xref: /openbmc/linux/arch/sparc/kernel/ldc.c (revision 33ac9dba)
1 /* ldc.c: Logical Domain Channel link-layer protocol driver.
2  *
3  * Copyright (C) 2007, 2008 David S. Miller <davem@davemloft.net>
4  */
5 
6 #include <linux/kernel.h>
7 #include <linux/export.h>
8 #include <linux/slab.h>
9 #include <linux/spinlock.h>
10 #include <linux/delay.h>
11 #include <linux/errno.h>
12 #include <linux/string.h>
13 #include <linux/scatterlist.h>
14 #include <linux/interrupt.h>
15 #include <linux/list.h>
16 #include <linux/init.h>
17 #include <linux/bitmap.h>
18 
19 #include <asm/hypervisor.h>
20 #include <asm/iommu.h>
21 #include <asm/page.h>
22 #include <asm/ldc.h>
23 #include <asm/mdesc.h>
24 
25 #define DRV_MODULE_NAME		"ldc"
26 #define PFX DRV_MODULE_NAME	": "
27 #define DRV_MODULE_VERSION	"1.1"
28 #define DRV_MODULE_RELDATE	"July 22, 2008"
29 
30 static char version[] =
31 	DRV_MODULE_NAME ".c:v" DRV_MODULE_VERSION " (" DRV_MODULE_RELDATE ")\n";
32 #define LDC_PACKET_SIZE		64
33 
34 /* Packet header layout for unreliable and reliable mode frames.
35  * When in RAW mode, packets are simply straight 64-byte payloads
36  * with no headers.
37  */
38 struct ldc_packet {
39 	u8			type;
40 #define LDC_CTRL		0x01
41 #define LDC_DATA		0x02
42 #define LDC_ERR			0x10
43 
44 	u8			stype;
45 #define LDC_INFO		0x01
46 #define LDC_ACK			0x02
47 #define LDC_NACK		0x04
48 
49 	u8			ctrl;
50 #define LDC_VERS		0x01 /* Link Version		*/
51 #define LDC_RTS			0x02 /* Request To Send		*/
52 #define LDC_RTR			0x03 /* Ready To Receive	*/
53 #define LDC_RDX			0x04 /* Ready for Data eXchange	*/
54 #define LDC_CTRL_MSK		0x0f
55 
56 	u8			env;
57 #define LDC_LEN			0x3f
58 #define LDC_FRAG_MASK		0xc0
59 #define LDC_START		0x40
60 #define LDC_STOP		0x80
61 
62 	u32			seqid;
63 
64 	union {
65 		u8		u_data[LDC_PACKET_SIZE - 8];
66 		struct {
67 			u32	pad;
68 			u32	ackid;
69 			u8	r_data[LDC_PACKET_SIZE - 8 - 8];
70 		} r;
71 	} u;
72 };
73 
74 struct ldc_version {
75 	u16 major;
76 	u16 minor;
77 };
78 
79 /* Ordered from largest major to lowest.  */
80 static struct ldc_version ver_arr[] = {
81 	{ .major = 1, .minor = 0 },
82 };
83 
84 #define LDC_DEFAULT_MTU			(4 * LDC_PACKET_SIZE)
85 #define LDC_DEFAULT_NUM_ENTRIES		(PAGE_SIZE / LDC_PACKET_SIZE)
86 
87 struct ldc_channel;
88 
89 struct ldc_mode_ops {
90 	int (*write)(struct ldc_channel *, const void *, unsigned int);
91 	int (*read)(struct ldc_channel *, void *, unsigned int);
92 };
93 
94 static const struct ldc_mode_ops raw_ops;
95 static const struct ldc_mode_ops nonraw_ops;
96 static const struct ldc_mode_ops stream_ops;
97 
98 int ldom_domaining_enabled;
99 
100 struct ldc_iommu {
101 	/* Protects arena alloc/free.  */
102 	spinlock_t			lock;
103 	struct iommu_arena		arena;
104 	struct ldc_mtable_entry		*page_table;
105 };
106 
107 struct ldc_channel {
108 	/* Protects all operations that depend upon channel state.  */
109 	spinlock_t			lock;
110 
111 	unsigned long			id;
112 
113 	u8				*mssbuf;
114 	u32				mssbuf_len;
115 	u32				mssbuf_off;
116 
117 	struct ldc_packet		*tx_base;
118 	unsigned long			tx_head;
119 	unsigned long			tx_tail;
120 	unsigned long			tx_num_entries;
121 	unsigned long			tx_ra;
122 
123 	unsigned long			tx_acked;
124 
125 	struct ldc_packet		*rx_base;
126 	unsigned long			rx_head;
127 	unsigned long			rx_tail;
128 	unsigned long			rx_num_entries;
129 	unsigned long			rx_ra;
130 
131 	u32				rcv_nxt;
132 	u32				snd_nxt;
133 
134 	unsigned long			chan_state;
135 
136 	struct ldc_channel_config	cfg;
137 	void				*event_arg;
138 
139 	const struct ldc_mode_ops	*mops;
140 
141 	struct ldc_iommu		iommu;
142 
143 	struct ldc_version		ver;
144 
145 	u8				hs_state;
146 #define LDC_HS_CLOSED			0x00
147 #define LDC_HS_OPEN			0x01
148 #define LDC_HS_GOTVERS			0x02
149 #define LDC_HS_SENTRTR			0x03
150 #define LDC_HS_GOTRTR			0x04
151 #define LDC_HS_COMPLETE			0x10
152 
153 	u8				flags;
154 #define LDC_FLAG_ALLOCED_QUEUES		0x01
155 #define LDC_FLAG_REGISTERED_QUEUES	0x02
156 #define LDC_FLAG_REGISTERED_IRQS	0x04
157 #define LDC_FLAG_RESET			0x10
158 
159 	u8				mss;
160 	u8				state;
161 
162 #define LDC_IRQ_NAME_MAX		32
163 	char				rx_irq_name[LDC_IRQ_NAME_MAX];
164 	char				tx_irq_name[LDC_IRQ_NAME_MAX];
165 
166 	struct hlist_head		mh_list;
167 
168 	struct hlist_node		list;
169 };
170 
171 #define ldcdbg(TYPE, f, a...) \
172 do {	if (lp->cfg.debug & LDC_DEBUG_##TYPE) \
173 		printk(KERN_INFO PFX "ID[%lu] " f, lp->id, ## a); \
174 } while (0)
175 
176 static const char *state_to_str(u8 state)
177 {
178 	switch (state) {
179 	case LDC_STATE_INVALID:
180 		return "INVALID";
181 	case LDC_STATE_INIT:
182 		return "INIT";
183 	case LDC_STATE_BOUND:
184 		return "BOUND";
185 	case LDC_STATE_READY:
186 		return "READY";
187 	case LDC_STATE_CONNECTED:
188 		return "CONNECTED";
189 	default:
190 		return "<UNKNOWN>";
191 	}
192 }
193 
194 static void ldc_set_state(struct ldc_channel *lp, u8 state)
195 {
196 	ldcdbg(STATE, "STATE (%s) --> (%s)\n",
197 	       state_to_str(lp->state),
198 	       state_to_str(state));
199 
200 	lp->state = state;
201 }
202 
203 static unsigned long __advance(unsigned long off, unsigned long num_entries)
204 {
205 	off += LDC_PACKET_SIZE;
206 	if (off == (num_entries * LDC_PACKET_SIZE))
207 		off = 0;
208 
209 	return off;
210 }
211 
212 static unsigned long rx_advance(struct ldc_channel *lp, unsigned long off)
213 {
214 	return __advance(off, lp->rx_num_entries);
215 }
216 
217 static unsigned long tx_advance(struct ldc_channel *lp, unsigned long off)
218 {
219 	return __advance(off, lp->tx_num_entries);
220 }
221 
222 static struct ldc_packet *handshake_get_tx_packet(struct ldc_channel *lp,
223 						  unsigned long *new_tail)
224 {
225 	struct ldc_packet *p;
226 	unsigned long t;
227 
228 	t = tx_advance(lp, lp->tx_tail);
229 	if (t == lp->tx_head)
230 		return NULL;
231 
232 	*new_tail = t;
233 
234 	p = lp->tx_base;
235 	return p + (lp->tx_tail / LDC_PACKET_SIZE);
236 }
237 
238 /* When we are in reliable or stream mode, have to track the next packet
239  * we haven't gotten an ACK for in the TX queue using tx_acked.  We have
240  * to be careful not to stomp over the queue past that point.  During
241  * the handshake, we don't have TX data packets pending in the queue
242  * and that's why handshake_get_tx_packet() need not be mindful of
243  * lp->tx_acked.
244  */
245 static unsigned long head_for_data(struct ldc_channel *lp)
246 {
247 	if (lp->cfg.mode == LDC_MODE_STREAM)
248 		return lp->tx_acked;
249 	return lp->tx_head;
250 }
251 
252 static int tx_has_space_for(struct ldc_channel *lp, unsigned int size)
253 {
254 	unsigned long limit, tail, new_tail, diff;
255 	unsigned int mss;
256 
257 	limit = head_for_data(lp);
258 	tail = lp->tx_tail;
259 	new_tail = tx_advance(lp, tail);
260 	if (new_tail == limit)
261 		return 0;
262 
263 	if (limit > new_tail)
264 		diff = limit - new_tail;
265 	else
266 		diff = (limit +
267 			((lp->tx_num_entries * LDC_PACKET_SIZE) - new_tail));
268 	diff /= LDC_PACKET_SIZE;
269 	mss = lp->mss;
270 
271 	if (diff * mss < size)
272 		return 0;
273 
274 	return 1;
275 }
276 
277 static struct ldc_packet *data_get_tx_packet(struct ldc_channel *lp,
278 					     unsigned long *new_tail)
279 {
280 	struct ldc_packet *p;
281 	unsigned long h, t;
282 
283 	h = head_for_data(lp);
284 	t = tx_advance(lp, lp->tx_tail);
285 	if (t == h)
286 		return NULL;
287 
288 	*new_tail = t;
289 
290 	p = lp->tx_base;
291 	return p + (lp->tx_tail / LDC_PACKET_SIZE);
292 }
293 
294 static int set_tx_tail(struct ldc_channel *lp, unsigned long tail)
295 {
296 	unsigned long orig_tail = lp->tx_tail;
297 	int limit = 1000;
298 
299 	lp->tx_tail = tail;
300 	while (limit-- > 0) {
301 		unsigned long err;
302 
303 		err = sun4v_ldc_tx_set_qtail(lp->id, tail);
304 		if (!err)
305 			return 0;
306 
307 		if (err != HV_EWOULDBLOCK) {
308 			lp->tx_tail = orig_tail;
309 			return -EINVAL;
310 		}
311 		udelay(1);
312 	}
313 
314 	lp->tx_tail = orig_tail;
315 	return -EBUSY;
316 }
317 
318 /* This just updates the head value in the hypervisor using
319  * a polling loop with a timeout.  The caller takes care of
320  * upating software state representing the head change, if any.
