1 /* ldc.c: Logical Domain Channel link-layer protocol driver.
3 * Copyright (C) 2007, 2008 David S. Miller <davem@davemloft.net>
6 #include <linux/kernel.h>
7 #include <linux/module.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>
18 #include <asm/hypervisor.h>
19 #include <asm/iommu.h>
22 #include <asm/mdesc.h>
24 #define DRV_MODULE_NAME "ldc"
25 #define PFX DRV_MODULE_NAME ": "
26 #define DRV_MODULE_VERSION "1.1"
27 #define DRV_MODULE_RELDATE "July 22, 2008"
29 static char version[] __devinitdata =
30 DRV_MODULE_NAME ".c:v" DRV_MODULE_VERSION " (" DRV_MODULE_RELDATE ")\n";
31 #define LDC_PACKET_SIZE 64
33 /* Packet header layout for unreliable and reliable mode frames.
34 * When in RAW mode, packets are simply straight 64-byte payloads
49 #define LDC_VERS 0x01 /* Link Version */
50 #define LDC_RTS 0x02 /* Request To Send */
51 #define LDC_RTR 0x03 /* Ready To Receive */
52 #define LDC_RDX 0x04 /* Ready for Data eXchange */
53 #define LDC_CTRL_MSK 0x0f
57 #define LDC_FRAG_MASK 0xc0
58 #define LDC_START 0x40
64 u8 u_data[LDC_PACKET_SIZE - 8];
68 u8 r_data[LDC_PACKET_SIZE - 8 - 8];
78 /* Ordered from largest major to lowest. */
79 static struct ldc_version ver_arr[] = {
80 { .major = 1, .minor = 0 },
83 #define LDC_DEFAULT_MTU (4 * LDC_PACKET_SIZE)
84 #define LDC_DEFAULT_NUM_ENTRIES (PAGE_SIZE / LDC_PACKET_SIZE)
89 int (*write)(struct ldc_channel *, const void *, unsigned int);
90 int (*read)(struct ldc_channel *, void *, unsigned int);
93 static const struct ldc_mode_ops raw_ops;
94 static const struct ldc_mode_ops nonraw_ops;
95 static const struct ldc_mode_ops stream_ops;
97 int ldom_domaining_enabled;
100 /* Protects arena alloc/free. */
102 struct iommu_arena arena;
103 struct ldc_mtable_entry *page_table;
107 /* Protects all operations that depend upon channel state. */
116 struct ldc_packet *tx_base;
117 unsigned long tx_head;
118 unsigned long tx_tail;
119 unsigned long tx_num_entries;
122 unsigned long tx_acked;
124 struct ldc_packet *rx_base;
125 unsigned long rx_head;
126 unsigned long rx_tail;
127 unsigned long rx_num_entries;
133 unsigned long chan_state;
135 struct ldc_channel_config cfg;
138 const struct ldc_mode_ops *mops;
140 struct ldc_iommu iommu;
142 struct ldc_version ver;
145 #define LDC_HS_CLOSED 0x00
146 #define LDC_HS_OPEN 0x01
147 #define LDC_HS_GOTVERS 0x02
148 #define LDC_HS_SENTRTR 0x03
149 #define LDC_HS_GOTRTR 0x04
150 #define LDC_HS_COMPLETE 0x10
153 #define LDC_FLAG_ALLOCED_QUEUES 0x01
154 #define LDC_FLAG_REGISTERED_QUEUES 0x02
155 #define LDC_FLAG_REGISTERED_IRQS 0x04
156 #define LDC_FLAG_RESET 0x10
161 #define LDC_IRQ_NAME_MAX 32
162 char rx_irq_name[LDC_IRQ_NAME_MAX];
163 char tx_irq_name[LDC_IRQ_NAME_MAX];
165 struct hlist_head mh_list;
167 struct hlist_node list;
170 #define ldcdbg(TYPE, f, a...) \
171 do { if (lp->cfg.debug & LDC_DEBUG_##TYPE) \
172 printk(KERN_INFO PFX "ID[%lu] " f, lp->id, ## a); \
175 static const char *state_to_str(u8 state)
178 case LDC_STATE_INVALID:
182 case LDC_STATE_BOUND:
184 case LDC_STATE_READY:
186 case LDC_STATE_CONNECTED:
193 static void ldc_set_state(struct ldc_channel *lp, u8 state)
195 ldcdbg(STATE, "STATE (%s) --> (%s)\n",
196 state_to_str(lp->state),
197 state_to_str(state));
202 static unsigned long __advance(unsigned long off, unsigned long num_entries)
204 off += LDC_PACKET_SIZE;
205 if (off == (num_entries * LDC_PACKET_SIZE))
211 static unsigned long rx_advance(struct ldc_channel *lp, unsigned long off)
213 return __advance(off, lp->rx_num_entries);
216 static unsigned long tx_advance(struct ldc_channel *lp, unsigned long off)
218 return __advance(off, lp->tx_num_entries);
221 static struct ldc_packet *handshake_get_tx_packet(struct ldc_channel *lp,
222 unsigned long *new_tail)
224 struct ldc_packet *p;
227 t = tx_advance(lp, lp->tx_tail);
228 if (t == lp->tx_head)
234 return p + (lp->tx_tail / LDC_PACKET_SIZE);
237 /* When we are in reliable or stream mode, have to track the next packet
238 * we haven't gotten an ACK for in the TX queue using tx_acked. We have
239 * to be careful not to stomp over the queue past that point. During
240 * the handshake, we don't have TX data packets pending in the queue
241 * and that's why handshake_get_tx_packet() need not be mindful of
244 static unsigned long head_for_data(struct ldc_channel *lp)
246 if (lp->cfg.mode == LDC_MODE_STREAM)
251 static int tx_has_space_for(struct ldc_channel *lp, unsigned int size)
253 unsigned long limit, tail, new_tail, diff;
