2 * Copyright (C) 2013 Intel Corporation
3 * Copyright (C) 2015 EMC Corporation
6 * Redistribution and use in source and binary forms, with or without
7 * modification, are permitted provided that the following conditions
9 * 1. Redistributions of source code must retain the above copyright
10 * notice, this list of conditions and the following disclaimer.
11 * 2. Redistributions in binary form must reproduce the above copyright
12 * notice, this list of conditions and the following disclaimer in the
13 * documentation and/or other materials provided with the distribution.
15 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
16 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
19 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
28 #include <sys/cdefs.h>
29 __FBSDID("$FreeBSD$");
31 #include <sys/param.h>
32 #include <sys/kernel.h>
33 #include <sys/systm.h>
34 #include <sys/bitset.h>
37 #include <sys/limits.h>
39 #include <sys/malloc.h>
40 #include <sys/module.h>
41 #include <sys/mutex.h>
42 #include <sys/queue.h>
43 #include <sys/socket.h>
44 #include <sys/sockio.h>
45 #include <sys/sysctl.h>
46 #include <sys/taskqueue.h>
49 #include <net/if_media.h>
50 #include <net/if_types.h>
51 #include <net/if_var.h>
53 #include <net/ethernet.h>
58 #include <machine/bus.h>
59 #include <machine/cpufunc.h>
60 #include <machine/pmap.h>
62 #include <netinet/in.h>
63 #include <netinet/ip.h>
65 #include "../ntb_hw/ntb_hw.h"
68 * The Non-Transparent Bridge (NTB) is a device on some Intel processors that
69 * allows you to connect two systems using a PCI-e link.
71 * This module contains a protocol for sending and receiving messages, and
72 * exposes that protocol through a simulated ethernet device called ntb.
74 * NOTE: Much of the code in this module is shared with Linux. Any patches may
75 * be picked up and redistributed in Linux with a dual GPL/BSD license.
79 BITSET_DEFINE(_qpset, QP_SETSIZE);
80 #define test_bit(pos, addr) BIT_ISSET(QP_SETSIZE, (pos), (addr))
81 #define set_bit(pos, addr) BIT_SET(QP_SETSIZE, (pos), (addr))
82 #define clear_bit(pos, addr) BIT_CLR(QP_SETSIZE, (pos), (addr))
83 #define ffs_bit(addr) BIT_FFS(QP_SETSIZE, (addr))
85 #define KTR_NTB KTR_SPARE3
87 #define NTB_TRANSPORT_VERSION 4
88 #define NTB_RX_MAX_PKTS 64
89 #define NTB_RXQ_SIZE 300
96 static SYSCTL_NODE(_hw, OID_AUTO, if_ntb, CTLFLAG_RW, 0, "if_ntb");
98 static unsigned g_if_ntb_debug_level;
99 TUNABLE_INT("hw.if_ntb.debug_level", &g_if_ntb_debug_level);
100 SYSCTL_UINT(_hw_if_ntb, OID_AUTO, debug_level, CTLFLAG_RWTUN,
101 &g_if_ntb_debug_level, 0, "if_ntb log level -- higher is more verbose");
102 #define ntb_printf(lvl, ...) do { \
103 if ((lvl) <= g_if_ntb_debug_level) { \
104 if_printf(nt->ifp, __VA_ARGS__); \
108 static unsigned transport_mtu = IP_MAXPACKET + ETHER_HDR_LEN + ETHER_CRC_LEN;
110 static uint64_t max_mw_size;
111 TUNABLE_QUAD("hw.if_ntb.max_mw_size", &max_mw_size);
112 SYSCTL_UQUAD(_hw_if_ntb, OID_AUTO, max_mw_size, CTLFLAG_RDTUN, &max_mw_size, 0,
113 "If enabled (non-zero), limit the size of large memory windows. "
114 "Both sides of the NTB MUST set the same value here.");
116 static unsigned max_num_clients;
117 TUNABLE_INT("hw.if_ntb.max_num_clients", &max_num_clients);
118 SYSCTL_UINT(_hw_if_ntb, OID_AUTO, max_num_clients, CTLFLAG_RDTUN,
119 &max_num_clients, 0, "Maximum number of NTB transport clients. "
120 "0 (default) - use all available NTB memory windows; "
121 "positive integer N - Limit to N memory windows.");
123 static unsigned enable_xeon_watchdog;
124 TUNABLE_INT("hw.if_ntb.enable_xeon_watchdog", &enable_xeon_watchdog);
125 SYSCTL_UINT(_hw_if_ntb, OID_AUTO, enable_xeon_watchdog, CTLFLAG_RDTUN,
126 &enable_xeon_watchdog, 0, "If non-zero, write a register every second to "
127 "keep a watchdog from tearing down the NTB link");
129 STAILQ_HEAD(ntb_queue_list, ntb_queue_entry);
131 typedef uint32_t ntb_q_idx_t;
133 struct ntb_queue_entry {
134 /* ntb_queue list reference */
135 STAILQ_ENTRY(ntb_queue_entry) entry;
137 /* info on data to be transferred */
143 struct ntb_transport_qp *qp;
144 struct ntb_payload_header *x_hdr;
152 struct ntb_transport_qp {
153 struct ntb_transport_ctx *transport;
154 struct ntb_softc *ntb;
159 volatile bool link_is_up;
160 uint8_t qp_num; /* Only 64 QPs are allowed. 0-63 */
162 struct ntb_rx_info *rx_info;
163 struct ntb_rx_info *remote_rx_info;
165 void (*tx_handler)(struct ntb_transport_qp *qp, void *qp_data,
166 void *data, int len);
167 struct ntb_queue_list tx_free_q;
168 struct mtx ntb_tx_free_q_lock;
170 bus_addr_t tx_mw_phys;
171 ntb_q_idx_t tx_index;
172 ntb_q_idx_t tx_max_entry;
173 uint64_t tx_max_frame;
175 void (*rx_handler)(struct ntb_transport_qp *qp, void *qp_data,
176 void *data, int len);
177 struct ntb_queue_list rx_post_q;
178 struct ntb_queue_list rx_pend_q;
179 /* ntb_rx_q_lock: synchronize access to rx_XXXX_q */
180 struct mtx ntb_rx_q_lock;
181 struct task rx_completion_task;
182 struct task rxc_db_work;
184 ntb_q_idx_t rx_index;
185 ntb_q_idx_t rx_max_entry;
186 uint64_t rx_max_frame;
188 void (*event_handler)(void *data, enum ntb_link_event status);
189 struct callout link_work;
190 struct callout queue_full;
191 struct callout rx_full;
193 uint64_t last_rx_no_buf;
198 uint64_t rx_ring_empty;
199 uint64_t rx_err_no_buf;
200 uint64_t rx_err_oflow;
204 uint64_t tx_ring_full;
205 uint64_t tx_err_no_buf;
208 struct ntb_queue_handlers {
209 void (*rx_handler)(struct ntb_transport_qp *qp, void *qp_data,
210 void *data, int len);
211 void (*tx_handler)(struct ntb_transport_qp *qp, void *qp_data,
212 void *data, int len);
213 void (*event_handler)(void *data, enum ntb_link_event status);
216 struct ntb_transport_mw {
217 vm_paddr_t phys_addr;
220 size_t xlat_align_size;
221 bus_addr_t addr_limit;
222 /* Tx buff is off vbase / phys_addr */
226 /* Rx buff is off virt_addr / dma_addr */
231 struct ntb_transport_ctx {
232 struct ntb_softc *ntb;
234 struct ntb_transport_mw mw_vec[NTB_MAX_NUM_MW];
235 struct ntb_transport_qp *qp_vec;
236 struct _qpset qp_bitmap;
237 struct _qpset qp_bitmap_free;
240 volatile bool link_is_up;
241 struct callout link_work;
242 struct callout link_watchdog;
243 struct task link_cleanup;
245 u_char eaddr[ETHER_ADDR_LEN];
249 /* The hardcoded single queuepair in ntb_setup_interface() */
250 struct ntb_transport_qp *qp;
253 static struct ntb_transport_ctx net_softc;
256 IF_NTB_DESC_DONE_FLAG = 1 << 0,
257 IF_NTB_LINK_DOWN_FLAG = 1 << 1,
260 struct ntb_payload_header {
268 * The order of this enum is part of the if_ntb remote protocol. Do
269 * not reorder without bumping protocol version (and it's probably best
270 * to keep the protocol in lock-step with the Linux NTB driver.