321  */
322 static int __set_rx_head(struct ldc_channel *lp, unsigned long head)
323 {
324 	int limit = 1000;
325 
326 	while (limit-- > 0) {
327 		unsigned long err;
328 
329 		err = sun4v_ldc_rx_set_qhead(lp->id, head);
330 		if (!err)
331 			return 0;
332 
333 		if (err != HV_EWOULDBLOCK)
334 			return -EINVAL;
335 
336 		udelay(1);
337 	}
338 
339 	return -EBUSY;
340 }
341 
342 static int send_tx_packet(struct ldc_channel *lp,
343 			  struct ldc_packet *p,
344 			  unsigned long new_tail)
345 {
346 	BUG_ON(p != (lp->tx_base + (lp->tx_tail / LDC_PACKET_SIZE)));
347 
348 	return set_tx_tail(lp, new_tail);
349 }
350 
351 static struct ldc_packet *handshake_compose_ctrl(struct ldc_channel *lp,
352 						 u8 stype, u8 ctrl,
353 						 void *data, int dlen,
354 						 unsigned long *new_tail)
355 {
356 	struct ldc_packet *p = handshake_get_tx_packet(lp, new_tail);
357 
358 	if (p) {
359 		memset(p, 0, sizeof(*p));
360 		p->type = LDC_CTRL;
361 		p->stype = stype;
362 		p->ctrl = ctrl;
363 		if (data)
364 			memcpy(p->u.u_data, data, dlen);
365 	}
366 	return p;
367 }
368 
369 static int start_handshake(struct ldc_channel *lp)
370 {
371 	struct ldc_packet *p;
372 	struct ldc_version *ver;
373 	unsigned long new_tail;
374 
375 	ver = &ver_arr[0];
376 
377 	ldcdbg(HS, "SEND VER INFO maj[%u] min[%u]\n",
378 	       ver->major, ver->minor);
379 
380 	p = handshake_compose_ctrl(lp, LDC_INFO, LDC_VERS,
381 				   ver, sizeof(*ver), &new_tail);
382 	if (p) {
383 		int err = send_tx_packet(lp, p, new_tail);
384 		if (!err)
385 			lp->flags &= ~LDC_FLAG_RESET;
386 		return err;
387 	}
388 	return -EBUSY;
389 }
390 
391 static int send_version_nack(struct ldc_channel *lp,
392 			     u16 major, u16 minor)
393 {
394 	struct ldc_packet *p;
395 	struct ldc_version ver;
396 	unsigned long new_tail;
397 
398 	ver.major = major;
399 	ver.minor = minor;
400 
401 	p = handshake_compose_ctrl(lp, LDC_NACK, LDC_VERS,
402 				   &ver, sizeof(ver), &new_tail);
403 	if (p) {
404 		ldcdbg(HS, "SEND VER NACK maj[%u] min[%u]\n",
405 		       ver.major, ver.minor);
406 
407 		return send_tx_packet(lp, p, new_tail);
408 	}
409 	return -EBUSY;
410 }
411 
412 static int send_version_ack(struct ldc_channel *lp,
413 			    struct ldc_version *vp)
414 {
415 	struct ldc_packet *p;
416 	unsigned long new_tail;
417 
418 	p = handshake_compose_ctrl(lp, LDC_ACK, LDC_VERS,
419 				   vp, sizeof(*vp), &new_tail);
420 	if (p) {
421 		ldcdbg(HS, "SEND VER ACK maj[%u] min[%u]\n",
422 		       vp->major, vp->minor);
423 
424 		return send_tx_packet(lp, p, new_tail);
425 	}
426 	return -EBUSY;
427 }
428 
429 static int send_rts(struct ldc_channel *lp)
430 {
431 	struct ldc_packet *p;
432 	unsigned long new_tail;
433 
434 	p = handshake_compose_ctrl(lp, LDC_INFO, LDC_RTS, NULL, 0,
435 				   &new_tail);
436 	if (p) {
437 		p->env = lp->cfg.mode;
438 		p->seqid = 0;
439 		lp->rcv_nxt = 0;
440 
441 		ldcdbg(HS, "SEND RTS env[0x%x] seqid[0x%x]\n",
442 		       p->env, p->seqid);
443 
444 		return send_tx_packet(lp, p, new_tail);
445 	}
446 	return -EBUSY;
447 }
448 
449 static int send_rtr(struct ldc_channel *lp)
450 {
451 	struct ldc_packet *p;
452 	unsigned long new_tail;
453 
454 	p = handshake_compose_ctrl(lp, LDC_INFO, LDC_RTR, NULL, 0,
455 				   &new_tail);
456 	if (p) {
457 		p->env = lp->cfg.mode;
458 		p->seqid = 0;
459 
460 		ldcdbg(HS, "SEND RTR env[0x%x] seqid[0x%x]\n",
461 		       p->env, p->seqid);
462 
463 		return send_tx_packet(lp, p, new_tail);
464 	}
465 	return -EBUSY;
466 }
467 
468 static int send_rdx(struct ldc_channel *lp)
469 {
470 	struct ldc_packet *p;
471 	unsigned long new_tail;
472 
473 	p = handshake_compose_ctrl(lp, LDC_INFO, LDC_RDX, NULL, 0,
474 				   &new_tail);
475 	if (p) {
476 		p->env = 0;
477 		p->seqid = ++lp->snd_nxt;
478 		p->u.r.ackid = lp->rcv_nxt;
479 
480 		ldcdbg(HS, "SEND RDX env[0x%x] seqid[0x%x] ackid[0x%x]\n",
481 		       p->env, p->seqid, p->u.r.ackid);
482 
483 		return send_tx_packet(lp, p, new_tail);
484 	}
485 	return -EBUSY;
486 }
487 
488 static int send_data_nack(struct ldc_channel *lp, struct ldc_packet *data_pkt)
489 {
490 	struct ldc_packet *p;
491 	unsigned long new_tail;
492 	int err;
493 
494 	p = data_get_tx_packet(lp, &new_tail);
495 	if (!p)
496 		return -EBUSY;
497 	memset(p, 0, sizeof(*p));
498 	p->type = data_pkt->type;
499 	p->stype = LDC_NACK;
500 	p->ctrl = data_pkt->ctrl & LDC_CTRL_MSK;
501 	p->seqid = lp->snd_nxt + 1;
502 	p->u.r.ackid = lp->rcv_nxt;
503 
504 	ldcdbg(HS, "SEND DATA NACK type[0x%x] ctl[0x%x] seq[0x%x] ack[0x%x]\n",
505 	       p->type, p->ctrl, p->seqid, p->u.r.ackid);
506 
507 	err = send_tx_packet(lp, p, new_tail);
508 	if (!err)
509 		lp->snd_nxt++;
510 
511 	return err;
512 }
513 
514 static int ldc_abort(struct ldc_channel *lp)
515 {
516 	unsigned long hv_err;
517 
518 	ldcdbg(STATE, "ABORT\n");
519 
520 	/* We report but do not act upon the hypervisor errors because
521 	 * there really isn't much we can do if they fail at this point.
522 	 */
523 	hv_err = sun4v_ldc_tx_qconf(lp->id, lp->tx_ra, lp->tx_num_entries);
524 	if (hv_err)
525 		printk(KERN_ERR PFX "ldc_abort: "
526 		       "sun4v_ldc_tx_qconf(%lx,%lx,%lx) failed, err=%lu\n",
527 		       lp->id, lp->tx_ra, lp->tx_num_entries, hv_err);
528 
529 	hv_err = sun4v_ldc_tx_get_state(lp->id,
530 					&lp->tx_head,
531 					&lp->tx_tail,
532 					&lp->chan_state);
533 	if (hv_err)
534 		printk(KERN_ERR PFX "ldc_abort: "
535 		       "sun4v_ldc_tx_get_state(%lx,...) failed, err=%lu\n",
536 		       lp->id, hv_err);
537 
538 	hv_err = sun4v_ldc_rx_qconf(lp->id, lp->rx_ra, lp->rx_num_entries);
539 	if (hv_err)
540 		printk(KERN_ERR PFX "ldc_abort: "
541 		       "sun4v_ldc_rx_qconf(%lx,%lx,%lx) failed, err=%lu\n",
542 		       lp->id, lp->rx_ra, lp->rx_num_entries, hv_err);
543 
544 	/* Refetch the RX queue state as well, because we could be invoked
545 	 * here in the queue processing context.
