2 * Copyright (c) 2016 Alexander Motin <mav@FreeBSD.org>
3 * Copyright (C) 2013 Intel Corporation
4 * Copyright (C) 2015 EMC Corporation
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
10 * 1. Redistributions of source code must retain the above copyright
11 * notice, this list of conditions and the following disclaimer.
12 * 2. Redistributions in binary form must reproduce the above copyright
13 * notice, this list of conditions and the following disclaimer in the
14 * documentation and/or other materials provided with the distribution.
16 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
17 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
20 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30 * The Non-Transparent Bridge (NTB) is a device that allows you to connect
31 * two or more systems using a PCI-e links, providing remote memory access.
33 * This module contains a transport for sending and receiving messages by
34 * writing to remote memory window(s) provided by underlying NTB device.
36 * NOTE: Much of the code in this module is shared with Linux. Any patches may
37 * be picked up and redistributed in Linux with a dual GPL/BSD license.
40 #include <sys/cdefs.h>
41 __FBSDID("$FreeBSD$");
43 #include <sys/param.h>
44 #include <sys/kernel.h>
45 #include <sys/systm.h>
48 #include <sys/limits.h>
50 #include <sys/malloc.h>
52 #include <sys/module.h>
53 #include <sys/mutex.h>
54 #include <sys/queue.h>
55 #include <sys/sysctl.h>
56 #include <sys/taskqueue.h>
61 #include <machine/bus.h>
64 #include "ntb_transport.h"
66 #define KTR_NTB KTR_SPARE3
68 #define NTB_TRANSPORT_VERSION 4
70 static SYSCTL_NODE(_hw, OID_AUTO, ntb_transport, CTLFLAG_RW, 0, "ntb_transport");
72 static unsigned g_ntb_transport_debug_level;
73 TUNABLE_INT("hw.ntb_transport.debug_level", &g_ntb_transport_debug_level);
74 SYSCTL_UINT(_hw_ntb_transport, OID_AUTO, debug_level, CTLFLAG_RWTUN,
75 &g_ntb_transport_debug_level, 0,
76 "ntb_transport log level -- higher is more verbose");
77 #define ntb_printf(lvl, ...) do { \
78 if ((lvl) <= g_ntb_transport_debug_level) { \
79 printf(__VA_ARGS__); \
83 static unsigned transport_mtu = 0x10000;
85 static uint64_t max_mw_size;
86 TUNABLE_QUAD("hw.ntb_transport.max_mw_size", &max_mw_size);
87 SYSCTL_UQUAD(_hw_ntb_transport, OID_AUTO, max_mw_size, CTLFLAG_RDTUN, &max_mw_size, 0,
88 "If enabled (non-zero), limit the size of large memory windows. "
89 "Both sides of the NTB MUST set the same value here.");
91 static unsigned enable_xeon_watchdog;
92 TUNABLE_INT("hw.ntb_transport.enable_xeon_watchdog", &enable_xeon_watchdog);
93 SYSCTL_UINT(_hw_ntb_transport, OID_AUTO, enable_xeon_watchdog, CTLFLAG_RDTUN,
94 &enable_xeon_watchdog, 0, "If non-zero, write a register every second to "
95 "keep a watchdog from tearing down the NTB link");
97 STAILQ_HEAD(ntb_queue_list, ntb_queue_entry);
99 typedef uint32_t ntb_q_idx_t;
101 struct ntb_queue_entry {
102 /* ntb_queue list reference */
103 STAILQ_ENTRY(ntb_queue_entry) entry;
105 /* info on data to be transferred */
111 struct ntb_transport_qp *qp;
112 struct ntb_payload_header *x_hdr;
120 struct ntb_transport_qp {
121 struct ntb_transport_ctx *transport;
127 volatile bool link_is_up;
128 uint8_t qp_num; /* Only 64 QPs are allowed. 0-63 */
130 struct ntb_rx_info *rx_info;
131 struct ntb_rx_info *remote_rx_info;
133 void (*tx_handler)(struct ntb_transport_qp *qp, void *qp_data,
134 void *data, int len);
135 struct ntb_queue_list tx_free_q;
136 struct mtx ntb_tx_free_q_lock;
138 bus_addr_t tx_mw_phys;
139 ntb_q_idx_t tx_index;
140 ntb_q_idx_t tx_max_entry;
141 uint64_t tx_max_frame;
143 void (*rx_handler)(struct ntb_transport_qp *qp, void *qp_data,
144 void *data, int len);
145 struct ntb_queue_list rx_post_q;
146 struct ntb_queue_list rx_pend_q;
147 /* ntb_rx_q_lock: synchronize access to rx_XXXX_q */
148 struct mtx ntb_rx_q_lock;
149 struct task rxc_db_work;
150 struct taskqueue *rxc_tq;
152 ntb_q_idx_t rx_index;
153 ntb_q_idx_t rx_max_entry;
154 uint64_t rx_max_frame;
156 void (*event_handler)(void *data, enum ntb_link_event status);
157 struct callout link_work;
158 struct callout rx_full;
160 uint64_t last_rx_no_buf;
165 uint64_t rx_ring_empty;
166 uint64_t rx_err_no_buf;
167 uint64_t rx_err_oflow;
171 uint64_t tx_ring_full;
172 uint64_t tx_err_no_buf;
177 struct ntb_transport_mw {
178 vm_paddr_t phys_addr;
181 size_t xlat_align_size;
182 bus_addr_t addr_limit;
183 /* Tx buff is off vbase / phys_addr */
187 /* Rx buff is off virt_addr / dma_addr */
192 struct ntb_transport_child {
196 struct ntb_transport_child *next;
199 struct ntb_transport_ctx {
201 struct ntb_transport_child *child;
202 struct ntb_transport_mw *mw_vec;
203 struct ntb_transport_qp *qp_vec;
207 volatile bool link_is_up;
208 struct callout link_work;
209 struct callout link_watchdog;
210 struct task link_cleanup;
214 NTBT_DESC_DONE_FLAG = 1 << 0,
215 NTBT_LINK_DOWN_FLAG = 1 << 1,
218 struct ntb_payload_header {
226 * The order of this enum is part of the remote protocol. Do not
227 * reorder without bumping protocol version (and it's probably best
228 * to keep the protocol in lock-step with the Linux NTB driver.
