc_wrqe(len, wrq); + if (wr == NULL) + return (ENOMEM); /* ok to just bail out */ + ulpmc = wrtod(wr); + + INIT_ULPTX_WR(ulpmc, len, 0, 0); + ulpmc->cmd = cmd; + ulpmc->dlen = htobe32(V_ULP_MEMIO_DATA_LEN(chunk / 32)); + ulpmc->len16 = htobe32(howmany(len - sizeof(ulpmc->wr), 16)); + ulpmc->lock_addr = htobe32(V_ULP_MEMIO_ADDR(ppod_addr >> 5)); + + ulpsc = (struct ulptx_idata *)(ulpmc + 1); + ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM)); + ulpsc->len = htobe32(chunk); + + ppod = (struct pagepod *)(ulpsc + 1); + for (j = 0; j < n; i++, j++, ppod++) { + ppod->vld_tid_pgsz_tag_color = htobe64(F_PPOD_VALID | + V_PPOD_TID(tid) | + (prsv->prsv_tag & ~V_PPOD_PGSZ(M_PPOD_PGSZ))); + ppod->len_offset = htobe64(V_PPOD_LEN(buflen) | + V_PPOD_OFST(offset)); + ppod->rsvd = 0; + + for (k = 0; k < nitems(ppod->addr); k++) { + if (pva > end_pva) + ppod->addr[k] = 0; + else { + pa = pmap_kextract(pva); + ppod->addr[k] = htobe64(pa); + pva += ddp_pgsz; + } +#if 0 + CTR5(KTR_CXGBE, + "%s: tid %d ppod[%d]->addr[%d] = %p", + __func__, tid, i, k, + htobe64(ppod->addr[k])); +#endif + } + + /* + * Walk back 1 segment so that the first address in the + * next pod is the same as the last one in the current + * pod. + */ + pva -= ddp_pgsz; + } + + t4_wrq_tx(sc, wr); + } + + MPASS(pva <= end_pva); + + return (0); +} + /* * Reuse, or allocate (and program the page pods for) a new DDP buffer. The * "pages" array is handed over to this function and should not be used in any * way by the caller after that. */ static int select_ddp_buffer(struct adapter *sc, struct toepcb *toep, vm_page_t *pages, int npages, int db_off, int db_len) { struct ddp_buffer *db; struct tom_data *td = sc->tom_softc; int i, empty_slot = -1; /* Try to reuse */ for (i = 0; i < nitems(toep->db); i++) { if (bufcmp(toep->db[i], pages, npages, db_off, db_len) == 0) { free(pages, M_CXGBE); return (i); /* pages still held */ } else if (toep->db[i] == NULL && empty_slot < 0) empty_slot = i; } /* Allocate new buffer, write its page pods. */ - db = alloc_ddp_buffer(td, pages, npages, db_off, db_len); + db = alloc_ddp_buffer(pages, npages, db_off, db_len); if (db == NULL) { vm_page_unhold_pages(pages, npages); free(pages, M_CXGBE); return (-1); } - if (write_page_pods(sc, toep, db) != 0) { + if (t4_alloc_page_pods_for_db(&td->pr, db)) { vm_page_unhold_pages(pages, npages); - free_ddp_buffer(td, db); + free_ddp_buffer(db); return (-1); } + if (t4_write_page_pods_for_db(sc, toep->ctrlq, toep->tid, db) != 0) { + vm_page_unhold_pages(pages, npages); + free_ddp_buffer(db); + return (-1); + } i = empty_slot; if (i < 0) { i = arc4random() % nitems(toep->db); - free_ddp_buffer(td, toep->db[i]); + free_ddp_buffer(toep->db[i]); } toep->db[i] = db; CTR5(KTR_CXGBE, "%s: tid %d, DDP buffer[%d] = %p (tag 0x%x)", - __func__, toep->tid, i, db, db->tag); + __func__, toep->tid, i, db, db->prsv.prsv_tag); return (i); } static void wire_ddp_buffer(struct ddp_buffer *db) { int i; vm_page_t p; for (i = 0; i < db->npages; i++) { p = db->pages[i]; vm_page_lock(p); vm_page_wire(p); vm_page_unhold(p); vm_page_unlock(p); } } static void unwire_ddp_buffer(struct ddp_buffer *db) { int i; vm_page_t p; for (i = 0; i < db->npages; i++) { p = db->pages[i]; vm_page_lock(p); vm_page_unwire(p, 0); vm_page_unlock(p); } } static int handle_ddp(struct socket *so, struct uio *uio, int flags, int error) { struct sockbuf *sb = &so->so_rcv; struct tcpcb *tp = so_sototcpcb(so); struct toepcb *toep = tp->t_toe; struct adapter *sc = td_adapter(toep->td); vm_page_t *pages; int npages, db_idx, rc, buf_flag; struct ddp_buffer *db; struct wrqe *wr; uint64_t ddp_flags; SOCKBUF_LOCK_ASSERT(sb); #if 0 if (sb->sb_cc + sc->tt.ddp_thres > uio->uio_resid) { CTR4(KTR_CXGBE, "%s: sb_cc %d, threshold %d, resid %d", __func__, sb->sb_cc, sc->tt.ddp_thres, uio->uio_resid); } #endif /* XXX: too eager to disable DDP, could handle NBIO better than this. */ if (sb->sb_cc >= uio->uio_resid || uio->uio_resid < sc->tt.ddp_thres || uio->uio_resid > MAX_DDP_BUFFER_SIZE || uio->uio_iovcnt > 1 || so->so_state & SS_NBIO || flags & (MSG_DONTWAIT | MSG_NBIO) || error || so->so_error || sb->sb_state & SBS_CANTRCVMORE) goto no_ddp; /* * Fault in and then hold the pages of the uio buffers. We'll wire them * a bit later if everything else works out. */ SOCKBUF_UNLOCK(sb); if (hold_uio(uio, &pages, &npages) != 0) { SOCKBUF_LOCK(sb); goto no_ddp; } SOCKBUF_LOCK(sb); if (__predict_false(so->so_error || sb->sb_state & SBS_CANTRCVMORE)) { vm_page_unhold_pages(pages, npages); free(pages, M_CXGBE); goto no_ddp; } /* * Figure out which one of the two DDP buffers to use this time. */ db_idx = select_ddp_buffer(sc, toep, pages, npages, (uintptr_t)uio->uio_iov->iov_base & PAGE_MASK, uio->uio_resid); pages = NULL; /* handed off to select_ddp_buffer */ if (db_idx < 0) goto no_ddp; db = toep->db[db_idx]; buf_flag = db_idx == 0 ? DDP_BUF0_ACTIVE : DDP_BUF1_ACTIVE; /* * Build the compound work request that tells the chip where to DMA the * payload. */ ddp_flags = select_ddp_flags(so, flags, db_idx); wr = mk_update_tcb_for_ddp(sc, toep, db_idx, sb->sb_cc, ddp_flags); if (wr == NULL) { /* * Just unhold the pages. The DDP buffer's software state is * left as-is in the toep. The page pods were written * successfully and we may have an opportunity to use it in the * future. */ vm_page_unhold_pages(db->pages, db->npages); goto no_ddp; } /* Wire (and then unhold) the pages, and give the chip the go-ahead. */ wire_ddp_buffer(db); t4_wrq_tx(sc, wr); sb->sb_flags &= ~SB_DDP_INDICATE; toep->ddp_flags |= buf_flag; /* * Wait for the DDP operation to complete and then unwire the pages. * The return code from the sbwait will be the final return code of this * function. But we do need to wait for DDP no matter what. */ rc = sbwait(sb); while (toep->ddp_flags & buf_flag) { sb->sb_flags |= SB_WAIT; msleep(&sb->sb_cc, &sb->sb_mtx, PSOCK , "sbwait", 0); } unwire_ddp_buffer(db); return (rc); no_ddp: disable_ddp(sc, toep); discourage_ddp(toep); sb->sb_flags &= ~SB_DDP_INDICATE; return (0); } -void -t4_init_ddp(struct adapter *sc, struct tom_data *td) +int +t4_init_ppod_region(struct ppod_region *pr, struct t4_range *r, u_int psz, + const char *name) { + int i; - td->ppod_start = sc->vres.ddp.start; - td->ppod_arena = vmem_create("DDP page pods", sc->vres.ddp.start, - sc->vres.ddp.size, 1, 32, M_FIRSTFIT | M_NOWAIT); + MPASS(pr != NULL); + MPASS(r->size > 0); + + pr->pr_start = r->start; + pr->pr_len = r->size; + pr->pr_page_shift[0] = 12 + G_HPZ0(psz); + pr->pr_page_shift[1] = 12 + G_HPZ1(psz); + pr->pr_page_shift[2] = 