546 	 */
547 	hv_err = sun4v_ldc_rx_get_state(lp->id,
548 					&lp->rx_head,
549 					&lp->rx_tail,
550 					&lp->chan_state);
551 	if (hv_err)
552 		printk(KERN_ERR PFX "ldc_abort: "
553 		       "sun4v_ldc_rx_get_state(%lx,...) failed, err=%lu\n",
554 		       lp->id, hv_err);
555 
556 	return -ECONNRESET;
557 }
558 
559 static struct ldc_version *find_by_major(u16 major)
560 {
561 	struct ldc_version *ret = NULL;
562 	int i;
563 
564 	for (i = 0; i < ARRAY_SIZE(ver_arr); i++) {
565 		struct ldc_version *v = &ver_arr[i];
566 		if (v->major <= major) {
567 			ret = v;
568 			break;
569 		}
570 	}
571 	return ret;
572 }
573 
574 static int process_ver_info(struct ldc_channel *lp, struct ldc_version *vp)
575 {
576 	struct ldc_version *vap;
577 	int err;
578 
579 	ldcdbg(HS, "GOT VERSION INFO major[%x] minor[%x]\n",
580 	       vp->major, vp->minor);
581 
582 	if (lp->hs_state == LDC_HS_GOTVERS) {
583 		lp->hs_state = LDC_HS_OPEN;
584 		memset(&lp->ver, 0, sizeof(lp->ver));
585 	}
586 
587 	vap = find_by_major(vp->major);
588 	if (!vap) {
589 		err = send_version_nack(lp, 0, 0);
590 	} else if (vap->major != vp->major) {
591 		err = send_version_nack(lp, vap->major, vap->minor);
592 	} else {
593 		struct ldc_version ver = *vp;
594 		if (ver.minor > vap->minor)
595 			ver.minor = vap->minor;
596 		err = send_version_ack(lp, &ver);
597 		if (!err) {
598 			lp->ver = ver;
599 			lp->hs_state = LDC_HS_GOTVERS;
600 		}
601 	}
602 	if (err)
603 		return ldc_abort(lp);
604 
605 	return 0;
606 }
607 
608 static int process_ver_ack(struct ldc_channel *lp, struct ldc_version *vp)
609 {
610 	ldcdbg(HS, "GOT VERSION ACK major[%x] minor[%x]\n",
611 	       vp->major, vp->minor);
612 
613 	if (lp->hs_state == LDC_HS_GOTVERS) {
614 		if (lp->ver.major != vp->major ||
615 		    lp->ver.minor != vp->minor)
616 			return ldc_abort(lp);
617 	} else {
618 		lp->ver = *vp;
619 		lp->hs_state = LDC_HS_GOTVERS;
620 	}
621 	if (send_rts(lp))
622 		return ldc_abort(lp);
623 	return 0;
624 }
625 
626 static int process_ver_nack(struct ldc_channel *lp, struct ldc_version *vp)
627 {
628 	struct ldc_version *vap;
629 	struct ldc_packet *p;
630 	unsigned long new_tail;
631 
632 	if (vp->major == 0 && vp->minor == 0)
633 		return ldc_abort(lp);
634 
635 	vap = find_by_major(vp->major);
636 	if (!vap)
637 		return ldc_abort(lp);
638 
639 	p = handshake_compose_ctrl(lp, LDC_INFO, LDC_VERS,
640 					   vap, sizeof(*vap),
641 					   &new_tail);
642 	if (!p)
643 		return ldc_abort(lp);
644 
645 	return send_tx_packet(lp, p, new_tail);
646 }
647 
648 static int process_version(struct ldc_channel *lp,
649 			   struct ldc_packet *p)
650 {
651 	struct ldc_version *vp;
652 
653 	vp = (struct ldc_version *) p->u.u_data;
654 
655 	switch (p->stype) {
656 	case LDC_INFO:
657 		return process_ver_info(lp, vp);
658 
659 	case LDC_ACK:
660 		return process_ver_ack(lp, vp);
661 
662 	case LDC_NACK:
663 		return process_ver_nack(lp, vp);
664 
665 	default:
666 		return ldc_abort(lp);
667 	}
668 }
669 
670 static int process_rts(struct ldc_channel *lp,
671 		       struct ldc_packet *p)
672 {
673 	ldcdbg(HS, "GOT RTS stype[%x] seqid[%x] env[%x]\n",
674 	       p->stype, p->seqid, p->env);
675 
676 	if (p->stype     != LDC_INFO	   ||
677 	    lp->hs_state != LDC_HS_GOTVERS ||
678 	    p->env       != lp->cfg.mode)
679 		return ldc_abort(lp);
680 
681 	lp->snd_nxt = p->seqid;
682 	lp->rcv_nxt = p->seqid;
683 	lp->hs_state = LDC_HS_SENTRTR;
684 	if (send_rtr(lp))
685 		return ldc_abort(lp);
686 
687 	return 0;
688 }
689 
690 static int process_rtr(struct ldc_channel *lp,
691 		       struct ldc_packet *p)
692 {
693 	ldcdbg(HS, "GOT RTR stype[%x] seqid[%x] env[%x]\n",
694 	       p->stype, p->seqid, p->env);
695 
696 	if (p->stype     != LDC_INFO ||
697 	    p->env       != lp->cfg.mode)
698 		return ldc_abort(lp);
699 
700 	lp->snd_nxt = p->seqid;
701 	lp->hs_state = LDC_HS_COMPLETE;
702 	ldc_set_state(lp, LDC_STATE_CONNECTED);
703 	send_rdx(lp);
704 
705 	return LDC_EVENT_UP;
706 }
707 
708 static int rx_seq_ok(struct ldc_channel *lp, u32 seqid)
709 {
710 	return lp->rcv_nxt + 1 == seqid;
711 }
712 
713 static int process_rdx(struct ldc_channel *lp,
714 		       struct ldc_packet *p)
715 {
716 	ldcdbg(HS, "GOT RDX stype[%x] seqid[%x] env[%x] ackid[%x]\n",
717 	       p->stype, p->seqid, p->env, p->u.r.ackid);
718 
719 	if (p->stype != LDC_INFO ||
720 	    !(rx_seq_ok(lp, p->seqid)))
721 		return ldc_abort(lp);
722 
723 	lp->rcv_nxt = p->seqid;
724 
725 	lp->hs_state = LDC_HS_COMPLETE;
726 	ldc_set_state(lp, LDC_STATE_CONNECTED);
727 
728 	return LDC_EVENT_UP;
729 }
730 
731 static int process_control_frame(struct ldc_channel *lp,
732 				 struct ldc_packet *p)
733 {
734 	switch (p->ctrl) {
735 	case LDC_VERS:
736 		return process_version(lp, p);
737 
738 	case LDC_RTS:
739 		return process_rts(lp, p);
740 
741 	case LDC_RTR:
742 		return process_rtr(lp, p);
743 
744 	case LDC_RDX:
745 		return process_rdx(lp, p);
746 
747 	default:
748 		return ldc_abort(lp);
749 	}
750 }
751 
752 static int process_error_frame(struct ldc_channel *lp,
753 			       struct ldc_packet *p)
754 {
755 	return ldc_abort(lp);
756 }
757 
758 static int process_data_ack(struct ldc_channel *lp,
759 			    struct ldc_packet *ack)
760 {
761 	unsigned long head = lp->tx_acked;
762 	u32 ackid = ack->u.r.ackid;
763 
764 	while (1) {
765 		struct ldc_packet *p = lp->tx_base + (head / LDC_PACKET_SIZE);
766 
767 		head = tx_advance(lp, head);
768 
769 		if (p->seqid == ackid) {
770 			lp->tx_acked = head;
771 			return 0;
772 		}
773 		if (head == lp->tx_tail)
774 			return ldc_abort(lp);
775 	}
776 
777 	return 0;
778 }
779 
780 static void send_events(struct ldc_channel *lp, unsigned int event_mask)
781 {
782 	if (event_mask & LDC_EVENT_RESET)
783 		lp->cfg.event(lp->event_arg, LDC_EVENT_RESET);
784 	if (event_mask & LDC_EVENT_UP)
785 		lp->cfg.event(lp->event_arg, LDC_EVENT_UP);
786 	if (event_mask & LDC_EVENT_DATA_READY)
787 		lp->cfg.event(lp->event_arg, LDC_EVENT_DATA_READY);
788 }
789 
790 static irqreturn_t ldc_rx(int irq, void *dev_id)
791 {
792 	struct ldc_channel *lp = dev_id;
793 	unsigned long orig_state, flags;
794 	unsigned int event_mask;
795 
796 	spin_lock_irqsave(&lp->lock, flags);
797 
798 	orig_state = lp->chan_state;
799 
800 	/* We should probably check for hypervisor errors here and
801 	 * reset the LDC channel if we get one.
802 	 */
803 	sun4v_ldc_rx_get_state(lp->id,
804 			       &lp->rx_head,
805 			       &lp->rx_tail,
806 			       &lp->chan_state);
807 
808 	ldcdbg(RX, "RX state[0x%02lx:0x%02lx] head[0x%04lx] tail[0x%04lx]\n",
809 	       orig_state, lp->chan_state, lp->rx_head, lp->rx_tail);
810 
811 	event_mask = 0;
812 
813 	if (lp->cfg.mode == LDC_MODE_RAW &&
814 	    lp->chan_state == LDC_CHANNEL_UP) {
815 		lp->hs_state = LDC_HS_COMPLETE;
816 		ldc_set_state(lp, LDC_STATE_CONNECTED);
817 
818 		event_mask |= LDC_EVENT_UP;
819 
820 		orig_state = lp->chan_state;
821 	}
822 
823 	/* If we are in reset state, flush the RX queue and ignore
824 	 * everything.
825 	 */
826 	if (lp->flags & LDC_FLAG_RESET) {
827 		(void) __set_rx_head(lp, lp->rx_tail);
828 		goto out;
829 	}
830 
831 	/* Once we finish the handshake, we let the ldc_read()
832 	 * paths do all of the control frame and state management.
833 	 * Just trigger the callback.
834 	 */
835 	if (lp->hs_state == LDC_HS_COMPLETE) {
836 handshake_complete:
837 		if (lp->chan_state != orig_state) {
838 			unsigned int event = LDC_EVENT_RESET;
839 
840 			if (lp->chan_state == LDC_CHANNEL_UP)
841 				event = LDC_EVENT_UP;
842 
843 			event_mask |= event;
844 		}
845 		if (lp->rx_head != lp->rx_tail)
846 			event_mask |= LDC_EVENT_DATA_READY;
847 
848 		goto out;
849 	}
850 
851 	if (lp->chan_state != orig_state)
852 		goto out;
853 
854 	while (lp->rx_head != lp->rx_tail) {
855 		struct ldc_packet *p;
856 		unsigned long new;
857 		int err;
858 
859 		p = lp->rx_base + (lp->rx_head / LDC_PACKET_SIZE);
860 
861 		switch (p->type) {
862 		case LDC_CTRL:
863 			err = process_control_frame(lp, p);
864 			if (err > 0)
865 				event_mask |= err;
866 			break;
867 
868 		case LDC_DATA:
869 			event_mask |= LDC_EVENT_DATA_READY;
870 			err = 0;
871 			break;
872 
873 		case LDC_ERR:
874 			err = process_error_frame(lp, p);
875 			break;
876 
877 		default:
878 			err = ldc_abort(lp);
879 			break;
880 		}
881 
882 		if (err < 0)
883 			break;
884 
885 		new = lp->rx_head;
886 		new += LDC_PACKET_SIZE;
887 		if (new == (lp->rx_num_entries * LDC_PACKET_SIZE))
888 			new = 0;
889 		lp->rx_head = new;
890 
891 		err = __set_rx_head(lp, new);
892 		if (err < 0) {
893 			(void) ldc_abort(lp);
894 			break;
895 		}
896 		if (lp->hs_state == LDC_HS_COMPLETE)
897 			goto handshake_complete;
898 	}
899 
900 out:
901 	spin_unlock_irqrestore(&lp->lock, flags);
902 
903 	send_events(lp, event_mask);
904 
905 	return IRQ_HANDLED;
906 }
907 
908 static irqreturn_t ldc_tx(int irq, void *dev_id)
909 {
910 	struct ldc_channel *lp = dev_id;
911 	unsigned long flags, orig_state;
912 	unsigned int event_mask = 0;
913 
914 	spin_lock_irqsave(&lp->lock, flags);
915 
916 	orig_state = lp->chan_state;
917 
918 	/* We should probably check for hypervisor errors here and
919 	 * reset the LDC channel if we get one.