235 * N.B.: transport_link_work assumes MW1 enums = MW0 + 2.
243 * Some NTB-using hardware have a watchdog to work around NTB hangs; if
244 * a register or doorbell isn't written every few seconds, the link is
245 * torn down. Write an otherwise unused register every few seconds to
246 * work around this watchdog.
248 NTBT_WATCHDOG_SPAD = 15
251 #define QP_TO_MW(nt, qp) ((qp) % nt->mw_count)
252 #define NTB_QP_DEF_NUM_ENTRIES 100
253 #define NTB_LINK_DOWN_TIMEOUT 10
255 static int ntb_transport_probe(device_t dev);
256 static int ntb_transport_attach(device_t dev);
257 static int ntb_transport_detach(device_t dev);
258 static void ntb_transport_init_queue(struct ntb_transport_ctx *nt,
259 unsigned int qp_num);
260 static int ntb_process_tx(struct ntb_transport_qp *qp,
261 struct ntb_queue_entry *entry);
262 static void ntb_transport_rxc_db(void *arg, int pending);
263 static int ntb_process_rxc(struct ntb_transport_qp *qp);
264 static void ntb_memcpy_rx(struct ntb_transport_qp *qp,
265 struct ntb_queue_entry *entry, void *offset);
266 static inline void ntb_rx_copy_callback(struct ntb_transport_qp *qp,
268 static void ntb_complete_rxc(struct ntb_transport_qp *qp);
269 static void ntb_transport_doorbell_callback(void *data, uint32_t vector);
270 static void ntb_transport_event_callback(void *data);
271 static void ntb_transport_link_work(void *arg);
272 static int ntb_set_mw(struct ntb_transport_ctx *, int num_mw, size_t size);
273 static void ntb_free_mw(struct ntb_transport_ctx *nt, int num_mw);
274 static int ntb_transport_setup_qp_mw(struct ntb_transport_ctx *nt,
275 unsigned int qp_num);
276 static void ntb_qp_link_work(void *arg);
277 static void ntb_transport_link_cleanup(struct ntb_transport_ctx *nt);
278 static void ntb_transport_link_cleanup_work(void *, int);
279 static void ntb_qp_link_down(struct ntb_transport_qp *qp);
280 static void ntb_qp_link_down_reset(struct ntb_transport_qp *qp);
281 static void ntb_qp_link_cleanup(struct ntb_transport_qp *qp);
282 static void ntb_send_link_down(struct ntb_transport_qp *qp);
283 static void ntb_list_add(struct mtx *lock, struct ntb_queue_entry *entry,
284 struct ntb_queue_list *list);
285 static struct ntb_queue_entry *ntb_list_rm(struct mtx *lock,
286 struct ntb_queue_list *list);
287 static struct ntb_queue_entry *ntb_list_mv(struct mtx *lock,
288 struct ntb_queue_list *from, struct ntb_queue_list *to);
289 static void xeon_link_watchdog_hb(void *);
291 static const struct ntb_ctx_ops ntb_transport_ops = {
292 .link_event = ntb_transport_event_callback,
293 .db_event = ntb_transport_doorbell_callback,
296 MALLOC_DEFINE(M_NTB_T, "ntb_transport", "ntb transport driver");
299 iowrite32(uint32_t val, void *addr)
302 bus_space_write_4(X86_BUS_SPACE_MEM, 0/* HACK */, (uintptr_t)addr,
306 /* Transport Init and teardown */
309 xeon_link_watchdog_hb(void *arg)
311 struct ntb_transport_ctx *nt;
314 ntb_spad_write(nt->dev, NTBT_WATCHDOG_SPAD, 0);
315 callout_reset(&nt->link_watchdog, 1 * hz, xeon_link_watchdog_hb, nt);
319 ntb_transport_probe(device_t dev)
322 device_set_desc(dev, "NTB Transport");
327 ntb_transport_attach(device_t dev)
329 struct ntb_transport_ctx *nt = device_get_softc(dev);
330 struct ntb_transport_child **cpp = &nt->child;
331 struct ntb_transport_child *nc;
332 struct ntb_transport_mw *mw;
334 int rc, i, db_count, spad_count, qp, qpu, qpo, qpt;
337 char *n, *np, *c, *name;
340 nt->mw_count = ntb_mw_count(dev);
341 spad_count = ntb_spad_count(dev);
342 db_bitmap = ntb_db_valid_mask(dev);
343 db_count = flsll(db_bitmap);
344 KASSERT(db_bitmap == (1 << db_count) - 1,
345 ("Doorbells are not sequential (%jx).\n", db_bitmap));
347 device_printf(dev, "%d memory windows, %d scratchpads, "
348 "%d doorbells\n", nt->mw_count, spad_count, db_count);
350 if (nt->mw_count == 0) {
351 device_printf(dev, "At least 1 memory window required.\n");
354 if (spad_count < 6) {
355 device_printf(dev, "At least 6 scratchpads required.\n");
358 if (spad_count < 4 + 2 * nt->mw_count) {
359 nt->mw_count = (spad_count - 4) / 2;
360 device_printf(dev, "Scratchpads enough only for %d "
361 "memory windows.\n", nt->mw_count);
363 if (db_bitmap == 0) {
364 device_printf(dev, "At least one doorbell required.\n");
368 nt->mw_vec = malloc(nt->mw_count * sizeof(*nt->mw_vec), M_NTB_T,
370 for (i = 0; i < nt->mw_count; i++) {
373 rc = ntb_mw_get_range(dev, i, &mw->phys_addr, &mw->vbase,
374 &mw->phys_size, &mw->xlat_align, &mw->xlat_align_size,
381 mw->virt_addr = NULL;
384 rc = ntb_mw_set_wc(dev, i, VM_MEMATTR_WRITE_COMBINING);
386 ntb_printf(0, "Unable to set mw%d caching\n", i);
390 qpo = imin(db_count, nt->mw_count);
393 snprintf(buf, sizeof(buf), "hint.%s.%d.config", device_get_name(dev),
394 device_get_unit(dev));
395 TUNABLE_STR_FETCH(buf, cfg, sizeof(cfg));
398 while ((c = strsep(&n, ",")) != NULL) {
400 name = strsep(&np, ":");
401 if (name != NULL && name[0] == 0)
403 qp = (np && np[0] != 0) ? strtol(np, NULL, 10) : qpo - qpu;