256 limit = head_for_data(lp);
258 new_tail = tx_advance(lp, tail);
259 if (new_tail == limit)
262 if (limit > new_tail)
263 diff = limit - new_tail;
266 ((lp->tx_num_entries * LDC_PACKET_SIZE) - new_tail));
267 diff /= LDC_PACKET_SIZE;
270 if (diff * mss < size)
276 static struct ldc_packet *data_get_tx_packet(struct ldc_channel *lp,
277 unsigned long *new_tail)
279 struct ldc_packet *p;
282 h = head_for_data(lp);
283 t = tx_advance(lp, lp->tx_tail);
290 return p + (lp->tx_tail / LDC_PACKET_SIZE);
293 static int set_tx_tail(struct ldc_channel *lp, unsigned long tail)
295 unsigned long orig_tail = lp->tx_tail;
299 while (limit-- > 0) {
302 err = sun4v_ldc_tx_set_qtail(lp->id, tail);
306 if (err != HV_EWOULDBLOCK) {
307 lp->tx_tail = orig_tail;
313 lp->tx_tail = orig_tail;
317 /* This just updates the head value in the hypervisor using
318 * a polling loop with a timeout. The caller takes care of
319 * upating software state representing the head change, if any.
321 static int __set_rx_head(struct ldc_channel *lp, unsigned long head)
325 while (limit-- > 0) {
328 err = sun4v_ldc_rx_set_qhead(lp->id, head);
332 if (err != HV_EWOULDBLOCK)
341 static int send_tx_packet(struct ldc_channel *lp,
342 struct ldc_packet *p,
343 unsigned long new_tail)
345 BUG_ON(p != (lp->tx_base + (lp->tx_tail / LDC_PACKET_SIZE)));
347 return set_tx_tail(lp, new_tail);
350 static struct ldc_packet *handshake_compose_ctrl(struct ldc_channel *lp,
352 void *data, int dlen,
353 unsigned long *new_tail)
355 struct ldc_packet *p = handshake_get_tx_packet(lp, new_tail);
358 memset(p, 0, sizeof(*p));
363 memcpy(p->u.u_data, data, dlen);
368 static int start_handshake(struct ldc_channel *lp)
370 struct ldc_packet *p;
371 struct ldc_version *ver;
372 unsigned long new_tail;
376 ldcdbg(HS, "SEND VER INFO maj[%u] min[%u]\n",
377 ver->major, ver->minor);
379 p = handshake_compose_ctrl(lp, LDC_INFO, LDC_VERS,
380 ver, sizeof(*ver), &new_tail);
382 int err = send_tx_packet(lp, p, new_tail);
384 lp->flags &= ~LDC_FLAG_RESET;
390 static int send_version_nack(struct ldc_channel *lp,
391 u16 major, u16 minor)
393 struct ldc_packet *p;
394 struct ldc_version ver;
395 unsigned long new_tail;
400 p = handshake_compose_ctrl(lp, LDC_NACK, LDC_VERS,
401 &ver, sizeof(ver), &new_tail);
403 ldcdbg(HS, "SEND VER NACK maj[%u] min[%u]\n",
404 ver.major, ver.minor);
406 return send_tx_packet(lp, p, new_tail);
411 static int send_version_ack(struct ldc_channel *lp,
412 struct ldc_version *vp)
414 struct ldc_packet *p;
415 unsigned long new_tail;
417 p = handshake_compose_ctrl(lp, LDC_ACK, LDC_VERS,
418 vp, sizeof(*vp), &new_tail);
420 ldcdbg(HS, "SEND VER ACK maj[%u] min[%u]\n",
421 vp->major, vp->minor);
423 return send_tx_packet(lp, p, new_tail);
428 static int send_rts(struct ldc_channel *lp)
430 struct ldc_packet *p;
431 unsigned long new_tail;
433 p = handshake_compose_ctrl(lp, LDC_INFO, LDC_RTS, NULL, 0,
436 p->env = lp->cfg.mode;
440 ldcdbg(HS, "SEND RTS env[0x%x] seqid[0x%x]\n",
443 return send_tx_packet(lp, p, new_tail);
448 static int send_rtr(struct ldc_channel *lp)
450 struct ldc_packet *p;
451 unsigned long new_tail;
453 p = handshake_compose_ctrl(lp, LDC_INFO, LDC_RTR, NULL, 0,
456 p->env = lp->cfg.mode;
459 ldcdbg(HS, "SEND RTR env[0x%x] seqid[0x%x]\n",
462 return send_tx_packet(lp, p, new_tail);
467 static int send_rdx(struct ldc_channel *lp)
469 struct ldc_packet *p;
470 unsigned long new_tail;
472 p = handshake_compose_ctrl(lp, LDC_INFO, LDC_RDX, NULL, 0,
476 p->seqid = ++lp->snd_nxt;
477 p->u.r.ackid = lp->rcv_nxt;
479 ldcdbg(HS, "SEND RDX env[0x%x] seqid[0x%x] ackid[0x%x]\n",
480 p->env, p->seqid, p->u.r.ackid);
482 return send_tx_packet(lp, p, new_tail);
487 static int send_data_nack(struct ldc_channel *lp, struct ldc_packet *data_pkt)
489 struct ldc_packet *p;
490 unsigned long new_tail;
493 p = data_get_tx_packet(lp, &new_tail);
496 memset(p, 0, sizeof(*p));
497 p->type = data_pkt->type;
499 p->ctrl = data_pkt->ctrl & LDC_CTRL_MSK;
500 p->seqid = lp->snd_nxt + 1;
501 p->u.r.ackid = lp->rcv_nxt;
503 ldcdbg(HS, "SEND DATA NACK type[0x%x] ctl[0x%x] seq[0x%x] ack[0x%x]\n",
504 p->type, p->ctrl, p->seqid, p->u.r.ackid);
506 err = send_tx_packet(lp, p, new_tail);
513 static int ldc_abort(struct ldc_channel *lp)
515 unsigned long hv_err;
517 ldcdbg(STATE, "ABORT\n");
519 /* We report but do not act upon the hypervisor errors because
520 * there really isn't much we can do if they fail at this point.
