xref: /freebsd/sys/dev/ntb/if_ntb/if_ntb.c (revision ee5cf11617a9b7f034d95c639bd4d27d1f09e848)
1 /*-
2  * Copyright (C) 2013 Intel Corporation
3  * Copyright (C) 2015 EMC Corporation
4  * All rights reserved.
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
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.
14  *
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
25  * SUCH DAMAGE.
26  */
27 
28 #include <sys/cdefs.h>
29 __FBSDID("$FreeBSD$");
30 
31 #include <sys/param.h>
32 #include <sys/kernel.h>
33 #include <sys/systm.h>
34 #include <sys/bitset.h>
35 #include <sys/bus.h>
36 #include <sys/ktr.h>
37 #include <sys/limits.h>
38 #include <sys/lock.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>
47 
48 #include <net/if.h>
49 #include <net/if_media.h>
50 #include <net/if_types.h>
51 #include <net/if_var.h>
52 #include <net/bpf.h>
53 #include <net/ethernet.h>
54 
55 #include <vm/vm.h>
56 #include <vm/pmap.h>
57 
58 #include <machine/bus.h>
59 #include <machine/cpufunc.h>
60 
61 #include <netinet/in.h>
62 #include <netinet/ip.h>
63 
64 #include "../ntb_hw/ntb_hw.h"
65 
66 /*
67  * The Non-Transparent Bridge (NTB) is a device on some Intel processors that
68  * allows you to connect two systems using a PCI-e link.
69  *
70  * This module contains a protocol for sending and receiving messages, and
71  * exposes that protocol through a simulated ethernet device called ntb.
72  *
73  * NOTE: Much of the code in this module is shared with Linux. Any patches may
74  * be picked up and redistributed in Linux with a dual GPL/BSD license.
75  */
76 
77 #define QP_SETSIZE	64
78 BITSET_DEFINE(_qpset, QP_SETSIZE);
79 #define test_bit(pos, addr)	BIT_ISSET(QP_SETSIZE, (pos), (addr))
80 #define set_bit(pos, addr)	BIT_SET(QP_SETSIZE, (pos), (addr))
81 #define clear_bit(pos, addr)	BIT_CLR(QP_SETSIZE, (pos), (addr))
82 #define ffs_bit(addr)		BIT_FFS(QP_SETSIZE, (addr))
83 
84 #define KTR_NTB KTR_SPARE3
85 
86 #define NTB_TRANSPORT_VERSION	4
87 #define NTB_RX_MAX_PKTS		64
88 #define	NTB_RXQ_SIZE		300
89 
90 enum ntb_link_event {
91 	NTB_LINK_DOWN = 0,
92 	NTB_LINK_UP,
93 };
94 
95 static SYSCTL_NODE(_hw, OID_AUTO, if_ntb, CTLFLAG_RW, 0, "if_ntb");
96 
97 static unsigned g_if_ntb_debug_level;
98 SYSCTL_UINT(_hw_if_ntb, OID_AUTO, debug_level, CTLFLAG_RWTUN,
99     &g_if_ntb_debug_level, 0, "if_ntb log level -- higher is more verbose");
100 #define ntb_printf(lvl, ...) do {			\
101 	if ((lvl) <= g_if_ntb_debug_level) {		\
102 		if_printf(nt->ifp, __VA_ARGS__);	\
103 	}						\
104 } while (0)
105 
106 static unsigned transport_mtu = IP_MAXPACKET + ETHER_HDR_LEN + ETHER_CRC_LEN;
107 
108 static uint64_t max_mw_size;
109 SYSCTL_UQUAD(_hw_if_ntb, OID_AUTO, max_mw_size, CTLFLAG_RDTUN, &max_mw_size, 0,
110     "If enabled (non-zero), limit the size of large memory windows. "
111     "Both sides of the NTB MUST set the same value here.");
112 
113 static unsigned max_num_clients;
114 SYSCTL_UINT(_hw_if_ntb, OID_AUTO, max_num_clients, CTLFLAG_RDTUN,
115     &max_num_clients, 0, "Maximum number of NTB transport clients.  "
116     "0 (default) - use all available NTB memory windows; "
117     "positive integer N - Limit to N memory windows.");
118 
119 static unsigned enable_xeon_watchdog;
120 SYSCTL_UINT(_hw_if_ntb, OID_AUTO, enable_xeon_watchdog, CTLFLAG_RDTUN,
121     &enable_xeon_watchdog, 0, "If non-zero, write a register every second to "
122     "keep a watchdog from tearing down the NTB link");
123 
124 STAILQ_HEAD(ntb_queue_list, ntb_queue_entry);
125 
126 typedef uint32_t ntb_q_idx_t;
127 
128 struct ntb_queue_entry {
129 	/* ntb_queue list reference */
130 	STAILQ_ENTRY(ntb_queue_entry) entry;
131 
132 	/* info on data to be transferred */
133 	void		*cb_data;
134 	void		*buf;
135 	uint32_t	len;
136 	uint32_t	flags;
137 
138 	struct ntb_transport_qp		*qp;
139 	struct ntb_payload_header	*x_hdr;
140 	ntb_q_idx_t	index;
141 };
142 
143 struct ntb_rx_info {
144 	ntb_q_idx_t	entry;
145 };
146 
147 struct ntb_transport_qp {
148 	struct ntb_transport_ctx	*transport;
149 	struct ntb_softc	*ntb;
150 
151 	void			*cb_data;
152 
153 	bool			client_ready;
154 	volatile bool		link_is_up;
155 	uint8_t			qp_num;	/* Only 64 QPs are allowed.  0-63 */
156 
157 	struct ntb_rx_info	*rx_info;
158 	struct ntb_rx_info	*remote_rx_info;
159 
160 	void (*tx_handler)(struct ntb_transport_qp *qp, void *qp_data,
161 	    void *data, int len);
162 	struct ntb_queue_list	tx_free_q;
163 	struct mtx		ntb_tx_free_q_lock;
164 	caddr_t			tx_mw;
165 	bus_addr_t		tx_mw_phys;
166 	ntb_q_idx_t		tx_index;
167 	ntb_q_idx_t		tx_max_entry;
168 	uint64_t		tx_max_frame;
169 
170 	void (*rx_handler)(struct ntb_transport_qp *qp, void *qp_data,
171 	    void *data, int len);
172 	struct ntb_queue_list	rx_post_q;
173 	struct ntb_queue_list	rx_pend_q;
174 	/* ntb_rx_q_lock: synchronize access to rx_XXXX_q */
175 	struct mtx		ntb_rx_q_lock;
176 	struct task		rx_completion_task;
177 	struct task		rxc_db_work;
178 	caddr_t			rx_buff;
179 	ntb_q_idx_t		rx_index;
180 	ntb_q_idx_t		rx_max_entry;
181 	uint64_t		rx_max_frame;
182 
183 	void (*event_handler)(void *data, enum ntb_link_event status);
184 	struct callout		link_work;
185 	struct callout		queue_full;
186 	struct callout		rx_full;
187 
188 	uint64_t		last_rx_no_buf;
189 
190 	/* Stats */
191 	uint64_t		rx_bytes;
192 	uint64_t		rx_pkts;
193 	uint64_t		rx_ring_empty;
194 	uint64_t		rx_err_no_buf;
195 	uint64_t		rx_err_oflow;
196 	uint64_t		rx_err_ver;
197 	uint64_t		tx_bytes;
198 	uint64_t		tx_pkts;
199 	uint64_t		tx_ring_full;
200 	uint64_t		tx_err_no_buf;
201 };
202 
203 struct ntb_queue_handlers {
204 	void (*rx_handler)(struct ntb_transport_qp *qp, void *qp_data,
205 	    void *data, int len);
206 	void (*tx_handler)(struct ntb_transport_qp *qp, void *qp_data,
207 	    void *data, int len);
208 	void (*event_handler)(void *data, enum ntb_link_event status);
209 };
210 
211 struct ntb_transport_mw {
212 	vm_paddr_t	phys_addr;
213 	size_t		phys_size;
214 	size_t		xlat_align;
215 	size_t		xlat_align_size;
216 	bus_addr_t	addr_limit;
217 	/* Tx buff is off vbase / phys_addr */
218 	caddr_t		vbase;
219 	size_t		xlat_size;
220 	size_t		buff_size;
221 	/* Rx buff is off virt_addr / dma_addr */
222 	caddr_t		virt_addr;
223 	bus_addr_t	dma_addr;
224 };
225 
226 struct ntb_transport_ctx {
227 	struct ntb_softc	*ntb;
228 	struct ifnet		*ifp;
229 	struct ntb_transport_mw	mw_vec[NTB_MAX_NUM_MW];
230 	struct ntb_transport_qp	*qp_vec;
231 	struct _qpset		qp_bitmap;
232 	struct _qpset		qp_bitmap_free;
233 	unsigned		mw_count;
234 	unsigned		qp_count;
235 	volatile bool		link_is_up;
236 	struct callout		link_work;
237 	struct callout		link_watchdog;
238 	struct task		link_cleanup;
239 	uint64_t		bufsize;
240 	u_char			eaddr[ETHER_ADDR_LEN];
241 	struct mtx		tx_lock;
242 	struct mtx		rx_lock;
243 
244 	/* The hardcoded single queuepair in ntb_setup_interface() */
245 	struct ntb_transport_qp *qp;
246 };
247 
248 static struct ntb_transport_ctx net_softc;
249 
250 enum {
251 	IF_NTB_DESC_DONE_FLAG = 1 << 0,
252 	IF_NTB_LINK_DOWN_FLAG = 1 << 1,
253 };
254 
255 struct ntb_payload_header {
256 	ntb_q_idx_t ver;
257 	uint32_t len;
258 	uint32_t flags;
259 };
260 
261 enum {
262 	/*
263 	 * The order of this enum is part of the if_ntb remote protocol.  Do
264 	 * not reorder without bumping protocol version (and it's probably best
265 	 * to keep the protocol in lock-step with the Linux NTB driver.