407 if (qp > qpt - qpu) {
408 device_printf(dev, "Not enough resources for config\n");
412 nc = malloc(sizeof(*nc), M_DEVBUF, M_WAITOK | M_ZERO);
415 nc->dev = device_add_child(dev, name, -1);
416 if (nc->dev == NULL) {
417 device_printf(dev, "Can not add child.\n");
420 device_set_ivars(nc->dev, nc);
425 device_printf(dev, "%d \"%s\": queues %d",
428 printf("-%d", qpu + qp - 1);
437 nt->qp_vec = malloc(nt->qp_count * sizeof(*nt->qp_vec), M_NTB_T,
440 for (i = 0; i < nt->qp_count; i++)
441 ntb_transport_init_queue(nt, i);
443 callout_init(&nt->link_work, 0);
444 callout_init(&nt->link_watchdog, 0);
445 TASK_INIT(&nt->link_cleanup, 0, ntb_transport_link_cleanup_work, nt);
447 rc = ntb_set_ctx(dev, nt, &ntb_transport_ops);
451 nt->link_is_up = false;
452 ntb_link_enable(dev, NTB_SPEED_AUTO, NTB_WIDTH_AUTO);
454 if (enable_xeon_watchdog != 0)
455 callout_reset(&nt->link_watchdog, 0, xeon_link_watchdog_hb, nt);
457 bus_generic_attach(dev);
461 free(nt->qp_vec, M_NTB_T);
462 free(nt->mw_vec, M_NTB_T);
467 ntb_transport_detach(device_t dev)
469 struct ntb_transport_ctx *nt = device_get_softc(dev);
470 struct ntb_transport_child **cpp = &nt->child;
471 struct ntb_transport_child *nc;
474 while ((nc = *cpp) != NULL) {
476 error = device_delete_child(dev, nc->dev);
481 KASSERT(nt->qp_bitmap == 0,
482 ("Some queues not freed on detach (%jx)", nt->qp_bitmap));
484 ntb_transport_link_cleanup(nt);
485 taskqueue_drain(taskqueue_swi, &nt->link_cleanup);
486 callout_drain(&nt->link_work);
487 callout_drain(&nt->link_watchdog);
489 ntb_link_disable(dev);
492 for (i = 0; i < nt->mw_count; i++)
495 free(nt->qp_vec, M_NTB_T);
496 free(nt->mw_vec, M_NTB_T);
501 ntb_transport_queue_count(device_t dev)
503 struct ntb_transport_child *nc = device_get_ivars(dev);
509 ntb_transport_init_queue(struct ntb_transport_ctx *nt, unsigned int qp_num)
511 struct ntb_transport_mw *mw;
512 struct ntb_transport_qp *qp;
514 uint64_t mw_size, qp_offset;
516 unsigned num_qps_mw, mw_num, mw_count;
518 mw_count = nt->mw_count;
519 mw_num = QP_TO_MW(nt, qp_num);
520 mw = &nt->mw_vec[mw_num];
522 qp = &nt->qp_vec[qp_num];
526 qp->client_ready = false;
527 qp->event_handler = NULL;
528 ntb_qp_link_down_reset(qp);
530 if (mw_num < nt->qp_count % mw_count)
531 num_qps_mw = nt->qp_count / mw_count + 1;
533 num_qps_mw = nt->qp_count / mw_count;
535 mw_base = mw->phys_addr;
536 mw_size = mw->phys_size;
538 tx_size = mw_size / num_qps_mw;
539 qp_offset = tx_size * (qp_num / mw_count);
541 qp->tx_mw = mw->vbase + qp_offset;
542 KASSERT(qp->tx_mw != NULL, ("uh oh?"));
544 /* XXX Assumes that a vm_paddr_t is equivalent to bus_addr_t */
545 qp->tx_mw_phys = mw_base + qp_offset;
546 KASSERT(qp->tx_mw_phys != 0, ("uh oh?"));
548 tx_size -= sizeof(struct ntb_rx_info);
549 qp->rx_info = (void *)(qp->tx_mw + tx_size);
551 /* Due to house-keeping, there must be at least 2 buffs */
552 qp->tx_max_frame = qmin(transport_mtu, tx_size / 2);
553 qp->tx_max_entry = tx_size / qp->tx_max_frame;
555 callout_init(&qp->link_work, 0);
556 callout_init(&qp->rx_full, 1);
558 mtx_init(&qp->ntb_rx_q_lock, "ntb rx q", NULL, MTX_SPIN);
559 mtx_init(&qp->ntb_tx_free_q_lock, "ntb tx free q", NULL, MTX_SPIN);
560 mtx_init(&qp->tx_lock, "ntb transport tx", NULL, MTX_DEF);
561 TASK_INIT(&qp->rxc_db_work, 0, ntb_transport_rxc_db, qp);
562 qp->rxc_tq = taskqueue_create("ntbt_rx", M_WAITOK,
563 taskqueue_thread_enqueue, &qp->rxc_tq);
564 taskqueue_start_threads(&qp->rxc_tq, 1, PI_NET, "%s rx%d",
565 device_get_nameunit(nt->dev), qp_num);
567 STAILQ_INIT(&qp->rx_post_q);
568 STAILQ_INIT(&qp->rx_pend_q);
569 STAILQ_INIT(&qp->tx_free_q);
573 ntb_transport_free_queue(struct ntb_transport_qp *qp)
575 struct ntb_transport_ctx *nt = qp->transport;
576 struct ntb_queue_entry *entry;
578 callout_drain(&qp->link_work);
580 ntb_db_set_mask(qp->dev, 1ull << qp->qp_num);
581 taskqueue_drain_all(qp->rxc_tq);
582 taskqueue_free(qp->rxc_tq);
585 qp->rx_handler = NULL;
586 qp->tx_handler = NULL;
587 qp->event_handler = NULL;
589 while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_pend_q)))
590 free(entry, M_NTB_T);
592 while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_post_q)))
593 free(entry, M_NTB_T);
595 while ((entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q)))
596 free(entry, M_NTB_T);
598 nt->qp_bitmap &= ~(1 << qp->qp_num);
602 * ntb_transport_create_queue - Create a new NTB transport layer queue
603 * @rx_handler: receive callback function
604 * @tx_handler: transmit callback function
605 * @event_handler: event callback function
607 * Create a new NTB transport layer queue and provide the queue with a callback
608 * routine for both transmit and receive. The receive callback routine will be
609 * used to pass up data when the transport has received it on the queue. The
610 * transmit callback routine will be called when the transport has completed the
611 * transmission of the data on the queue and the data is ready to be freed.