920 	 */
921 	sun4v_ldc_tx_get_state(lp->id,
922 			       &lp->tx_head,
923 			       &lp->tx_tail,
924 			       &lp->chan_state);
925 
926 	ldcdbg(TX, " TX state[0x%02lx:0x%02lx] head[0x%04lx] tail[0x%04lx]\n",
927 	       orig_state, lp->chan_state, lp->tx_head, lp->tx_tail);
928 
929 	if (lp->cfg.mode == LDC_MODE_RAW &&
930 	    lp->chan_state == LDC_CHANNEL_UP) {
931 		lp->hs_state = LDC_HS_COMPLETE;
932 		ldc_set_state(lp, LDC_STATE_CONNECTED);
933 
934 		event_mask |= LDC_EVENT_UP;
935 	}
936 
937 	spin_unlock_irqrestore(&lp->lock, flags);
938 
939 	send_events(lp, event_mask);
940 
941 	return IRQ_HANDLED;
942 }
943 
944 /* XXX ldc_alloc() and ldc_free() needs to run under a mutex so
945  * XXX that addition and removal from the ldc_channel_list has
946  * XXX atomicity, otherwise the __ldc_channel_exists() check is
947  * XXX totally pointless as another thread can slip into ldc_alloc()
948  * XXX and add a channel with the same ID.  There also needs to be
949  * XXX a spinlock for ldc_channel_list.
950  */
951 static HLIST_HEAD(ldc_channel_list);
952 
953 static int __ldc_channel_exists(unsigned long id)
954 {
955 	struct ldc_channel *lp;
956 
957 	hlist_for_each_entry(lp, &ldc_channel_list, list) {
958 		if (lp->id == id)
959 			return 1;
960 	}
961 	return 0;
962 }
963 
964 static int alloc_queue(const char *name, unsigned long num_entries,
965 		       struct ldc_packet **base, unsigned long *ra)
966 {
967 	unsigned long size, order;
968 	void *q;
969 
970 	size = num_entries * LDC_PACKET_SIZE;
971 	order = get_order(size);
972 
973 	q = (void *) __get_free_pages(GFP_KERNEL, order);
974 	if (!q) {
975 		printk(KERN_ERR PFX "Alloc of %s queue failed with "
976 		       "size=%lu order=%lu\n", name, size, order);
977 		return -ENOMEM;
978 	}
979 
980 	memset(q, 0, PAGE_SIZE << order);
981 
982 	*base = q;
983 	*ra = __pa(q);
984 
985 	return 0;
986 }
987 
988 static void free_queue(unsigned long num_entries, struct ldc_packet *q)
989 {
990 	unsigned long size, order;
991 
992 	if (!q)
993 		return;
994 
995 	size = num_entries * LDC_PACKET_SIZE;
996 	order = get_order(size);
997 
998 	free_pages((unsigned long)q, order);
999 }
1000 
1001 /* XXX Make this configurable... XXX */
1002 #define LDC_IOTABLE_SIZE	(8 * 1024)
1003 
1004 static int ldc_iommu_init(struct ldc_channel *lp)
1005 {
1006 	unsigned long sz, num_tsb_entries, tsbsize, order;
1007 	struct ldc_iommu *iommu = &lp->iommu;
1008 	struct ldc_mtable_entry *table;
1009 	unsigned long hv_err;
1010 	int err;
1011 
1012 	num_tsb_entries = LDC_IOTABLE_SIZE;
1013 	tsbsize = num_tsb_entries * sizeof(struct ldc_mtable_entry);
1014 
1015 	spin_lock_init(&iommu->lock);
1016 
1017 	sz = num_tsb_entries / 8;
1018 	sz = (sz + 7UL) & ~7UL;
1019 	iommu->arena.map = kzalloc(sz, GFP_KERNEL);
1020 	if (!iommu->arena.map) {
1021 		printk(KERN_ERR PFX "Alloc of arena map failed, sz=%lu\n", sz);
1022 		return -ENOMEM;
1023 	}
1024 
1025 	iommu->arena.limit = num_tsb_entries;
1026 
1027 	order = get_order(tsbsize);
1028 
1029 	table = (struct ldc_mtable_entry *)
1030 		__get_free_pages(GFP_KERNEL, order);
1031 	err = -ENOMEM;
1032 	if (!table) {
1033 		printk(KERN_ERR PFX "Alloc of MTE table failed, "
1034 		       "size=%lu order=%lu\n", tsbsize, order);
1035 		goto out_free_map;
1036 	}
1037 
1038 	memset(table, 0, PAGE_SIZE << order);
1039 
1040 	iommu->page_table = table;
1041 
1042 	hv_err = sun4v_ldc_set_map_table(lp->id, __pa(table),
1043 					 num_tsb_entries);
1044 	err = -EINVAL;
1045 	if (hv_err)
1046 		goto out_free_table;
1047 
1048 	return 0;
1049 
1050 out_free_table:
1051 	free_pages((unsigned long) table, order);
1052 	iommu->page_table = NULL;
1053 
1054 out_free_map:
1055 	kfree(iommu->arena.map);
1056 	iommu->arena.map = NULL;
1057 
1058 	return err;
1059 }
1060 
1061 static void ldc_iommu_release(struct ldc_channel *lp)
1062 {
1063 	struct ldc_iommu *iommu = &lp->iommu;
1064 	unsigned long num_tsb_entries, tsbsize, order;
1065 
1066 	(void) sun4v_ldc_set_map_table(lp->id, 0, 0);
1067 
1068 	num_tsb_entries = iommu->arena.limit;
1069 	tsbsize = num_tsb_entries * sizeof(struct ldc_mtable_entry);
1070 	order = get_order(tsbsize);
1071 
1072 	free_pages((unsigned long) iommu->page_table, order);
1073 	iommu->page_table = NULL;
1074 
1075 	kfree(iommu->arena.map);
1076 	iommu->arena.map = NULL;
1077 }
1078 
1079 struct ldc_channel *ldc_alloc(unsigned long id,
1080 			      const struct ldc_channel_config *cfgp,
1081 			      void *event_arg)
1082 {
1083 	struct ldc_channel *lp;
1084 	const struct ldc_mode_ops *mops;
1085 	unsigned long dummy1, dummy2, hv_err;
1086 	u8 mss, *mssbuf;
1087 	int err;
1088 
1089 	err = -ENODEV;
1090 	if (!ldom_domaining_enabled)
1091 		goto out_err;
1092 
1093 	err = -EINVAL;
1094 	if (!cfgp)
1095 		goto out_err;
1096 
1097 	switch (cfgp->mode) {
1098 	case LDC_MODE_RAW:
1099 		mops = &raw_ops;
1100 		mss = LDC_PACKET_SIZE;
1101 		break;
1102 
1103 	case LDC_MODE_UNRELIABLE:
1104 		mops = &nonraw_ops;
1105 		mss = LDC_PACKET_SIZE - 8;
1106 		break;
1107 
1108 	case LDC_MODE_STREAM:
1109 		mops = &stream_ops;
1110 		mss = LDC_PACKET_SIZE - 8 - 8;
1111 		break;
1112 
1113 	default:
1114 		goto out_err;
1115 	}
1116 
1117 	if (!cfgp->event || !event_arg || !cfgp->rx_irq || !cfgp->tx_irq)
1118 		goto out_err;
1119 
1120 	hv_err = sun4v_ldc_tx_qinfo(id, &dummy1, &dummy2);
1121 	err = -ENODEV;
1122 	if (hv_err == HV_ECHANNEL)
1123 		goto out_err;
1124 
1125 	err = -EEXIST;
1126 	if (__ldc_channel_exists(id))
1127 		goto out_err;
1128 
1129 	mssbuf = NULL;
1130 
1131 	lp = kzalloc(sizeof(*lp), GFP_KERNEL);
1132 	err = -ENOMEM;
1133 	if (!lp)
1134 		goto out_err;
1135 
1136 	spin_lock_init(&lp->lock);
1137 
1138 	lp->id = id;
1139 
1140 	err = ldc_iommu_init(lp);
1141 	if (err)
1142 		goto out_free_ldc;
1143 
1144 	lp->mops = mops;
1145 	lp->mss = mss;
1146 
1147 	lp->cfg = *cfgp;
1148 	if (!lp->cfg.mtu)
1149 		lp->cfg.mtu = LDC_DEFAULT_MTU;
1150 
1151 	if (lp->cfg.mode == LDC_MODE_STREAM) {
1152 		mssbuf = kzalloc(lp->cfg.mtu, GFP_KERNEL);
1153 		if (!mssbuf) {
1154 			err = -ENOMEM;
1155 			goto out_free_iommu;
1156 		}
1157 		lp->mssbuf = mssbuf;
1158 	}
1159 
1160 	lp->event_arg = event_arg;
1161 
1162 	/* XXX allow setting via ldc_channel_config to override defaults
1163 	 * XXX or use some formula based upon mtu
1164 	 */
1165 	lp->tx_num_entries = LDC_DEFAULT_NUM_ENTRIES;
1166 	lp->rx_num_entries = LDC_DEFAULT_NUM_ENTRIES;
1167 
1168 	err = alloc_queue("TX", lp->tx_num_entries,
1169 			  &lp->tx_base, &lp->tx_ra);
1170 	if (err)
1171 		goto out_free_mssbuf;
1172 
1173 	err = alloc_queue("RX", lp->rx_num_entries,
1174 			  &lp->rx_base, &lp->rx_ra);
1175 	if (err)
1176 		goto out_free_txq;
1177 
1178 	lp->flags |= LDC_FLAG_ALLOCED_QUEUES;
1179 
1180 	lp->hs_state = LDC_HS_CLOSED;
1181 	ldc_set_state(lp, LDC_STATE_INIT);
1182 
1183 	INIT_HLIST_NODE(&lp->list);
1184 	hlist_add_head(&lp->list, &ldc_channel_list);
1185 
1186 	INIT_HLIST_HEAD(&lp->mh_list);
1187 
1188 	return lp;
1189 
1190 out_free_txq:
1191 	free_queue(lp->tx_num_entries, lp->tx_base);
1192 
1193 out_free_mssbuf:
1194 	kfree(mssbuf);
1195 
1196 out_free_iommu:
1197 	ldc_iommu_release(lp);
1198 
1199 out_free_ldc:
1200 	kfree(lp);
1201 
1202 out_err:
1203 	return ERR_PTR(err);
1204 }
1205 EXPORT_SYMBOL(ldc_alloc);
1206 
1207 void ldc_free(struct ldc_channel *lp)
1208 {
1209 	if (lp->flags & LDC_FLAG_REGISTERED_IRQS) {
1210 		free_irq(lp->cfg.rx_irq, lp);
1211 		free_irq(lp->cfg.tx_irq, lp);
1212 	}
1213 
1214 	if (lp->flags & LDC_FLAG_REGISTERED_QUEUES) {
1215 		sun4v_ldc_tx_qconf(lp->id, 0, 0);
1216 		sun4v_ldc_rx_qconf(lp->id, 0, 0);
1217 		lp->flags &= ~LDC_FLAG_REGISTERED_QUEUES;
1218 	}
1219 	if (lp->flags & LDC_FLAG_ALLOCED_QUEUES) {
1220 		free_queue(lp->tx_num_entries, lp->tx_base);
1221 		free_queue(lp->rx_num_entries, lp->rx_base);
1222 		lp->flags &= ~LDC_FLAG_ALLOCED_QUEUES;
1223 	}
1224 
1225 	hlist_del(&lp->list);
1226 
1227 	kfree(lp->mssbuf);
1228 
1229 	ldc_iommu_release(lp);
1230 
1231 	kfree(lp);
1232 }
1233 EXPORT_SYMBOL(ldc_free);
1234 
1235 /* Bind the channel.  This registers the LDC queues with
1236  * the hypervisor and puts the channel into a pseudo-listening
1237  * state.  This does not initiate a handshake, ldc_connect() does
1238  * that.