277 * N.B.: transport_link_work assumes MW1 enums = MW0 + 2.
286 * Some NTB-using hardware have a watchdog to work around NTB hangs; if
287 * a register or doorbell isn't written every few seconds, the link is
288 * torn down. Write an otherwise unused register every few seconds to
289 * work around this watchdog.
291 IF_NTB_WATCHDOG_SPAD = 15
293 CTASSERT(IF_NTB_WATCHDOG_SPAD < XEON_SPAD_COUNT &&
294 IF_NTB_WATCHDOG_SPAD < ATOM_SPAD_COUNT);
296 #define QP_TO_MW(nt, qp) ((qp) % nt->mw_count)
297 #define NTB_QP_DEF_NUM_ENTRIES 100
298 #define NTB_LINK_DOWN_TIMEOUT 10
300 static int ntb_handle_module_events(struct module *m, int what, void *arg);
301 static int ntb_setup_interface(void);
302 static int ntb_teardown_interface(void);
303 static void ntb_net_init(void *arg);
304 static int ntb_ioctl(struct ifnet *ifp, u_long command, caddr_t data);
305 static void ntb_start(struct ifnet *ifp);
306 static void ntb_net_tx_handler(struct ntb_transport_qp *qp, void *qp_data,
307 void *data, int len);
308 static void ntb_net_rx_handler(struct ntb_transport_qp *qp, void *qp_data,
309 void *data, int len);
310 static void ntb_net_event_handler(void *data, enum ntb_link_event status);
311 static int ntb_transport_probe(struct ntb_softc *ntb);
312 static void ntb_transport_free(struct ntb_transport_ctx *);
313 static void ntb_transport_init_queue(struct ntb_transport_ctx *nt,
314 unsigned int qp_num);
315 static void ntb_transport_free_queue(struct ntb_transport_qp *qp);
316 static struct ntb_transport_qp *ntb_transport_create_queue(void *data,
317 struct ntb_softc *pdev, const struct ntb_queue_handlers *handlers);
318 static void ntb_transport_link_up(struct ntb_transport_qp *qp);
319 static int ntb_transport_tx_enqueue(struct ntb_transport_qp *qp, void *cb,
320 void *data, unsigned int len);
321 static int ntb_process_tx(struct ntb_transport_qp *qp,
322 struct ntb_queue_entry *entry);
323 static void ntb_memcpy_tx(struct ntb_transport_qp *qp,
324 struct ntb_queue_entry *entry, void *offset);
325 static void ntb_qp_full(void *arg);
326 static void ntb_transport_rxc_db(void *arg, int pending);
327 static int ntb_process_rxc(struct ntb_transport_qp *qp);
328 static void ntb_memcpy_rx(struct ntb_transport_qp *qp,
329 struct ntb_queue_entry *entry, void *offset);
330 static inline void ntb_rx_copy_callback(struct ntb_transport_qp *qp,
332 static void ntb_complete_rxc(void *arg, int pending);
333 static void ntb_transport_doorbell_callback(void *data, uint32_t vector);
334 static void ntb_transport_event_callback(void *data);
335 static void ntb_transport_link_work(void *arg);
336 static int ntb_set_mw(struct ntb_transport_ctx *, int num_mw, size_t size);
337 static void ntb_free_mw(struct ntb_transport_ctx *nt, int num_mw);
338 static int ntb_transport_setup_qp_mw(struct ntb_transport_ctx *nt,
339 unsigned int qp_num);
340 static void ntb_qp_link_work(void *arg);
341 static void ntb_transport_link_cleanup(struct ntb_transport_ctx *nt);
342 static void ntb_transport_link_cleanup_work(void *, int);
343 static void ntb_qp_link_down(struct ntb_transport_qp *qp);
344 static void ntb_qp_link_down_reset(struct ntb_transport_qp *qp);
345 static void ntb_qp_link_cleanup(struct ntb_transport_qp *qp);
346 static void ntb_transport_link_down(struct ntb_transport_qp *qp);
347 static void ntb_send_link_down(struct ntb_transport_qp *qp);
348 static void ntb_list_add(struct mtx *lock, struct ntb_queue_entry *entry,
349 struct ntb_queue_list *list);
350 static struct ntb_queue_entry *ntb_list_rm(struct mtx *lock,
351 struct ntb_queue_list *list);
352 static struct ntb_queue_entry *ntb_list_mv(struct mtx *lock,
353 struct ntb_queue_list *from, struct ntb_queue_list *to);
354 static void create_random_local_eui48(u_char *eaddr);
355 static unsigned int ntb_transport_max_size(struct ntb_transport_qp *qp);
356 static void xeon_link_watchdog_hb(void *);
358 static const struct ntb_ctx_ops ntb_transport_ops = {
359 .link_event = ntb_transport_event_callback,
360 .db_event = ntb_transport_doorbell_callback,
363 MALLOC_DEFINE(M_NTB_IF, "if_ntb", "ntb network driver");
366 iowrite32(uint32_t val, void *addr)
369 bus_space_write_4(X86_BUS_SPACE_MEM, 0/* HACK */, (uintptr_t)addr,
373 /* Module setup and teardown */
375 ntb_handle_module_events(struct module *m, int what, void *arg)
381 err = ntb_setup_interface();
384 err = ntb_teardown_interface();
393 static moduledata_t if_ntb_mod = {
395 ntb_handle_module_events,
399 DECLARE_MODULE(if_ntb, if_ntb_mod, SI_SUB_KLD, SI_ORDER_ANY);
400 MODULE_DEPEND(if_ntb, ntb_hw, 1, 1, 1);
403 ntb_setup_interface(void)
406 struct ntb_queue_handlers handlers = { ntb_net_rx_handler,