613 * RETURNS: pointer to newly created ntb_queue, NULL on error.
615 struct ntb_transport_qp *
616 ntb_transport_create_queue(device_t dev, int q,
617 const struct ntb_queue_handlers *handlers, void *data)
619 struct ntb_transport_child *nc = device_get_ivars(dev);
620 struct ntb_transport_ctx *nt = device_get_softc(device_get_parent(dev));
621 struct ntb_queue_entry *entry;
622 struct ntb_transport_qp *qp;
625 if (q < 0 || q >= nc->qpcnt)
628 qp = &nt->qp_vec[nc->qpoff + q];
629 nt->qp_bitmap |= (1 << qp->qp_num);
631 qp->rx_handler = handlers->rx_handler;
632 qp->tx_handler = handlers->tx_handler;
633 qp->event_handler = handlers->event_handler;
635 for (i = 0; i < NTB_QP_DEF_NUM_ENTRIES; i++) {
636 entry = malloc(sizeof(*entry), M_NTB_T, M_WAITOK | M_ZERO);
637 entry->cb_data = data;
639 entry->len = transport_mtu;
641 ntb_list_add(&qp->ntb_rx_q_lock, entry, &qp->rx_pend_q);
644 for (i = 0; i < NTB_QP_DEF_NUM_ENTRIES; i++) {
645 entry = malloc(sizeof(*entry), M_NTB_T, M_WAITOK | M_ZERO);
647 ntb_list_add(&qp->ntb_tx_free_q_lock, entry, &qp->tx_free_q);
650 ntb_db_clear(dev, 1ull << qp->qp_num);
655 * ntb_transport_link_up - Notify NTB transport of client readiness to use queue
656 * @qp: NTB transport layer queue to be enabled
658 * Notify NTB transport layer of client readiness to use queue
661 ntb_transport_link_up(struct ntb_transport_qp *qp)
663 struct ntb_transport_ctx *nt = qp->transport;
665 qp->client_ready = true;
667 ntb_printf(2, "qp %d client ready\n", qp->qp_num);
670 callout_reset(&qp->link_work, 0, ntb_qp_link_work, qp);
678 * ntb_transport_tx_enqueue - Enqueue a new NTB queue entry
679 * @qp: NTB transport layer queue the entry is to be enqueued on
680 * @cb: per buffer pointer for callback function to use
681 * @data: pointer to data buffer that will be sent
682 * @len: length of the data buffer
684 * Enqueue a new transmit buffer onto the transport queue from which a NTB
685 * payload will be transmitted. This assumes that a lock is being held to
686 * serialize access to the qp.
688 * RETURNS: An appropriate ERRNO error value on error, or zero for success.
691 ntb_transport_tx_enqueue(struct ntb_transport_qp *qp, void *cb, void *data,
694 struct ntb_queue_entry *entry;
697 if (!qp->link_is_up || len == 0) {
698 CTR0(KTR_NTB, "TX: link not up");
702 entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q);
704 CTR0(KTR_NTB, "TX: could not get entry from tx_free_q");
708 CTR1(KTR_NTB, "TX: got entry %p from tx_free_q", entry);
715 mtx_lock(&qp->tx_lock);
716 rc = ntb_process_tx(qp, entry);
717 mtx_unlock(&qp->tx_lock);
719 ntb_list_add(&qp->ntb_tx_free_q_lock, entry, &qp->tx_free_q);
721 "TX: process_tx failed. Returning entry %p to tx_free_q",
728 ntb_tx_copy_callback(void *data)
730 struct ntb_queue_entry *entry = data;
731 struct ntb_transport_qp *qp = entry->qp;
732 struct ntb_payload_header *hdr = entry->x_hdr;
734 iowrite32(entry->flags | NTBT_DESC_DONE_FLAG, &hdr->flags);
735 CTR1(KTR_NTB, "TX: hdr %p set DESC_DONE", hdr);
737 ntb_peer_db_set(qp->dev, 1ull << qp->qp_num);
740 * The entry length can only be zero if the packet is intended to be a
741 * "link down" or similar. Since no payload is being sent in these
742 * cases, there is nothing to add to the completion queue.