266 	 */
267 	IF_NTB_VERSION = 0,
268 	IF_NTB_QP_LINKS,
269 	IF_NTB_NUM_QPS,
270 	IF_NTB_NUM_MWS,
271 	/*
272 	 * N.B.: transport_link_work assumes MW1 enums = MW0 + 2.
273 	 */
274 	IF_NTB_MW0_SZ_HIGH,
275 	IF_NTB_MW0_SZ_LOW,
276 	IF_NTB_MW1_SZ_HIGH,
277 	IF_NTB_MW1_SZ_LOW,
278 	IF_NTB_MAX_SPAD,
279 
280 	/*
281 	 * Some NTB-using hardware have a watchdog to work around NTB hangs; if
282 	 * a register or doorbell isn't written every few seconds, the link is
283 	 * torn down.  Write an otherwise unused register every few seconds to
284 	 * work around this watchdog.
285 	 */
286 	IF_NTB_WATCHDOG_SPAD = 15
287 };
288 CTASSERT(IF_NTB_WATCHDOG_SPAD < XEON_SPAD_COUNT &&
289     IF_NTB_WATCHDOG_SPAD < ATOM_SPAD_COUNT);
290 
291 #define QP_TO_MW(nt, qp)	((qp) % nt->mw_count)
292 #define NTB_QP_DEF_NUM_ENTRIES	100
293 #define NTB_LINK_DOWN_TIMEOUT	10
294 
295 static int ntb_handle_module_events(struct module *m, int what, void *arg);
296 static int ntb_setup_interface(void);
297 static int ntb_teardown_interface(void);
298 static void ntb_net_init(void *arg);
299 static int ntb_ioctl(struct ifnet *ifp, u_long command, caddr_t data);
300 static void ntb_start(struct ifnet *ifp);
301 static void ntb_net_tx_handler(struct ntb_transport_qp *qp, void *qp_data,
302     void *data, int len);
303 static void ntb_net_rx_handler(struct ntb_transport_qp *qp, void *qp_data,
304     void *data, int len);
305 static void ntb_net_event_handler(void *data, enum ntb_link_event status);
306 static int ntb_transport_probe(struct ntb_softc *ntb);
307 static void ntb_transport_free(struct ntb_transport_ctx *);
308 static void ntb_transport_init_queue(struct ntb_transport_ctx *nt,
309     unsigned int qp_num);
310 static void ntb_transport_free_queue(struct ntb_transport_qp *qp);
311 static struct ntb_transport_qp *ntb_transport_create_queue(void *data,
312     struct ntb_softc *pdev, const struct ntb_queue_handlers *handlers);
313 static void ntb_transport_link_up(struct ntb_transport_qp *qp);
314 static int ntb_transport_tx_enqueue(struct ntb_transport_qp *qp, void *cb,
315     void *data, unsigned int len);
316 static int ntb_process_tx(struct ntb_transport_qp *qp,
317     struct ntb_queue_entry *entry);
318 static void ntb_memcpy_tx(struct ntb_transport_qp *qp,
319     struct ntb_queue_entry *entry, void *offset);
320 static void ntb_qp_full(void *arg);
321 static void ntb_transport_rxc_db(void *arg, int pending);
322 static int ntb_process_rxc(struct ntb_transport_qp *qp);
323 static void ntb_memcpy_rx(struct ntb_transport_qp *qp,
324     struct ntb_queue_entry *entry, void *offset);
325 static inline void ntb_rx_copy_callback(struct ntb_transport_qp *qp,
326     void *data);
327 static void ntb_complete_rxc(void *arg, int pending);
328 static void ntb_transport_doorbell_callback(void *data, uint32_t vector);
329 static void ntb_transport_event_callback(void *data);
330 static void ntb_transport_link_work(void *arg);
331 static int ntb_set_mw(struct ntb_transport_ctx *, int num_mw, size_t size);
332 static void ntb_free_mw(struct ntb_transport_ctx *nt, int num_mw);
333 static int ntb_transport_setup_qp_mw(struct ntb_transport_ctx *nt,
334     unsigned int qp_num);
335 static void ntb_qp_link_work(void *arg);
336 static void ntb_transport_link_cleanup(struct ntb_transport_ctx *nt);
337 static void ntb_transport_link_cleanup_work(void *, int);
338 static void ntb_qp_link_down(struct ntb_transport_qp *qp);
339 static void ntb_qp_link_down_reset(struct ntb_transport_qp *qp);
340 static void ntb_qp_link_cleanup(struct ntb_transport_qp *qp);
341 static void ntb_transport_link_down(struct ntb_transport_qp *qp);
342 static void ntb_send_link_down(struct ntb_transport_qp *qp);
343 static void ntb_list_add(struct mtx *lock, struct ntb_queue_entry *entry,
344     struct ntb_queue_list *list);
345 static struct ntb_queue_entry *ntb_list_rm(struct mtx *lock,
346     struct ntb_queue_list *list);
347 static struct ntb_queue_entry *ntb_list_mv(struct mtx *lock,
348     struct ntb_queue_list *from, struct ntb_queue_list *to);
349 static void create_random_local_eui48(u_char *eaddr);
350 static unsigned int ntb_transport_max_size(struct ntb_transport_qp *qp);
351 static void xeon_link_watchdog_hb(void *);
352 
353 static const struct ntb_ctx_ops ntb_transport_ops = {
354 	.link_event = ntb_transport_event_callback,
355 	.db_event = ntb_transport_doorbell_callback,
356 };
357 
358 MALLOC_DEFINE(M_NTB_IF, "if_ntb", "ntb network driver");
359 
360 static inline void
361 iowrite32(uint32_t val, void *addr)
362 {
363 
364 	bus_space_write_4(X86_BUS_SPACE_MEM, 0/* HACK */, (uintptr_t)addr,
365 	    val);
366 }
367 
368 /* Module setup and teardown */
369 static int
370 ntb_handle_module_events(struct module *m, int what, void *arg)
371 {
372 	int err = 0;
373 
374 	switch (what) {
375 	case MOD_LOAD:
376 		err = ntb_setup_interface();
377 		break;
378 	case MOD_UNLOAD:
379 		err = ntb_teardown_interface();
380 		break;
381 	default:
382 		err = EOPNOTSUPP;
383 		break;
384 	}
385 	return (err);
386 }
387 
388 static moduledata_t if_ntb_mod = {
389 	"if_ntb",
390 	ntb_handle_module_events,
391 	NULL
392 };
393 
394 DECLARE_MODULE(if_ntb, if_ntb_mod, SI_SUB_KLD, SI_ORDER_ANY);
395 MODULE_DEPEND(if_ntb, ntb_hw, 1, 1, 1);
396 
397 static int
398 ntb_setup_interface(void)
399 {
400 	struct ifnet *ifp;
401 	struct ntb_queue_handlers handlers = { ntb_net_rx_handler,
402 	    ntb_net_tx_handler, ntb_net_event_handler };
403 	int rc;
404 
405 	net_softc.ntb = devclass_get_softc(devclass_find("ntb_hw"), 0);
406 	if (net_softc.ntb == NULL) {
407 		printf("ntb: Cannot find devclass\n");
408 		return (ENXIO);
409 	}
410 
411 	ifp = net_softc.ifp = if_alloc(IFT_ETHER);
412 	if (ifp == NULL) {
413 		ntb_transport_free(&net_softc);
414 		printf("ntb: Cannot allocate ifnet structure\n");
415 		return (ENOMEM);
416 	}
417 	if_initname(ifp, "ntb", 0);
418 
419 	rc = ntb_transport_probe(net_softc.ntb);
420 	if (rc != 0) {
421 		printf("ntb: Cannot init transport: %d\n", rc);
422 		if_free(net_softc.ifp);
423 		return (rc);
424 	}
425 
426 	net_softc.qp = ntb_transport_create_queue(ifp, net_softc.ntb,
427 	    &handlers);
428 	ifp->if_init = ntb_net_init;
429 	ifp->if_softc = &net_softc;
430 	ifp->if_flags = IFF_BROADCAST | IFF_SIMPLEX;
431 	ifp->if_ioctl = ntb_ioctl;
432 	ifp->if_start = ntb_start;
433 	IFQ_SET_MAXLEN(&ifp->if_snd, IFQ_MAXLEN);
434 	ifp->if_snd.ifq_drv_maxlen = IFQ_MAXLEN;
435 	IFQ_SET_READY(&ifp->if_snd);
436 	create_random_local_eui48(net_softc.eaddr);
437 	ether_ifattach(ifp, net_softc.eaddr);