522 hv_err = sun4v_ldc_tx_qconf(lp->id, lp->tx_ra, lp->tx_num_entries);
524 printk(KERN_ERR PFX "ldc_abort: "
525 "sun4v_ldc_tx_qconf(%lx,%lx,%lx) failed, err=%lu\n",
526 lp->id, lp->tx_ra, lp->tx_num_entries, hv_err);
528 hv_err = sun4v_ldc_tx_get_state(lp->id,
533 printk(KERN_ERR PFX "ldc_abort: "
534 "sun4v_ldc_tx_get_state(%lx,...) failed, err=%lu\n",
537 hv_err = sun4v_ldc_rx_qconf(lp->id, lp->rx_ra, lp->rx_num_entries);
539 printk(KERN_ERR PFX "ldc_abort: "
540 "sun4v_ldc_rx_qconf(%lx,%lx,%lx) failed, err=%lu\n",
541 lp->id, lp->rx_ra, lp->rx_num_entries, hv_err);
543 /* Refetch the RX queue state as well, because we could be invoked
544 * here in the queue processing context.
546 hv_err = sun4v_ldc_rx_get_state(lp->id,
551 printk(KERN_ERR PFX "ldc_abort: "
552 "sun4v_ldc_rx_get_state(%lx,...) failed, err=%lu\n",
558 static struct ldc_version *find_by_major(u16 major)
560 struct ldc_version *ret = NULL;
563 for (i = 0; i < ARRAY_SIZE(ver_arr); i++) {
564 struct ldc_version *v = &ver_arr[i];
565 if (v->major <= major) {
573 static int process_ver_info(struct ldc_channel *lp, struct ldc_version *vp)
575 struct ldc_version *vap;
578 ldcdbg(HS, "GOT VERSION INFO major[%x] minor[%x]\n",
579 vp->major, vp->minor);
581 if (lp->hs_state == LDC_HS_GOTVERS) {
582 lp->hs_state = LDC_HS_OPEN;
583 memset(&lp->ver, 0, sizeof(lp->ver));
586 vap = find_by_major(vp->major);
588 err = send_version_nack(lp, 0, 0);
589 } else if (vap->major != vp->major) {
590 err = send_version_nack(lp, vap->major, vap->minor);
592 struct ldc_version ver = *vp;
593 if (ver.minor > vap->minor)
594 ver.minor = vap->minor;
595 err = send_version_ack(lp, &ver);
598 lp->hs_state = LDC_HS_GOTVERS;
602 return ldc_abort(lp);
607 static int process_ver_ack(struct ldc_channel *lp, struct ldc_version *vp)
609 ldcdbg(HS, "GOT VERSION ACK major[%x] minor[%x]\n",
610 vp->major, vp->minor);
612 if (lp->hs_state == LDC_HS_GOTVERS) {
613 if (lp->ver.major != vp->major ||
614 lp->ver.minor != vp->minor)
615 return ldc_abort(lp);
618 lp->hs_state = LDC_HS_GOTVERS;
621 return ldc_abort(lp);
625 static int process_ver_nack(struct ldc_channel *lp, struct ldc_version *vp)
627 struct ldc_version *vap;
628 struct ldc_packet *p;
629 unsigned long new_tail;
631 if (vp->major == 0 && vp->minor == 0)
632 return ldc_abort(lp);
634 vap = find_by_major(vp->major);
636 return ldc_abort(lp);
638 p = handshake_compose_ctrl(lp, LDC_INFO, LDC_VERS,
642 return ldc_abort(lp);
644 return send_tx_packet(lp, p, new_tail);
647 static int process_version(struct ldc_channel *lp,
648 struct ldc_packet *p)
650 struct ldc_version *vp;
652 vp = (struct ldc_version *) p->u.u_data;
656 return process_ver_info(lp, vp);
659 return process_ver_ack(lp, vp);
662 return process_ver_nack(lp, vp);
665 return ldc_abort(lp);
669 static int process_rts(struct ldc_channel *lp,
670 struct ldc_packet *p)
672 ldcdbg(HS, "GOT RTS stype[%x] seqid[%x] env[%x]\n",
673 p->stype, p->seqid, p->env);
675 if (p->stype != LDC_INFO ||
676 lp->hs_state != LDC_HS_GOTVERS ||