744 if (entry->len > 0) {
745 qp->tx_bytes += entry->len;
748 qp->tx_handler(qp, qp->cb_data, entry->buf,
756 "TX: entry %p sent. hdr->ver = %u, hdr->flags = 0x%x, Returning "
757 "to tx_free_q", entry, hdr->ver, hdr->flags);
758 ntb_list_add(&qp->ntb_tx_free_q_lock, entry, &qp->tx_free_q);
762 ntb_memcpy_tx(struct ntb_queue_entry *entry, void *offset)
765 CTR2(KTR_NTB, "TX: copying %d bytes to offset %p", entry->len, offset);
766 if (entry->buf != NULL) {
767 m_copydata((struct mbuf *)entry->buf, 0, entry->len, offset);
770 * Ensure that the data is fully copied before setting the
776 ntb_tx_copy_callback(entry);
780 ntb_async_tx(struct ntb_transport_qp *qp, struct ntb_queue_entry *entry)
782 struct ntb_payload_header *hdr;
785 offset = qp->tx_mw + qp->tx_max_frame * qp->tx_index;
786 hdr = (struct ntb_payload_header *)((char *)offset + qp->tx_max_frame -
787 sizeof(struct ntb_payload_header));
790 iowrite32(entry->len, &hdr->len);
791 iowrite32(qp->tx_pkts, &hdr->ver);
793 ntb_memcpy_tx(entry, offset);
797 ntb_process_tx(struct ntb_transport_qp *qp, struct ntb_queue_entry *entry)
801 "TX: process_tx: tx_pkts=%lu, tx_index=%u, remote entry=%u",
802 qp->tx_pkts, qp->tx_index, qp->remote_rx_info->entry);
803 if (qp->tx_index == qp->remote_rx_info->entry) {
804 CTR0(KTR_NTB, "TX: ring full");
809 if (entry->len > qp->tx_max_frame - sizeof(struct ntb_payload_header)) {
810 if (qp->tx_handler != NULL)
811 qp->tx_handler(qp, qp->cb_data, entry->buf,
817 ntb_list_add(&qp->ntb_tx_free_q_lock, entry, &qp->tx_free_q);
819 "TX: frame too big. returning entry %p to tx_free_q",
823 CTR2(KTR_NTB, "TX: copying entry %p to index %u", entry, qp->tx_index);
824 ntb_async_tx(qp, entry);
827 qp->tx_index %= qp->tx_max_entry;
836 ntb_transport_rxc_db(void *arg, int pending __unused)
838 struct ntb_transport_qp *qp = arg;
841 CTR0(KTR_NTB, "RX: transport_rx");
843 while ((rc = ntb_process_rxc(qp)) == 0)
845 CTR1(KTR_NTB, "RX: process_rxc returned %d", rc);
847 if ((ntb_db_read(qp->dev) & (1ull << qp->qp_num)) != 0) {
848 /* If db is set, clear it and check queue once more. */
849 ntb_db_clear(qp->dev, 1ull << qp->qp_num);
855 ntb_process_rxc(struct ntb_transport_qp *qp)
857 struct ntb_payload_header *hdr;
858 struct ntb_queue_entry *entry;
861 offset = qp->rx_buff + qp->rx_max_frame * qp->rx_index;
862 hdr = (void *)(offset + qp->rx_max_frame -
863 sizeof(struct ntb_payload_header));
865 CTR1(KTR_NTB, "RX: process_rxc rx_index = %u", qp->rx_index);
866 if ((hdr->flags & NTBT_DESC_DONE_FLAG) == 0) {
867 CTR0(KTR_NTB, "RX: hdr not done");
872 if ((hdr->flags & NTBT_LINK_DOWN_FLAG) != 0) {
873 CTR0(KTR_NTB, "RX: link down");
874 ntb_qp_link_down(qp);
879 if (hdr->ver != (uint32_t)qp->rx_pkts) {
880 CTR2(KTR_NTB,"RX: ver != rx_pkts (%x != %lx). "
881 "Returning entry to rx_pend_q", hdr->ver, qp->rx_pkts);
886 entry = ntb_list_mv(&qp->ntb_rx_q_lock, &qp->rx_pend_q, &qp->rx_post_q);
889 CTR0(KTR_NTB, "RX: No entries in rx_pend_q");
892 callout_stop(&qp->rx_full);
893 CTR1(KTR_NTB, "RX: rx entry %p from rx_pend_q", entry);
896 entry->index = qp->rx_index;
898 if (hdr->len > entry->len) {
899 CTR2(KTR_NTB, "RX: len too long. Wanted %ju got %ju",
900 (uintmax_t)hdr->len, (uintmax_t)entry->len);
904 entry->flags |= NTBT_DESC_DONE_FLAG;
906 ntb_complete_rxc(qp);
908 qp->rx_bytes += hdr->len;
911 CTR1(KTR_NTB, "RX: received %ld rx_pkts", qp->rx_pkts);
913 entry->len = hdr->len;
915 ntb_memcpy_rx(qp, entry, offset);
919 qp->rx_index %= qp->rx_max_entry;
924 ntb_memcpy_rx(struct ntb_transport_qp *qp, struct ntb_queue_entry *entry,
927 struct ifnet *ifp = entry->cb_data;
928 unsigned int len = entry->len;
930 CTR2(KTR_NTB, "RX: copying %d bytes from offset %p", len, offset);
932 entry->buf = (void *)m_devget(offset, len, 0, ifp, NULL);
933 if (entry->buf == NULL)
934 entry->len = -ENOMEM;
936 /* Ensure that the data is globally visible before clearing the flag */
939 CTR2(KTR_NTB, "RX: copied entry %p to mbuf %p.", entry, entry->buf);
940 ntb_rx_copy_callback(qp, entry);
944 ntb_rx_copy_callback(struct ntb_transport_qp *qp, void *data)
946 struct ntb_queue_entry *entry;
949 entry->flags |= NTBT_DESC_DONE_FLAG;
950 ntb_complete_rxc(qp);
954 ntb_complete_rxc(struct ntb_transport_qp *qp)
956 struct ntb_queue_entry *entry;
960 CTR0(KTR_NTB, "RX: rx_completion_task");
962 mtx_lock_spin(&qp->ntb_rx_q_lock);
964 while (!STAILQ_EMPTY(&qp->rx_post_q)) {
965 entry = STAILQ_FIRST(&qp->rx_post_q);
966 if ((entry->flags & NTBT_DESC_DONE_FLAG) == 0)
969 entry->x_hdr->flags = 0;
970 iowrite32(entry->index, &qp->rx_info->entry);
972 STAILQ_REMOVE_HEAD(&qp->rx_post_q, entry);
978 * Re-initialize queue_entry for reuse; rx_handler takes
979 * ownership of the mbuf.