12 + G_HPZ2(psz); + pr->pr_page_shift[3] = 12 + G_HPZ3(psz); + + /* The SGL -> page pod algorithm requires the sizes to be in order. */ + for (i = 1; i < nitems(pr->pr_page_shift); i++) { + if (pr->pr_page_shift[i] <= pr->pr_page_shift[i - 1]) + return (ENXIO); + } + + pr->pr_tag_mask = ((1 << fls(r->size)) - 1) & V_PPOD_TAG(M_PPOD_TAG); + pr->pr_alias_mask = V_PPOD_TAG(M_PPOD_TAG) & ~pr->pr_tag_mask; + if (pr->pr_tag_mask == 0 || pr->pr_alias_mask == 0) + return (ENXIO); + pr->pr_alias_shift = fls(pr->pr_tag_mask); + pr->pr_invalid_bit = 1 << (pr->pr_alias_shift - 1); + + pr->pr_arena = vmem_create(name, 0, pr->pr_len, PPOD_SIZE, 0, + M_FIRSTFIT | M_NOWAIT); + if (pr->pr_arena == NULL) + return (ENOMEM); + + return (0); } void -t4_uninit_ddp(struct adapter *sc __unused, struct tom_data *td) +t4_free_ppod_region(struct ppod_region *pr) { - if (td->ppod_arena != NULL) { - vmem_destroy(td->ppod_arena); - td->ppod_arena = NULL; - } + MPASS(pr != NULL); + + if (pr->pr_arena) + vmem_destroy(pr->pr_arena); + bzero(pr, sizeof(*pr)); } #define VNET_SO_ASSERT(so) \ VNET_ASSERT(curvnet != NULL, \ ("%s:%d curvnet is NULL, so=%p", __func__, __LINE__, (so))); #define SBLOCKWAIT(f) (((f) & MSG_DONTWAIT) ? 0 : SBL_WAIT) static int soreceive_rcvoob(struct socket *so, struct uio *uio, int flags) { CXGBE_UNIMPLEMENTED(__func__); } static char ddp_magic_str[] = "nothing to see here"; static struct mbuf * get_ddp_mbuf(int len) { struct mbuf *m; m = m_get(M_NOWAIT, MT_DATA); if (m == NULL) CXGBE_UNIMPLEMENTED("mbuf alloc failure"); m->m_len = len; m->m_data = &ddp_magic_str[0]; return (m); } static inline int is_ddp_mbuf(struct mbuf *m) { return (m->m_data == &ddp_magic_str[0]); } /* * Copy an mbuf chain into a uio limited by len if set. */ static int m_mbuftouio_ddp(struct uio *uio, struct mbuf *m, int len) { int error, length, total; int progress = 0; if (len > 0) total = min(uio->uio_resid, len); else total = uio->uio_resid; /* Fill the uio with data from the mbufs. */ for (; m != NULL; m = m->m_next) { length = min(m->m_len, total - progress); if (is_ddp_mbuf(m)) { enum uio_seg segflag = uio->uio_segflg; uio->uio_segflg = UIO_NOCOPY; error = uiomove(mtod(m, void *), length, uio); uio->uio_segflg = segflag; } else error = uiomove(mtod(m, void *), length, uio); if (error) return (error); progress += length; } return (0); } /* * Based on soreceive_stream() in uipc_socket.c */ int t4_soreceive_ddp(struct socket *so, struct sockaddr **psa, struct uio *uio, struct mbuf **mp0, struct mbuf **controlp, int *flagsp) { int len = 0, error = 0, flags, oresid, ddp_handled = 0; struct sockbuf *sb; struct mbuf *m, *n = NULL; /* We only do stream sockets. */ if (so->so_type != SOCK_STREAM) return (EINVAL); if (psa != NULL) *psa = NULL; if (controlp != NULL) return (EINVAL); if (flagsp != NULL) flags = *flagsp &~ MSG_EOR; else flags = 0; if (flags & MSG_OOB) return (soreceive_rcvoob(so, uio, flags)); if (mp0 != NULL) *mp0 = NULL; sb = &so->so_rcv; /* Prevent other readers from entering the socket. */ error = sblock(sb, SBLOCKWAIT(flags)); SOCKBUF_LOCK(sb); if (error) goto out; /* Easy one, no space to copyout anything. */ if (uio->uio_resid == 0) { error = EINVAL; goto out; } oresid = uio->uio_resid; /* We will never ever get anything unless we are or were connected. */ if (!(so->so_state & (SS_ISCONNECTED|SS_ISDISCONNECTED))) { error = ENOTCONN; goto out; } restart: SOCKBUF_LOCK_ASSERT(&so->so_rcv); if (sb->sb_flags & SB_DDP_INDICATE && !ddp_handled) { /* uio should be just as it was at entry */ KASSERT(oresid == uio->uio_resid, ("%s: oresid = %d, uio_resid = %zd, sb_cc = %d", __func__, oresid, uio->uio_resid, sb->sb_cc)); error = handle_ddp(so, uio, flags, 0); ddp_handled = 1; if (error) goto out; } /* Abort if socket has reported problems. */ if (so->so_error) { if (sb->sb_cc > 0) goto deliver; if (oresid > uio->uio_resid) goto out; error = so->so_error; if (!(flags & MSG_PEEK)) so->so_error = 0; goto out; } /* Door is closed. Deliver what is left, if any. */ if (sb->sb_state & SBS_CANTRCVMORE) { if (sb->sb_cc > 0) goto deliver; else goto out; } /* Socket buffer is empty and we shall not block. */ if (sb->sb_cc == 0 && ((so->so_state & SS_NBIO) || (flags & (MSG_DONTWAIT|MSG_NBIO)))) { error = EAGAIN; goto out; } /* Socket buffer got some data that we shall deliver now. */ if (sb->sb_cc > 0 && !(flags & MSG_WAITALL) && ((so->so_state & SS_NBIO) || (flags & (MSG_DONTWAIT|MSG_NBIO)) || sb->sb_cc >= sb->sb_lowat || sb->sb_cc >= uio->uio_resid || sb->sb_cc >= sb->sb_hiwat) ) { goto deliver; } /* On MSG_WAITALL we must wait until all data or error arrives. */ if ((flags & MSG_WAITALL) && (sb->sb_cc >= uio->uio_resid || sb->sb_cc >= sb->sb_lowat)) goto deliver; /* * Wait and block until (more) data comes in. * NB: Drops the sockbuf lock during wait. */ error = sbwait(sb); if (error) { if (sb->sb_flags & SB_DDP_INDICATE && !ddp_handled) { (void) handle_ddp(so, uio, flags, 1); ddp_handled = 1; } goto out; } goto restart; deliver: SOCKBUF_LOCK_ASSERT(&so->so_rcv); KASSERT(sb->sb_cc > 0, ("%s: sockbuf empty", __func__)); KASSERT(sb->sb_mb != NULL, ("%s: sb_mb == NULL", __func__)); if (sb->sb_flags & SB_DDP_INDICATE && !ddp_handled) goto restart; /* Statistics. */ if (uio->uio_td) uio->uio_td->td_ru.ru_msgrcv++; /* Fill uio until full or current end of socket buffer is reached. */ len = min(uio->uio_resid, sb->sb_cc); if (mp0 != NULL) { /* Dequeue as many mbufs as possible. */ if (!(flags & MSG_PEEK) && len >= sb->sb_mb->m_len) { for (*mp0 = m = sb->sb_mb; m != NULL && m->m_len <= len; m = m->m_next) { len -= m->m_len; uio->uio_resid -= m->m_len; sbfree(sb, m); n = m; } sb->sb_mb = m; if (sb->sb_mb == NULL) SB_EMPTY_FIXUP(sb); n->m_next = NULL; } /* Copy the remainder. */ if (len > 0) { KASSERT(sb->sb_mb != NULL, ("%s: len > 0 && sb->sb_mb empty", __func__)); m = m_copym(sb->sb_mb, 0, len, M_NOWAIT); if (m == NULL) len = 0; /* Don't flush data from sockbuf. */ else uio->uio_resid -= m->m_len; if (*mp0 != NULL) n->m_next = m; else *mp0 = m; if (*mp0 == NULL) { error = ENOBUFS; goto out; } } } else { /* NB: Must unlock socket buffer as uiomove may sleep. */ SOCKBUF_UNLOCK(sb); error = m_mbuftouio_ddp(uio, sb->sb_mb, len); SOCKBUF_LOCK(sb); if (error) goto out; } SBLASTRECORDCHK(sb); SBLASTMBUFCHK(sb); /* * Remove the delivered data from the socket buffer unless we * were only peeking. */ if (!(flags & MSG_PEEK)) { if (len > 0) sbdrop_locked(sb, len); /* Notify protocol that we drained some data. */ if ((so->so_proto->pr_flags & PR_WANTRCVD) && (((flags & MSG_WAITALL) && uio->uio_resid > 0) || !