1239  */
1240 int ldc_bind(struct ldc_channel *lp, const char *name)
1241 {
1242 	unsigned long hv_err, flags;
1243 	int err = -EINVAL;
1244 
1245 	if (!name ||
1246 	    (lp->state != LDC_STATE_INIT))
1247 		return -EINVAL;
1248 
1249 	snprintf(lp->rx_irq_name, LDC_IRQ_NAME_MAX, "%s RX", name);
1250 	snprintf(lp->tx_irq_name, LDC_IRQ_NAME_MAX, "%s TX", name);
1251 
1252 	err = request_irq(lp->cfg.rx_irq, ldc_rx, 0,
1253 			  lp->rx_irq_name, lp);
1254 	if (err)
1255 		return err;
1256 
1257 	err = request_irq(lp->cfg.tx_irq, ldc_tx, 0,
1258 			  lp->tx_irq_name, lp);
1259 	if (err) {
1260 		free_irq(lp->cfg.rx_irq, lp);
1261 		return err;
1262 	}
1263 
1264 
1265 	spin_lock_irqsave(&lp->lock, flags);
1266 
1267 	enable_irq(lp->cfg.rx_irq);
1268 	enable_irq(lp->cfg.tx_irq);
1269 
1270 	lp->flags |= LDC_FLAG_REGISTERED_IRQS;
1271 
1272 	err = -ENODEV;
1273 	hv_err = sun4v_ldc_tx_qconf(lp->id, 0, 0);
1274 	if (hv_err)
1275 		goto out_free_irqs;
1276 
1277 	hv_err = sun4v_ldc_tx_qconf(lp->id, lp->tx_ra, lp->tx_num_entries);
1278 	if (hv_err)
1279 		goto out_free_irqs;
1280 
1281 	hv_err = sun4v_ldc_rx_qconf(lp->id, 0, 0);
1282 	if (hv_err)
1283 		goto out_unmap_tx;
1284 
1285 	hv_err = sun4v_ldc_rx_qconf(lp->id, lp->rx_ra, lp->rx_num_entries);
1286 	if (hv_err)
1287 		goto out_unmap_tx;
1288 
1289 	lp->flags |= LDC_FLAG_REGISTERED_QUEUES;
1290 
1291 	hv_err = sun4v_ldc_tx_get_state(lp->id,
1292 					&lp->tx_head,
1293 					&lp->tx_tail,
1294 					&lp->chan_state);
1295 	err = -EBUSY;
1296 	if (hv_err)
1297 		goto out_unmap_rx;
1298 
1299 	lp->tx_acked = lp->tx_head;
1300 
1301 	lp->hs_state = LDC_HS_OPEN;
1302 	ldc_set_state(lp, LDC_STATE_BOUND);
1303 
1304 	spin_unlock_irqrestore(&lp->lock, flags);
1305 
1306 	return 0;
1307 
1308 out_unmap_rx:
1309 	lp->flags &= ~LDC_FLAG_REGISTERED_QUEUES;
1310 	sun4v_ldc_rx_qconf(lp->id, 0, 0);
1311 
1312 out_unmap_tx:
1313 	sun4v_ldc_tx_qconf(lp->id, 0, 0);
1314 
1315 out_free_irqs:
1316 	lp->flags &= ~LDC_FLAG_REGISTERED_IRQS;
1317 	free_irq(lp->cfg.tx_irq, lp);
1318 	free_irq(lp->cfg.rx_irq, lp);
1319 
1320 	spin_unlock_irqrestore(&lp->lock, flags);
1321 
1322 	return err;
1323 }
1324 EXPORT_SYMBOL(ldc_bind);
1325 
1326 int ldc_connect(struct ldc_channel *lp)
1327 {
1328 	unsigned long flags;
1329 	int err;
1330 
1331 	if (lp->cfg.mode == LDC_MODE_RAW)
1332 		return -EINVAL;
1333 
1334 	spin_lock_irqsave(&lp->lock, flags);
1335 
1336 	if (!(lp->flags & LDC_FLAG_ALLOCED_QUEUES) ||
1337 	    !(lp->flags & LDC_FLAG_REGISTERED_QUEUES) ||
1338 	    lp->hs_state != LDC_HS_OPEN)
1339 		err = ((lp->hs_state > LDC_HS_OPEN) ? 0 : -EINVAL);
1340 	else
1341 		err = start_handshake(lp);
1342 
1343 	spin_unlock_irqrestore(&lp->lock, flags);
1344 
1345 	return err;
1346 }
1347 EXPORT_SYMBOL(ldc_connect);
1348 
1349 int ldc_disconnect(struct ldc_channel *lp)
1350 {
1351 	unsigned long hv_err, flags;
1352 	int err;
1353 
1354 	if (lp->cfg.mode == LDC_MODE_RAW)
1355 		return -EINVAL;
1356 
1357 	if (!(lp->flags & LDC_FLAG_ALLOCED_QUEUES) ||
1358 	    !(lp->flags & LDC_FLAG_REGISTERED_QUEUES))
1359 		return -EINVAL;
1360 
1361 	spin_lock_irqsave(&lp->lock, flags);
1362 
1363 	err = -ENODEV;
1364 	hv_err = sun4v_ldc_tx_qconf(lp->id, 0, 0);
1365 	if (hv_err)
1366 		goto out_err;
1367 
1368 	hv_err = sun4v_ldc_tx_qconf(lp->id, lp->tx_ra, lp->tx_num_entries);
1369 	if (hv_err)
1370 		goto out_err;
1371 
1372 	hv_err = sun4v_ldc_rx_qconf(lp->id, 0, 0);
1373 	if (hv_err)
1374 		goto out_err;
1375 
1376 	hv_err = sun4v_ldc_rx_qconf(lp->id, lp->rx_ra, lp->rx_num_entries);
1377 	if (hv_err)
1378 		goto out_err;
1379 
1380 	ldc_set_state(lp, LDC_STATE_BOUND);
1381 	lp->hs_state = LDC_HS_OPEN;
1382 	lp->flags |= LDC_FLAG_RESET;
1383 
1384 	spin_unlock_irqrestore(&lp->lock, flags);
1385 
1386 	return 0;
1387 
1388 out_err:
1389 	sun4v_ldc_tx_qconf(lp->id, 0, 0);
1390 	sun4v_ldc_rx_qconf(lp->id, 0, 0);
1391 	free_irq(lp->cfg.tx_irq, lp);
1392 	free_irq(lp->cfg.rx_irq, lp);
1393 	lp->flags &= ~(LDC_FLAG_REGISTERED_IRQS |
1394 		       LDC_FLAG_REGISTERED_QUEUES);
1395 	ldc_set_state(lp, LDC_STATE_INIT);
1396 
1397 	spin_unlock_irqrestore(&lp->lock, flags);
1398 
1399 	return err;
1400 }
1401 EXPORT_SYMBOL(ldc_disconnect);
1402 
1403 int ldc_state(struct ldc_channel *lp)
1404 {
1405 	return lp->state;
1406 }
1407 EXPORT_SYMBOL(ldc_state);
1408 
1409 static int write_raw(struct ldc_channel *lp, const void *buf, unsigned int size)
1410 {
1411 	struct ldc_packet *p;
1412 	unsigned long new_tail;
1413 	int err;
1414 
1415 	if (size > LDC_PACKET_SIZE)
1416 		return -EMSGSIZE;
1417 
1418 	p = data_get_tx_packet(lp, &new_tail);
1419 	if (!p)
1420 		return -EAGAIN;
1421 
1422 	memcpy(p, buf, size);
1423 
1424 	err = send_tx_packet(lp, p, new_tail);
1425 	if (!err)
1426 		err = size;
1427 
1428 	return err;
1429 }
1430 
1431 static int read_raw(struct ldc_channel *lp, void *buf, unsigned int size)
1432 {
1433 	struct ldc_packet *p;
1434 	unsigned long hv_err, new;
1435 	int err;
1436 
1437 	if (size < LDC_PACKET_SIZE)
1438 		return -EINVAL;
1439 
1440 	hv_err = sun4v_ldc_rx_get_state(lp->id,
1441 					&lp->rx_head,
1442 					&lp->rx_tail,
1443 					&lp->chan_state);
1444 	if (hv_err)
1445 		return ldc_abort(lp);
1446 
1447 	if (lp->chan_state == LDC_CHANNEL_DOWN ||
1448 	    lp->chan_state == LDC_CHANNEL_RESETTING)
1449 		return -ECONNRESET;
1450 
1451 	if (lp->rx_head == lp->rx_tail)
1452 		return 0;
1453 
1454 	p = lp->rx_base + (lp->rx_head / LDC_PACKET_SIZE);
1455 	memcpy(buf, p, LDC_PACKET_SIZE);
1456 
1457 	new = rx_advance(lp, lp->rx_head);
1458 	lp->rx_head = new;
1459 
1460 	err = __set_rx_head(lp, new);
1461 	if (err < 0)
1462 		err = -ECONNRESET;
1463 	else
1464 		err = LDC_PACKET_SIZE;
1465 
1466 	return err;
1467 }
1468 
1469 static const struct ldc_mode_ops raw_ops = {
1470 	.write		=	write_raw,
1471 	.read		=	read_raw,
1472 };
1473 
1474 static int write_nonraw(struct ldc_channel *lp, const void *buf,
1475 			unsigned int size)
1476 {
1477 	unsigned long hv_err, tail;
1478 	unsigned int copied;
1479 	u32 seq;
1480 	int err;
1481 
1482 	hv_err = sun4v_ldc_tx_get_state(lp->id, &lp->tx_head, &lp->tx_tail,
1483 					&lp->chan_state);
1484 	if (unlikely(hv_err))
1485 		return -EBUSY;
1486 
1487 	if (unlikely(lp->chan_state != LDC_CHANNEL_UP))
1488 		return ldc_abort(lp);
1489 
1490 	if (!tx_has_space_for(lp, size))
1491 		return -EAGAIN;
1492 
1493 	seq = lp->snd_nxt;
1494 	copied = 0;
1495 	tail = lp->tx_tail;
1496 	while (copied < size) {
1497 		struct ldc_packet *p = lp->tx_base + (tail / LDC_PACKET_SIZE);
1498 		u8 *data = ((lp->cfg.mode == LDC_MODE_UNRELIABLE) ?