982 entry->len = transport_mtu;
983 entry->cb_data = qp->cb_data;
985 STAILQ_INSERT_TAIL(&qp->rx_pend_q, entry, entry);
987 mtx_unlock_spin(&qp->ntb_rx_q_lock);
989 CTR2(KTR_NTB, "RX: completing entry %p, mbuf %p", entry, m);
990 if (qp->rx_handler != NULL && qp->client_ready)
991 qp->rx_handler(qp, qp->cb_data, m, len);
995 mtx_lock_spin(&qp->ntb_rx_q_lock);
998 mtx_unlock_spin(&qp->ntb_rx_q_lock);
1002 ntb_transport_doorbell_callback(void *data, uint32_t vector)
1004 struct ntb_transport_ctx *nt = data;
1005 struct ntb_transport_qp *qp;
1009 vec_mask = ntb_db_vector_mask(nt->dev, vector);
1010 vec_mask &= nt->qp_bitmap;
1011 if ((vec_mask & (vec_mask - 1)) != 0)
1012 vec_mask &= ntb_db_read(nt->dev);
1013 while (vec_mask != 0) {
1014 qp_num = ffsll(vec_mask) - 1;
1016 qp = &nt->qp_vec[qp_num];
1018 taskqueue_enqueue(qp->rxc_tq, &qp->rxc_db_work);
1020 vec_mask &= ~(1ull << qp_num);
1024 /* Link Event handler */
1026 ntb_transport_event_callback(void *data)
1028 struct ntb_transport_ctx *nt = data;
1030 if (ntb_link_is_up(nt->dev, NULL, NULL)) {
1031 ntb_printf(1, "HW link up\n");
1032 callout_reset(&nt->link_work, 0, ntb_transport_link_work, nt);
1034 ntb_printf(1, "HW link down\n");
1035 taskqueue_enqueue(taskqueue_swi, &nt->link_cleanup);
1041 ntb_transport_link_work(void *arg)
1043 struct ntb_transport_ctx *nt = arg;
1044 device_t dev = nt->dev;
1045 struct ntb_transport_qp *qp;
1046 uint64_t val64, size;
1051 /* send the local info, in the opposite order of the way we read it */
1052 for (i = 0; i < nt->mw_count; i++) {
1053 size = nt->mw_vec[i].phys_size;
1055 if (max_mw_size != 0 && size > max_mw_size)
1058 ntb_peer_spad_write(dev, NTBT_MW0_SZ_HIGH + (i * 2),
1060 ntb_peer_spad_write(dev, NTBT_MW0_SZ_LOW + (i * 2), size);
1062 ntb_peer_spad_write(dev, NTBT_NUM_MWS, nt->mw_count);
1063 ntb_peer_spad_write(dev, NTBT_NUM_QPS, nt->qp_count);
1064 ntb_peer_spad_write(dev, NTBT_QP_LINKS, 0);
1065 ntb_peer_spad_write(dev, NTBT_VERSION, NTB_TRANSPORT_VERSION);
1067 /* Query the remote side for its info */
1069 ntb_spad_read(dev, NTBT_VERSION, &val);
1070 if (val != NTB_TRANSPORT_VERSION)
1073 ntb_spad_read(dev, NTBT_NUM_QPS, &val);
1074 if (val != nt->qp_count)
1077 ntb_spad_read(dev, NTBT_NUM_MWS, &val);
1078 if (val != nt->mw_count)
1081 for (i = 0; i < nt->mw_count; i++) {
1082 ntb_spad_read(dev, NTBT_MW0_SZ_HIGH + (i * 2), &val);
1083 val64 = (uint64_t)val << 32;
1085 ntb_spad_read(dev, NTBT_MW0_SZ_LOW + (i * 2), &val);
1088 rc = ntb_set_mw(nt, i, val64);
1093 nt->link_is_up = true;
1094 ntb_printf(1, "transport link up\n");
1096 for (i = 0; i < nt->qp_count; i++) {
1097 qp = &nt->qp_vec[i];
1099 ntb_transport_setup_qp_mw(nt, i);
1101 if (qp->client_ready)
1102 callout_reset(&qp->link_work, 0, ntb_qp_link_work, qp);
1108 for (i = 0; i < nt->mw_count; i++)
1111 if (ntb_link_is_up(dev, NULL, NULL))
1112 callout_reset(&nt->link_work,
1113 NTB_LINK_DOWN_TIMEOUT * hz / 1000, ntb_transport_link_work, nt);
1117 ntb_set_mw(struct ntb_transport_ctx *nt, int num_mw, size_t size)
1119 struct ntb_transport_mw *mw = &nt->mw_vec[num_mw];
1120 size_t xlat_size, buff_size;
1126 xlat_size = roundup(size, mw->xlat_align_size);
1127 buff_size = xlat_size;
1129 /* No need to re-setup */
1130 if (mw->xlat_size == xlat_size)
1133 if (mw->buff_size != 0)
1134 ntb_free_mw(nt, num_mw);
1136 /* Alloc memory for receiving data. Must be aligned */
1137 mw->xlat_size = xlat_size;
1138 mw->buff_size = buff_size;
1140 mw->virt_addr = contigmalloc(mw->buff_size, M_NTB_T, M_ZERO, 0,
1141 mw->addr_limit, mw->xlat_align, 0);
1142 if (mw->virt_addr == NULL) {
1143 ntb_printf(0, "Unable to allocate MW buffer of size %zu/%zu\n",
1144 mw->buff_size, mw->xlat_size);
1149 /* TODO: replace with bus_space_* functions */
1150 mw->dma_addr = vtophys(mw->virt_addr);
1153 * Ensure that the allocation from contigmalloc is aligned as
1154 * requested. XXX: This may not be needed -- brought in for parity
1155 * with the Linux driver.