677 p->env != lp->cfg.mode)
678 return ldc_abort(lp);
680 lp->snd_nxt = p->seqid;
681 lp->rcv_nxt = p->seqid;
682 lp->hs_state = LDC_HS_SENTRTR;
684 return ldc_abort(lp);
689 static int process_rtr(struct ldc_channel *lp,
690 struct ldc_packet *p)
692 ldcdbg(HS, "GOT RTR stype[%x] seqid[%x] env[%x]\n",
693 p->stype, p->seqid, p->env);
695 if (p->stype != LDC_INFO ||
696 p->env != lp->cfg.mode)
697 return ldc_abort(lp);
699 lp->snd_nxt = p->seqid;
700 lp->hs_state = LDC_HS_COMPLETE;
701 ldc_set_state(lp, LDC_STATE_CONNECTED);
707 static int rx_seq_ok(struct ldc_channel *lp, u32 seqid)
709 return lp->rcv_nxt + 1 == seqid;
712 static int process_rdx(struct ldc_channel *lp,
713 struct ldc_packet *p)
715 ldcdbg(HS, "GOT RDX stype[%x] seqid[%x] env[%x] ackid[%x]\n",
716 p->stype, p->seqid, p->env, p->u.r.ackid);
718 if (p->stype != LDC_INFO ||
719 !(rx_seq_ok(lp, p->seqid)))
720 return ldc_abort(lp);
722 lp->rcv_nxt = p->seqid;
724 lp->hs_state = LDC_HS_COMPLETE;
725 ldc_set_state(lp, LDC_STATE_CONNECTED);
730 static int process_control_frame(struct ldc_channel *lp,
731 struct ldc_packet *p)
735 return process_version(lp, p);
738 return process_rts(lp, p);
741 return process_rtr(lp, p);
744 return process_rdx(lp, p);
747 return ldc_abort(lp);
751 static int process_error_frame(struct ldc_channel *lp,
752 struct ldc_packet *p)
754 return ldc_abort(lp);
757 static int process_data_ack(struct ldc_channel *lp,
758 struct ldc_packet *ack)
760 unsigned long head = lp->tx_acked;
761 u32 ackid = ack->u.r.ackid;
764 struct ldc_packet *p = lp->tx_base + (head / LDC_PACKET_SIZE);
766 head = tx_advance(lp, head);
768 if (p->seqid == ackid) {
772 if (head == lp->tx_tail)
773 return ldc_abort(lp);
779 static void send_events(struct ldc_channel *lp, unsigned int event_mask)
781 if (event_mask & LDC_EVENT_RESET)
782 lp->cfg.event(lp->event_arg, LDC_EVENT_RESET);
783 if (event_mask & LDC_EVENT_UP)
784 lp->cfg.event(lp->event_arg, LDC_EVENT_UP);
785 if (event_mask & LDC_EVENT_DATA_READY)
786 lp->cfg.event(lp->event_arg, LDC_EVENT_DATA_READY);
789 static irqreturn_t ldc_rx(int irq, void *dev_id)
791 struct ldc_channel *lp = dev_id;
792 unsigned long orig_state, hv_err, flags;
793 unsigned int event_mask;
795 spin_lock_irqsave(&lp->lock, flags);
797 orig_state = lp->chan_state;
798 hv_err = sun4v_ldc_rx_get_state(lp->id,
803 ldcdbg(RX, "RX state[0x%02lx:0x%02lx] head[0x%04lx] tail[0x%04lx]\n",
804 orig_state, lp->chan_state, lp->rx_head, lp->rx_tail);
808 if (lp->cfg.mode == LDC_MODE_RAW &&
809 lp->chan_state == LDC_CHANNEL_UP) {
810 lp->hs_state = LDC_HS_COMPLETE;
811 ldc_set_state(lp, LDC_STATE_CONNECTED);
813 event_mask |= LDC_EVENT_UP;
815 orig_state = lp->chan_state;
818 /* If we are in reset state, flush the RX queue and ignore
821 if (lp->flags & LDC_FLAG_RESET) {
822 (void) __set_rx_head(lp, lp->rx_tail);
826 /* Once we finish the handshake, we let the ldc_read()
827 * paths do all of the control frame and state management.
828 * Just trigger the callback.