407 ntb_net_tx_handler, ntb_net_event_handler };
410 net_softc.ntb = devclass_get_softc(devclass_find("ntb_hw"), 0);
411 if (net_softc.ntb == NULL) {
412 printf("ntb: Cannot find devclass\n");
416 ifp = net_softc.ifp = if_alloc(IFT_ETHER);
418 ntb_transport_free(&net_softc);
419 printf("ntb: Cannot allocate ifnet structure\n");
422 if_initname(ifp, "ntb", 0);
424 rc = ntb_transport_probe(net_softc.ntb);
426 printf("ntb: Cannot init transport: %d\n", rc);
427 if_free(net_softc.ifp);
431 net_softc.qp = ntb_transport_create_queue(ifp, net_softc.ntb,
433 ifp->if_init = ntb_net_init;
434 ifp->if_softc = &net_softc;
435 ifp->if_flags = IFF_BROADCAST | IFF_SIMPLEX;
436 ifp->if_ioctl = ntb_ioctl;
437 ifp->if_start = ntb_start;
438 IFQ_SET_MAXLEN(&ifp->if_snd, IFQ_MAXLEN);
439 ifp->if_snd.ifq_drv_maxlen = IFQ_MAXLEN;
440 IFQ_SET_READY(&ifp->if_snd);
441 create_random_local_eui48(net_softc.eaddr);
442 ether_ifattach(ifp, net_softc.eaddr);
443 ifp->if_capabilities = IFCAP_HWCSUM | IFCAP_JUMBO_MTU;
444 ifp->if_capenable = ifp->if_capabilities;
445 ifp->if_mtu = ntb_transport_max_size(net_softc.qp) - ETHER_HDR_LEN -
448 ntb_transport_link_up(net_softc.qp);
449 net_softc.bufsize = ntb_transport_max_size(net_softc.qp) +
450 sizeof(struct ether_header);
455 ntb_teardown_interface(void)
458 if (net_softc.qp != NULL) {
459 ntb_transport_link_down(net_softc.qp);
461 ntb_transport_free_queue(net_softc.qp);
462 ntb_transport_free(&net_softc);
465 if (net_softc.ifp != NULL) {
466 ether_ifdetach(net_softc.ifp);
467 if_free(net_softc.ifp);
468 net_softc.ifp = NULL;
474 /* Network device interface */
477 ntb_net_init(void *arg)
479 struct ntb_transport_ctx *ntb_softc = arg;
480 struct ifnet *ifp = ntb_softc->ifp;
482 ifp->if_drv_flags |= IFF_DRV_RUNNING;
483 ifp->if_drv_flags &= ~IFF_DRV_OACTIVE;
484 ifp->if_flags |= IFF_UP;
485 if_link_state_change(ifp, LINK_STATE_UP);
489 ntb_ioctl(struct ifnet *ifp, u_long command, caddr_t data)
491 struct ntb_transport_ctx *nt = ifp->if_softc;
492 struct ifreq *ifr = (struct ifreq *)data;
498 if (ifr->ifr_mtu > ntb_transport_max_size(nt->qp) -
499 ETHER_HDR_LEN - ETHER_CRC_LEN) {
504 ifp->if_mtu = ifr->ifr_mtu;
508 error = ether_ioctl(ifp, command, data);
517 ntb_start(struct ifnet *ifp)
520 struct ntb_transport_ctx *nt = ifp->if_softc;
523 mtx_lock(&nt->tx_lock);
524 ifp->if_drv_flags &= ~IFF_DRV_OACTIVE;
525 CTR0(KTR_NTB, "TX: ntb_start");
526 while (!IFQ_DRV_IS_EMPTY(&ifp->if_snd)) {
527 IFQ_DRV_DEQUEUE(&ifp->if_snd, m_head);
528 CTR1(KTR_NTB, "TX: start mbuf %p", m_head);
529 rc = ntb_transport_tx_enqueue(nt->qp, m_head, m_head,
530 m_length(m_head, NULL));
533 "TX: could not tx mbuf %p. Returning to snd q",
536 ifp->if_drv_flags |= IFF_DRV_OACTIVE;
537 IFQ_DRV_PREPEND(&ifp->if_snd, m_head);
538 callout_reset(&nt->qp->queue_full, hz / 1000,
545 mtx_unlock(&nt->tx_lock);
548 /* Network Device Callbacks */
550 ntb_net_tx_handler(struct ntb_transport_qp *qp, void *qp_data, void *data,
555 CTR1(KTR_NTB, "TX: tx_handler freeing mbuf %p", data);
559 ntb_net_rx_handler(struct ntb_transport_qp *qp, void *qp_data, void *data,
562 struct mbuf *m = data;
563 struct ifnet *ifp = qp_data;
565 CTR0(KTR_NTB, "RX: rx handler");
566 (*ifp->if_input)(ifp, m);
570 ntb_net_event_handler(void *data, enum ntb_link_event status)
577 /* XXX The Linux driver munges with the carrier status here. */
585 panic("Bogus ntb_link_event %u\n", status);
589 /* Transport Init and teardown */
592 xeon_link_watchdog_hb(void *arg)
594 struct ntb_transport_ctx *nt;
597 ntb_spad_write(nt->ntb, IF_NTB_WATCHDOG_SPAD, 0);
598 callout_reset(&nt->link_watchdog, 1 * hz, xeon_link_watchdog_hb, nt);
602 ntb_transport_probe(struct ntb_softc *ntb)
604 struct ntb_transport_ctx *nt = &net_softc;
605 struct ntb_transport_mw *mw;
610 nt->mw_count = ntb_mw_count(ntb);
611 for (i = 0; i < nt->mw_count; i++) {
614 rc = ntb_mw_get_range(ntb, i, &mw->phys_addr, &mw->vbase,
615 &mw->phys_size, &mw->xlat_align, &mw->xlat_align_size,
622 mw->virt_addr = NULL;
625 rc = ntb_mw_set_wc(nt->ntb, i, VM_MEMATTR_WRITE_COMBINING);
627 ntb_printf(0, "Unable to set mw%d caching\n", i);
630 qp_bitmap = ntb_db_valid_mask(ntb);
631 nt->qp_count = flsll(qp_bitmap);
632 KASSERT(nt->qp_count != 0, ("bogus db bitmap"));
635 if (max_num_clients != 0 && max_num_clients < nt->qp_count)
636 nt->qp_count = max_num_clients;
637 else if (nt->mw_count < nt->qp_count)
638 nt->qp_count = nt->mw_count;
639 KASSERT(nt->qp_count <= QP_SETSIZE, ("invalid qp_count"));
641 mtx_init(&nt->tx_lock, "ntb transport tx", NULL, MTX_DEF);
642 mtx_init(&nt->rx_lock, "ntb transport rx", NULL, MTX_DEF);