(flags & MSG_SOCALLBCK))) { SOCKBUF_UNLOCK(sb); VNET_SO_ASSERT(so); (*so->so_proto->pr_usrreqs->pru_rcvd)(so, flags); SOCKBUF_LOCK(sb); } } /* * For MSG_WAITALL we may have to loop again and wait for * more data to come in. */ if ((flags & MSG_WAITALL) && uio->uio_resid > 0) goto restart; out: SOCKBUF_LOCK_ASSERT(sb); SBLASTRECORDCHK(sb); SBLASTMBUFCHK(sb); SOCKBUF_UNLOCK(sb); sbunlock(sb); return (error); } int t4_ddp_mod_load(void) { t4_register_cpl_handler(CPL_RX_DATA_DDP, do_rx_data_ddp); t4_register_cpl_handler(CPL_RX_DDP_COMPLETE, do_rx_ddp_complete); return (0); } void t4_ddp_mod_unload(void) { t4_register_cpl_handler(CPL_RX_DATA_DDP, NULL); t4_register_cpl_handler(CPL_RX_DDP_COMPLETE, NULL); } #endif Index: stable/10/sys/dev/cxgbe/tom/t4_tom.c =================================================================== --- stable/10/sys/dev/cxgbe/tom/t4_tom.c (revision 309555) +++ stable/10/sys/dev/cxgbe/tom/t4_tom.c (revision 309556) @@ -1,1241 +1,1245 @@ /*- * Copyright (c) 2012 Chelsio Communications, Inc. * All rights reserved. * Written by: Navdeep Parhar * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ #include __FBSDID("$FreeBSD$"); #include "opt_inet.h" #include "opt_inet6.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #define TCPSTATES #include #include #ifdef TCP_OFFLOAD #include "common/common.h" #include "common/t4_msg.h" #include "common/t4_regs.h" #include "common/t4_regs_values.h" #include "common/t4_tcb.h" #include "tom/t4_tom_l2t.h" #include "tom/t4_tom.h" static struct protosw ddp_protosw; static struct pr_usrreqs ddp_usrreqs; static struct protosw ddp6_protosw; static struct pr_usrreqs ddp6_usrreqs; /* Module ops */ static int t4_tom_mod_load(void); static int t4_tom_mod_unload(void); static int t4_tom_modevent(module_t, int, void *); /* ULD ops and helpers */ static int t4_tom_activate(struct adapter *); static int t4_tom_deactivate(struct adapter *); static struct uld_info tom_uld_info = { .uld_id = ULD_TOM, .activate = t4_tom_activate, .deactivate = t4_tom_deactivate, }; static void queue_tid_release(struct adapter *, int); static void release_offload_resources(struct toepcb *); static int alloc_tid_tabs(struct tid_info *); static void free_tid_tabs(struct tid_info *); static int add_lip(struct adapter *, struct in6_addr *); static int delete_lip(struct adapter *, struct in6_addr *); static struct clip_entry *search_lip(struct tom_data *, struct in6_addr *); static void init_clip_table(struct adapter *, struct tom_data *); static void update_clip(struct adapter *, void *); static void t4_clip_task(void *, int); static void update_clip_table(struct adapter *, struct tom_data *); static void destroy_clip_table(struct adapter *, struct tom_data *); static void free_tom_data(struct adapter *, struct tom_data *); static void reclaim_wr_resources(void *, int); static int in6_ifaddr_gen; static eventhandler_tag ifaddr_evhandler; static struct timeout_task clip_task; static void mbufq_init(struct mbufq *q, int limit) { q->head = q->tail = NULL; } static void mbufq_drain(struct mbufq *q) { struct mbuf *m; while ((m = q->head) != NULL) { q->head = m->m_nextpkt; m_freem(m); } q->tail = NULL; } #ifdef INVARIANTS static inline int mbufq_len(const struct mbufq *q) { struct mbuf *m; int len; len = 0; for (m = q->head; m != NULL; m = m->m_nextpkt) len++; return (len); } #endif struct toepcb * alloc_toepcb(struct vi_info *vi, int txqid, int rxqid, int flags) { struct port_info *pi = vi->pi; struct adapter *sc = pi->adapter; struct toepcb *toep; int tx_credits, txsd_total, len; /* * The firmware counts tx work request credits in units of 16 bytes * each. Reserve room for an ABORT_REQ so the driver never has to worry * about tx credits if it wants to abort a connection. */ tx_credits = sc->params.ofldq_wr_cred; tx_credits -= howmany(sizeof(struct cpl_abort_req), 16); /* * Shortest possible tx work request is a fw_ofld_tx_data_wr + 1 byte * immediate payload, and firmware counts tx work request credits in * units of 16 byte. Calculate the maximum work requests possible. */ txsd_total = tx_credits / howmany(sizeof(struct fw_ofld_tx_data_wr) + 1, 16); if (txqid < 0) txqid = (arc4random() % vi->nofldtxq) + vi->first_ofld_txq; KASSERT(txqid >= vi->first_ofld_txq && txqid < vi->first_ofld_txq + vi->nofldtxq, ("%s: txqid %d for vi %p (first %d, n %d)", __func__, txqid, vi, vi->first_ofld_txq, vi->nofldtxq)); if (rxqid < 0) rxqid = (arc4random() % vi->nofldrxq) + vi->first_ofld_rxq; KASSERT(rxqid >= vi->first_ofld_rxq && rxqid < vi->first_ofld_rxq + vi->nofldrxq, ("%s: rxqid %d for vi %p (first %d, n %d)", __func__, rxqid, vi, vi->first_ofld_rxq, vi->nofldrxq)); len = offsetof(struct toepcb, txsd) + txsd_total * sizeof(struct ofld_tx_sdesc); toep = malloc(len, M_CXGBE, M_ZERO | flags); if (toep == NULL) return (NULL); toep->td = sc->tom_softc; toep->vi = vi; toep->tx_total = tx_credits; toep->tx_credits = tx_credits; toep->ofld_txq = &sc->sge.ofld_txq[txqid]; toep->ofld_rxq = &sc->sge.ofld_rxq[rxqid]; toep->ctrlq = &sc->sge.ctrlq[pi->port_id]; mbufq_init(&toep->ulp_pduq, INT_MAX); mbufq_init(&toep->ulp_pdu_reclaimq, INT_MAX); toep->txsd_total = txsd_total; toep->txsd_avail = txsd_total; toep->txsd_pidx = 0; toep->txsd_cidx = 0; return (toep); } void free_toepcb(struct toepcb *toep) { KASSERT(!(toep->flags & TPF_ATTACHED), ("%s: attached to an inpcb", __func__)); KASSERT(!(toep->flags & TPF_CPL_PENDING), ("%s: CPL pending", __func__)); free(toep, M_CXGBE); } /* * Set up the socket for TCP offload. */ void offload_socket(struct socket *so, struct toepcb *toep) { struct tom_data *td = toep->td; struct inpcb *inp = sotoinpcb(so); struct tcpcb *tp = intotcpcb(inp); struct sockbuf *sb; INP_WLOCK_ASSERT(inp); /* Update socket */ sb = &so->so_snd; SOCKBUF_LOCK(sb); sb->sb_flags |= SB_NOCOALESCE; SOCKBUF_UNLOCK(sb); sb = &so->so_rcv; SOCKBUF_LOCK(sb); sb->sb_flags |= SB_NOCOALESCE; if (toep->ulp_mode == ULP_MODE_TCPDDP) { if (inp->inp_vflag & INP_IPV6) so->so_proto = &ddp6_protosw; else so->so_proto = &ddp_protosw; } SOCKBUF_UNLOCK(sb); /* Update TCP PCB */ tp->tod = &td->tod; tp->t_toe = toep; tp->t_flags |= TF_TOE; /* Install an extra hold on inp */ toep->inp = inp; toep->flags |= TPF_ATTACHED; in_pcbref(inp); /* Add the TOE PCB to the active list */ mtx_lock(&td->toep_list_lock); TAILQ_INSERT_HEAD(&td->toep_list, toep, link); mtx_unlock(&td->toep_list_lock); } /* This is _not_ the normal way to "unoffload" a socket. */ void undo_offload_socket(struct socket *so) { struct inpcb *inp = sotoinpcb(so); struct tcpcb *tp = intotcpcb(inp); struct toepcb *toep = tp->t_toe; struct tom_data *td = toep->td; struct sockbuf *sb; INP_WLOCK_ASSERT(inp); sb = &so->so_snd; SOCKBUF_LOCK(sb); sb->sb_flags &= ~SB_NOCOALESCE; SOCKBUF_UNLOCK(sb); sb = &so->so_rcv; SOCKBUF_LOCK(sb); sb->sb_flags &= ~SB_NOCOALESCE; SOCKBUF_UNLOCK(sb); tp->tod = NULL; tp->t_toe = NULL; tp->t_flags &= ~TF_TOE; toep->inp = NULL; toep->flags &= ~TPF_ATTACHED; if (in_pcbrele_wlocked(inp)) panic("%s: inp freed.", __func__); mtx_lock(&td->toep_list_lock); TAILQ_REMOVE(&td->toep_list, toep, link); mtx_unlock(&td->toep_list_lock); } static void release_offload_resources(struct toepcb *toep) { struct tom_data *td = toep->td; struct adapter *sc = td_adapter(td); int tid = toep->tid; KASSERT(!