438 	ifp->if_capabilities = IFCAP_HWCSUM | IFCAP_JUMBO_MTU;
439 	ifp->if_capenable = ifp->if_capabilities;
440 	ifp->if_mtu = ntb_transport_max_size(net_softc.qp) - ETHER_HDR_LEN -
441 	    ETHER_CRC_LEN;
442 
443 	ntb_transport_link_up(net_softc.qp);
444 	net_softc.bufsize = ntb_transport_max_size(net_softc.qp) +
445 	    sizeof(struct ether_header);
446 	return (0);
447 }
448 
449 static int
450 ntb_teardown_interface(void)
451 {
452 
453 	if (net_softc.qp != NULL) {
454 		ntb_transport_link_down(net_softc.qp);
455 
456 		ntb_transport_free_queue(net_softc.qp);
457 		ntb_transport_free(&net_softc);
458 	}
459 
460 	if (net_softc.ifp != NULL) {
461 		ether_ifdetach(net_softc.ifp);
462 		if_free(net_softc.ifp);
463 		net_softc.ifp = NULL;
464 	}
465 
466 	return (0);
467 }
468 
469 /* Network device interface */
470 
471 static void
472 ntb_net_init(void *arg)
473 {
474 	struct ntb_transport_ctx *ntb_softc = arg;
475 	struct ifnet *ifp = ntb_softc->ifp;
476 
477 	ifp->if_drv_flags |= IFF_DRV_RUNNING;
478 	ifp->if_drv_flags &= ~IFF_DRV_OACTIVE;
479 	ifp->if_flags |= IFF_UP;
480 	if_link_state_change(ifp, LINK_STATE_UP);
481 }
482 
483 static int
484 ntb_ioctl(struct ifnet *ifp, u_long command, caddr_t data)
485 {
486 	struct ntb_transport_ctx *nt = ifp->if_softc;
487 	struct ifreq *ifr = (struct ifreq *)data;
488 	int error = 0;
489 
490 	switch (command) {
491 	case SIOCSIFMTU:
492 	    {
493 		if (ifr->ifr_mtu > ntb_transport_max_size(nt->qp) -
494 		    ETHER_HDR_LEN - ETHER_CRC_LEN) {
495 			error = EINVAL;
496 			break;
497 		}
498 
499 		ifp->if_mtu = ifr->ifr_mtu;
500 		break;
501 	    }
502 	default:
503 		error = ether_ioctl(ifp, command, data);
504 		break;
505 	}
506 
507 	return (error);
508 }
509 
510 
511 static void
512 ntb_start(struct ifnet *ifp)
513 {
514 	struct mbuf *m_head;
515 	struct ntb_transport_ctx *nt = ifp->if_softc;
516 	int rc;
517 
518 	mtx_lock(&nt->tx_lock);
519 	ifp->if_drv_flags &= ~IFF_DRV_OACTIVE;
520 	CTR0(KTR_NTB, "TX: ntb_start");
521 	while (!IFQ_DRV_IS_EMPTY(&ifp->if_snd)) {
522 		IFQ_DRV_DEQUEUE(&ifp->if_snd, m_head);
523 		CTR1(KTR_NTB, "TX: start mbuf %p", m_head);
524 		rc = ntb_transport_tx_enqueue(nt->qp, m_head, m_head,
525 			     m_length(m_head, NULL));
526 		if (rc != 0) {
527 			CTR1(KTR_NTB,
528 			    "TX: could not tx mbuf %p. Returning to snd q",
529 			    m_head);
530 			if (rc == EAGAIN) {
531 				ifp->if_drv_flags |= IFF_DRV_OACTIVE;
532 				IFQ_DRV_PREPEND(&ifp->if_snd, m_head);
533 				callout_reset(&nt->qp->queue_full, hz / 1000,
534 				    ntb_qp_full, ifp);
535 			}
536 			break;
537 		}
538 
539 	}
540 	mtx_unlock(&nt->tx_lock);
541 }
542 
543 /* Network Device Callbacks */
544 static void
545 ntb_net_tx_handler(struct ntb_transport_qp *qp, void *qp_data, void *data,
546     int len)
547 {
548 
549 	m_freem(data);
550 	CTR1(KTR_NTB, "TX: tx_handler freeing mbuf %p", data);
551 }
552 
553 static void
554 ntb_net_rx_handler(struct ntb_transport_qp *qp, void *qp_data, void *data,
555     int len)
556 {
557 	struct mbuf *m = data;
558 	struct ifnet *ifp = qp_data;
559 
560 	CTR0(KTR_NTB, "RX: rx handler");
561 	(*ifp->if_input)(ifp, m);
562 }
563 
564 static void
565 ntb_net_event_handler(void *data, enum ntb_link_event status)
566 {
567 	struct ifnet *ifp;
568 
569 	ifp = data;
570 	(void)ifp;
571 
572 	/* XXX The Linux driver munges with the carrier status here. */
573 
574 	switch (status) {
575 	case NTB_LINK_DOWN:
576 		break;
577 	case NTB_LINK_UP:
578 		break;
579 	default:
580 		panic("Bogus ntb_link_event %u\n", status);
581 	}
582 }
583 
584 /* Transport Init and teardown */
585 
586 static void
587 xeon_link_watchdog_hb(void *arg)
588 {
589 	struct ntb_transport_ctx *nt;
590 
591 	nt = arg;
592 	ntb_spad_write(nt->ntb, IF_NTB_WATCHDOG_SPAD, 0);
593 	callout_reset(&nt->link_watchdog, 1 * hz, xeon_link_watchdog_hb, nt);
594 }
595 
596 static int
597 ntb_transport_probe(struct ntb_softc *ntb)
598 {
599 	struct ntb_transport_ctx *nt = &net_softc;
600 	struct ntb_transport_mw *mw;
601 	uint64_t qp_bitmap;
602 	int rc;
603 	unsigned i;
604 
605 	nt->mw_count = ntb_mw_count(ntb);
606 	for (i = 0; i < nt->mw_count; i++) {
607 		mw = &nt->mw_vec[i];
608 
609 		rc = ntb_mw_get_range(ntb, i, &mw->phys_addr, &mw->vbase,
610 		    &mw->phys_size, &mw->xlat_align, &mw->xlat_align_size,
611 		    &mw->addr_limit);
612 		if (rc != 0)
613 			goto err;
614 
615 		mw->buff_size = 0;
616 		mw->xlat_size = 0;
617 		mw->virt_addr = NULL;
618 		mw->dma_addr = 0;
619 
620 		rc = ntb_mw_set_wc(nt->ntb, i, VM_MEMATTR_WRITE_COMBINING);
621 		if (rc)
622 			ntb_printf(0, "Unable to set mw%d caching\n", i);
623 	}
624 
625 	qp_bitmap = ntb_db_valid_mask(ntb);
626 	nt->qp_count = flsll(qp_bitmap);
627 	KASSERT(nt->qp_count != 0, ("bogus db bitmap"));
628 	nt->qp_count -= 1;
629 
630 	if (max_num_clients != 0 && max_num_clients < nt->qp_count)
631 		nt->qp_count = max_num_clients;
632 	else if (nt->mw_count < nt->qp_count)
633 		nt->qp_count = nt->mw_count;
634 	KASSERT(nt->qp_count <= QP_SETSIZE, ("invalid qp_count"));
635 
636 	mtx_init(&nt->tx_lock, "ntb transport tx", NULL, MTX_DEF);
637 	mtx_init(&nt->rx_lock, "ntb transport rx", NULL, MTX_DEF);
638 
639 	nt->qp_vec = malloc(nt->qp_count * sizeof(*nt->qp_vec), M_NTB_IF,
640 	    M_WAITOK | M_ZERO);
641 
642 	for (i = 0; i < nt->qp_count; i++) {
643 		set_bit(i, &nt->qp_bitmap);
644 		set_bit(i, &nt->qp_bitmap_free);
645 		ntb_transport_init_queue(nt, i);
646 	}
647 
648 	callout_init(&nt->link_work, 0);
649 	callout_init(&nt->link_watchdog, 0);
650 	TASK_INIT(&nt->link_cleanup, 0, ntb_transport_link_cleanup_work, nt);
651 
652 	rc = ntb_set_ctx(ntb, nt, &ntb_transport_ops);
653 	if (rc != 0)
654 		goto err;
655 
656 	nt->link_is_up = false;
657 	ntb_link_enable(ntb, NTB_SPEED_AUTO, NTB_WIDTH_AUTO);
658 	ntb_link_event(ntb);
659 
660 	callout_reset(&nt->link_work, 0, ntb_transport_link_work, nt);
661 	if (enable_xeon_watchdog != 0)
662 		callout_reset(&nt->link_watchdog, 0, xeon_link_watchdog_hb, nt);
663 	return (0);
664 
665 err:
666 	free(nt->qp_vec, M_NTB_IF);
667 	nt->qp_vec = NULL;
668 	return (rc);
669 }
670 
671 static void
672 ntb_transport_free(struct ntb_transport_ctx *nt)
673 {
674 	struct ntb_softc *ntb = nt->ntb;
675 	struct _qpset qp_bitmap_alloc;
676 	uint8_t i;
677 
678 	ntb_transport_link_cleanup(nt);
679 	taskqueue_drain(taskqueue_swi, &nt->link_cleanup);
680 	callout_drain(&nt->link_work);
681 	callout_drain(&nt->link_watchdog);
682 
683 	BIT_COPY(QP_SETSIZE, &nt->qp_bitmap, &qp_bitmap_alloc);
684 	BIT_NAND(QP_SETSIZE, &qp_bitmap_alloc, &nt->qp_bitmap_free);
685 