830 if (lp->hs_state == LDC_HS_COMPLETE) {
832 if (lp->chan_state != orig_state) {
833 unsigned int event = LDC_EVENT_RESET;
835 if (lp->chan_state == LDC_CHANNEL_UP)
836 event = LDC_EVENT_UP;
840 if (lp->rx_head != lp->rx_tail)
841 event_mask |= LDC_EVENT_DATA_READY;
846 if (lp->chan_state != orig_state)
849 while (lp->rx_head != lp->rx_tail) {
850 struct ldc_packet *p;
854 p = lp->rx_base + (lp->rx_head / LDC_PACKET_SIZE);
858 err = process_control_frame(lp, p);
864 event_mask |= LDC_EVENT_DATA_READY;
869 err = process_error_frame(lp, p);
881 new += LDC_PACKET_SIZE;
882 if (new == (lp->rx_num_entries * LDC_PACKET_SIZE))
886 err = __set_rx_head(lp, new);
888 (void) ldc_abort(lp);
891 if (lp->hs_state == LDC_HS_COMPLETE)
892 goto handshake_complete;
896 spin_unlock_irqrestore(&lp->lock, flags);
898 send_events(lp, event_mask);
903 static irqreturn_t ldc_tx(int irq, void *dev_id)
905 struct ldc_channel *lp = dev_id;
906 unsigned long flags, hv_err, orig_state;
907 unsigned int event_mask = 0;
909 spin_lock_irqsave(&lp->lock, flags);
911 orig_state = lp->chan_state;
912 hv_err = sun4v_ldc_tx_get_state(lp->id,
917 ldcdbg(TX, " TX state[0x%02lx:0x%02lx] head[0x%04lx] tail[0x%04lx]\n",
918 orig_state, lp->chan_state, lp->tx_head, lp->tx_tail);
920 if (lp->cfg.mode == LDC_MODE_RAW &&
921 lp->chan_state == LDC_CHANNEL_UP) {
922 lp->hs_state = LDC_HS_COMPLETE;
923 ldc_set_state(lp, LDC_STATE_CONNECTED);
925 event_mask |= LDC_EVENT_UP;
928 spin_unlock_irqrestore(&lp->lock, flags);
930 send_events(lp, event_mask);
935 /* XXX ldc_alloc() and ldc_free() needs to run under a mutex so
936 * XXX that addition and removal from the ldc_channel_list has
937 * XXX atomicity, otherwise the __ldc_channel_exists() check is
938 * XXX totally pointless as another thread can slip into ldc_alloc()
939 * XXX and add a channel with the same ID. There also needs to be
940 * XXX a spinlock for ldc_channel_list.
942 static HLIST_HEAD(ldc_channel_list);
944 static int __ldc_channel_exists(unsigned long id)
946 struct ldc_channel *lp;
947 struct hlist_node *n;
949 hlist_for_each_entry(lp, n, &ldc_channel_list, list) {
956 static int alloc_queue(const char *name, unsigned long num_entries,
957 struct ldc_packet **base, unsigned long *ra)
959 unsigned long size, order;
962 size = num_entries * LDC_PACKET_SIZE;
963 order = get_order(size);
965 q = (void *) __get_free_pages(GFP_KERNEL, order);
967 printk(KERN_ERR PFX "Alloc of %s queue failed with "
968 "size=%lu order=%lu\n", name, size, order);
972 memset(q, 0, PAGE_SIZE << order);
980 static void free_queue(unsigned long num_entries, struct ldc_packet *q)
982 unsigned long size, order;
987 size = num_entries * LDC_PACKET_SIZE;
988 order = get_order(size);
990 free_pages((unsigned long)q, order);
993 /* XXX Make this configurable... XXX */
994 #define LDC_IOTABLE_SIZE (8 * 1024)
996 static int ldc_iommu_init(struct ldc_channel *lp)
998 unsigned long sz, num_tsb_entries, tsbsize, order;
999 struct ldc_iommu *iommu = &lp->iommu;
1000 struct ldc_mtable_entry *table;
1001 unsigned long hv_err;
1004 num_tsb_entries = LDC_IOTABLE_SIZE;
1005 tsbsize = num_tsb_entries * sizeof(struct ldc_mtable_entry);
1007 spin_lock_init(&iommu->lock);
1009 sz = num_tsb_entries / 8;
1010 sz = (sz + 7UL) & ~7UL;
1011 iommu->arena.map = kzalloc(sz, GFP_KERNEL);
1012 if (!iommu->arena.map) {
1013 printk(KERN_ERR PFX "Alloc of arena map failed, sz=%lu\n", sz);
1017 iommu->arena.limit = num_tsb_entries;
1019 order = get_order(tsbsize);
1021 table = (struct ldc_mtable_entry *)
1022 __get_free_pages(GFP_KERNEL, order);
1025 printk(KERN_ERR PFX "Alloc of MTE table failed, "
1026 "size=%lu order=%lu\n", tsbsize, order);
1030 memset(table, 0, PAGE_SIZE << order);
1032 iommu->page_table = table;
1034 hv_err = sun4v_ldc_set_map_table(lp->id, __pa(table),