1499 			    p->u.u_data :
1500 			    p->u.r.r_data);
1501 		int data_len;
1502 
1503 		p->type = LDC_DATA;
1504 		p->stype = LDC_INFO;
1505 		p->ctrl = 0;
1506 
1507 		data_len = size - copied;
1508 		if (data_len > lp->mss)
1509 			data_len = lp->mss;
1510 
1511 		BUG_ON(data_len > LDC_LEN);
1512 
1513 		p->env = (data_len |
1514 			  (copied == 0 ? LDC_START : 0) |
1515 			  (data_len == size - copied ? LDC_STOP : 0));
1516 
1517 		p->seqid = ++seq;
1518 
1519 		ldcdbg(DATA, "SENT DATA [%02x:%02x:%02x:%02x:%08x]\n",
1520 		       p->type,
1521 		       p->stype,
1522 		       p->ctrl,
1523 		       p->env,
1524 		       p->seqid);
1525 
1526 		memcpy(data, buf, data_len);
1527 		buf += data_len;
1528 		copied += data_len;
1529 
1530 		tail = tx_advance(lp, tail);
1531 	}
1532 
1533 	err = set_tx_tail(lp, tail);
1534 	if (!err) {
1535 		lp->snd_nxt = seq;
1536 		err = size;
1537 	}
1538 
1539 	return err;
1540 }
1541 
1542 static int rx_bad_seq(struct ldc_channel *lp, struct ldc_packet *p,
1543 		      struct ldc_packet *first_frag)
1544 {
1545 	int err;
1546 
1547 	if (first_frag)
1548 		lp->rcv_nxt = first_frag->seqid - 1;
1549 
1550 	err = send_data_nack(lp, p);
1551 	if (err)
1552 		return err;
1553 
1554 	err = __set_rx_head(lp, lp->rx_tail);
1555 	if (err < 0)
1556 		return ldc_abort(lp);
1557 
1558 	return 0;
1559 }
1560 
1561 static int data_ack_nack(struct ldc_channel *lp, struct ldc_packet *p)
1562 {
1563 	if (p->stype & LDC_ACK) {
1564 		int err = process_data_ack(lp, p);
1565 		if (err)
1566 			return err;
1567 	}
1568 	if (p->stype & LDC_NACK)
1569 		return ldc_abort(lp);
1570 
1571 	return 0;
1572 }
1573 
1574 static int rx_data_wait(struct ldc_channel *lp, unsigned long cur_head)
1575 {
1576 	unsigned long dummy;
1577 	int limit = 1000;
1578 
1579 	ldcdbg(DATA, "DATA WAIT cur_head[%lx] rx_head[%lx] rx_tail[%lx]\n",
1580 	       cur_head, lp->rx_head, lp->rx_tail);
1581 	while (limit-- > 0) {
1582 		unsigned long hv_err;
1583 
1584 		hv_err = sun4v_ldc_rx_get_state(lp->id,
1585 						&dummy,
1586 						&lp->rx_tail,
1587 						&lp->chan_state);
1588 		if (hv_err)
1589 			return ldc_abort(lp);
1590 
1591 		if (lp->chan_state == LDC_CHANNEL_DOWN ||
1592 		    lp->chan_state == LDC_CHANNEL_RESETTING)
1593 			return -ECONNRESET;
1594 
1595 		if (cur_head != lp->rx_tail) {
1596 			ldcdbg(DATA, "DATA WAIT DONE "
1597 			       "head[%lx] tail[%lx] chan_state[%lx]\n",
1598 			       dummy, lp->rx_tail, lp->chan_state);
1599 			return 0;
1600 		}
1601 
1602 		udelay(1);
1603 	}
1604 	return -EAGAIN;
1605 }
1606 
1607 static int rx_set_head(struct ldc_channel *lp, unsigned long head)
1608 {
1609 	int err = __set_rx_head(lp, head);
1610 
1611 	if (err < 0)
1612 		return ldc_abort(lp);
1613 
1614 	lp->rx_head = head;
1615 	return 0;
1616 }
1617 
1618 static void send_data_ack(struct ldc_channel *lp)
1619 {
1620 	unsigned long new_tail;
1621 	struct ldc_packet *p;
1622 
1623 	p = data_get_tx_packet(lp, &new_tail);
1624 	if (likely(p)) {
1625 		int err;
1626 
1627 		memset(p, 0, sizeof(*p));
1628 		p->type = LDC_DATA;
1629 		p->stype = LDC_ACK;
1630 		p->ctrl = 0;
1631 		p->seqid = lp->snd_nxt + 1;
1632 		p->u.r.ackid = lp->rcv_nxt;
1633 
1634 		err = send_tx_packet(lp, p, new_tail);
1635 		if (!err)
1636 			lp->snd_nxt++;
1637 	}
1638 }
1639 
1640 static int read_nonraw(struct ldc_channel *lp, void *buf, unsigned int size)
1641 {
1642 	struct ldc_packet *first_frag;
1643 	unsigned long hv_err, new;
1644 	int err, copied;
1645 
1646 	hv_err = sun4v_ldc_rx_get_state(lp->id,
1647 					&lp->rx_head,
1648 					&lp->rx_tail,
1649 					&lp->chan_state);
1650 	if (hv_err)
1651 		return ldc_abort(lp);
1652 
1653 	if (lp->chan_state == LDC_CHANNEL_DOWN ||
1654 	    lp->chan_state == LDC_CHANNEL_RESETTING)
1655 		return -ECONNRESET;
1656 
1657 	if (lp->rx_head == lp->rx_tail)
1658 		return 0;
1659 
1660 	first_frag = NULL;
1661 	copied = err = 0;
1662 	new = lp->rx_head;
1663 	while (1) {
1664 		struct ldc_packet *p;
1665 		int pkt_len;
1666 
1667 		BUG_ON(new == lp->rx_tail);
1668 		p = lp->rx_base + (new / LDC_PACKET_SIZE);
1669 
1670 		ldcdbg(RX, "RX read pkt[%02x:%02x:%02x:%02x:%08x:%08x] "
1671 		       "rcv_nxt[%08x]\n",
1672 		       p->type,
1673 		       p->stype,
1674 		       p->ctrl,
1675 		       p->env,
1676 		       p->seqid,
1677 		       p->u.r.ackid,
1678 		       lp->rcv_nxt);
1679 
1680 		if (unlikely(!rx_seq_ok(lp, p->seqid))) {
1681 			err = rx_bad_seq(lp, p, first_frag);
1682 			copied = 0;
1683 			break;
1684 		}
1685 
1686 		if (p->type & LDC_CTRL) {
1687 			err = process_control_frame(lp, p);
1688 			if (err < 0)
1689 				break;
1690 			err = 0;
1691 		}
1692 
1693 		lp->rcv_nxt = p->seqid;
1694 
1695 		if (!(p->type & LDC_DATA)) {
1696 			new = rx_advance(lp, new);
1697 			goto no_data;
1698 		}
1699 		if (p->stype & (LDC_ACK | LDC_NACK)) {
1700 			err = data_ack_nack(lp, p);
1701 			if (err)
1702 				break;
1703 		}
1704 		if (!(p->stype & LDC_INFO)) {
1705 			new = rx_advance(lp, new);
1706 			err = rx_set_head(lp, new);
1707 			if (err)
1708 				break;
1709 			goto no_data;
1710 		}
1711 
1712 		pkt_len = p->env & LDC_LEN;
1713 
1714 		/* Every initial packet starts with the START bit set.
1715 		 *
1716 		 * Singleton packets will have both START+STOP set.
1717 		 *
1718 		 * Fragments will have START set in the first frame, STOP
1719 		 * set in the last frame, and neither bit set in middle
1720 		 * frames of the packet.
1721 		 *
1722 		 * Therefore if we are at the beginning of a packet and
1723 		 * we don't see START, or we are in the middle of a fragmented
1724 		 * packet and do see START, we are unsynchronized and should
1725 		 * flush the RX queue.
1726 		 */
1727 		if ((first_frag == NULL && !(p->env & LDC_START)) ||
1728 		    (first_frag != NULL &&  (p->env & LDC_START))) {
1729 			if (!first_frag)
1730 				new = rx_advance(lp, new);
1731 
1732 			err = rx_set_head(lp, new);
1733 			if (err)
1734 				break;
1735 
1736 			if (!first_frag)
1737 				goto no_data;
1738 		}
1739 		if (!first_frag)
1740 			first_frag = p;
1741 
1742 		if (pkt_len > size - copied) {
1743 			/* User didn't give us a big enough buffer,
1744 			 * what to do?  This is a pretty serious error.
1745 			 *
1746 			 * Since we haven't updated the RX ring head to
1747 			 * consume any of the packets, signal the error
1748 			 * to the user and just leave the RX ring alone.
1749 			 *
1750 			 * This seems the best behavior because this allows
1751 			 * a user of the LDC layer to start with a small
1752 			 * RX buffer for ldc_read() calls and use -EMSGSIZE
1753 			 * as a cue to enlarge it's read buffer.
1754 			 */
1755 			err = -EMSGSIZE;
1756 			break;
1757 		}
1758 
1759 		/* Ok, we are gonna eat this one.  */
1760 		new = rx_advance(lp, new);
1761 
1762 		memcpy(buf,
1763 		       (lp->cfg.mode == LDC_MODE_UNRELIABLE ?