644 nt->qp_vec = malloc(nt->qp_count * sizeof(*nt->qp_vec), M_NTB_IF,
647 for (i = 0; i < nt->qp_count; i++) {
648 set_bit(i, &nt->qp_bitmap);
649 set_bit(i, &nt->qp_bitmap_free);
650 ntb_transport_init_queue(nt, i);
653 callout_init(&nt->link_work, 0);
654 callout_init(&nt->link_watchdog, 0);
655 TASK_INIT(&nt->link_cleanup, 0, ntb_transport_link_cleanup_work, nt);
657 rc = ntb_set_ctx(ntb, nt, &ntb_transport_ops);
661 nt->link_is_up = false;
662 ntb_link_enable(ntb, NTB_SPEED_AUTO, NTB_WIDTH_AUTO);
665 callout_reset(&nt->link_work, 0, ntb_transport_link_work, nt);
666 if (enable_xeon_watchdog != 0)
667 callout_reset(&nt->link_watchdog, 0, xeon_link_watchdog_hb, nt);
671 free(nt->qp_vec, M_NTB_IF);
677 ntb_transport_free(struct ntb_transport_ctx *nt)
679 struct ntb_softc *ntb = nt->ntb;
680 struct _qpset qp_bitmap_alloc;
683 ntb_transport_link_cleanup(nt);
684 taskqueue_drain(taskqueue_swi, &nt->link_cleanup);
685 callout_drain(&nt->link_work);
686 callout_drain(&nt->link_watchdog);
688 BIT_COPY(QP_SETSIZE, &nt->qp_bitmap, &qp_bitmap_alloc);
689 BIT_NAND(QP_SETSIZE, &qp_bitmap_alloc, &nt->qp_bitmap_free);
691 /* Verify that all the QPs are freed */
692 for (i = 0; i < nt->qp_count; i++)
693 if (test_bit(i, &qp_bitmap_alloc))
694 ntb_transport_free_queue(&nt->qp_vec[i]);
696 ntb_link_disable(ntb);
699 for (i = 0; i < nt->mw_count; i++)
702 free(nt->qp_vec, M_NTB_IF);
706 ntb_transport_init_queue(struct ntb_transport_ctx *nt, unsigned int qp_num)
708 struct ntb_transport_mw *mw;
709 struct ntb_transport_qp *qp;
711 uint64_t mw_size, qp_offset;
713 unsigned num_qps_mw, mw_num, mw_count;
715 mw_count = nt->mw_count;
716 mw_num = QP_TO_MW(nt, qp_num);
717 mw = &nt->mw_vec[mw_num];
719 qp = &nt->qp_vec[qp_num];
723 qp->client_ready = false;
724 qp->event_handler = NULL;
725 ntb_qp_link_down_reset(qp);
727 if (nt->qp_count % mw_count && mw_num + 1 < nt->qp_count / mw_count)
728 num_qps_mw = nt->qp_count / mw_count + 1;
730 num_qps_mw = nt->qp_count / mw_count;
732 mw_base = mw->phys_addr;
733 mw_size = mw->phys_size;
735 tx_size = mw_size / num_qps_mw;
736 qp_offset = tx_size * (qp_num / mw_count);
738 qp->tx_mw = mw->vbase + qp_offset;
739 KASSERT(qp->tx_mw != NULL, ("uh oh?"));
741 /* XXX Assumes that a vm_paddr_t is equivalent to bus_addr_t */
742 qp->tx_mw_phys = mw_base + qp_offset;
743 KASSERT(qp->tx_mw_phys != 0, ("uh oh?"));
745 tx_size -= sizeof(struct ntb_rx_info);
746 qp->rx_info = (void *)(qp->tx_mw + tx_size);
748 /* Due to house-keeping, there must be at least 2 buffs */
749 qp->tx_max_frame = qmin(tx_size / 2,
750 transport_mtu + sizeof(struct ntb_payload_header));
751 qp->tx_max_entry = tx_size / qp->tx_max_frame;
753 callout_init(&qp->link_work, 0);
754 callout_init(&qp->queue_full, CALLOUT_MPSAFE);
755 callout_init(&qp->rx_full, CALLOUT_MPSAFE);
757 mtx_init(&qp->ntb_rx_q_lock, "ntb rx q", NULL, MTX_SPIN);
758 mtx_init(&qp->ntb_tx_free_q_lock, "ntb tx free q", NULL, MTX_SPIN);
759 TASK_INIT(&qp->rx_completion_task, 0, ntb_complete_rxc, qp);
760 TASK_INIT(&qp->rxc_db_work, 0, ntb_transport_rxc_db, qp);
762 STAILQ_INIT(&qp->rx_post_q);
763 STAILQ_INIT(&qp->rx_pend_q);
764 STAILQ_INIT(&qp->tx_free_q);
766 callout_reset(&qp->link_work, 0, ntb_qp_link_work, qp);
770 ntb_transport_free_queue(struct ntb_transport_qp *qp)
772 struct ntb_queue_entry *entry;
777 callout_drain(&qp->link_work);
779 ntb_db_set_mask(qp->ntb, 1ull << qp->qp_num);
780 taskqueue_drain(taskqueue_swi, &qp->rxc_db_work);
781 taskqueue_drain(taskqueue_swi, &qp->rx_completion_task);
784 qp->rx_handler = NULL;
785 qp->tx_handler = NULL;
786 qp->event_handler = NULL;
788 while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_pend_q)))
789 free(entry, M_NTB_IF);
791 while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_post_q)))
792 free(entry, M_NTB_IF);
794 while ((entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q)))
795 free(entry, M_NTB_IF);
797 set_bit(qp->qp_num, &qp->transport->qp_bitmap_free);
801 * ntb_transport_create_queue - Create a new NTB transport layer queue
802 * @rx_handler: receive callback function
803 * @tx_handler: transmit callback function
804 * @event_handler: event callback function
806 * Create a new NTB transport layer queue and provide the queue with a callback
807 * routine for both transmit and receive. The receive callback routine will be
808 * used to pass up data when the transport has received it on the queue. The
809 * transmit callback routine will be called when the transport has completed the
810 * transmission of the data on the queue and the data is ready to be freed.
812 * RETURNS: pointer to newly created ntb_queue, NULL on error.