(toep->flags & TPF_CPL_PENDING), ("%s: %p has CPL pending.", __func__, toep)); KASSERT(!(toep->flags & TPF_ATTACHED), ("%s: %p is still attached.", __func__, toep)); CTR5(KTR_CXGBE, "%s: toep %p (tid %d, l2te %p, ce %p)", __func__, toep, tid, toep->l2te, toep->ce); /* * These queues should have been emptied at approximately the same time * that a normal connection's socket's so_snd would have been purged or * drained. Do _not_ clean up here. */ MPASS(mbufq_len(&toep->ulp_pduq) == 0); MPASS(mbufq_len(&toep->ulp_pdu_reclaimq) == 0); if (toep->ulp_mode == ULP_MODE_TCPDDP) release_ddp_resources(toep); if (toep->l2te) t4_l2t_release(toep->l2te); if (tid >= 0) { remove_tid(sc, tid); release_tid(sc, tid, toep->ctrlq); } if (toep->ce) release_lip(td, toep->ce); mtx_lock(&td->toep_list_lock); TAILQ_REMOVE(&td->toep_list, toep, link); mtx_unlock(&td->toep_list_lock); free_toepcb(toep); } /* * The kernel is done with the TCP PCB and this is our opportunity to unhook the * toepcb hanging off of it. If the TOE driver is also done with the toepcb (no * pending CPL) then it is time to release all resources tied to the toepcb. * * Also gets called when an offloaded active open fails and the TOM wants the * kernel to take the TCP PCB back. */ static void t4_pcb_detach(struct toedev *tod __unused, struct tcpcb *tp) { #if defined(KTR) || defined(INVARIANTS) struct inpcb *inp = tp->t_inpcb; #endif struct toepcb *toep = tp->t_toe; INP_WLOCK_ASSERT(inp); KASSERT(toep != NULL, ("%s: toep is NULL", __func__)); KASSERT(toep->flags & TPF_ATTACHED, ("%s: not attached", __func__)); #ifdef KTR if (tp->t_state == TCPS_SYN_SENT) { CTR6(KTR_CXGBE, "%s: atid %d, toep %p (0x%x), inp %p (0x%x)", __func__, toep->tid, toep, toep->flags, inp, inp->inp_flags); } else { CTR6(KTR_CXGBE, "t4_pcb_detach: tid %d (%s), toep %p (0x%x), inp %p (0x%x)", toep->tid, tcpstates[tp->t_state], toep, toep->flags, inp, inp->inp_flags); } #endif tp->t_toe = NULL; tp->t_flags &= ~TF_TOE; toep->flags &= ~TPF_ATTACHED; if (!(toep->flags & TPF_CPL_PENDING)) release_offload_resources(toep); } /* * setsockopt handler. */ static void t4_ctloutput(struct toedev *tod, struct tcpcb *tp, int dir, int name) { struct adapter *sc = tod->tod_softc; struct toepcb *toep = tp->t_toe; if (dir == SOPT_GET) return; CTR4(KTR_CXGBE, "%s: tp %p, dir %u, name %u", __func__, tp, dir, name); switch (name) { case TCP_NODELAY: t4_set_tcb_field(sc, toep->ctrlq, toep->tid, W_TCB_T_FLAGS, V_TF_NAGLE(1), V_TF_NAGLE(tp->t_flags & TF_NODELAY ? 0 : 1), 0, 0, toep->ofld_rxq->iq.abs_id); break; default: break; } } /* * The TOE driver will not receive any more CPLs for the tid associated with the * toepcb; release the hold on the inpcb. */ void final_cpl_received(struct toepcb *toep) { struct inpcb *inp = toep->inp; KASSERT(inp != NULL, ("%s: inp is NULL", __func__)); INP_WLOCK_ASSERT(inp); KASSERT(toep->flags & TPF_CPL_PENDING, ("%s: CPL not pending already?", __func__)); CTR6(KTR_CXGBE, "%s: tid %d, toep %p (0x%x), inp %p (0x%x)", __func__, toep->tid, toep, toep->flags, inp, inp->inp_flags); toep->inp = NULL; toep->flags &= ~TPF_CPL_PENDING; mbufq_drain(&toep->ulp_pdu_reclaimq); if (!(toep->flags & TPF_ATTACHED)) release_offload_resources(toep); if (!in_pcbrele_wlocked(inp)) INP_WUNLOCK(inp); } void insert_tid(struct adapter *sc, int tid, void *ctx) { struct tid_info *t = &sc->tids; t->tid_tab[tid] = ctx; atomic_add_int(&t->tids_in_use, 1); } void * lookup_tid(struct adapter *sc, int tid) { struct tid_info *t = &sc->tids; return (t->tid_tab[tid]); } void update_tid(struct adapter *sc, int tid, void *ctx) { struct tid_info *t = &sc->tids; t->tid_tab[tid] = ctx; } void remove_tid(struct adapter *sc, int tid) { struct tid_info *t = &sc->tids; t->tid_tab[tid] = NULL; atomic_subtract_int(&t->tids_in_use, 1); } void release_tid(struct adapter *sc, int tid, struct sge_wrq *ctrlq) { struct wrqe *wr; struct cpl_tid_release *req; wr = alloc_wrqe(sizeof(*req), ctrlq); if (wr == NULL) { queue_tid_release(sc, tid); /* defer */ return; } req = wrtod(wr); INIT_TP_WR_MIT_CPL(req, CPL_TID_RELEASE, tid); t4_wrq_tx(sc, wr); } static void queue_tid_release(struct adapter *sc, int tid) { CXGBE_UNIMPLEMENTED("deferred tid release"); } /* * What mtu_idx to use, given a 4-tuple and/or an MSS cap */ int find_best_mtu_idx(struct adapter *sc, struct in_conninfo *inc, int pmss) { unsigned short *mtus = &sc->params.mtus[0]; int i, mss, n; KASSERT(inc != NULL || pmss > 0, ("%s: at least one of inc/pmss must be specified", __func__)); mss = inc ? tcp_mssopt(inc) : pmss; if (pmss > 0 && mss > pmss) mss = pmss; if (inc->inc_flags & INC_ISIPV6) n = sizeof(struct ip6_hdr) + sizeof(struct tcphdr); else n = sizeof(struct ip) + sizeof(struct tcphdr); for (i = 0; i < NMTUS - 1 && mtus[i + 1] <= mss + n; i++) continue; return (i); } /* * Determine the receive window size for a socket. */ u_long select_rcv_wnd(struct socket *so) { unsigned long wnd; SOCKBUF_LOCK_ASSERT(&so->so_rcv); wnd = sbspace(&so->so_rcv); if (wnd < MIN_RCV_WND) wnd = MIN_RCV_WND; return min(wnd, MAX_RCV_WND); } int select_rcv_wscale(void) { int wscale = 0; unsigned long space = sb_max; if (space > MAX_RCV_WND) space = MAX_RCV_WND; while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < space) wscale++; return (wscale); } extern int always_keepalive; #define VIID_SMACIDX(v) (((unsigned int)(v) & 0x7f) << 1) /* * socket so could be a listening socket too. */ uint64_t calc_opt0(struct socket *so, struct vi_info *vi, struct l2t_entry *e, int mtu_idx, int rscale, int rx_credits, int ulp_mode) { uint64_t opt0; KASSERT(rx_credits <= M_RCV_BUFSIZ, ("%s: rcv_bufsiz too high", __func__)); opt0 = F_TCAM_BYPASS | V_WND_SCALE(rscale) | V_MSS_IDX(mtu_idx) | V_ULP_MODE(ulp_mode) | V_RCV_BUFSIZ(rx_credits); if (so != NULL) { struct inpcb *inp = sotoinpcb(so); struct tcpcb *tp = intotcpcb(inp); int keepalive = always_keepalive || so_options_get(so) & SO_KEEPALIVE; opt0 |= V_NAGLE((tp->t_flags & TF_NODELAY) == 0); opt0 |= V_KEEP_ALIVE(keepalive != 0); } if (e != NULL) opt0 |= V_L2T_IDX(e->idx); if (vi != NULL) { opt0 |= V_SMAC_SEL(VIID_SMACIDX(vi->viid)); opt0 |= V_TX_CHAN(vi->pi->tx_chan); } return htobe64(opt0); } uint64_t select_ntuple(struct vi_info *vi, struct l2t_entry *e) { struct adapter *sc = vi->pi->adapter; struct tp_params *tp = &sc->params.tp; uint16_t viid = vi->viid; uint64_t ntuple = 0; /* * Initialize each of the fields which we care about which are present * in the Compressed Filter Tuple. */ if (tp->vlan_shift >= 0 && e->vlan != CPL_L2T_VLAN_NONE) ntuple |= (uint64_t)(F_FT_VLAN_VLD | e->vlan) << tp->vlan_shift; if (tp->port_shift >= 0) ntuple |= (uint64_t)e->lport << tp->port_shift; if (tp->protocol_shift >= 0) ntuple |= (uint64_t)IPPROTO_TCP << tp->protocol_shift; if (tp->vnic_shift >= 0) { uint32_t vf = G_FW_VIID_VIN(viid); uint32_t pf = G_FW_VIID_PFN(viid); uint32_t vld = G_FW_VIID_VIVLD(viid); ntuple |= (uint64_t)(V_FT_VNID_ID_VF(vf) | V_FT_VNID_ID_PF(pf) | V_FT_VNID_ID_VLD(vld)) << tp->vnic_shift; } if (is_t4(sc)) return (htobe32((uint32_t)ntuple)); else return (htobe64(V_FILTER_TUPLE(ntuple))); } void set_tcpddp_ulp_mode(struct toepcb *toep) { toep->ulp_mode = ULP_MODE_TCPDDP; toep->ddp_flags = DDP_OK; toep->ddp_score = DDP_LOW_SCORE; } int negative_advice(int status) { return (status == CPL_ERR_RTX_NEG_ADVICE || status == CPL_ERR_PERSIST_NEG_ADVICE || status == CPL_ERR_KEEPALV_NEG_ADVICE); } static int alloc_tid_tabs(struct tid_info *t) { size_t size; unsigned int i; size = t->ntids * sizeof(*t->tid_tab) + t->natids * sizeof(*t->atid_tab) + t->nstids * sizeof(*t->stid_tab); t->tid_tab = malloc(size, M_CXGBE, M_ZERO | M_NOWAIT); if (t->tid_tab == NULL) return (ENOMEM); mtx_init(&t->atid_lock, "atid lock", NULL, MTX_DEF); t->atid_tab = (union aopen_entry *)&t->tid_tab[t->ntids]; t->afree = t->atid_tab; t->atids_in_use = 0; for (i = 1; i < t->natids; i++) t->atid_tab[i - 1].next = &t->atid_tab[i]; t->atid_tab[t->natids - 1].next = NULL; mtx_init(&t->stid_lock, "stid lock", NULL, MTX_DEF); t->stid_tab = (struct listen_ctx **)&t->atid_tab[t->natids]; t->stids_in_use = 0; TAILQ_INIT(&t->stids); t->nstids_free_head = t->nstids; atomic_store_rel_int(&t->tids_in_use, 0); return (0); } static void free_tid_tabs(struct tid_info *t) { KASSERT(t->tids_in_use == 0, ("%s: %d tids still in use.", __func__, t->tids_in_use)); KASSERT(t->atids_in_use == 0, ("%s: %d atids still in use.", __func__, t->atids_in_use)); KASSERT(t->stids_in_use == 0, ("%s: %d tids still in use.", __func__, t->stids_in_use)); free(t->tid_tab, M_CXGBE); t->tid_tab = NULL; if (mtx_initialized(&t->atid_lock)) mtx_destroy(&t->atid_lock); if (mtx_initialized(&t->stid_lock)) mtx_destroy(&t->stid_lock); } static int add_lip(struct adapter *sc, struct in6_addr *lip) { struct fw_clip_cmd c; ASSERT_SYNCHRONIZED_OP(sc); /* mtx_assert(&td->clip_table_lock, MA_OWNED); */ memset(&c, 0, sizeof(c)); c.op_to_write = htonl(V_FW_CMD_OP(FW_CLIP_CMD) | F_FW_CMD_REQUEST | F_FW_CMD_WRITE); c.alloc_to_len16 = htonl(F_FW_CLIP_CMD_ALLOC | FW_LEN16(c)); c.ip_hi = *(uint64_t *)&lip->s6_addr[0]; c.ip_lo = *(uint64_t *)&lip->s6_addr[8]; return (-t4_wr_mbox_ns(sc, sc->mbox, &c, sizeof(c), &c)); } static int delete_lip(struct adapter *sc, struct in6_addr *lip) { struct fw_clip_cmd c; ASSERT_SYNCHRONIZED_OP(sc); /* mtx_assert(&td->clip_table_lock, MA_OWNED); */ memset(&c, 0, sizeof(c)); c.op_to_write = htonl(V_FW_CMD_OP(FW_CLIP_CMD) | F_FW_CMD_REQUEST | F_FW_CMD_READ); c.alloc_to_len16 = htonl(F_FW_CLIP_CMD_FREE | FW_LEN16(c)); c.ip_hi = *(uint64_t *)&lip->s6_addr[0]; c.ip_lo = *(uint64_t *)&lip->s6_addr[8]; return (-t4_wr_mbox_ns(sc, sc->mbox, &c, sizeof(c), &c)); } static struct clip_entry * search_lip(struct tom_data *td, struct in6_addr *lip) { struct clip_entry *ce; mtx_assert(&td->clip_table_lock, MA_OWNED); TAILQ_FOREACH(ce, &td->clip_table, link) { if (IN6_ARE_ADDR_EQUAL(&ce->lip, lip)) return (ce); } return (NULL); } struct clip_entry * hold_lip(struct tom_data *td, struct in6_addr *lip) { struct clip_entry *ce; mtx_lock(&td->clip_table_lock); ce = search_lip(td, lip); if (ce != NULL) ce->refcount++; mtx_unlock(&td->clip_table_lock); return (ce); } void release_lip(struct tom_data *td, struct clip_entry *ce) { mtx_lock(&td->clip_table_lock); KASSERT(search_lip(td, &ce->lip) == ce, ("%s: CLIP entry %p p not in CLIP table.", __func__, ce)); KASSERT(ce->refcount > 0, ("%s: CLIP entry %p has refcount 0", __func__, ce)); --ce->refcount; mtx_unlock(&td->clip_table_lock); } static void init_clip_table(struct adapter *sc, struct tom_data *td) { ASSERT_SYNCHRONIZED_OP(sc); mtx_init(&td->clip_table_lock, "CLIP table lock", NULL, MTX_DEF); TAILQ_INIT(&td->clip_table); td->clip_gen = -1; update_clip_table(sc, td); } static void update_clip(struct adapter *sc, void *arg __unused) { if (begin_synchronized_op(sc, NULL, HOLD_LOCK, "t4tomuc")) return; if (uld_active(sc, ULD_TOM)) update_clip_table(sc, sc->tom_softc); end_synchronized_op(sc, LOCK_HELD); } static void t4_clip_task(void *arg, int count) { t4_iterate(update_clip, NULL); } static void update_clip_table(struct adapter *sc, struct tom_data *td) { struct in6_ifaddr *ia; struct in6_addr *lip, tlip; struct clip_head stale; struct clip_entry *ce, *ce_temp; int rc, gen = atomic_load_acq_int(&in6_ifaddr_gen); ASSERT_SYNCHRONIZED_OP(sc); IN6_IFADDR_RLOCK(); mtx_lock(&td->clip_table_lock); if (gen == td->clip_gen) goto done; TAILQ_INIT(&stale); TAILQ_CONCAT(&stale, &td->clip_table, link); TAILQ_FOREACH(ia, &V_in6_ifaddrhead, ia_link) { lip = &ia->ia_addr.sin6_addr; KASSERT(!IN6_IS_ADDR_MULTICAST(lip), ("%s: mcast address in in6_ifaddr list", __func__)); if (IN6_IS_ADDR_LOOPBACK(lip)) continue; if (IN6_IS_SCOPE_EMBED(lip)) { /* Remove the embedded scope */ tlip = *lip; lip = &tlip; in6_clearscope(lip); } /* * XXX: how to weed out the link local address for the loopback * interface? It's fe80::1 usually (always?). */ /* * If it's in the main list then we already know it's not stale. */ TAILQ_FOREACH(ce, &td->clip_table, link) { if (IN6_ARE_ADDR_EQUAL(&ce->lip, lip)) goto next; } /* * If it's in the stale list we should move it to the main list. */ TAILQ_FOREACH(ce, &stale, link) { if (IN6_ARE_ADDR_EQUAL(&ce->lip, lip)) { TAILQ_REMOVE(&stale, ce, link); TAILQ_INSERT_TAIL(&td->clip_table, ce, link); goto next; } } /* A new IP6 address; add it to the CLIP table */ ce = malloc(sizeof(*ce), M_CXGBE, M_NOWAIT); memcpy(&ce->lip, lip, sizeof(ce->lip)); ce->refcount = 0; rc = add_lip(sc, lip); if (rc == 0) TAILQ_INSERT_TAIL(&td->clip_table, ce, link); else { char ip[INET6_ADDRSTRLEN]; inet_ntop(AF_INET6, &ce->lip, &ip[0], sizeof(ip)); log(LOG_ERR, "%s: could not add %s (%d)\n", __func__, ip, rc); free(ce, M_CXGBE); } next: continue; } /* * Remove stale addresses (those no longer in V_in6_ifaddrhead) that are * no longer referenced by the driver. */ TAILQ_FOREACH_SAFE(ce, &stale, link, ce_temp) { if (ce->refcount == 0) { rc = delete_lip(sc, &ce->lip); if (rc == 0) { TAILQ_REMOVE(&stale, ce, link); free(ce, M_CXGBE); } else { char ip[INET6_ADDRSTRLEN]; inet_ntop(AF_INET6, &ce->lip, &ip[0], sizeof(ip)); log(LOG_ERR, "%s: could not delete %s (%d)\n", __func__, ip, rc); } } } /* The ones that are still referenced need to stay in the CLIP table */ TAILQ_CONCAT(&td->clip_table, &stale, link); td->clip_gen = gen; done: mtx_unlock(&td->clip_table_lock); IN6_IFADDR_RUNLOCK(); } static void destroy_clip_table(struct adapter *sc, struct tom_data *td) { struct clip_entry *ce, *ce_temp; if (mtx_initialized(&td->clip_table_lock)) { mtx_lock(&td->clip_table_lock); TAILQ_FOREACH_SAFE(ce, &td->clip_table, link, ce_temp) { KASSERT(ce->refcount == 0, ("%s: CLIP entry %p still in use (%d)", __func__, ce, ce->refcount)); TAILQ_REMOVE(&td->clip_table, ce, link); delete_lip(sc, &ce->lip); free(ce, M_CXGBE); } mtx_unlock(&td->clip_table_lock); mtx_destroy(&td->clip_table_lock); } } static void free_tom_data(struct adapter *sc, struct tom_data *td) { ASSERT_SYNCHRONIZED_OP(sc); KASSERT(TAILQ_EMPTY(&td->toep_list), ("%s: TOE PCB list is not empty.", __func__)); KASSERT(td->lctx_count == 0, ("%s: lctx hash table is not empty.", __func__)); - t4_uninit_ddp(sc, td); + t4_free_ppod_region(&td->pr); destroy_clip_table(sc, td); if (td->listen_mask != 0) hashdestroy(td->listen_hash, M_CXGBE, td->listen_mask); if (mtx_initialized(&td->unsent_wr_lock)) mtx_destroy(&td->unsent_wr_lock); if (mtx_initialized(&td->lctx_hash_lock)) mtx_destroy(&td->lctx_hash_lock); if (mtx_initialized(&td->toep_list_lock)) mtx_destroy(&td->toep_list_lock); free_tid_tabs(&sc->tids); free(td, M_CXGBE); } static void reclaim_wr_resources(void *arg, int count) { struct tom_data *td = arg; STAILQ_HEAD(, wrqe) twr_list = STAILQ_HEAD_INITIALIZER(twr_list); struct cpl_act_open_req *cpl; u_int opcode, atid; struct wrqe *wr; struct adapter *sc; mtx_lock(&td->unsent_wr_lock); STAILQ_SWAP(&td->unsent_wr_list, &twr_list, wrqe); mtx_unlock(&td->unsent_wr_lock); while ((wr = STAILQ_FIRST(&twr_list)) != NULL) { STAILQ_REMOVE_HEAD(&twr_list, link); cpl = wrtod(wr); opcode = GET_OPCODE(cpl); switch (opcode) { case CPL_ACT_OPEN_REQ: case CPL_ACT_OPEN_REQ6: atid = G_TID_TID(be32toh(OPCODE_TID(cpl))); sc = td_adapter(td); CTR2(KTR_CXGBE, "%s: atid %u ", __func__, atid); act_open_failure_cleanup(sc, atid, EHOSTUNREACH); free(wr, M_CXGBE); break; default: log(LOG_ERR, "%s: leaked work request %p, wr_len %d, " "opcode %x\n", __func__, wr, wr->wr_len, opcode); /* WR not freed here; go look at it with a debugger. */ } } } /* * Ground control to Major TOM * Commencing countdown, engines on */ static int t4_tom_activate(struct adapter *sc) { struct tom_data *td; struct toedev *tod; struct vi_info *vi; struct sge_ofld_rxq *ofld_rxq; int i, j, rc, v; ASSERT_SYNCHRONIZED_OP(sc); /* per-adapter softc for TOM */ td = malloc(sizeof(*td), M_CXGBE, M_ZERO | M_NOWAIT); if (td == NULL) return (ENOMEM); /* List of TOE PCBs and associated lock */ mtx_init(&td->toep_list_lock, "PCB list lock", NULL, MTX_DEF); TAILQ_INIT(&td->toep_list); /* Listen context */ mtx_init(&td->lctx_hash_lock, "lctx hash lock", NULL, MTX_DEF); td->listen_hash = hashinit_flags(LISTEN_HASH_SIZE, M_CXGBE, &td->listen_mask, HASH_NOWAIT); /* List of WRs for which L2 resolution failed */ mtx_init(&td->unsent_wr_lock, "Unsent WR list lock", NULL, MTX_DEF); STAILQ_INIT(&td->unsent_wr_list); TASK_INIT(&td->reclaim_wr_resources, 0, reclaim_wr_resources, td); /* TID tables */ rc = alloc_tid_tabs(&sc->tids); if (rc != 0) goto done; - /* DDP page pods and CPL handlers */ - t4_init_ddp(sc, td); + rc = t4_init_ppod_region(&td->pr, &sc->vres.ddp, + t4_read_reg(sc, A_ULP_RX_TDDP_PSZ), "TDDP page pods"); + if (rc != 0) + goto done; + t4_set_reg_field(sc, A_ULP_RX_TDDP_TAGMASK, + V_TDDPTAGMASK(M_TDDPTAGMASK), td->pr.pr_tag_mask); /* CLIP table for IPv6 offload */ init_clip_table(sc, td); /* toedev ops */ tod = &td->tod; init_toedev(tod); tod->tod_softc = sc; tod->tod_connect = t4_connect; tod->tod_listen_start = t4_listen_start; tod->tod_listen_stop = t4_listen_stop; tod->tod_rcvd = t4_rcvd; tod->tod_output = t4_tod_output; tod->tod_send_rst = t4_send_rst; tod->tod_send_fin = t4_send_fin; tod->tod_pcb_detach = t4_pcb_detach; tod->tod_l2_update = t4_l2_update; tod->tod_syncache_added = t4_syncache_added; tod->tod_syncache_removed = t4_syncache_removed; tod->tod_syncache_respond = t4_syncache_respond; tod->tod_offload_socket = t4_offload_socket; tod->tod_ctloutput = t4_ctloutput; for_each_port(sc, i) { for_each_vi(sc->port[i], v, vi) { TOEDEV(vi->ifp) = &td->tod; for_each_ofld_rxq(vi, j, ofld_rxq) { ofld_rxq->iq.set_tcb_rpl = do_set_tcb_rpl; ofld_rxq->iq.l2t_write_rpl = do_l2t_write_rpl2; } } } sc->tom_softc = td; register_toedev(sc->tom_softc); done: if (rc != 0) free_tom_data(sc, td); return (rc); } static int t4_tom_deactivate(struct adapter *sc) { int rc = 0; struct tom_data *td = sc->tom_softc; ASSERT_SYNCHRONIZED_OP(sc); if (td == NULL) return (0); /* XXX. KASSERT? */ if (sc->offload_map != 0) return (EBUSY); /* at least one port has IFCAP_TOE enabled */ if (uld_active(sc, ULD_IWARP) || uld_active(sc, ULD_ISCSI)) return (EBUSY); /* both iWARP and iSCSI rely on the TOE. */ mtx_lock(&td->toep_list_lock); if (!TAILQ_EMPTY(&td->toep_list)) rc = EBUSY; mtx_unlock(&td->toep_list_lock); mtx_lock(&td->lctx_hash_lock); if (td->lctx_count > 0) rc = EBUSY; mtx_unlock(&td->lctx_hash_lock); taskqueue_drain(taskqueue_thread, &td->reclaim_wr_resources); mtx_lock(&td->unsent_wr_lock); if (!STAILQ_EMPTY(&td->unsent_wr_list)) rc = EBUSY; mtx_unlock(&td->unsent_wr_lock); if (rc == 0) { unregister_toedev(sc->tom_softc); free_tom_data(sc, td); sc->tom_softc = NULL; } return (rc); } static void t4_tom_ifaddr_event(void *arg __unused, struct ifnet *ifp) { atomic_add_rel_int(&in6_ifaddr_gen, 1); taskqueue_enqueue_timeout(taskqueue_thread, &clip_task, -hz / 4); } static int t4_tom_mod_load(void) { int rc; struct protosw *tcp_protosw, *tcp6_protosw; /* CPL handlers */ t4_init_connect_cpl_handlers(); t4_init_listen_cpl_handlers(); t4_init_cpl_io_handlers(); rc = t4_ddp_mod_load(); if (rc != 0) return (rc); tcp_protosw = pffindproto(PF_INET, IPPROTO_TCP, SOCK_STREAM); if (tcp_protosw == NULL) return (ENOPROTOOPT); bcopy(tcp_protosw, &ddp_protosw, sizeof(ddp_protosw)); bcopy(tcp_protosw->pr_usrreqs, &ddp_usrreqs, sizeof(ddp_usrreqs)); ddp_usrreqs.pru_soreceive = t4_soreceive_ddp; ddp_protosw.pr_usrreqs = &ddp_usrreqs; tcp6_protosw = pffindproto(PF_INET6, IPPROTO_TCP, SOCK_STREAM); if (tcp6_protosw == NULL) return (ENOPROTOOPT); bcopy(tcp6_protosw, &ddp6_protosw, sizeof(ddp6_protosw)); bcopy(tcp6_protosw->pr_usrreqs, &ddp6_usrreqs, sizeof(ddp6_usrreqs)); ddp6_usrreqs.pru_soreceive = t4_soreceive_ddp; ddp6_protosw.pr_usrreqs = &ddp6_usrreqs; TIMEOUT_TASK_INIT(taskqueue_thread, &clip_task, 0, t4_clip_task, NULL); ifaddr_evhandler = EVENTHANDLER_REGISTER(ifaddr_event, t4_tom_ifaddr_event, NULL, EVENTHANDLER_PRI_ANY); rc = t4_register_uld(&tom_uld_info); if (rc != 0) t4_tom_mod_unload(); return (rc); } static void tom_uninit(struct adapter *sc, void *arg __unused) { if (begin_synchronized_op(sc, NULL, SLEEP_OK | INTR_OK, "t4tomun")) return; /* Try to free resources (works only if no port has IFCAP_TOE) */ if (uld_active(sc, ULD_TOM)) t4_deactivate_uld(sc, ULD_TOM); end_synchronized_op(sc, 0); } static int t4_tom_mod_unload(void) { t4_iterate(tom_uninit, NULL); if (t4_unregister_uld(&tom_uld_info) == EBUSY) return (EBUSY); if (ifaddr_evhandler) { EVENTHANDLER_DEREGISTER(ifaddr_event, ifaddr_evhandler); taskqueue_cancel_timeout(taskqueue_thread, &clip_task, NULL); } t4_ddp_mod_unload(); return (0); } #endif /* TCP_OFFLOAD */ static int t4_tom_modevent(module_t mod, int cmd, void *arg) { int rc = 0; #ifdef TCP_OFFLOAD switch (cmd) { case MOD_LOAD: rc = t4_tom_mod_load(); break; case MOD_UNLOAD: rc = t4_tom_mod_unload(); break; default: rc = EINVAL; } #else printf("t4_tom: compiled without TCP_OFFLOAD support.\n"); rc = EOPNOTSUPP; #endif return (rc); } static moduledata_t t4_tom_moddata= { "t4_tom", t4_tom_modevent, 0 }; MODULE_VERSION(t4_tom, 1); MODULE_DEPEND(t4_tom, toecore, 1, 1, 1); MODULE_DEPEND(t4_tom, t4nex, 1, 1, 1); DECLARE_MODULE(t4_tom, t4_tom_moddata, SI_SUB_EXEC, SI_ORDER_ANY); Index: stable/10/sys/dev/cxgbe/tom/t4_tom.h =================================================================== --- stable/10/sys/dev/cxgbe/tom/t4_tom.h (revision 309555) +++ stable/10/sys/dev/cxgbe/tom/t4_tom.h (revision 309556) @@ -1,329 +1,355 @@ /*- * Copyright (c) 2012, 2015 Chelsio Communications, Inc. * All rights reserved. * Written by: Navdeep Parhar * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * $FreeBSD$ * */ #ifndef __T4_TOM_H__ #define __T4_TOM_H__ #include /* * Inline version of mbufq for use on 10.x. Borrowed from * sys/cam/ctl/ctl_ha.c. */ struct mbufq { struct mbuf *head; struct mbuf *tail; }; #define LISTEN_HASH_SIZE 32 /* * Min receive window. We want it to be large enough to accommodate receive * coalescing, handle jumbo frames, and not trigger sender SWS avoidance. */ #define MIN_RCV_WND (24 * 1024U) /* * Max receive window supported by HW in bytes. Only a small part of it can * be set through option0, the rest needs to be set through RX_DATA_ACK. */ #define MAX_RCV_WND ((1U << 27) - 1) #define DDP_RSVD_WIN (16 * 1024U) #define SB_DDP_INDICATE SB_IN_TOE /* soreceive must respond to indicate */ #define USE_DDP_RX_FLOW_CONTROL +#define PPOD_SZ(n) ((n) * sizeof(struct pagepod)) +#define PPOD_SIZE (PPOD_SZ(1)) + /* TOE PCB flags */ enum { TPF_ATTACHED = (1 << 0), /* a tcpcb refers to this toepcb */ TPF_FLOWC_WR_SENT = (1 << 1), /* firmware flow context WR sent */ TPF_TX_DATA_SENT = (1 << 2), /* some data sent */ TPF_TX_SUSPENDED = (1 << 3), /* tx suspended for lack of resources */ TPF_SEND_FIN = (1 << 4), /* send FIN after all pending data */ TPF_FIN_SENT = (1 << 5), /* FIN has been sent */ TPF_ABORT_SHUTDOWN = (1 << 6), /* connection abort is in progress */ TPF_CPL_PENDING = (1 << 7), /* haven't received the last CPL */ TPF_SYNQE = (1 << 8), /* synq_entry, not really a toepcb */ TPF_SYNQE_NEEDFREE = (1 << 9), /* synq_entry was malloc'd separately */ TPF_SYNQE_TCPDDP = (1 << 10), /* ulp_mode TCPDDP in toepcb */ TPF_SYNQE_EXPANDED = (1 << 11), /* toepcb ready, tid context updated */ TPF_SYNQE_HAS_L2TE = (1 << 12), /* we've replied to PASS_ACCEPT_REQ */ }; enum { DDP_OK = (1 << 0), /* OK to turn on DDP */ DDP_SC_REQ = (1 << 1), /* state change (on/off) requested */ DDP_ON = (1 << 2), /* DDP is turned on */ DDP_BUF0_ACTIVE = (1 << 3), /* buffer 0 in use (not invalidated) */ DDP_BUF1_ACTIVE = (1 << 4), /* buffer 1 in use (not invalidated) */ }; struct ofld_tx_sdesc { uint32_t plen; /* payload length */ uint8_t tx_credits; /* firmware tx credits (unit is 16B) */ }; +struct ppod_region { + u_int pr_start; + u_int pr_len; + u_int pr_page_shift[4]; + uint32_t pr_tag_mask; /* hardware tagmask for this region. */ + uint32_t pr_invalid_bit; /* OR with this to invalidate tag. */ + uint32_t pr_alias_mask; /* AND with tag to get alias bits. */ + u_int pr_alias_shift; /* shift this much for first alias bit. */ + vmem_t *pr_arena; +}; + +struct ppod_reservation { + struct ppod_region *prsv_pr; + uint32_t prsv_tag; /* Full tag: pgsz, alias, tag, color */ + u_int prsv_nppods; +}; + struct ddp_buffer { - uint32_t tag; /* includes color, page pod addr, and DDP page size */ - u_int ppod_addr; - int nppods; int offset; int len; int npages; vm_page_t *pages; + struct ppod_reservation prsv; }; struct toepcb { TAILQ_ENTRY(toepcb) link; /* toep_list */ u_int flags; /* miscellaneous flags */ struct tom_data *td; struct inpcb *inp; /* backpointer to host stack's PCB */ struct vi_info *vi; /* virtual interface */ struct sge_wrq *ofld_txq; struct sge_ofld_rxq *ofld_rxq; struct sge_wrq *ctrlq; struct l2t_entry *l2te; /* L2 table entry used by this connection */ struct clip_entry *ce; /* CLIP table entry used by this tid */ int tid; /* Connection identifier */ /* tx credit handling */ u_int tx_total; /* total tx WR credits (in 16B units) */ u_int tx_credits; /* tx WR credits (in 16B units) available */ u_int tx_nocompl; /* tx WR credits since last compl request */ u_int plen_nocompl; /* payload since last compl request */ /* rx credit handling */ u_int sb_cc; /* last noted value of so_rcv->sb_cc */ int rx_credits; /* rx credits (in bytes) to be returned to hw */ u_int ulp_mode; /* ULP mode */ void *ulpcb; void *ulpcb2; struct mbufq ulp_pduq; /* PDUs waiting to be sent out. */ struct mbufq ulp_pdu_reclaimq; u_int ddp_flags; struct ddp_buffer *db[2]; time_t ddp_disabled; uint8_t ddp_score; /* Tx software descriptor */ uint8_t txsd_total; uint8_t