1764 			p->u.u_data : p->u.r.r_data), pkt_len);
1765 		buf += pkt_len;
1766 		copied += pkt_len;
1767 
1768 		if (p->env & LDC_STOP)
1769 			break;
1770 
1771 no_data:
1772 		if (new == lp->rx_tail) {
1773 			err = rx_data_wait(lp, new);
1774 			if (err)
1775 				break;
1776 		}
1777 	}
1778 
1779 	if (!err)
1780 		err = rx_set_head(lp, new);
1781 
1782 	if (err && first_frag)
1783 		lp->rcv_nxt = first_frag->seqid - 1;
1784 
1785 	if (!err) {
1786 		err = copied;
1787 		if (err > 0 && lp->cfg.mode != LDC_MODE_UNRELIABLE)
1788 			send_data_ack(lp);
1789 	}
1790 
1791 	return err;
1792 }
1793 
1794 static const struct ldc_mode_ops nonraw_ops = {
1795 	.write		=	write_nonraw,
1796 	.read		=	read_nonraw,
1797 };
1798 
1799 static int write_stream(struct ldc_channel *lp, const void *buf,
1800 			unsigned int size)
1801 {
1802 	if (size > lp->cfg.mtu)
1803 		size = lp->cfg.mtu;
1804 	return write_nonraw(lp, buf, size);
1805 }
1806 
1807 static int read_stream(struct ldc_channel *lp, void *buf, unsigned int size)
1808 {
1809 	if (!lp->mssbuf_len) {
1810 		int err = read_nonraw(lp, lp->mssbuf, lp->cfg.mtu);
1811 		if (err < 0)
1812 			return err;
1813 
1814 		lp->mssbuf_len = err;
1815 		lp->mssbuf_off = 0;
1816 	}
1817 
1818 	if (size > lp->mssbuf_len)
1819 		size = lp->mssbuf_len;
1820 	memcpy(buf, lp->mssbuf + lp->mssbuf_off, size);
1821 
1822 	lp->mssbuf_off += size;
1823 	lp->mssbuf_len -= size;
1824 
1825 	return size;
1826 }
1827 
1828 static const struct ldc_mode_ops stream_ops = {
1829 	.write		=	write_stream,
1830 	.read		=	read_stream,
1831 };
1832 
1833 int ldc_write(struct ldc_channel *lp, const void *buf, unsigned int size)
1834 {
1835 	unsigned long flags;
1836 	int err;
1837 
1838 	if (!buf)
1839 		return -EINVAL;
1840 
1841 	if (!size)
1842 		return 0;
1843 
1844 	spin_lock_irqsave(&lp->lock, flags);
1845 
1846 	if (lp->hs_state != LDC_HS_COMPLETE)
1847 		err = -ENOTCONN;
1848 	else
1849 		err = lp->mops->write(lp, buf, size);
1850 
1851 	spin_unlock_irqrestore(&lp->lock, flags);
1852 
1853 	return err;
1854 }
1855 EXPORT_SYMBOL(ldc_write);
1856 
1857 int ldc_read(struct ldc_channel *lp, void *buf, unsigned int size)
1858 {
1859 	unsigned long flags;
1860 	int err;
1861 
1862 	if (!buf)
1863 		return -EINVAL;
1864 
1865 	if (!size)
1866 		return 0;
1867 
1868 	spin_lock_irqsave(&lp->lock, flags);
1869 
1870 	if (lp->hs_state != LDC_HS_COMPLETE)
1871 		err = -ENOTCONN;
1872 	else
1873 		err = lp->mops->read(lp, buf, size);
1874 
1875 	spin_unlock_irqrestore(&lp->lock, flags);
1876 
1877 	return err;
1878 }
1879 EXPORT_SYMBOL(ldc_read);
1880 
1881 static long arena_alloc(struct ldc_iommu *iommu, unsigned long npages)
1882 {
1883 	struct iommu_arena *arena = &iommu->arena;
1884 	unsigned long n, start, end, limit;
1885 	int pass;
1886 
1887 	limit = arena->limit;
1888 	start = arena->hint;
1889 	pass = 0;
1890 
1891 again:
1892 	n = bitmap_find_next_zero_area(arena->map, limit, start, npages, 0);
1893 	end = n + npages;
1894 	if (unlikely(end >= limit)) {
1895 		if (likely(pass < 1)) {
1896 			limit = start;
1897 			start = 0;
1898 			pass++;
1899 			goto again;
1900 		} else {
1901 			/* Scanned the whole thing, give up. */
1902 			return -1;
1903 		}
1904 	}
1905 	bitmap_set(arena->map, n, npages);
1906 
1907 	arena->hint = end;
1908 
1909 	return n;
1910 }
1911 
1912 #define COOKIE_PGSZ_CODE	0xf000000000000000ULL
1913 #define COOKIE_PGSZ_CODE_SHIFT	60ULL
1914 
1915 static u64 pagesize_code(void)
1916 {
1917 	switch (PAGE_SIZE) {
1918 	default:
1919 	case (8ULL * 1024ULL):
1920 		return 0;
1921 	case (64ULL * 1024ULL):
1922 		return 1;
1923 	case (512ULL * 1024ULL):
1924 		return 2;
1925 	case (4ULL * 1024ULL * 1024ULL):
1926 		return 3;
1927 	case (32ULL * 1024ULL * 1024ULL):
1928 		return 4;
1929 	case (256ULL * 1024ULL * 1024ULL):
1930 		return 5;
1931 	}
1932 }
1933 
1934 static u64 make_cookie(u64 index, u64 pgsz_code, u64 page_offset)
1935 {
1936 	return ((pgsz_code << COOKIE_PGSZ_CODE_SHIFT) |
1937 		(index << PAGE_SHIFT) |
1938 		page_offset);
1939 }
1940 
1941 static u64 cookie_to_index(u64 cookie, unsigned long *shift)
1942 {
1943 	u64 szcode = cookie >> COOKIE_PGSZ_CODE_SHIFT;
1944 
1945 	cookie &= ~COOKIE_PGSZ_CODE;
1946 
1947 	*shift = szcode * 3;
1948 
1949 	return (cookie >> (13ULL + (szcode * 3ULL)));
1950 }
1951 
1952 static struct ldc_mtable_entry *alloc_npages(struct ldc_iommu *iommu,
1953 					     unsigned long npages)
1954 {
1955 	long entry;
1956 
1957 	entry = arena_alloc(iommu, npages);
1958 	if (unlikely(entry < 0))
1959 		return NULL;
1960 
1961 	return iommu->page_table + entry;
1962 }
1963 
1964 static u64 perm_to_mte(unsigned int map_perm)
1965 {
1966 	u64 mte_base;
1967 
1968 	mte_base = pagesize_code();
1969 
1970 	if (map_perm & LDC_MAP_SHADOW) {
1971 		if (map_perm & LDC_MAP_R)
1972 			mte_base |= LDC_MTE_COPY_R;
1973 		if (map_perm & LDC_MAP_W)
1974 			mte_base |= LDC_MTE_COPY_W;
1975 	}
1976 	if (map_perm & LDC_MAP_DIRECT) {
1977 		if (map_perm & LDC_MAP_R)
1978 			mte_base |= LDC_MTE_READ;
1979 		if (map_perm & LDC_MAP_W)
1980 			mte_base |= LDC_MTE_WRITE;
1981 		if (map_perm & LDC_MAP_X)
1982 			mte_base |= LDC_MTE_EXEC;
1983 	}
1984 	if (map_perm & LDC_MAP_IO) {
1985 		if (map_perm & LDC_MAP_R)
1986 			mte_base |= LDC_MTE_IOMMU_R;
1987 		if (map_perm & LDC_MAP_W)
1988 			mte_base |= LDC_MTE_IOMMU_W;
1989 	}
1990 
1991 	return mte_base;
1992 }
1993 
1994 static int pages_in_region(unsigned long base, long len)
1995 {
1996 	int count = 0;
1997 
1998 	do {
1999 		unsigned long new = (base + PAGE_SIZE) & PAGE_MASK;
2000 
2001 		len -= (new - base);
2002 		base = new;
2003 		count++;
2004 	} while (len > 0);
2005 
2006 	return count;
2007 }
2008 
2009 struct cookie_state {
2010 	struct ldc_mtable_entry		*page_table;
2011 	struct ldc_trans_cookie		*cookies;
2012 	u64				mte_base;
2013 	u64				prev_cookie;
2014 	u32				pte_idx;
2015 	u32				nc;
2016 };
2017 
2018 static void fill_cookies(struct cookie_state *sp, unsigned long pa,
2019 			 unsigned long off, unsigned long len)
2020 {
2021 	do {
2022 		unsigned long tlen, new = pa + PAGE_SIZE;
2023 		u64 this_cookie;
2024 
2025 		sp->page_table[sp->pte_idx].mte = sp->mte_base | pa;
2026 
2027 		tlen = PAGE_SIZE;
2028 		if (off)
2029 			tlen = PAGE_SIZE - off;
2030 		if (tlen > len)
2031 			tlen = len;
2032 
2033 		this_cookie = make_cookie(sp->pte_idx,
2034 					  pagesize_code(), off);
2035 
2036 		off = 0;
2037 
2038 		if (this_cookie == sp->prev_cookie) {
2039 			sp->cookies[sp->nc - 1].cookie_size += tlen;
2040 		} else {
2041 			sp->cookies[sp->nc].cookie_addr = this_cookie;
2042 			sp->cookies[sp->nc].cookie_size = tlen;
2043 			sp->nc++;
2044 		}
2045 		sp->prev_cookie = this_cookie + tlen;
2046 
2047 		sp->pte_idx++;
2048 
2049 		len -= tlen;
2050 		pa = new;
2051 	} while (len > 0);
2052 }
2053 
2054 static int sg_count_one(struct scatterlist *sg)
2055 {
2056 	unsigned long base = page_to_pfn(sg_page(sg)) << PAGE_SHIFT;
2057 	long len = sg->length;
2058 
2059 	if ((sg->offset | len) & (8UL - 1))
2060 		return -EFAULT;
2061 
2062 	return pages_in_region(base + sg->offset, len);
2063 }
2064 
2065 static int sg_count_pages(struct scatterlist *sg, int num_sg)
2066 {
2067 	int count;
2068 	int i;
2069 
2070 	count = 0;
2071 	for (i = 0; i < num_sg; i++) {
2072 		int err = sg_count_one(sg + i);
2073 		if (err < 0)
2074 			return err;
2075 		count += err;
2076 	}
2077 
2078 	return count;
2079 }
2080 
2081 int ldc_map_sg(struct ldc_channel *lp,
2082 	       struct scatterlist *sg, int num_sg,
2083 	       struct ldc_trans_cookie *cookies, int ncookies,
2084 	       unsigned int map_perm)
2085 {
2086 	unsigned long i, npages, flags;
2087 	struct ldc_mtable_entry *base;
2088 	struct cookie_state state;