686 	/* Verify that all the QPs are freed */
687 	for (i = 0; i < nt->qp_count; i++)
688 		if (test_bit(i, &qp_bitmap_alloc))
689 			ntb_transport_free_queue(&nt->qp_vec[i]);
690 
691 	ntb_link_disable(ntb);
692 	ntb_clear_ctx(ntb);
693 
694 	for (i = 0; i < nt->mw_count; i++)
695 		ntb_free_mw(nt, i);
696 
697 	free(nt->qp_vec, M_NTB_IF);
698 }
699 
700 static void
701 ntb_transport_init_queue(struct ntb_transport_ctx *nt, unsigned int qp_num)
702 {
703 	struct ntb_transport_mw *mw;
704 	struct ntb_transport_qp *qp;
705 	vm_paddr_t mw_base;
706 	uint64_t mw_size, qp_offset;
707 	size_t tx_size;
708 	unsigned num_qps_mw, mw_num, mw_count;
709 
710 	mw_count = nt->mw_count;
711 	mw_num = QP_TO_MW(nt, qp_num);
712 	mw = &nt->mw_vec[mw_num];
713 
714 	qp = &nt->qp_vec[qp_num];
715 	qp->qp_num = qp_num;
716 	qp->transport = nt;
717 	qp->ntb = nt->ntb;
718 	qp->client_ready = false;
719 	qp->event_handler = NULL;
720 	ntb_qp_link_down_reset(qp);
721 
722 	if (nt->qp_count % mw_count && mw_num + 1 < nt->qp_count / mw_count)
723 		num_qps_mw = nt->qp_count / mw_count + 1;
724 	else
725 		num_qps_mw = nt->qp_count / mw_count;
726 
727 	mw_base = mw->phys_addr;
728 	mw_size = mw->phys_size;
729 
730 	tx_size = mw_size / num_qps_mw;
731 	qp_offset = tx_size * (qp_num / mw_count);
732 
733 	qp->tx_mw = mw->vbase + qp_offset;
734 	KASSERT(qp->tx_mw != NULL, ("uh oh?"));
735 
736 	/* XXX Assumes that a vm_paddr_t is equivalent to bus_addr_t */
737 	qp->tx_mw_phys = mw_base + qp_offset;
738 	KASSERT(qp->tx_mw_phys != 0, ("uh oh?"));
739 
740 	tx_size -= sizeof(struct ntb_rx_info);
741 	qp->rx_info = (void *)(qp->tx_mw + tx_size);
742 
743 	/* Due to house-keeping, there must be at least 2 buffs */
744 	qp->tx_max_frame = qmin(tx_size / 2,
745 	    transport_mtu + sizeof(struct ntb_payload_header));
746 	qp->tx_max_entry = tx_size / qp->tx_max_frame;
747 
748 	callout_init(&qp->link_work, 0);
749 	callout_init(&qp->queue_full, 1);
750 	callout_init(&qp->rx_full, 1);
751 
752 	mtx_init(&qp->ntb_rx_q_lock, "ntb rx q", NULL, MTX_SPIN);
753 	mtx_init(&qp->ntb_tx_free_q_lock, "ntb tx free q", NULL, MTX_SPIN);
754 	TASK_INIT(&qp->rx_completion_task, 0, ntb_complete_rxc, qp);
755 	TASK_INIT(&qp->rxc_db_work, 0, ntb_transport_rxc_db, qp);
756 
757 	STAILQ_INIT(&qp->rx_post_q);
758 	STAILQ_INIT(&qp->rx_pend_q);
759 	STAILQ_INIT(&qp->tx_free_q);
760 
761 	callout_reset(&qp->link_work, 0, ntb_qp_link_work, qp);
762 }
763 
764 static void
765 ntb_transport_free_queue(struct ntb_transport_qp *qp)
766 {
767 	struct ntb_queue_entry *entry;
768 
769 	if (qp == NULL)
770 		return;
771 
772 	callout_drain(&qp->link_work);
773 
774 	ntb_db_set_mask(qp->ntb, 1ull << qp->qp_num);
775 	taskqueue_drain(taskqueue_swi, &qp->rxc_db_work);
776 	taskqueue_drain(taskqueue_swi, &qp->rx_completion_task);
777 
778 	qp->cb_data = NULL;
779 	qp->rx_handler = NULL;
780 	qp->tx_handler = NULL;
781 	qp->event_handler = NULL;
782 
783 	while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_pend_q)))
784 		free(entry, M_NTB_IF);
785 
786 	while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_post_q)))
787 		free(entry, M_NTB_IF);
788 
789 	while ((entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q)))
790 		free(entry, M_NTB_IF);
791 
792 	set_bit(qp->qp_num, &qp->transport->qp_bitmap_free);
793 }
794 
795 /**
796  * ntb_transport_create_queue - Create a new NTB transport layer queue
797  * @rx_handler: receive callback function
798  * @tx_handler: transmit callback function
799  * @event_handler: event callback function
800  *
801  * Create a new NTB transport layer queue and provide the queue with a callback
802  * routine for both transmit and receive.  The receive callback routine will be
803  * used to pass up data when the transport has received it on the queue.   The
804  * transmit callback routine will be called when the transport has completed the
805  * transmission of the data on the queue and the data is ready to be freed.
806  *
807  * RETURNS: pointer to newly created ntb_queue, NULL on error.
808  */
809 static struct ntb_transport_qp *
810 ntb_transport_create_queue(void *data, struct ntb_softc *ntb,
811     const struct ntb_queue_handlers *handlers)
812 {
813 	struct ntb_queue_entry *entry;
814 	struct ntb_transport_qp *qp;
815 	struct ntb_transport_ctx *nt;
816 	unsigned int free_queue;
817 	int i;
818 
819 	nt = ntb_get_ctx(ntb, NULL);
820 	KASSERT(nt != NULL, ("bogus"));
821 
822 	free_queue = ffs_bit(&nt->qp_bitmap);
823 	if (free_queue == 0)
824 		return (NULL);
825 
826 	/* decrement free_queue to make it zero based */
827 	free_queue--;
828 
829 	qp = &nt->qp_vec[free_queue];
830 	clear_bit(qp->qp_num, &nt->qp_bitmap_free);
831 	qp->cb_data = data;
832 	qp->rx_handler = handlers->rx_handler;
833 	qp->tx_handler = handlers->tx_handler;
834 	qp->event_handler = handlers->event_handler;
835 
836 	for (i = 0; i < NTB_QP_DEF_NUM_ENTRIES; i++) {
837 		entry = malloc(sizeof(*entry), M_NTB_IF, M_WAITOK | M_ZERO);
838 		entry->cb_data = nt->ifp;
839 		entry->buf = NULL;
840 		entry->len = transport_mtu;
841 		ntb_list_add(&qp->ntb_rx_q_lock, entry, &qp->rx_pend_q);
842 	}
843 
844 	for (i = 0; i < NTB_QP_DEF_NUM_ENTRIES; i++) {
845 		entry = malloc(sizeof(*entry), M_NTB_IF, M_WAITOK | M_ZERO);
846 		ntb_list_add(&qp->ntb_tx_free_q_lock, entry, &qp->tx_free_q);
847 	}
848 
849 	ntb_db_clear(ntb, 1ull << qp->qp_num);
850 	ntb_db_clear_mask(ntb, 1ull << qp->qp_num);
851 	return (qp);
852 }
853 
854 /**
855  * ntb_transport_link_up - Notify NTB transport of client readiness to use queue
856  * @qp: NTB transport layer queue to be enabled
857  *
858  * Notify NTB transport layer of client readiness to use queue
859  */
860 static void
861 ntb_transport_link_up(struct ntb_transport_qp *qp)
862 {
863 	struct ntb_transport_ctx *nt;
864 
865 	if (qp == NULL)
866 		return;
867 
868 	qp->client_ready = true;
869 
870 	nt = qp->transport;
871 	ntb_printf(2, "qp client ready\n");
872 
873 	if (qp->transport->link_is_up)
874 		callout_reset(&qp->link_work, 0, ntb_qp_link_work, qp);
875 }
876 
877 
878 
879 /* Transport Tx */
880 
881 /**
882  * ntb_transport_tx_enqueue - Enqueue a new NTB queue entry
883  * @qp: NTB transport layer queue the entry is to be enqueued on
884  * @cb: per buffer pointer for callback function to use
885  * @data: pointer to data buffer that will be sent
886  * @len: length of the data buffer
887  *
888  * Enqueue a new transmit buffer onto the transport queue from which a NTB
889  * payload will be transmitted.  This assumes that a lock is being held to
890  * serialize access to the qp.