1157 if (mw->dma_addr % mw->xlat_align != 0) {
1159 "DMA memory 0x%jx not aligned to BAR size 0x%zx\n",
1160 (uintmax_t)mw->dma_addr, size);
1161 ntb_free_mw(nt, num_mw);
1165 /* Notify HW the memory location of the receive buffer */
1166 rc = ntb_mw_set_trans(nt->dev, num_mw, mw->dma_addr, mw->xlat_size);
1168 ntb_printf(0, "Unable to set mw%d translation\n", num_mw);
1169 ntb_free_mw(nt, num_mw);
1177 ntb_free_mw(struct ntb_transport_ctx *nt, int num_mw)
1179 struct ntb_transport_mw *mw = &nt->mw_vec[num_mw];
1181 if (mw->virt_addr == NULL)
1184 ntb_mw_clear_trans(nt->dev, num_mw);
1185 contigfree(mw->virt_addr, mw->xlat_size, M_NTB_T);
1188 mw->virt_addr = NULL;
1192 ntb_transport_setup_qp_mw(struct ntb_transport_ctx *nt, unsigned int qp_num)
1194 struct ntb_transport_qp *qp = &nt->qp_vec[qp_num];
1195 struct ntb_transport_mw *mw;
1199 unsigned num_qps_mw, mw_num, mw_count;
1201 mw_count = nt->mw_count;
1202 mw_num = QP_TO_MW(nt, qp_num);
1203 mw = &nt->mw_vec[mw_num];
1205 if (mw->virt_addr == NULL)
1208 if (mw_num < nt->qp_count % mw_count)
1209 num_qps_mw = nt->qp_count / mw_count + 1;
1211 num_qps_mw = nt->qp_count / mw_count;
1213 rx_size = mw->xlat_size / num_qps_mw;
1214 qp->rx_buff = mw->virt_addr + rx_size * (qp_num / mw_count);
1215 rx_size -= sizeof(struct ntb_rx_info);
1217 qp->remote_rx_info = (void*)(qp->rx_buff + rx_size);
1219 /* Due to house-keeping, there must be at least 2 buffs */
1220 qp->rx_max_frame = qmin(transport_mtu, rx_size / 2);
1221 qp->rx_max_entry = rx_size / qp->rx_max_frame;
1224 qp->remote_rx_info->entry = qp->rx_max_entry - 1;
1226 /* Set up the hdr offsets with 0s */
1227 for (i = 0; i < qp->rx_max_entry; i++) {
1228 offset = (void *)(qp->rx_buff + qp->rx_max_frame * (i + 1) -
1229 sizeof(struct ntb_payload_header));
1230 memset(offset, 0, sizeof(struct ntb_payload_header));
1241 ntb_qp_link_work(void *arg)
1243 struct ntb_transport_qp *qp = arg;
1244 device_t dev = qp->dev;
1245 struct ntb_transport_ctx *nt = qp->transport;
1249 /* Report queues that are up on our side */
1250 for (i = 0, val = 0; i < nt->qp_count; i++) {
1251 if (nt->qp_vec[i].client_ready)
1254 ntb_peer_spad_write(dev, NTBT_QP_LINKS, val);
1256 /* See if the remote side is up */
1257 ntb_spad_read(dev, NTBT_QP_LINKS, &val);
1258 if ((val & (1ull << qp->qp_num)) != 0) {
1259 ntb_printf(2, "qp %d link up\n", qp->qp_num);
1260 qp->link_is_up = true;
1262 if (qp->event_handler != NULL)
1263 qp->event_handler(qp->cb_data, NTB_LINK_UP);
1265 ntb_db_clear_mask(dev, 1ull << qp->qp_num);
1266 } else if (nt->link_is_up)
1267 callout_reset(&qp->link_work,
1268 NTB_LINK_DOWN_TIMEOUT * hz / 1000, ntb_qp_link_work, qp);
1271 /* Link down event*/
1273 ntb_transport_link_cleanup(struct ntb_transport_ctx *nt)
1275 struct ntb_transport_qp *qp;
1278 /* Pass along the info to any clients */
1279 for (i = 0; i < nt->qp_count; i++) {
1280 if ((nt->qp_bitmap & (1 << i)) != 0) {
1281 qp = &nt->qp_vec[i];
1282 ntb_qp_link_cleanup(qp);
1283 callout_drain(&qp->link_work);
1287 if (!nt->link_is_up)
1288 callout_drain(&nt->link_work);
1291 * The scratchpad registers keep the values if the remote side
1292 * goes down, blast them now to give them a sane value the next
1293 * time they are accessed
1295 ntb_spad_clear(nt->dev);
1299 ntb_transport_link_cleanup_work(void *arg, int pending __unused)
1302 ntb_transport_link_cleanup(arg);
1306 ntb_qp_link_down(struct ntb_transport_qp *qp)
1309 ntb_qp_link_cleanup(qp);
1313 ntb_qp_link_down_reset(struct ntb_transport_qp *qp)
1316 qp->link_is_up = false;
1317 ntb_db_set_mask(qp->dev, 1ull << qp->qp_num);
1319 qp->tx_index = qp->rx_index = 0;
1320 qp->tx_bytes = qp->rx_bytes = 0;
1321 qp->tx_pkts = qp->rx_pkts = 0;
1323 qp->rx_ring_empty = 0;
1324 qp->tx_ring_full = 0;
1326 qp->rx_err_no_buf = qp->tx_err_no_buf = 0;
1327 qp->rx_err_oflow = qp->rx_err_ver = 0;
1331 ntb_qp_link_cleanup(struct ntb_transport_qp *qp)
1334 callout_drain(&qp->link_work);
1335 ntb_qp_link_down_reset(qp);
1337 if (qp->event_handler != NULL)
1338 qp->event_handler(qp->cb_data, NTB_LINK_DOWN);
1341 /* Link commanded down */
1343 * ntb_transport_link_down - Notify NTB transport to no longer enqueue data
1344 * @qp: NTB transport layer queue to be disabled
1346 * Notify NTB transport layer of client's desire to no longer receive data on
1347 * transport queue specified. It is the client's responsibility to ensure all
1348 * entries on queue are purged or otherwise handled appropriately.