814 static struct ntb_transport_qp *
815 ntb_transport_create_queue(void *data, struct ntb_softc *ntb,
816 const struct ntb_queue_handlers *handlers)
818 struct ntb_queue_entry *entry;
819 struct ntb_transport_qp *qp;
820 struct ntb_transport_ctx *nt;
821 unsigned int free_queue;
824 nt = ntb_get_ctx(ntb, NULL);
825 KASSERT(nt != NULL, ("bogus"));
827 free_queue = ffs_bit(&nt->qp_bitmap);
831 /* decrement free_queue to make it zero based */
834 qp = &nt->qp_vec[free_queue];
835 clear_bit(qp->qp_num, &nt->qp_bitmap_free);
837 qp->rx_handler = handlers->rx_handler;
838 qp->tx_handler = handlers->tx_handler;
839 qp->event_handler = handlers->event_handler;
841 for (i = 0; i < NTB_QP_DEF_NUM_ENTRIES; i++) {
842 entry = malloc(sizeof(*entry), M_NTB_IF, M_WAITOK | M_ZERO);
843 entry->cb_data = nt->ifp;
845 entry->len = transport_mtu;
846 ntb_list_add(&qp->ntb_rx_q_lock, entry, &qp->rx_pend_q);
849 for (i = 0; i < NTB_QP_DEF_NUM_ENTRIES; i++) {
850 entry = malloc(sizeof(*entry), M_NTB_IF, M_WAITOK | M_ZERO);
851 ntb_list_add(&qp->ntb_tx_free_q_lock, entry, &qp->tx_free_q);
854 ntb_db_clear(ntb, 1ull << qp->qp_num);
855 ntb_db_clear_mask(ntb, 1ull << qp->qp_num);
860 * ntb_transport_link_up - Notify NTB transport of client readiness to use queue
861 * @qp: NTB transport layer queue to be enabled
863 * Notify NTB transport layer of client readiness to use queue
866 ntb_transport_link_up(struct ntb_transport_qp *qp)
868 struct ntb_transport_ctx *nt;
873 qp->client_ready = true;
876 ntb_printf(2, "qp client ready\n");
878 if (qp->transport->link_is_up)
879 callout_reset(&qp->link_work, 0, ntb_qp_link_work, qp);
887 * ntb_transport_tx_enqueue - Enqueue a new NTB queue entry
888 * @qp: NTB transport layer queue the entry is to be enqueued on
889 * @cb: per buffer pointer for callback function to use
890 * @data: pointer to data buffer that will be sent
891 * @len: length of the data buffer
893 * Enqueue a new transmit buffer onto the transport queue from which a NTB
894 * payload will be transmitted. This assumes that a lock is being held to
895 * serialize access to the qp.
897 * RETURNS: An appropriate ERRNO error value on error, or zero for success.
900 ntb_transport_tx_enqueue(struct ntb_transport_qp *qp, void *cb, void *data,
903 struct ntb_queue_entry *entry;
906 if (qp == NULL || !qp->link_is_up || len == 0) {
907 CTR0(KTR_NTB, "TX: link not up");
911 entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q);
913 CTR0(KTR_NTB, "TX: could not get entry from tx_free_q");
917 CTR1(KTR_NTB, "TX: got entry %p from tx_free_q", entry);
924 rc = ntb_process_tx(qp, entry);
926 ntb_list_add(&qp->ntb_tx_free_q_lock, entry, &qp->tx_free_q);
928 "TX: process_tx failed. Returning entry %p to tx_free_q",
935 ntb_process_tx(struct ntb_transport_qp *qp, struct ntb_queue_entry *entry)
939 offset = qp->tx_mw + qp->tx_max_frame * qp->tx_index;
941 "TX: process_tx: tx_pkts=%lu, tx_index=%u, remote entry=%u",
942 qp->tx_pkts, qp->tx_index, qp->remote_rx_info->entry);
943 if (qp->tx_index == qp->remote_rx_info->entry) {
944 CTR0(KTR_NTB, "TX: ring full");
949 if (entry->len > qp->tx_max_frame - sizeof(struct ntb_payload_header)) {
950 if (qp->tx_handler != NULL)
951 qp->tx_handler(qp, qp->cb_data, entry->buf,
957 ntb_list_add(&qp->ntb_tx_free_q_lock, entry, &qp->tx_free_q);
959 "TX: frame too big. returning entry %p to tx_free_q",
963 CTR2(KTR_NTB, "TX: copying entry %p to offset %p", entry, offset);
964 ntb_memcpy_tx(qp, entry, offset);
967 qp->tx_index %= qp->tx_max_entry;
975 ntb_memcpy_tx(struct ntb_transport_qp *qp, struct ntb_queue_entry *entry,
978 struct ntb_payload_header *hdr;
980 /* This piece is from Linux' ntb_async_tx() */
981 hdr = (struct ntb_payload_header *)((char *)offset + qp->tx_max_frame -
982 sizeof(struct ntb_payload_header));
984 iowrite32(entry->len, &hdr->len);
985 iowrite32(qp->tx_pkts, &hdr->ver);
987 /* This piece is ntb_memcpy_tx() */
988 CTR2(KTR_NTB, "TX: copying %d bytes to offset %p", entry->len, offset);
989 if (entry->buf != NULL) {
990 m_copydata((struct mbuf *)entry->buf, 0, entry->len, offset);
993 * Ensure that the data is fully copied before setting the
999 /* The rest is ntb_tx_copy_callback() */
1000 iowrite32(entry->flags | IF_NTB_DESC_DONE_FLAG, &hdr->flags);
1001 CTR1(KTR_NTB, "TX: hdr %p set DESC_DONE", hdr);
1003 ntb_peer_db_set(qp->ntb, 1ull << qp->qp_num);
1006 * The entry length can only be zero if the packet is intended to be a
1007 * "link down" or similar. Since no payload is being sent in these
1008 * cases, there is nothing to add to the completion queue.
1010 if (entry->len > 0) {
1011 qp->tx_bytes += entry->len;
1014 qp->tx_handler(qp, qp->cb_data, entry->buf,
1017 m_freem(entry->buf);
1022 "TX: entry %p sent. hdr->ver = %u, hdr->flags = 0x%x, Returning "
1023 "to tx_free_q", entry, hdr->ver, hdr->flags);
1024 ntb_list_add(&qp->ntb_tx_free_q_lock, entry, &qp->tx_free_q);
1028 ntb_qp_full(void *arg)
1031 CTR0(KTR_NTB, "TX: qp_full callout");
1037 ntb_transport_rxc_db(void *arg, int pending __unused)
1039 struct ntb_transport_qp *qp = arg;
1044 * Limit the number of packets processed in a single interrupt to
1045 * provide fairness to others
1047 CTR0(KTR_NTB, "RX: transport_rx");
1048 mtx_lock(&qp->transport->rx_lock);
1049 for (i = 0; i < qp->rx_max_entry; i++) {
1050 rc = ntb_process_rxc(qp);
1052 CTR0(KTR_NTB, "RX: process_rxc failed");
1056 mtx_unlock(&qp->transport->rx_lock);
1058 if (i == qp->rx_max_entry)
1059 taskqueue_enqueue(taskqueue_swi, &qp->rxc_db_work);
1060 else if ((ntb_db_read(qp->ntb) & (1ull << qp->qp_num)) != 0) {
1061 /* If db is set, clear it and read it back to commit clear. */
1062 ntb_db_clear(qp->ntb, 1ull << qp->qp_num);
1063 (void)ntb_db_read(qp->ntb);
1066 * An interrupt may have arrived between finishing
1067 * ntb_process_rxc and clearing the doorbell bit: there might
1068 * be some more work to do.