891  *
892  * RETURNS: An appropriate ERRNO error value on error, or zero for success.
893  */
894 static int
895 ntb_transport_tx_enqueue(struct ntb_transport_qp *qp, void *cb, void *data,
896     unsigned int len)
897 {
898 	struct ntb_queue_entry *entry;
899 	int rc;
900 
901 	if (qp == NULL || !qp->link_is_up || len == 0) {
902 		CTR0(KTR_NTB, "TX: link not up");
903 		return (EINVAL);
904 	}
905 
906 	entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q);
907 	if (entry == NULL) {
908 		CTR0(KTR_NTB, "TX: could not get entry from tx_free_q");
909 		qp->tx_err_no_buf++;
910 		return (EBUSY);
911 	}
912 	CTR1(KTR_NTB, "TX: got entry %p from tx_free_q", entry);
913 
914 	entry->cb_data = cb;
915 	entry->buf = data;
916 	entry->len = len;
917 	entry->flags = 0;
918 
919 	rc = ntb_process_tx(qp, entry);
920 	if (rc != 0) {
921 		ntb_list_add(&qp->ntb_tx_free_q_lock, entry, &qp->tx_free_q);
922 		CTR1(KTR_NTB,
923 		    "TX: process_tx failed. Returning entry %p to tx_free_q",
924 		    entry);
925 	}
926 	return (rc);
927 }
928 
929 static int
930 ntb_process_tx(struct ntb_transport_qp *qp, struct ntb_queue_entry *entry)
931 {
932 	void *offset;
933 
934 	offset = qp->tx_mw + qp->tx_max_frame * qp->tx_index;
935 	CTR3(KTR_NTB,
936 	    "TX: process_tx: tx_pkts=%lu, tx_index=%u, remote entry=%u",
937 	    qp->tx_pkts, qp->tx_index, qp->remote_rx_info->entry);
938 	if (qp->tx_index == qp->remote_rx_info->entry) {
939 		CTR0(KTR_NTB, "TX: ring full");
940 		qp->tx_ring_full++;
941 		return (EAGAIN);
942 	}
943 
944 	if (entry->len > qp->tx_max_frame - sizeof(struct ntb_payload_header)) {
945 		if (qp->tx_handler != NULL)
946 			qp->tx_handler(qp, qp->cb_data, entry->buf,
947 			    EIO);
948 		else
949 			m_freem(entry->buf);
950 
951 		entry->buf = NULL;
952 		ntb_list_add(&qp->ntb_tx_free_q_lock, entry, &qp->tx_free_q);
953 		CTR1(KTR_NTB,
954 		    "TX: frame too big. returning entry %p to tx_free_q",
955 		    entry);
956 		return (0);
957 	}
958 	CTR2(KTR_NTB, "TX: copying entry %p to offset %p", entry, offset);
959 	ntb_memcpy_tx(qp, entry, offset);
960 
961 	qp->tx_index++;
962 	qp->tx_index %= qp->tx_max_entry;
963 
964 	qp->tx_pkts++;
965 
966 	return (0);
967 }
968 
969 static void
970 ntb_memcpy_tx(struct ntb_transport_qp *qp, struct ntb_queue_entry *entry,
971     void *offset)
972 {
973 	struct ntb_payload_header *hdr;
974 
975 	/* This piece is from Linux' ntb_async_tx() */
976 	hdr = (struct ntb_payload_header *)((char *)offset + qp->tx_max_frame -
977 	    sizeof(struct ntb_payload_header));
978 	entry->x_hdr = hdr;
979 	iowrite32(entry->len, &hdr->len);
980 	iowrite32(qp->tx_pkts, &hdr->ver);
981 
982 	/* This piece is ntb_memcpy_tx() */
983 	CTR2(KTR_NTB, "TX: copying %d bytes to offset %p", entry->len, offset);
984 	if (entry->buf != NULL) {
985 		m_copydata((struct mbuf *)entry->buf, 0, entry->len, offset);
986 
987 		/*
988 		 * Ensure that the data is fully copied before setting the
989 		 * flags
990 		 */
991 		wmb();
992 	}
993 
994 	/* The rest is ntb_tx_copy_callback() */
995 	iowrite32(entry->flags | IF_NTB_DESC_DONE_FLAG, &hdr->flags);
996 	CTR1(KTR_NTB, "TX: hdr %p set DESC_DONE", hdr);
997 
998 	ntb_peer_db_set(qp->ntb, 1ull << qp->qp_num);
999 
1000 	/*
1001 	 * The entry length can only be zero if the packet is intended to be a
1002 	 * "link down" or similar.  Since no payload is being sent in these
1003 	 * cases, there is nothing to add to the completion queue.
1004 	 */
1005 	if (entry->len > 0) {
1006 		qp->tx_bytes += entry->len;
1007 
1008 		if (qp->tx_handler)
1009 			qp->tx_handler(qp, qp->cb_data, entry->buf,
1010 			    entry->len);
1011 		else
1012 			m_freem(entry->buf);
1013 		entry->buf = NULL;
1014 	}
1015 
1016 	CTR3(KTR_NTB,
1017 	    "TX: entry %p sent. hdr->ver = %u, hdr->flags = 0x%x, Returning "
1018 	    "to tx_free_q", entry, hdr->ver, hdr->flags);
1019 	ntb_list_add(&qp->ntb_tx_free_q_lock, entry, &qp->tx_free_q);
1020 }
1021 
1022 static void
1023 ntb_qp_full(void *arg)
1024 {
1025 
1026 	CTR0(KTR_NTB, "TX: qp_full callout");
1027 	ntb_start(arg);
1028 }
1029 
1030 /* Transport Rx */
1031 static void
1032 ntb_transport_rxc_db(void *arg, int pending __unused)
1033 {
1034 	struct ntb_transport_qp *qp = arg;
1035 	ntb_q_idx_t i;
1036 	int rc;
1037 
1038 	/*
1039 	 * Limit the number of packets processed in a single interrupt to
1040 	 * provide fairness to others
1041 	 */
1042 	CTR0(KTR_NTB, "RX: transport_rx");
1043 	mtx_lock(&qp->transport->rx_lock);
1044 	for (i = 0; i < qp->rx_max_entry; i++) {
1045 		rc = ntb_process_rxc(qp);
1046 		if (rc != 0) {
1047 			CTR0(KTR_NTB, "RX: process_rxc failed");
1048 			break;
1049 		}
1050 	}
1051 	mtx_unlock(&qp->transport->rx_lock);
1052 
1053 	if (i == qp->rx_max_entry)
1054 		taskqueue_enqueue(taskqueue_swi, &qp->rxc_db_work);
1055 	else if ((ntb_db_read(qp->ntb) & (1ull << qp->qp_num)) != 0) {
1056 		/* If db is set, clear it and read it back to commit clear. */
1057 		ntb_db_clear(qp->ntb, 1ull << qp->qp_num);
1058 		(void)ntb_db_read(qp->ntb);
1059 
1060 		/*
1061 		 * An interrupt may have arrived between finishing
1062 		 * ntb_process_rxc and clearing the doorbell bit: there might
1063 		 * be some more work to do.