1038 goto out_free_table;
1043 free_pages((unsigned long) table, order);
1044 iommu->page_table = NULL;
1047 kfree(iommu->arena.map);
1048 iommu->arena.map = NULL;
1053 static void ldc_iommu_release(struct ldc_channel *lp)
1055 struct ldc_iommu *iommu = &lp->iommu;
1056 unsigned long num_tsb_entries, tsbsize, order;
1058 (void) sun4v_ldc_set_map_table(lp->id, 0, 0);
1060 num_tsb_entries = iommu->arena.limit;
1061 tsbsize = num_tsb_entries * sizeof(struct ldc_mtable_entry);
1062 order = get_order(tsbsize);
1064 free_pages((unsigned long) iommu->page_table, order);
1065 iommu->page_table = NULL;
1067 kfree(iommu->arena.map);
1068 iommu->arena.map = NULL;
1071 struct ldc_channel *ldc_alloc(unsigned long id,
1072 const struct ldc_channel_config *cfgp,
1075 struct ldc_channel *lp;
1076 const struct ldc_mode_ops *mops;
1077 unsigned long dummy1, dummy2, hv_err;
1082 if (!ldom_domaining_enabled)
1089 switch (cfgp->mode) {
1092 mss = LDC_PACKET_SIZE;
1095 case LDC_MODE_UNRELIABLE:
1097 mss = LDC_PACKET_SIZE - 8;
1100 case LDC_MODE_STREAM:
1102 mss = LDC_PACKET_SIZE - 8 - 8;
1109 if (!cfgp->event || !event_arg || !cfgp->rx_irq || !cfgp->tx_irq)
1112 hv_err = sun4v_ldc_tx_qinfo(id, &dummy1, &dummy2);
1114 if (hv_err == HV_ECHANNEL)
1118 if (__ldc_channel_exists(id))
1123 lp = kzalloc(sizeof(*lp), GFP_KERNEL);
1128 spin_lock_init(&lp->lock);
1132 err = ldc_iommu_init(lp);
1141 lp->cfg.mtu = LDC_DEFAULT_MTU;
1143 if (lp->cfg.mode == LDC_MODE_STREAM) {
1144 mssbuf = kzalloc(lp->cfg.mtu, GFP_KERNEL);
1147 goto out_free_iommu;
1149 lp->mssbuf = mssbuf;
1152 lp->event_arg = event_arg;
1154 /* XXX allow setting via ldc_channel_config to override defaults
1155 * XXX or use some formula based upon mtu
1157 lp->tx_num_entries = LDC_DEFAULT_NUM_ENTRIES;
1158 lp->rx_num_entries = LDC_DEFAULT_NUM_ENTRIES;
1160 err = alloc_queue("TX", lp->tx_num_entries,
1161 &lp->tx_base, &lp->tx_ra);
1163 goto out_free_mssbuf;
1165 err = alloc_queue("RX", lp->rx_num_entries,
1166 &lp->rx_base, &lp->rx_ra);
1170 lp->flags |= LDC_FLAG_ALLOCED_QUEUES;
1172 lp->hs_state = LDC_HS_CLOSED;
1173 ldc_set_state(lp, LDC_STATE_INIT);
1175 INIT_HLIST_NODE(&lp->list);
1176 hlist_add_head(&lp->list, &ldc_channel_list);
1178 INIT_HLIST_HEAD(&lp->mh_list);
1183 free_queue(lp->tx_num_entries, lp->tx_base);
1189 ldc_iommu_release(lp);
1195 return ERR_PTR(err);
1197 EXPORT_SYMBOL(ldc_alloc);
1199 void ldc_free(struct ldc_channel *lp)
1201 if (lp->flags & LDC_FLAG_REGISTERED_IRQS) {
1202 free_irq(lp->cfg.rx_irq, lp);
1203 free_irq(lp->cfg.tx_irq, lp);
1206 if (lp->flags & LDC_FLAG_REGISTERED_QUEUES) {
1207 sun4v_ldc_tx_qconf(lp->id, 0, 0);
1208 sun4v_ldc_rx_qconf(lp->id, 0, 0);
1209 lp->flags &= ~LDC_FLAG_REGISTERED_QUEUES;
1211 if (lp->flags & LDC_FLAG_ALLOCED_QUEUES) {
1212 free_queue(lp->tx_num_entries, lp->tx_base);
1213 free_queue(lp->rx_num_entries, lp->rx_base);
1214 lp->flags &= ~LDC_FLAG_ALLOCED_QUEUES;
1217 hlist_del(&lp->list);
1221 ldc_iommu_release(lp);
1225 EXPORT_SYMBOL(ldc_free);
1227 /* Bind the channel. This registers the LDC queues with
1228 * the hypervisor and puts the channel into a pseudo-listening
1229 * state. This does not initiate a handshake, ldc_connect() does
1232 int ldc_bind(struct ldc_channel *lp, const char *name)
1234 unsigned long hv_err, flags;
1238 (lp->state != LDC_STATE_INIT))
1241 snprintf(lp->rx_irq_name, LDC_IRQ_NAME_MAX, "%s RX", name);
1242 snprintf(lp->tx_irq_name, LDC_IRQ_NAME_MAX, "%s TX", name);
1244 err = request_irq(lp->cfg.rx_irq, ldc_rx,
1245 IRQF_SAMPLE_RANDOM | IRQF_SHARED,
1246 lp->rx_irq_name, lp);
1250 err = request_irq(lp->cfg.tx_irq, ldc_tx,
1251 IRQF_SAMPLE_RANDOM | IRQF_SHARED,
1252 lp->tx_irq_name, lp);
1254 free_irq(lp->cfg.rx_irq, lp);
1259 spin_lock_irqsave(&lp->lock, flags);
1261 enable_irq(lp->cfg.rx_irq);