1070 taskqueue_enqueue(taskqueue_swi, &qp->rxc_db_work);
1075 ntb_process_rxc(struct ntb_transport_qp *qp)
1077 struct ntb_payload_header *hdr;
1078 struct ntb_queue_entry *entry;
1081 offset = qp->rx_buff + qp->rx_max_frame * qp->rx_index;
1082 hdr = (void *)(offset + qp->rx_max_frame -
1083 sizeof(struct ntb_payload_header));
1085 CTR1(KTR_NTB, "RX: process_rxc rx_index = %u", qp->rx_index);
1086 if ((hdr->flags & IF_NTB_DESC_DONE_FLAG) == 0) {
1087 CTR0(KTR_NTB, "RX: hdr not done");
1088 qp->rx_ring_empty++;
1092 if ((hdr->flags & IF_NTB_LINK_DOWN_FLAG) != 0) {
1093 CTR0(KTR_NTB, "RX: link down");
1094 ntb_qp_link_down(qp);
1099 if (hdr->ver != (uint32_t)qp->rx_pkts) {
1100 CTR2(KTR_NTB,"RX: ver != rx_pkts (%x != %lx). "
1101 "Returning entry to rx_pend_q", hdr->ver, qp->rx_pkts);
1106 entry = ntb_list_mv(&qp->ntb_rx_q_lock, &qp->rx_pend_q, &qp->rx_post_q);
1107 if (entry == NULL) {
1108 qp->rx_err_no_buf++;
1109 CTR0(KTR_NTB, "RX: No entries in rx_pend_q");
1112 callout_stop(&qp->rx_full);
1113 CTR1(KTR_NTB, "RX: rx entry %p from rx_pend_q", entry);
1116 entry->index = qp->rx_index;
1118 if (hdr->len > entry->len) {
1119 CTR2(KTR_NTB, "RX: len too long. Wanted %ju got %ju",
1120 (uintmax_t)hdr->len, (uintmax_t)entry->len);
1124 entry->flags |= IF_NTB_DESC_DONE_FLAG;
1126 taskqueue_enqueue(taskqueue_swi, &qp->rx_completion_task);
1128 qp->rx_bytes += hdr->len;
1131 CTR1(KTR_NTB, "RX: received %ld rx_pkts", qp->rx_pkts);
1133 entry->len = hdr->len;
1135 ntb_memcpy_rx(qp, entry, offset);
1139 qp->rx_index %= qp->rx_max_entry;
1144 ntb_memcpy_rx(struct ntb_transport_qp *qp, struct ntb_queue_entry *entry,
1147 struct ifnet *ifp = entry->cb_data;
1148 unsigned int len = entry->len;
1151 CTR2(KTR_NTB, "RX: copying %d bytes from offset %p", len, offset);
1152 m = m_devget(offset, len, 0, ifp, NULL);
1153 m->m_pkthdr.csum_flags = CSUM_IP_CHECKED | CSUM_IP_VALID;
1155 entry->buf = (void *)m;
1157 /* Ensure that the data is globally visible before clearing the flag */
1160 CTR2(KTR_NTB, "RX: copied entry %p to mbuf %p.", entry, m);
1161 ntb_rx_copy_callback(qp, entry);
1165 ntb_rx_copy_callback(struct ntb_transport_qp *qp, void *data)
1167 struct ntb_queue_entry *entry;
1170 entry->flags |= IF_NTB_DESC_DONE_FLAG;
1171 taskqueue_enqueue(taskqueue_swi, &qp->rx_completion_task);
1175 ntb_complete_rxc(void *arg, int pending)
1177 struct ntb_transport_qp *qp = arg;
1178 struct ntb_queue_entry *entry;
1182 CTR0(KTR_NTB, "RX: rx_completion_task");
1184 mtx_lock_spin(&qp->ntb_rx_q_lock);
1186 while (!STAILQ_EMPTY(&qp->rx_post_q)) {
1187 entry = STAILQ_FIRST(&qp->rx_post_q);
1188 if ((entry->flags & IF_NTB_DESC_DONE_FLAG) == 0)
1191 entry->x_hdr->flags = 0;
1192 iowrite32(entry->index, &qp->rx_info->entry);
1194 STAILQ_REMOVE_HEAD(&qp->rx_post_q, entry);
1200 * Re-initialize queue_entry for reuse; rx_handler takes
1201 * ownership of the mbuf.
1204 entry->len = transport_mtu;
1205 entry->cb_data = qp->transport->ifp;
1207 STAILQ_INSERT_TAIL(&qp->rx_pend_q, entry, entry);
1209 mtx_unlock_spin(&qp->ntb_rx_q_lock);
1211 CTR2(KTR_NTB, "RX: completing entry %p, mbuf %p", entry, m);
1212 if (qp->rx_handler != NULL && qp->client_ready)
1213 qp->rx_handler(qp, qp->cb_data, m, len);
1217 mtx_lock_spin(&qp->ntb_rx_q_lock);
1220 mtx_unlock_spin(&qp->ntb_rx_q_lock);
1224 ntb_transport_doorbell_callback(void *data, uint32_t vector)
1226 struct ntb_transport_ctx *nt = data;
1227 struct ntb_transport_qp *qp;
1228 struct _qpset db_bits;
1232 BIT_COPY(QP_SETSIZE, &nt->qp_bitmap, &db_bits);
1233 BIT_NAND(QP_SETSIZE, &db_bits, &nt->qp_bitmap_free);
1235 vec_mask = ntb_db_vector_mask(nt->ntb, vector);
1236 while (vec_mask != 0) {
1237 qp_num = ffsll(vec_mask) - 1;
1239 if (test_bit(qp_num, &db_bits)) {
1240 qp = &nt->qp_vec[qp_num];
1241 taskqueue_enqueue(taskqueue_swi, &qp->rxc_db_work);
1244 vec_mask &= ~(1ull << qp_num);
1248 /* Link Event handler */
1250 ntb_transport_event_callback(void *data)
1252 struct ntb_transport_ctx *nt = data;
1254 if (ntb_link_is_up(nt->ntb, NULL, NULL)) {
1255 ntb_printf(1, "HW link up\n");
1256 callout_reset(&nt->link_work, 0, ntb_transport_link_work, nt);
1258 ntb_printf(1, "HW link down\n");
1259 taskqueue_enqueue(taskqueue_swi, &nt->link_cleanup);
1265 ntb_transport_link_work(void *arg)
1267 struct ntb_transport_ctx *nt = arg;
1268 struct ntb_softc *ntb = nt->ntb;
1269 struct ntb_transport_qp *qp;
1270 uint64_t val64, size;
1275 /* send the local info, in the opposite order of the way we read it */
1276 for (i = 0; i < nt->mw_count; i++) {
1277 size = nt->mw_vec[i].phys_size;
1279 if (max_mw_size != 0 && size > max_mw_size)
1282 ntb_peer_spad_write(ntb, IF_NTB_MW0_SZ_HIGH + (i * 2),
1284 ntb_peer_spad_write(ntb, IF_NTB_MW0_SZ_LOW + (i * 2), size);