1064 		 */
1065 		taskqueue_enqueue(taskqueue_swi, &qp->rxc_db_work);
1066 	}
1067 }
1068 
1069 static int
1070 ntb_process_rxc(struct ntb_transport_qp *qp)
1071 {
1072 	struct ntb_payload_header *hdr;
1073 	struct ntb_queue_entry *entry;
1074 	caddr_t offset;
1075 
1076 	offset = qp->rx_buff + qp->rx_max_frame * qp->rx_index;
1077 	hdr = (void *)(offset + qp->rx_max_frame -
1078 	    sizeof(struct ntb_payload_header));
1079 
1080 	CTR1(KTR_NTB, "RX: process_rxc rx_index = %u", qp->rx_index);
1081 	if ((hdr->flags & IF_NTB_DESC_DONE_FLAG) == 0) {
1082 		CTR0(KTR_NTB, "RX: hdr not done");
1083 		qp->rx_ring_empty++;
1084 		return (EAGAIN);
1085 	}
1086 
1087 	if ((hdr->flags & IF_NTB_LINK_DOWN_FLAG) != 0) {
1088 		CTR0(KTR_NTB, "RX: link down");
1089 		ntb_qp_link_down(qp);
1090 		hdr->flags = 0;
1091 		return (EAGAIN);
1092 	}
1093 
1094 	if (hdr->ver != (uint32_t)qp->rx_pkts) {
1095 		CTR2(KTR_NTB,"RX: ver != rx_pkts (%x != %lx). "
1096 		    "Returning entry to rx_pend_q", hdr->ver, qp->rx_pkts);
1097 		qp->rx_err_ver++;
1098 		return (EIO);
1099 	}
1100 
1101 	entry = ntb_list_mv(&qp->ntb_rx_q_lock, &qp->rx_pend_q, &qp->rx_post_q);
1102 	if (entry == NULL) {
1103 		qp->rx_err_no_buf++;
1104 		CTR0(KTR_NTB, "RX: No entries in rx_pend_q");
1105 		return (EAGAIN);
1106 	}
1107 	callout_stop(&qp->rx_full);
1108 	CTR1(KTR_NTB, "RX: rx entry %p from rx_pend_q", entry);
1109 
1110 	entry->x_hdr = hdr;
1111 	entry->index = qp->rx_index;
1112 
1113 	if (hdr->len > entry->len) {
1114 		CTR2(KTR_NTB, "RX: len too long. Wanted %ju got %ju",
1115 		    (uintmax_t)hdr->len, (uintmax_t)entry->len);
1116 		qp->rx_err_oflow++;
1117 
1118 		entry->len = -EIO;
1119 		entry->flags |= IF_NTB_DESC_DONE_FLAG;
1120 
1121 		taskqueue_enqueue(taskqueue_swi, &qp->rx_completion_task);
1122 	} else {
1123 		qp->rx_bytes += hdr->len;
1124 		qp->rx_pkts++;
1125 
1126 		CTR1(KTR_NTB, "RX: received %ld rx_pkts", qp->rx_pkts);
1127 
1128 		entry->len = hdr->len;
1129 
1130 		ntb_memcpy_rx(qp, entry, offset);
1131 	}
1132 
1133 	qp->rx_index++;
1134 	qp->rx_index %= qp->rx_max_entry;
1135 	return (0);
1136 }
1137 
1138 static void
1139 ntb_memcpy_rx(struct ntb_transport_qp *qp, struct ntb_queue_entry *entry,
1140     void *offset)
1141 {
1142 	struct ifnet *ifp = entry->cb_data;
1143 	unsigned int len = entry->len;
1144 	struct mbuf *m;
1145 
1146 	CTR2(KTR_NTB, "RX: copying %d bytes from offset %p", len, offset);
1147 	m = m_devget(offset, len, 0, ifp, NULL);
1148 	m->m_pkthdr.csum_flags = CSUM_IP_CHECKED | CSUM_IP_VALID;
1149 
1150 	entry->buf = (void *)m;
1151 
1152 	/* Ensure that the data is globally visible before clearing the flag */
1153 	wmb();
1154 
1155 	CTR2(KTR_NTB, "RX: copied entry %p to mbuf %p.", entry, m);
1156 	ntb_rx_copy_callback(qp, entry);
1157 }
1158 
1159 static inline void
1160 ntb_rx_copy_callback(struct ntb_transport_qp *qp, void *data)
1161 {
1162 	struct ntb_queue_entry *entry;
1163 
1164 	entry = data;
1165 	entry->flags |= IF_NTB_DESC_DONE_FLAG;
1166 	taskqueue_enqueue(taskqueue_swi, &qp->rx_completion_task);
1167 }
1168 
1169 static void
1170 ntb_complete_rxc(void *arg, int pending)
1171 {
1172 	struct ntb_transport_qp *qp = arg;
1173 	struct ntb_queue_entry *entry;
1174 	struct mbuf *m;
1175 	unsigned len;
1176 
1177 	CTR0(KTR_NTB, "RX: rx_completion_task");
1178 
1179 	mtx_lock_spin(&qp->ntb_rx_q_lock);
1180 
1181 	while (!STAILQ_EMPTY(&qp->rx_post_q)) {
1182 		entry = STAILQ_FIRST(&qp->rx_post_q);
1183 		if ((entry->flags & IF_NTB_DESC_DONE_FLAG) == 0)
1184 			break;
1185 
1186 		entry->x_hdr->flags = 0;
1187 		iowrite32(entry->index, &qp->rx_info->entry);
1188 
1189 		STAILQ_REMOVE_HEAD(&qp->rx_post_q, entry);
1190 
1191 		len = entry->len;
1192 		m = entry->buf;
1193 
1194 		/*
1195 		 * Re-initialize queue_entry for reuse; rx_handler takes
1196 		 * ownership of the mbuf.
1197 		 */
1198 		entry->buf = NULL;
1199 		entry->len = transport_mtu;
1200 		entry->cb_data = qp->transport->ifp;
1201 
1202 		STAILQ_INSERT_TAIL(&qp->rx_pend_q, entry, entry);
1203 
1204 		mtx_unlock_spin(&qp->ntb_rx_q_lock);
1205 
1206 		CTR2(KTR_NTB, "RX: completing entry %p, mbuf %p", entry, m);
1207 		if (qp->rx_handler != NULL && qp->client_ready)
1208 			qp->rx_handler(qp, qp->cb_data, m, len);
1209 		else
1210 			m_freem(m);
1211 
1212 		mtx_lock_spin(&qp->ntb_rx_q_lock);
1213 	}
1214 
1215 	mtx_unlock_spin(&qp->ntb_rx_q_lock);
1216 }
1217 
1218 static void
1219 ntb_transport_doorbell_callback(void *data, uint32_t vector)
1220 {
1221 	struct ntb_transport_ctx *nt = data;
1222 	struct ntb_transport_qp *qp;
1223 	struct _qpset db_bits;
1224 	uint64_t vec_mask;
1225 	unsigned qp_num;
1226 
1227 	BIT_COPY(QP_SETSIZE, &nt->qp_bitmap, &db_bits);
1228 	BIT_NAND(QP_SETSIZE, &db_bits, &nt->qp_bitmap_free);
1229 
1230 	vec_mask = ntb_db_vector_mask(nt->ntb, vector);
1231 	while (vec_mask != 0) {
1232 		qp_num = ffsll(vec_mask) - 1;
1233 
1234 		if (test_bit(qp_num, &db_bits)) {
1235 			qp = &nt->qp_vec[qp_num];
1236 			taskqueue_enqueue(taskqueue_swi, &qp->rxc_db_work);
1237 		}
1238 
1239 		vec_mask &= ~(1ull << qp_num);
1240 	}
1241 }
1242 
1243 /* Link Event handler */
1244 static void
1245 ntb_transport_event_callback(void *data)
1246 {
1247 	struct ntb_transport_ctx *nt = data;
1248 
1249 	if (ntb_link_is_up(nt->ntb, NULL, NULL)) {
1250 		ntb_printf(1, "HW link up\n");
1251 		callout_reset(&nt->link_work, 0, ntb_transport_link_work, nt);
1252 	} else {
1253 		ntb_printf(1, "HW link down\n");
1254 		taskqueue_enqueue(taskqueue_swi, &nt->link_cleanup);
1255 	}
1256 }
1257 
1258 /* Link bring up */
1259 static void
1260 ntb_transport_link_work(void *arg)
1261 {
1262 	struct ntb_transport_ctx *nt = arg;
1263 	struct ntb_softc *ntb = nt->ntb;
1264 	struct ntb_transport_qp *qp;
1265 	uint64_t val64, size;
1266 	uint32_t val;
1267 	unsigned i;
1268 	int rc;
1269 
1270 	/* send the local info, in the opposite order of the way we read it */
1271 	for (i = 0; i < nt->mw_count; i++) {
1272 		size = nt->mw_vec[i].phys_size;
1273 
1274 		if (max_mw_size != 0 && size > max_mw_size)
1275 			size = max_mw_size;
1276 
1277 		ntb_peer_spad_write(ntb, IF_NTB_MW0_SZ_HIGH + (i * 2),