1262 enable_irq(lp->cfg.tx_irq);
1264 lp->flags |= LDC_FLAG_REGISTERED_IRQS;
1267 hv_err = sun4v_ldc_tx_qconf(lp->id, 0, 0);
1271 hv_err = sun4v_ldc_tx_qconf(lp->id, lp->tx_ra, lp->tx_num_entries);
1275 hv_err = sun4v_ldc_rx_qconf(lp->id, 0, 0);
1279 hv_err = sun4v_ldc_rx_qconf(lp->id, lp->rx_ra, lp->rx_num_entries);
1283 lp->flags |= LDC_FLAG_REGISTERED_QUEUES;
1285 hv_err = sun4v_ldc_tx_get_state(lp->id,
1293 lp->tx_acked = lp->tx_head;
1295 lp->hs_state = LDC_HS_OPEN;
1296 ldc_set_state(lp, LDC_STATE_BOUND);
1298 spin_unlock_irqrestore(&lp->lock, flags);
1303 lp->flags &= ~LDC_FLAG_REGISTERED_QUEUES;
1304 sun4v_ldc_rx_qconf(lp->id, 0, 0);
1307 sun4v_ldc_tx_qconf(lp->id, 0, 0);
1310 lp->flags &= ~LDC_FLAG_REGISTERED_IRQS;
1311 free_irq(lp->cfg.tx_irq, lp);
1312 free_irq(lp->cfg.rx_irq, lp);
1314 spin_unlock_irqrestore(&lp->lock, flags);
1318 EXPORT_SYMBOL(ldc_bind);
1320 int ldc_connect(struct ldc_channel *lp)
1322 unsigned long flags;
1325 if (lp->cfg.mode == LDC_MODE_RAW)
1328 spin_lock_irqsave(&lp->lock, flags);
1330 if (!(lp->flags & LDC_FLAG_ALLOCED_QUEUES) ||
1331 !(lp->flags & LDC_FLAG_REGISTERED_QUEUES) ||
1332 lp->hs_state != LDC_HS_OPEN)
1335 err = start_handshake(lp);
1337 spin_unlock_irqrestore(&lp->lock, flags);
1341 EXPORT_SYMBOL(ldc_connect);
1343 int ldc_disconnect(struct ldc_channel *lp)
1345 unsigned long hv_err, flags;
1348 if (lp->cfg.mode == LDC_MODE_RAW)
1351 if (!(lp->flags & LDC_FLAG_ALLOCED_QUEUES) ||
1352 !(lp->flags & LDC_FLAG_REGISTERED_QUEUES))
1355 spin_lock_irqsave(&lp->lock, flags);
1358 hv_err = sun4v_ldc_tx_qconf(lp->id, 0, 0);
1362 hv_err = sun4v_ldc_tx_qconf(lp->id, lp->tx_ra, lp->tx_num_entries);
1366 hv_err = sun4v_ldc_rx_qconf(lp->id, 0, 0);
1370 hv_err = sun4v_ldc_rx_qconf(lp->id, lp->rx_ra, lp->rx_num_entries);
1374 ldc_set_state(lp, LDC_STATE_BOUND);
1375 lp->hs_state = LDC_HS_OPEN;
1376 lp->flags |= LDC_FLAG_RESET;
1378 spin_unlock_irqrestore(&lp->lock, flags);
1383 sun4v_ldc_tx_qconf(lp->id, 0, 0);
1384 sun4v_ldc_rx_qconf(lp->id, 0, 0);
1385 free_irq(lp->cfg.tx_irq, lp);
1386 free_irq(lp->cfg.rx_irq, lp);
1387 lp->flags &= ~(LDC_FLAG_REGISTERED_IRQS |
1388 LDC_FLAG_REGISTERED_QUEUES);
1389 ldc_set_state(lp, LDC_STATE_INIT);
1391 spin_unlock_irqrestore(&lp->lock, flags);
1395 EXPORT_SYMBOL(ldc_disconnect);
1397 int ldc_state(struct ldc_channel *lp)
1401 EXPORT_SYMBOL(ldc_state);
1403 static int write_raw(struct ldc_channel *lp, const void *buf, unsigned int size)
1405 struct ldc_packet *p;
1406 unsigned long new_tail;
1409 if (size > LDC_PACKET_SIZE)
1412 p = data_get_tx_packet(lp, &new_tail);
1416 memcpy(p, buf, size);
1418 err = send_tx_packet(lp, p, new_tail);
1425 static int read_raw(struct ldc_channel *lp, void *buf, unsigned int size)
1427 struct ldc_packet *p;
1428 unsigned long hv_err, new;
1431 if (size < LDC_PACKET_SIZE)
1434 hv_err = sun4v_ldc_rx_get_state(lp->id,
1439 return ldc_abort(lp);
1441 if (lp->chan_state == LDC_CHANNEL_DOWN ||
1442 lp->chan_state == LDC_CHANNEL_RESETTING)
1445 if (lp->rx_head == lp->rx_tail)
1448 p = lp->rx_base + (lp->rx_head / LDC_PACKET_SIZE);
1449 memcpy(buf, p, LDC_PACKET_SIZE);
1451 new = rx_advance(lp, lp->rx_head);
1454 err = __set_rx_head(lp, new);
1458 err = LDC_PACKET_SIZE;
1463 static const struct ldc_mode_ops raw_ops = {
1468 static int write_nonraw(struct ldc_channel *lp, const void *buf,
1471 unsigned long hv_err, tail;
1472 unsigned int copied;
1476 hv_err = sun4v_ldc_tx_get_state(lp->id, &lp->tx_head, &lp->tx_tail,
1478 if (unlikely(hv_err))
1481 if (unlikely(lp->chan_state != LDC_CHANNEL_UP))
1482 return ldc_abort(lp);
1484 if (!tx_has_space_for(lp, size))
1490 while (copied < size) {
1491 struct ldc_packet *p = lp->tx_base + (tail / LDC_PACKET_SIZE);
1492 u8 *data = ((lp->cfg.mode == LDC_MODE_UNRELIABLE) ?