1287 ntb_peer_spad_write(ntb, IF_NTB_NUM_MWS, nt->mw_count);
1289 ntb_peer_spad_write(ntb, IF_NTB_NUM_QPS, nt->qp_count);
1291 ntb_peer_spad_write(ntb, IF_NTB_VERSION, NTB_TRANSPORT_VERSION);
1293 /* Query the remote side for its info */
1295 ntb_spad_read(ntb, IF_NTB_VERSION, &val);
1296 if (val != NTB_TRANSPORT_VERSION)
1299 ntb_spad_read(ntb, IF_NTB_NUM_QPS, &val);
1300 if (val != nt->qp_count)
1303 ntb_spad_read(ntb, IF_NTB_NUM_MWS, &val);
1304 if (val != nt->mw_count)
1307 for (i = 0; i < nt->mw_count; i++) {
1308 ntb_spad_read(ntb, IF_NTB_MW0_SZ_HIGH + (i * 2), &val);
1309 val64 = (uint64_t)val << 32;
1311 ntb_spad_read(ntb, IF_NTB_MW0_SZ_LOW + (i * 2), &val);
1314 rc = ntb_set_mw(nt, i, val64);
1319 nt->link_is_up = true;
1320 ntb_printf(1, "transport link up\n");
1322 for (i = 0; i < nt->qp_count; i++) {
1323 qp = &nt->qp_vec[i];
1325 ntb_transport_setup_qp_mw(nt, i);
1327 if (qp->client_ready)
1328 callout_reset(&qp->link_work, 0, ntb_qp_link_work, qp);
1334 for (i = 0; i < nt->mw_count; i++)
1337 if (ntb_link_is_up(ntb, NULL, NULL))
1338 callout_reset(&nt->link_work,
1339 NTB_LINK_DOWN_TIMEOUT * hz / 1000, ntb_transport_link_work, nt);
1343 ntb_set_mw(struct ntb_transport_ctx *nt, int num_mw, size_t size)
1345 struct ntb_transport_mw *mw = &nt->mw_vec[num_mw];
1346 size_t xlat_size, buff_size;
1352 xlat_size = roundup(size, mw->xlat_align_size);
1353 buff_size = xlat_size;
1355 /* No need to re-setup */
1356 if (mw->xlat_size == xlat_size)
1359 if (mw->buff_size != 0)
1360 ntb_free_mw(nt, num_mw);
1362 /* Alloc memory for receiving data. Must be aligned */
1363 mw->xlat_size = xlat_size;
1364 mw->buff_size = buff_size;
1366 mw->virt_addr = contigmalloc(mw->buff_size, M_NTB_IF, M_ZERO, 0,
1367 mw->addr_limit, mw->xlat_align, 0);
1368 if (mw->virt_addr == NULL) {
1369 ntb_printf(0, "Unable to allocate MW buffer of size %zu/%zu\n",
1370 mw->buff_size, mw->xlat_size);
1375 /* TODO: replace with bus_space_* functions */
1376 mw->dma_addr = vtophys(mw->virt_addr);
1379 * Ensure that the allocation from contigmalloc is aligned as
1380 * requested. XXX: This may not be needed -- brought in for parity
1381 * with the Linux driver.
1383 if (mw->dma_addr % mw->xlat_align != 0) {
1385 "DMA memory 0x%jx not aligned to BAR size 0x%zx\n",
1386 (uintmax_t)mw->dma_addr, size);
1387 ntb_free_mw(nt, num_mw);
1391 /* Notify HW the memory location of the receive buffer */
1392 rc = ntb_mw_set_trans(nt->ntb, num_mw, mw->dma_addr, mw->xlat_size);
1394 ntb_printf(0, "Unable to set mw%d translation\n", num_mw);
1395 ntb_free_mw(nt, num_mw);
1403 ntb_free_mw(struct ntb_transport_ctx *nt, int num_mw)
1405 struct ntb_transport_mw *mw = &nt->mw_vec[num_mw];
1407 if (mw->virt_addr == NULL)
1410 ntb_mw_clear_trans(nt->ntb, num_mw);
1411 contigfree(mw->virt_addr, mw->xlat_size, M_NTB_IF);
1414 mw->virt_addr = NULL;
1418 ntb_transport_setup_qp_mw(struct ntb_transport_ctx *nt, unsigned int qp_num)
1420 struct ntb_transport_qp *qp = &nt->qp_vec[qp_num];
1421 struct ntb_transport_mw *mw;
1425 unsigned num_qps_mw, mw_num, mw_count;
1427 mw_count = nt->mw_count;
1428 mw_num = QP_TO_MW(nt, qp_num);
1429 mw = &nt->mw_vec[mw_num];
1431 if (mw->virt_addr == NULL)
1434 if (nt->qp_count % mw_count && mw_num + 1 < nt->qp_count / mw_count)
1435 num_qps_mw = nt->qp_count / mw_count + 1;
1437 num_qps_mw = nt->qp_count / mw_count;
1439 rx_size = mw->xlat_size / num_qps_mw;
1440 qp->rx_buff = mw->virt_addr + rx_size * (qp_num / mw_count);
1441 rx_size -= sizeof(struct ntb_rx_info);
1443 qp->remote_rx_info = (void*)(qp->rx_buff + rx_size);
1445 /* Due to house-keeping, there must be at least 2 buffs */
1446 qp->rx_max_frame = qmin(rx_size / 2,
1447 transport_mtu + sizeof(struct ntb_payload_header));
1448 qp->rx_max_entry = rx_size / qp->rx_max_frame;
1451 qp->remote_rx_info->entry = qp->rx_max_entry - 1;
1453 /* Set up the hdr offsets with 0s */
1454 for (i = 0; i < qp->rx_max_entry; i++) {
1455 offset = (void *)(qp->rx_buff + qp->rx_max_frame * (i + 1) -
1456 sizeof(struct ntb_payload_header));
1457 memset(offset, 0, sizeof(struct ntb_payload_header));
1468 ntb_qp_link_work(void *arg)
1470 struct ntb_transport_qp *qp = arg;
1471 struct ntb_softc *ntb = qp->ntb;
1472 struct ntb_transport_ctx *nt = qp->transport;
1473 uint32_t val, dummy;
1475 ntb_spad_read(ntb, IF_NTB_QP_LINKS, &val);
1477 ntb_peer_spad_write(ntb, IF_NTB_QP_LINKS, val | (1ull << qp->qp_num));
1479 /* query remote spad for qp ready bits */
1480 ntb_peer_spad_read(ntb, IF_NTB_QP_LINKS, &dummy);
1482 /* See if the remote side is up */
1483 if ((val & (1ull << qp->qp_num)) != 0) {
1484 ntb_printf(2, "qp link up\n");
1485 qp->link_is_up = true;
1487 if (qp->event_handler != NULL)
1488 qp->event_handler(qp->cb_data, NTB_LINK_UP);