txsd_pidx; uint8_t txsd_cidx; uint8_t txsd_avail; struct ofld_tx_sdesc txsd[]; }; struct flowc_tx_params { uint32_t snd_nxt; uint32_t rcv_nxt; unsigned int snd_space; unsigned int mss; }; #define DDP_RETRY_WAIT 5 /* seconds to wait before re-enabling DDP */ #define DDP_LOW_SCORE 1 #define DDP_HIGH_SCORE 3 /* * Compressed state for embryonic connections for a listener. Barely fits in * 64B, try not to grow it further. */ struct synq_entry { TAILQ_ENTRY(synq_entry) link; /* listen_ctx's synq link */ int flags; /* same as toepcb's tp_flags */ int tid; struct listen_ctx *lctx; /* backpointer to listen ctx */ struct mbuf *syn; uint32_t iss; uint32_t ts; volatile uintptr_t wr; volatile u_int refcnt; uint16_t l2e_idx; uint16_t rcv_bufsize; }; /* listen_ctx flags */ #define LCTX_RPL_PENDING 1 /* waiting for a CPL_PASS_OPEN_RPL */ struct listen_ctx { LIST_ENTRY(listen_ctx) link; /* listen hash linkage */ volatile int refcount; int stid; struct stid_region stid_region; int flags; struct inpcb *inp; /* listening socket's inp */ struct sge_wrq *ctrlq; struct sge_ofld_rxq *ofld_rxq; struct clip_entry *ce; TAILQ_HEAD(, synq_entry) synq; }; struct clip_entry { TAILQ_ENTRY(clip_entry) link; struct in6_addr lip; /* local IPv6 address */ u_int refcount; }; TAILQ_HEAD(clip_head, clip_entry); struct tom_data { struct toedev tod; /* toepcb's associated with this TOE device */ struct mtx toep_list_lock; TAILQ_HEAD(, toepcb) toep_list; struct mtx lctx_hash_lock; LIST_HEAD(, listen_ctx) *listen_hash; u_long listen_mask; int lctx_count; /* # of lctx in the hash table */ - u_int ppod_start; - vmem_t *ppod_arena; + struct ppod_region pr; struct mtx clip_table_lock; struct clip_head clip_table; int clip_gen; /* WRs that will not be sent to the chip because L2 resolution failed */ struct mtx unsent_wr_lock; STAILQ_HEAD(, wrqe) unsent_wr_list; struct task reclaim_wr_resources; }; static inline struct tom_data * tod_td(struct toedev *tod) { return (__containerof(tod, struct tom_data, tod)); } static inline struct adapter * td_adapter(struct tom_data *td) { return (td->tod.tod_softc); } /* * XXX: Don't define these for the iWARP driver on 10 due to differences * in LinuxKPI. */ #ifndef _LINUX_TYPES_H_ static inline void set_mbuf_ulp_submode(struct mbuf *m, uint8_t ulp_submode) { M_ASSERTPKTHDR(m); m->m_pkthdr.PH_per.eigth[0] = ulp_submode; } static inline uint8_t mbuf_ulp_submode(struct mbuf *m) { M_ASSERTPKTHDR(m); return (m->m_pkthdr.PH_per.eigth[0]); } #endif /* t4_tom.c */ struct toepcb *alloc_toepcb(struct vi_info *, int, int, int); void free_toepcb(struct toepcb *); void offload_socket(struct socket *, struct toepcb *); void undo_offload_socket(struct socket *); void final_cpl_received(struct toepcb *); void insert_tid(struct adapter *, int, void *); void *lookup_tid(struct adapter *, int); void update_tid(struct adapter *, int, void *); void remove_tid(struct adapter *, int); void release_tid(struct adapter *, int, struct sge_wrq *); int find_best_mtu_idx(struct adapter *, struct in_conninfo *, int); u_long select_rcv_wnd(struct socket *); int select_rcv_wscale(void); uint64_t calc_opt0(struct socket *, struct vi_info *, struct l2t_entry *, int, int, int, int); uint64_t select_ntuple(struct vi_info *, struct l2t_entry *); void set_tcpddp_ulp_mode(struct toepcb *); int negative_advice(int); struct clip_entry *hold_lip(struct tom_data *, struct in6_addr *); void release_lip(struct tom_data *, struct clip_entry *); /* t4_connect.c */ void t4_init_connect_cpl_handlers(void); int t4_connect(struct toedev *, struct socket *, struct rtentry *, struct sockaddr *); void act_open_failure_cleanup(struct adapter *, u_int, u_int); /* t4_listen.c */ void t4_init_listen_cpl_handlers(void); int t4_listen_start(struct toedev *, struct tcpcb *); int t4_listen_stop(struct toedev *, struct tcpcb *); void t4_syncache_added(struct toedev *, void *); void t4_syncache_removed(struct toedev *, void *); int t4_syncache_respond(struct toedev *, void *, struct mbuf *); int do_abort_req_synqe(struct sge_iq *, const struct rss_header *, struct mbuf *); int do_abort_rpl_synqe(struct sge_iq *, const struct rss_header *, struct mbuf *); void t4_offload_socket(struct toedev *, void *, struct socket *); /* t4_cpl_io.c */ void t4_init_cpl_io_handlers(void); void t4_uninit_cpl_io_handlers(void); void send_abort_rpl(struct adapter *, struct sge_wrq *, int , int); void send_flowc_wr(struct toepcb *, struct flowc_tx_params *); void send_reset(struct adapter *, struct toepcb *, uint32_t); void make_established(struct toepcb *, uint32_t, uint32_t, uint16_t); void t4_rcvd(struct toedev *, struct tcpcb *); int t4_tod_output(struct toedev *, struct tcpcb *); int t4_send_fin(struct toedev *, struct tcpcb *); int t4_send_rst(struct toedev *, struct tcpcb *); void t4_set_tcb_field(struct adapter *, struct sge_wrq *, int, uint16_t, uint64_t, uint64_t, int, int, int); void t4_push_frames(struct adapter *sc, struct toepcb *toep, int drop); void t4_push_pdus(struct adapter *sc, struct toepcb *toep, int drop); int do_set_tcb_rpl(struct sge_iq *, const struct rss_header *, struct mbuf *); /* t4_ddp.c */ -void t4_init_ddp(struct adapter *, struct tom_data *); -void t4_uninit_ddp(struct adapter *, struct tom_data *); +int t4_init_ppod_region(struct ppod_region *, struct t4_range *, u_int, + const char *); +void t4_free_ppod_region(struct ppod_region *); +int t4_alloc_page_pods_for_db(struct ppod_region *, struct ddp_buffer *); +int t4_alloc_page_pods_for_buf(struct ppod_region *, vm_offset_t, int, + struct ppod_reservation *); +int t4_write_page_pods_for_db(struct adapter *, struct sge_wrq *, int, + struct ddp_buffer *); +int t4_write_page_pods_for_buf(struct adapter *, struct sge_wrq *, int tid, + struct ppod_reservation *, vm_offset_t, int); +void t4_free_page_pods(struct ppod_reservation *); int t4_soreceive_ddp(struct socket *, struct sockaddr **, struct uio *, struct mbuf **, struct mbuf **, int *); int t4_ddp_mod_load(void); void t4_ddp_mod_unload(void); void enable_ddp(struct adapter *, struct toepcb *toep); void release_ddp_resources(struct toepcb *toep); void handle_ddp_close(struct toepcb *, struct tcpcb *, struct sockbuf *, uint32_t); void insert_ddp_data(struct toepcb *, uint32_t); #endif Index: stable/10 =================================================================== --- stable/10 (revision 309555) +++ stable/10 (revision 309556) Property changes on: stable/10 ___________________________________________________________________ Modified: svn:mergeinfo ## -0,0 +0,1 ## Merged /head:r303688,303750,305166-305167