1278 		    size >> 32);
1279 		ntb_peer_spad_write(ntb, IF_NTB_MW0_SZ_LOW + (i * 2), size);
1280 	}
1281 
1282 	ntb_peer_spad_write(ntb, IF_NTB_NUM_MWS, nt->mw_count);
1283 
1284 	ntb_peer_spad_write(ntb, IF_NTB_NUM_QPS, nt->qp_count);
1285 
1286 	ntb_peer_spad_write(ntb, IF_NTB_VERSION, NTB_TRANSPORT_VERSION);
1287 
1288 	/* Query the remote side for its info */
1289 	val = 0;
1290 	ntb_spad_read(ntb, IF_NTB_VERSION, &val);
1291 	if (val != NTB_TRANSPORT_VERSION)
1292 		goto out;
1293 
1294 	ntb_spad_read(ntb, IF_NTB_NUM_QPS, &val);
1295 	if (val != nt->qp_count)
1296 		goto out;
1297 
1298 	ntb_spad_read(ntb, IF_NTB_NUM_MWS, &val);
1299 	if (val != nt->mw_count)
1300 		goto out;
1301 
1302 	for (i = 0; i < nt->mw_count; i++) {
1303 		ntb_spad_read(ntb, IF_NTB_MW0_SZ_HIGH + (i * 2), &val);
1304 		val64 = (uint64_t)val << 32;
1305 
1306 		ntb_spad_read(ntb, IF_NTB_MW0_SZ_LOW + (i * 2), &val);
1307 		val64 |= val;
1308 
1309 		rc = ntb_set_mw(nt, i, val64);
1310 		if (rc != 0)
1311 			goto free_mws;
1312 	}
1313 
1314 	nt->link_is_up = true;
1315 	ntb_printf(1, "transport link up\n");
1316 
1317 	for (i = 0; i < nt->qp_count; i++) {
1318 		qp = &nt->qp_vec[i];
1319 
1320 		ntb_transport_setup_qp_mw(nt, i);
1321 
1322 		if (qp->client_ready)
1323 			callout_reset(&qp->link_work, 0, ntb_qp_link_work, qp);
1324 	}
1325 
1326 	return;
1327 
1328 free_mws:
1329 	for (i = 0; i < nt->mw_count; i++)
1330 		ntb_free_mw(nt, i);
1331 out:
1332 	if (ntb_link_is_up(ntb, NULL, NULL))
1333 		callout_reset(&nt->link_work,
1334 		    NTB_LINK_DOWN_TIMEOUT * hz / 1000, ntb_transport_link_work, nt);
1335 }
1336 
1337 static int
1338 ntb_set_mw(struct ntb_transport_ctx *nt, int num_mw, size_t size)
1339 {
1340 	struct ntb_transport_mw *mw = &nt->mw_vec[num_mw];
1341 	size_t xlat_size, buff_size;
1342 	int rc;
1343 
1344 	if (size == 0)
1345 		return (EINVAL);
1346 
1347 	xlat_size = roundup(size, mw->xlat_align_size);
1348 	buff_size = xlat_size;
1349 
1350 	/* No need to re-setup */
1351 	if (mw->xlat_size == xlat_size)
1352 		return (0);
1353 
1354 	if (mw->buff_size != 0)
1355 		ntb_free_mw(nt, num_mw);
1356 
1357 	/* Alloc memory for receiving data.  Must be aligned */
1358 	mw->xlat_size = xlat_size;
1359 	mw->buff_size = buff_size;
1360 
1361 	mw->virt_addr = contigmalloc(mw->buff_size, M_NTB_IF, M_ZERO, 0,
1362 	    mw->addr_limit, mw->xlat_align, 0);
1363 	if (mw->virt_addr == NULL) {
1364 		ntb_printf(0, "Unable to allocate MW buffer of size %zu/%zu\n",
1365 		    mw->buff_size, mw->xlat_size);
1366 		mw->xlat_size = 0;
1367 		mw->buff_size = 0;
1368 		return (ENOMEM);
1369 	}
1370 	/* TODO: replace with bus_space_* functions */
1371 	mw->dma_addr = vtophys(mw->virt_addr);
1372 
1373 	/*
1374 	 * Ensure that the allocation from contigmalloc is aligned as
1375 	 * requested.  XXX: This may not be needed -- brought in for parity
1376 	 * with the Linux driver.
1377 	 */
1378 	if (mw->dma_addr % mw->xlat_align != 0) {
1379 		ntb_printf(0,
1380 		    "DMA memory 0x%jx not aligned to BAR size 0x%zx\n",
1381 		    (uintmax_t)mw->dma_addr, size);
1382 		ntb_free_mw(nt, num_mw);
1383 		return (ENOMEM);
1384 	}
1385 
1386 	/* Notify HW the memory location of the receive buffer */
1387 	rc = ntb_mw_set_trans(nt->ntb, num_mw, mw->dma_addr, mw->xlat_size);
1388 	if (rc) {
1389 		ntb_printf(0, "Unable to set mw%d translation\n", num_mw);
1390 		ntb_free_mw(nt, num_mw);
1391 		return (rc);
1392 	}
1393 
1394 	return (0);
1395 }
1396 
1397 static void
1398 ntb_free_mw(struct ntb_transport_ctx *nt, int num_mw)
1399 {
1400 	struct ntb_transport_mw *mw = &nt->mw_vec[num_mw];
1401 
1402 	if (mw->virt_addr == NULL)
1403 		return;
1404 
1405 	ntb_mw_clear_trans(nt->ntb, num_mw);
1406 	contigfree(mw->virt_addr, mw->xlat_size, M_NTB_IF);
1407 	mw->xlat_size = 0;
1408 	mw->buff_size = 0;
1409 	mw->virt_addr = NULL;
1410 }
1411 
1412 static int
1413 ntb_transport_setup_qp_mw(struct ntb_transport_ctx *nt, unsigned int qp_num)
1414 {
1415 	struct ntb_transport_qp *qp = &nt->qp_vec[qp_num];
1416 	struct ntb_transport_mw *mw;
1417 	void *offset;
1418 	ntb_q_idx_t i;
1419 	size_t rx_size;
1420 	unsigned num_qps_mw, mw_num, mw_count;
1421 
1422 	mw_count = nt->mw_count;
1423 	mw_num = QP_TO_MW(nt, qp_num);
1424 	mw = &nt->mw_vec[mw_num];
1425 
1426 	if (mw->virt_addr == NULL)
1427 		return (ENOMEM);
1428 
1429 	if (nt->qp_count % mw_count && mw_num + 1 < nt->qp_count / mw_count)
1430 		num_qps_mw = nt->qp_count / mw_count + 1;
1431 	else
1432 		num_qps_mw = nt->qp_count / mw_count;
1433 
1434 	rx_size = mw->xlat_size / num_qps_mw;
1435 	qp->rx_buff = mw->virt_addr + rx_size * (qp_num / mw_count);
1436 	rx_size -= sizeof(struct ntb_rx_info);
1437 
1438 	qp->remote_rx_info = (void*)(qp->rx_buff + rx_size);
1439 
1440 	/* Due to house-keeping, there must be at least 2 buffs */
1441 	qp->rx_max_frame = qmin(rx_size / 2,
1442 	    transport_mtu + sizeof(struct ntb_payload_header));
1443 	qp->rx_max_entry = rx_size / qp->rx_max_frame;
1444 	qp->rx_index = 0;
1445 
1446 	qp->remote_rx_info->entry = qp->rx_max_entry - 1;
1447 
1448 	/* Set up the hdr offsets with 0s */
1449 	for (i = 0; i < qp->rx_max_entry; i++) {
1450 		offset = (void *)(qp->rx_buff + qp->rx_max_frame * (i + 1) -
1451 		    sizeof(struct ntb_payload_header));
1452 		memset(offset, 0, sizeof(struct ntb_payload_header));
1453 	}
1454 
1455 	qp->rx_pkts = 0;
1456 	qp->tx_pkts = 0;
1457 	qp->tx_index = 0;
1458 
1459 	return (0);
1460 }
1461 
1462 static void
1463 ntb_qp_link_work(void *arg)
1464 {
1465 	struct ntb_transport_qp *qp = arg;
1466 	struct ntb_softc *ntb = qp->ntb;
1467 	struct ntb_transport_ctx *nt = qp->transport;
1468 	uint32_t val, dummy;
1469 
1470 	ntb_spad_read(ntb, IF_NTB_QP_LINKS, &val);
1471 
1472 	ntb_peer_spad_write(ntb, IF_NTB_QP_LINKS, val | (1ull << qp->qp_num));
1473 
1474 	/* query remote spad for qp ready bits */
1475 	ntb_peer_spad_read(ntb, IF_NTB_QP_LINKS, &dummy);
1476 
1477 	/* See if the remote side is up */
1478 	if ((val & (1ull << qp->qp_num)) != 0) {
1479 		ntb_printf(2, "qp link up\n");
1480 		qp->link_is_up = true;
1481 
1482 		if (qp->event_handler != NULL)
1483 			qp->event_handler(qp->cb_data, NTB_LINK_UP);