1490 taskqueue_enqueue(taskqueue_swi, &qp->rxc_db_work);
1491 } else if (nt->link_is_up)
1492 callout_reset(&qp->link_work,
1493 NTB_LINK_DOWN_TIMEOUT * hz / 1000, ntb_qp_link_work, qp);
1496 /* Link down event*/
1498 ntb_transport_link_cleanup(struct ntb_transport_ctx *nt)
1500 struct ntb_transport_qp *qp;
1501 struct _qpset qp_bitmap_alloc;
1504 BIT_COPY(QP_SETSIZE, &nt->qp_bitmap, &qp_bitmap_alloc);
1505 BIT_NAND(QP_SETSIZE, &qp_bitmap_alloc, &nt->qp_bitmap_free);
1507 /* Pass along the info to any clients */
1508 for (i = 0; i < nt->qp_count; i++)
1509 if (test_bit(i, &qp_bitmap_alloc)) {
1510 qp = &nt->qp_vec[i];
1511 ntb_qp_link_cleanup(qp);
1512 callout_drain(&qp->link_work);
1515 if (!nt->link_is_up)
1516 callout_drain(&nt->link_work);
1519 * The scratchpad registers keep the values if the remote side
1520 * goes down, blast them now to give them a sane value the next
1521 * time they are accessed
1523 for (i = 0; i < IF_NTB_MAX_SPAD; i++)
1524 ntb_spad_write(nt->ntb, i, 0);
1528 ntb_transport_link_cleanup_work(void *arg, int pending __unused)
1531 ntb_transport_link_cleanup(arg);
1535 ntb_qp_link_down(struct ntb_transport_qp *qp)
1538 ntb_qp_link_cleanup(qp);
1542 ntb_qp_link_down_reset(struct ntb_transport_qp *qp)
1545 qp->link_is_up = false;
1547 qp->tx_index = qp->rx_index = 0;
1548 qp->tx_bytes = qp->rx_bytes = 0;
1549 qp->tx_pkts = qp->rx_pkts = 0;
1551 qp->rx_ring_empty = 0;
1552 qp->tx_ring_full = 0;
1554 qp->rx_err_no_buf = qp->tx_err_no_buf = 0;
1555 qp->rx_err_oflow = qp->rx_err_ver = 0;
1559 ntb_qp_link_cleanup(struct ntb_transport_qp *qp)
1561 struct ntb_transport_ctx *nt = qp->transport;
1563 callout_drain(&qp->link_work);
1564 ntb_qp_link_down_reset(qp);
1566 if (qp->event_handler != NULL)
1567 qp->event_handler(qp->cb_data, NTB_LINK_DOWN);
1570 callout_reset(&qp->link_work,
1571 NTB_LINK_DOWN_TIMEOUT * hz / 1000, ntb_qp_link_work, qp);
1574 /* Link commanded down */
1576 * ntb_transport_link_down - Notify NTB transport to no longer enqueue data
1577 * @qp: NTB transport layer queue to be disabled
1579 * Notify NTB transport layer of client's desire to no longer receive data on
1580 * transport queue specified. It is the client's responsibility to ensure all
1581 * entries on queue are purged or otherwise handled appropriately.
1584 ntb_transport_link_down(struct ntb_transport_qp *qp)
1591 qp->client_ready = false;
1593 ntb_spad_read(qp->ntb, IF_NTB_QP_LINKS, &val);
1595 ntb_peer_spad_write(qp->ntb, IF_NTB_QP_LINKS,
1596 val & ~(1 << qp->qp_num));
1599 ntb_send_link_down(qp);
1601 callout_drain(&qp->link_work);
1605 ntb_send_link_down(struct ntb_transport_qp *qp)
1607 struct ntb_queue_entry *entry;
1610 if (!qp->link_is_up)
1613 for (i = 0; i < NTB_LINK_DOWN_TIMEOUT; i++) {
1614 entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q);
1617 pause("NTB Wait for link down", hz / 10);
1623 entry->cb_data = NULL;
1626 entry->flags = IF_NTB_LINK_DOWN_FLAG;
1628 mtx_lock(&qp->transport->tx_lock);
1629 rc = ntb_process_tx(qp, entry);
1631 printf("ntb: Failed to send link down\n");
1632 mtx_unlock(&qp->transport->tx_lock);
1634 ntb_qp_link_down_reset(qp);
1638 /* List Management */
1641 ntb_list_add(struct mtx *lock, struct ntb_queue_entry *entry,
1642 struct ntb_queue_list *list)
1645 mtx_lock_spin(lock);
1646 STAILQ_INSERT_TAIL(list, entry, entry);
1647 mtx_unlock_spin(lock);
1650 static struct ntb_queue_entry *
1651 ntb_list_rm(struct mtx *lock, struct ntb_queue_list *list)
1653 struct ntb_queue_entry *entry;
1655 mtx_lock_spin(lock);
1656 if (STAILQ_EMPTY(list)) {
1660 entry = STAILQ_FIRST(list);
1661 STAILQ_REMOVE_HEAD(list, entry);
1663 mtx_unlock_spin(lock);
1668 static struct ntb_queue_entry *
1669 ntb_list_mv(struct mtx *lock, struct ntb_queue_list *from,
1670 struct ntb_queue_list *to)
1672 struct ntb_queue_entry *entry;
1674 mtx_lock_spin(lock);
1675 if (STAILQ_EMPTY(from)) {
1679 entry = STAILQ_FIRST(from);
1680 STAILQ_REMOVE_HEAD(from, entry);
1681 STAILQ_INSERT_TAIL(to, entry, entry);
1684 mtx_unlock_spin(lock);
1688 /* Helper functions */
1689 /* TODO: This too should really be part of the kernel */
1690 #define EUI48_MULTICAST 1 << 0
1691 #define EUI48_LOCALLY_ADMINISTERED 1 << 1
1693 create_random_local_eui48(u_char *eaddr)
1695 static uint8_t counter = 0;
1696 uint32_t seed = ticks;
1698 eaddr[0] = EUI48_LOCALLY_ADMINISTERED;
1699 memcpy(&eaddr[1], &seed, sizeof(uint32_t));
1700 eaddr[5] = counter++;
1704 * ntb_transport_max_size - Query the max payload size of a qp
1705 * @qp: NTB transport layer queue to be queried
1707 * Query the maximum payload size permissible on the given qp
1709 * RETURNS: the max payload size of a qp
1712 ntb_transport_max_size(struct ntb_transport_qp *qp)
1718 return (qp->tx_max_frame - sizeof(struct ntb_payload_header));