2089 	struct ldc_iommu *iommu;
2090 	int err;
2091 
2092 	if (map_perm & ~LDC_MAP_ALL)
2093 		return -EINVAL;
2094 
2095 	err = sg_count_pages(sg, num_sg);
2096 	if (err < 0)
2097 		return err;
2098 
2099 	npages = err;
2100 	if (err > ncookies)
2101 		return -EMSGSIZE;
2102 
2103 	iommu = &lp->iommu;
2104 
2105 	spin_lock_irqsave(&iommu->lock, flags);
2106 	base = alloc_npages(iommu, npages);
2107 	spin_unlock_irqrestore(&iommu->lock, flags);
2108 
2109 	if (!base)
2110 		return -ENOMEM;
2111 
2112 	state.page_table = iommu->page_table;
2113 	state.cookies = cookies;
2114 	state.mte_base = perm_to_mte(map_perm);
2115 	state.prev_cookie = ~(u64)0;
2116 	state.pte_idx = (base - iommu->page_table);
2117 	state.nc = 0;
2118 
2119 	for (i = 0; i < num_sg; i++)
2120 		fill_cookies(&state, page_to_pfn(sg_page(&sg[i])) << PAGE_SHIFT,
2121 			     sg[i].offset, sg[i].length);
2122 
2123 	return state.nc;
2124 }
2125 EXPORT_SYMBOL(ldc_map_sg);
2126 
2127 int ldc_map_single(struct ldc_channel *lp,
2128 		   void *buf, unsigned int len,
2129 		   struct ldc_trans_cookie *cookies, int ncookies,
2130 		   unsigned int map_perm)
2131 {
2132 	unsigned long npages, pa, flags;
2133 	struct ldc_mtable_entry *base;
2134 	struct cookie_state state;
2135 	struct ldc_iommu *iommu;
2136 
2137 	if ((map_perm & ~LDC_MAP_ALL) || (ncookies < 1))
2138 		return -EINVAL;
2139 
2140 	pa = __pa(buf);
2141 	if ((pa | len) & (8UL - 1))
2142 		return -EFAULT;
2143 
2144 	npages = pages_in_region(pa, len);
2145 
2146 	iommu = &lp->iommu;
2147 
2148 	spin_lock_irqsave(&iommu->lock, flags);
2149 	base = alloc_npages(iommu, npages);
2150 	spin_unlock_irqrestore(&iommu->lock, flags);
2151 
2152 	if (!base)
2153 		return -ENOMEM;
2154 
2155 	state.page_table = iommu->page_table;
2156 	state.cookies = cookies;
2157 	state.mte_base = perm_to_mte(map_perm);
2158 	state.prev_cookie = ~(u64)0;
2159 	state.pte_idx = (base - iommu->page_table);
2160 	state.nc = 0;
2161 	fill_cookies(&state, (pa & PAGE_MASK), (pa & ~PAGE_MASK), len);
2162 	BUG_ON(state.nc != 1);
2163 
2164 	return state.nc;
2165 }
2166 EXPORT_SYMBOL(ldc_map_single);
2167 
2168 static void free_npages(unsigned long id, struct ldc_iommu *iommu,
2169 			u64 cookie, u64 size)
2170 {
2171 	struct iommu_arena *arena = &iommu->arena;
2172 	unsigned long i, shift, index, npages;
2173 	struct ldc_mtable_entry *base;
2174 
2175 	npages = PAGE_ALIGN(((cookie & ~PAGE_MASK) + size)) >> PAGE_SHIFT;
2176 	index = cookie_to_index(cookie, &shift);
2177 	base = iommu->page_table + index;
2178 
2179 	BUG_ON(index > arena->limit ||
2180 	       (index + npages) > arena->limit);
2181 
2182 	for (i = 0; i < npages; i++) {
2183 		if (base->cookie)
2184 			sun4v_ldc_revoke(id, cookie + (i << shift),
2185 					 base->cookie);
2186 		base->mte = 0;
2187 		__clear_bit(index + i, arena->map);
2188 	}
2189 }
2190 
2191 void ldc_unmap(struct ldc_channel *lp, struct ldc_trans_cookie *cookies,
2192 	       int ncookies)
2193 {
2194 	struct ldc_iommu *iommu = &lp->iommu;
2195 	unsigned long flags;
2196 	int i;
2197 
2198 	spin_lock_irqsave(&iommu->lock, flags);
2199 	for (i = 0; i < ncookies; i++) {
2200 		u64 addr = cookies[i].cookie_addr;
2201 		u64 size = cookies[i].cookie_size;
2202 
2203 		free_npages(lp->id, iommu, addr, size);
2204 	}
2205 	spin_unlock_irqrestore(&iommu->lock, flags);
2206 }
2207 EXPORT_SYMBOL(ldc_unmap);
2208 
2209 int ldc_copy(struct ldc_channel *lp, int copy_dir,
2210 	     void *buf, unsigned int len, unsigned long offset,
2211 	     struct ldc_trans_cookie *cookies, int ncookies)
2212 {
2213 	unsigned int orig_len;
2214 	unsigned long ra;
2215 	int i;
2216 
2217 	if (copy_dir != LDC_COPY_IN && copy_dir != LDC_COPY_OUT) {
2218 		printk(KERN_ERR PFX "ldc_copy: ID[%lu] Bad copy_dir[%d]\n",
2219 		       lp->id, copy_dir);
2220 		return -EINVAL;
2221 	}
2222 
2223 	ra = __pa(buf);
2224 	if ((ra | len | offset) & (8UL - 1)) {
2225 		printk(KERN_ERR PFX "ldc_copy: ID[%lu] Unaligned buffer "
2226 		       "ra[%lx] len[%x] offset[%lx]\n",
2227 		       lp->id, ra, len, offset);
2228 		return -EFAULT;
2229 	}
2230 
2231 	if (lp->hs_state != LDC_HS_COMPLETE ||
2232 	    (lp->flags & LDC_FLAG_RESET)) {
2233 		printk(KERN_ERR PFX "ldc_copy: ID[%lu] Link down hs_state[%x] "
2234 		       "flags[%x]\n", lp->id, lp->hs_state, lp->flags);
2235 		return -ECONNRESET;
2236 	}
2237 
2238 	orig_len = len;
2239 	for (i = 0; i < ncookies; i++) {
2240 		unsigned long cookie_raddr = cookies[i].cookie_addr;
2241 		unsigned long this_len = cookies[i].cookie_size;
2242 		unsigned long actual_len;
2243 
2244 		if (unlikely(offset)) {
2245 			unsigned long this_off = offset;
2246 
2247 			if (this_off > this_len)
2248 				this_off = this_len;
2249 
2250 			offset -= this_off;
2251 			this_len -= this_off;
2252 			if (!this_len)
2253 				continue;
2254 			cookie_raddr += this_off;
2255 		}
2256 
2257 		if (this_len > len)
2258 			this_len = len;
2259 
2260 		while (1) {
2261 			unsigned long hv_err;
2262 
2263 			hv_err = sun4v_ldc_copy(lp->id, copy_dir,
2264 						cookie_raddr, ra,
2265 						this_len, &actual_len);
2266 			if (unlikely(hv_err)) {
2267 				printk(KERN_ERR PFX "ldc_copy: ID[%lu] "
2268 				       "HV error %lu\n",
2269 				       lp->id, hv_err);
2270 				if (lp->hs_state != LDC_HS_COMPLETE ||
2271 				    (lp->flags & LDC_FLAG_RESET))
2272 					return -ECONNRESET;
2273 				else
2274 					return -EFAULT;
2275 			}
2276 
2277 			cookie_raddr += actual_len;
2278 			ra += actual_len;
2279 			len -= actual_len;
2280 			if (actual_len == this_len)
2281 				break;
2282 
2283 			this_len -= actual_len;
2284 		}
2285 
2286 		if (!len)
2287 			break;
2288 	}
2289 
2290 	/* It is caller policy what to do about short copies.
2291 	 * For example, a networking driver can declare the
2292 	 * packet a runt and drop it.
2293 	 */
2294 
2295 	return orig_len - len;
2296 }
2297 EXPORT_SYMBOL(ldc_copy);
2298 
2299 void *ldc_alloc_exp_dring(struct ldc_channel *lp, unsigned int len,
2300 			  struct ldc_trans_cookie *cookies, int *ncookies,
2301 			  unsigned int map_perm)
2302 {
2303 	void *buf;
2304 	int err;
2305 
2306 	if (len & (8UL - 1))
2307 		return ERR_PTR(-EINVAL);
2308 
2309 	buf = kzalloc(len, GFP_KERNEL);
2310 	if (!buf)
2311 		return ERR_PTR(-ENOMEM);
2312 
2313 	err = ldc_map_single(lp, buf, len, cookies, *ncookies, map_perm);
2314 	if (err < 0) {
2315 		kfree(buf);
2316 		return ERR_PTR(err);
2317 	}
2318 	*ncookies = err;
2319 
2320 	return buf;
2321 }
2322 EXPORT_SYMBOL(ldc_alloc_exp_dring);
2323 
2324 void ldc_free_exp_dring(struct ldc_channel *lp, void *buf, unsigned int len,
2325 			struct ldc_trans_cookie *cookies, int ncookies)
2326 {
2327 	ldc_unmap(lp, cookies, ncookies);
2328 	kfree(buf);
2329 }
2330 EXPORT_SYMBOL(ldc_free_exp_dring);
2331 
2332 static int __init ldc_init(void)
2333 {
2334 	unsigned long major, minor;
2335 	struct mdesc_handle *hp;
2336 	const u64 *v;
2337 	int err;
2338 	u64 mp;
2339 
2340 	hp = mdesc_grab();
2341 	if (!hp)
2342 		return -ENODEV;
2343 
2344 	mp = mdesc_node_by_name(hp, MDESC_NODE_NULL, "platform");
2345 	err = -ENODEV;
2346 	if (mp == MDESC_NODE_NULL)
2347 		goto out;
2348 
2349 	v = mdesc_get_property(hp, mp, "domaining-enabled", NULL);
2350 	if (!v)
2351 		goto out;
2352 
2353 	major = 1;
2354 	minor = 0;
2355 	if (sun4v_hvapi_register(HV_GRP_LDOM, major, &minor)) {
2356 		printk(KERN_INFO PFX "Could not register LDOM hvapi.\n");
2357 		goto out;
2358 	}
2359 
2360 	printk(KERN_INFO "%s", version);
2361 
2362 	if (!*v) {
2363 		printk(KERN_INFO PFX "Domaining disabled.\n");
2364 		goto out;
2365 	}
2366 	ldom_domaining_enabled = 1;
2367 	err = 0;
2368 
2369 out:
2370 	mdesc_release(hp);
2371 	return err;
2372 }
2373 
2374 core_initcall(ldc_init);
2375