1484 
1485 		taskqueue_enqueue(taskqueue_swi, &qp->rxc_db_work);
1486 	} else if (nt->link_is_up)
1487 		callout_reset(&qp->link_work,
1488 		    NTB_LINK_DOWN_TIMEOUT * hz / 1000, ntb_qp_link_work, qp);
1489 }
1490 
1491 /* Link down event*/
1492 static void
1493 ntb_transport_link_cleanup(struct ntb_transport_ctx *nt)
1494 {
1495 	struct ntb_transport_qp *qp;
1496 	struct _qpset qp_bitmap_alloc;
1497 	unsigned i;
1498 
1499 	BIT_COPY(QP_SETSIZE, &nt->qp_bitmap, &qp_bitmap_alloc);
1500 	BIT_NAND(QP_SETSIZE, &qp_bitmap_alloc, &nt->qp_bitmap_free);
1501 
1502 	/* Pass along the info to any clients */
1503 	for (i = 0; i < nt->qp_count; i++)
1504 		if (test_bit(i, &qp_bitmap_alloc)) {
1505 			qp = &nt->qp_vec[i];
1506 			ntb_qp_link_cleanup(qp);
1507 			callout_drain(&qp->link_work);
1508 		}
1509 
1510 	if (!nt->link_is_up)
1511 		callout_drain(&nt->link_work);
1512 
1513 	/*
1514 	 * The scratchpad registers keep the values if the remote side
1515 	 * goes down, blast them now to give them a sane value the next
1516 	 * time they are accessed
1517 	 */
1518 	for (i = 0; i < IF_NTB_MAX_SPAD; i++)
1519 		ntb_spad_write(nt->ntb, i, 0);
1520 }
1521 
1522 static void
1523 ntb_transport_link_cleanup_work(void *arg, int pending __unused)
1524 {
1525 
1526 	ntb_transport_link_cleanup(arg);
1527 }
1528 
1529 static void
1530 ntb_qp_link_down(struct ntb_transport_qp *qp)
1531 {
1532 
1533 	ntb_qp_link_cleanup(qp);
1534 }
1535 
1536 static void
1537 ntb_qp_link_down_reset(struct ntb_transport_qp *qp)
1538 {
1539 
1540 	qp->link_is_up = false;
1541 
1542 	qp->tx_index = qp->rx_index = 0;
1543 	qp->tx_bytes = qp->rx_bytes = 0;
1544 	qp->tx_pkts = qp->rx_pkts = 0;
1545 
1546 	qp->rx_ring_empty = 0;
1547 	qp->tx_ring_full = 0;
1548 
1549 	qp->rx_err_no_buf = qp->tx_err_no_buf = 0;
1550 	qp->rx_err_oflow = qp->rx_err_ver = 0;
1551 }
1552 
1553 static void
1554 ntb_qp_link_cleanup(struct ntb_transport_qp *qp)
1555 {
1556 	struct ntb_transport_ctx *nt = qp->transport;
1557 
1558 	callout_drain(&qp->link_work);
1559 	ntb_qp_link_down_reset(qp);
1560 
1561 	if (qp->event_handler != NULL)
1562 		qp->event_handler(qp->cb_data, NTB_LINK_DOWN);
1563 
1564 	if (nt->link_is_up)
1565 		callout_reset(&qp->link_work,
1566 		    NTB_LINK_DOWN_TIMEOUT * hz / 1000, ntb_qp_link_work, qp);
1567 }
1568 
1569 /* Link commanded down */
1570 /**
1571  * ntb_transport_link_down - Notify NTB transport to no longer enqueue data
1572  * @qp: NTB transport layer queue to be disabled
1573  *
1574  * Notify NTB transport layer of client's desire to no longer receive data on
1575  * transport queue specified.  It is the client's responsibility to ensure all
1576  * entries on queue are purged or otherwise handled appropriately.
1577  */
1578 static void
1579 ntb_transport_link_down(struct ntb_transport_qp *qp)
1580 {
1581 	uint32_t val;
1582 
1583 	if (qp == NULL)
1584 		return;
1585 
1586 	qp->client_ready = false;
1587 
1588 	ntb_spad_read(qp->ntb, IF_NTB_QP_LINKS, &val);
1589 
1590 	ntb_peer_spad_write(qp->ntb, IF_NTB_QP_LINKS,
1591 	   val & ~(1 << qp->qp_num));
1592 
1593 	if (qp->link_is_up)
1594 		ntb_send_link_down(qp);
1595 	else
1596 		callout_drain(&qp->link_work);
1597 }
1598 
1599 static void
1600 ntb_send_link_down(struct ntb_transport_qp *qp)
1601 {
1602 	struct ntb_queue_entry *entry;
1603 	int i, rc;
1604 
1605 	if (!qp->link_is_up)
1606 		return;
1607 
1608 	for (i = 0; i < NTB_LINK_DOWN_TIMEOUT; i++) {
1609 		entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q);
1610 		if (entry != NULL)
1611 			break;
1612 		pause("NTB Wait for link down", hz / 10);
1613 	}
1614 
1615 	if (entry == NULL)
1616 		return;
1617 
1618 	entry->cb_data = NULL;
1619 	entry->buf = NULL;
1620 	entry->len = 0;
1621 	entry->flags = IF_NTB_LINK_DOWN_FLAG;
1622 
1623 	mtx_lock(&qp->transport->tx_lock);
1624 	rc = ntb_process_tx(qp, entry);
1625 	if (rc != 0)
1626 		printf("ntb: Failed to send link down\n");
1627 	mtx_unlock(&qp->transport->tx_lock);
1628 
1629 	ntb_qp_link_down_reset(qp);
1630 }
1631 
1632 
1633 /* List Management */
1634 
1635 static void
1636 ntb_list_add(struct mtx *lock, struct ntb_queue_entry *entry,
1637     struct ntb_queue_list *list)
1638 {
1639 
1640 	mtx_lock_spin(lock);
1641 	STAILQ_INSERT_TAIL(list, entry, entry);
1642 	mtx_unlock_spin(lock);
1643 }
1644 
1645 static struct ntb_queue_entry *
1646 ntb_list_rm(struct mtx *lock, struct ntb_queue_list *list)
1647 {
1648 	struct ntb_queue_entry *entry;
1649 
1650 	mtx_lock_spin(lock);
1651 	if (STAILQ_EMPTY(list)) {
1652 		entry = NULL;
1653 		goto out;
1654 	}
1655 	entry = STAILQ_FIRST(list);
1656 	STAILQ_REMOVE_HEAD(list, entry);
1657 out:
1658 	mtx_unlock_spin(lock);
1659 
1660 	return (entry);
1661 }
1662 
1663 static struct ntb_queue_entry *
1664 ntb_list_mv(struct mtx *lock, struct ntb_queue_list *from,
1665     struct ntb_queue_list *to)
1666 {
1667 	struct ntb_queue_entry *entry;
1668 
1669 	mtx_lock_spin(lock);
1670 	if (STAILQ_EMPTY(from)) {
1671 		entry = NULL;
1672 		goto out;
1673 	}
1674 	entry = STAILQ_FIRST(from);
1675 	STAILQ_REMOVE_HEAD(from, entry);
1676 	STAILQ_INSERT_TAIL(to, entry, entry);
1677 
1678 out:
1679 	mtx_unlock_spin(lock);
1680 	return (entry);
1681 }
1682 
1683 /* Helper functions */
1684 /* TODO: This too should really be part of the kernel */
1685 #define EUI48_MULTICAST			1 << 0
1686 #define EUI48_LOCALLY_ADMINISTERED	1 << 1
1687 static void
1688 create_random_local_eui48(u_char *eaddr)
1689 {
1690 	static uint8_t counter = 0;
1691 	uint32_t seed = ticks;
1692 
1693 	eaddr[0] = EUI48_LOCALLY_ADMINISTERED;
1694 	memcpy(&eaddr[1], &seed, sizeof(uint32_t));
1695 	eaddr[5] = counter++;
1696 }
1697 
1698 /**
1699  * ntb_transport_max_size - Query the max payload size of a qp
1700  * @qp: NTB transport layer queue to be queried
1701  *
1702  * Query the maximum payload size permissible on the given qp
1703  *
1704  * RETURNS: the max payload size of a qp
1705  */
1706 static unsigned int
1707 ntb_transport_max_size(struct ntb_transport_qp *qp)
1708 {
1709 
1710 	if (qp == NULL)
1711 		return (0);
1712 
1713 	return (qp->tx_max_frame - sizeof(struct ntb_payload_header));
1714 }
1715