1351 ntb_transport_link_down(struct ntb_transport_qp *qp)
1353 struct ntb_transport_ctx *nt = qp->transport;
1357 qp->client_ready = false;
1358 for (i = 0, val = 0; i < nt->qp_count; i++) {
1359 if (nt->qp_vec[i].client_ready)
1362 ntb_peer_spad_write(qp->dev, NTBT_QP_LINKS, val);
1365 ntb_send_link_down(qp);
1367 callout_drain(&qp->link_work);
1371 * ntb_transport_link_query - Query transport link state
1372 * @qp: NTB transport layer queue to be queried
1374 * Query connectivity to the remote system of the NTB transport queue
1376 * RETURNS: true for link up or false for link down
1379 ntb_transport_link_query(struct ntb_transport_qp *qp)
1382 return (qp->link_is_up);
1386 ntb_send_link_down(struct ntb_transport_qp *qp)
1388 struct ntb_queue_entry *entry;
1391 if (!qp->link_is_up)
1394 for (i = 0; i < NTB_LINK_DOWN_TIMEOUT; i++) {
1395 entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q);
1398 pause("NTB Wait for link down", hz / 10);
1404 entry->cb_data = NULL;
1407 entry->flags = NTBT_LINK_DOWN_FLAG;
1409 mtx_lock(&qp->tx_lock);
1410 rc = ntb_process_tx(qp, entry);
1411 mtx_unlock(&qp->tx_lock);
1413 printf("ntb: Failed to send link down\n");
1415 ntb_qp_link_down_reset(qp);
1419 /* List Management */
1422 ntb_list_add(struct mtx *lock, struct ntb_queue_entry *entry,
1423 struct ntb_queue_list *list)
1426 mtx_lock_spin(lock);
1427 STAILQ_INSERT_TAIL(list, entry, entry);
1428 mtx_unlock_spin(lock);
1431 static struct ntb_queue_entry *
1432 ntb_list_rm(struct mtx *lock, struct ntb_queue_list *list)
1434 struct ntb_queue_entry *entry;
1436 mtx_lock_spin(lock);
1437 if (STAILQ_EMPTY(list)) {
1441 entry = STAILQ_FIRST(list);
1442 STAILQ_REMOVE_HEAD(list, entry);
1444 mtx_unlock_spin(lock);
1449 static struct ntb_queue_entry *
1450 ntb_list_mv(struct mtx *lock, struct ntb_queue_list *from,
1451 struct ntb_queue_list *to)
1453 struct ntb_queue_entry *entry;
1455 mtx_lock_spin(lock);
1456 if (STAILQ_EMPTY(from)) {
1460 entry = STAILQ_FIRST(from);
1461 STAILQ_REMOVE_HEAD(from, entry);
1462 STAILQ_INSERT_TAIL(to, entry, entry);
1465 mtx_unlock_spin(lock);
1470 * ntb_transport_qp_num - Query the qp number
1471 * @qp: NTB transport layer queue to be queried
1473 * Query qp number of the NTB transport queue
1475 * RETURNS: a zero based number specifying the qp number
1477 unsigned char ntb_transport_qp_num(struct ntb_transport_qp *qp)
1480 return (qp->qp_num);
1484 * ntb_transport_max_size - Query the max payload size of a qp
1485 * @qp: NTB transport layer queue to be queried
1487 * Query the maximum payload size permissible on the given qp
1489 * RETURNS: the max payload size of a qp
1492 ntb_transport_max_size(struct ntb_transport_qp *qp)
1495 return (qp->tx_max_frame - sizeof(struct ntb_payload_header));
1499 ntb_transport_tx_free_entry(struct ntb_transport_qp *qp)
1501 unsigned int head = qp->tx_index;
1502 unsigned int tail = qp->remote_rx_info->entry;
1504 return (tail >= head ? tail - head : qp->tx_max_entry + tail - head);
1507 static device_method_t ntb_transport_methods[] = {
1508 /* Device interface */
1509 DEVMETHOD(device_probe, ntb_transport_probe),
1510 DEVMETHOD(device_attach, ntb_transport_attach),
1511 DEVMETHOD(device_detach, ntb_transport_detach),
1515 devclass_t ntb_transport_devclass;
1516 static DEFINE_CLASS_0(ntb_transport, ntb_transport_driver,
1517 ntb_transport_methods, sizeof(struct ntb_transport_ctx));
1518 DRIVER_MODULE(ntb_transport, ntb_hw, ntb_transport_driver,
1519 ntb_transport_devclass, NULL, NULL);
1520 MODULE_DEPEND(ntb_transport, ntb, 1, 1, 1);
1521 MODULE_VERSION(ntb_transport, 1);