1498 p->stype = LDC_INFO;
1501 data_len = size - copied;
1502 if (data_len > lp->mss)
1505 BUG_ON(data_len > LDC_LEN);
1507 p->env = (data_len |
1508 (copied == 0 ? LDC_START : 0) |
1509 (data_len == size - copied ? LDC_STOP : 0));
1513 ldcdbg(DATA, "SENT DATA [%02x:%02x:%02x:%02x:%08x]\n",
1520 memcpy(data, buf, data_len);
1524 tail = tx_advance(lp, tail);
1527 err = set_tx_tail(lp, tail);
1536 static int rx_bad_seq(struct ldc_channel *lp, struct ldc_packet *p,
1537 struct ldc_packet *first_frag)
1542 lp->rcv_nxt = first_frag->seqid - 1;
1544 err = send_data_nack(lp, p);
1548 err = __set_rx_head(lp, lp->rx_tail);
1550 return ldc_abort(lp);
1555 static int data_ack_nack(struct ldc_channel *lp, struct ldc_packet *p)
1557 if (p->stype & LDC_ACK) {
1558 int err = process_data_ack(lp, p);
1562 if (p->stype & LDC_NACK)
1563 return ldc_abort(lp);
1568 static int rx_data_wait(struct ldc_channel *lp, unsigned long cur_head)
1570 unsigned long dummy;
1573 ldcdbg(DATA, "DATA WAIT cur_head[%lx] rx_head[%lx] rx_tail[%lx]\n",
1574 cur_head, lp->rx_head, lp->rx_tail);
1575 while (limit-- > 0) {
1576 unsigned long hv_err;
1578 hv_err = sun4v_ldc_rx_get_state(lp->id,
1583 return ldc_abort(lp);
1585 if (lp->chan_state == LDC_CHANNEL_DOWN ||
1586 lp->chan_state == LDC_CHANNEL_RESETTING)
1589 if (cur_head != lp->rx_tail) {
1590 ldcdbg(DATA, "DATA WAIT DONE "
1591 "head[%lx] tail[%lx] chan_state[%lx]\n",
1592 dummy, lp->rx_tail, lp->chan_state);
1601 static int rx_set_head(struct ldc_channel *lp, unsigned long head)
1603 int err = __set_rx_head(lp, head);
1606 return ldc_abort(lp);
1612 static void send_data_ack(struct ldc_channel *lp)
1614 unsigned long new_tail;
1615 struct ldc_packet *p;
1617 p = data_get_tx_packet(lp, &new_tail);
1621 memset(p, 0, sizeof(*p));
1625 p->seqid = lp->snd_nxt + 1;
1626 p->u.r.ackid = lp->rcv_nxt;
1628 err = send_tx_packet(lp, p, new_tail);
1634 static int read_nonraw(struct ldc_channel *lp, void *buf, unsigned int size)
1636 struct ldc_packet *first_frag;
1637 unsigned long hv_err, new;
1640 hv_err = sun4v_ldc_rx_get_state(lp->id,
1645 return ldc_abort(lp);
1647 if (lp->chan_state == LDC_CHANNEL_DOWN ||
1648 lp->chan_state == LDC_CHANNEL_RESETTING)
1651 if (lp->rx_head == lp->rx_tail)
1658 struct ldc_packet *p;
1661 BUG_ON(new == lp->rx_tail);
1662 p = lp->rx_base + (new / LDC_PACKET_SIZE);
1664 ldcdbg(RX, "RX read pkt[%02x:%02x:%02x:%02x:%08x:%08x] "
1674 if (unlikely(!rx_seq_ok(lp, p->seqid))) {
1675 err = rx_bad_seq(lp, p, first_frag);
1680 if (p->type & LDC_CTRL) {
1681 err = process_control_frame(lp, p);
1687 lp->rcv_nxt = p->seqid;
1689 if (!(p->type & LDC_DATA)) {
1690 new = rx_advance(lp, new);
1693 if (p->stype & (LDC_ACK | LDC_NACK)) {
1694 err = data_ack_nack(lp, p);
1698 if (!(p->stype & LDC_INFO)) {
1699 new = rx_advance(lp, new);
1700 err = rx_set_head(lp, new);
1706 pkt_len = p->env & LDC_LEN;
1708 /* Every initial packet starts with the START bit set.
1710 * Singleton packets will have both START+STOP set.
1712 * Fragments will have START set in the first frame, STOP
1713 * set in the last frame, and neither bit set in middle
1714 * frames of the packet.
1716 * Therefore if we are at the beginning of a packet and
1717 * we don't see START, or we are in the middle of a fragmented
1718 * packet and do see START, we are unsynchronized and should
1719 * flush the RX queue.
1721 if ((first_frag == NULL && !(p->env & LDC_START)) ||
1722 (first_frag != NULL && (p->env & LDC_START))) {
1724 new = rx_advance(lp, new);
1726 err = rx_set_head(lp, new);
1736 if (pkt_len > size - copied) {
1737 /* User didn't give us a big enough buffer,
1738 * what to do? This is a pretty serious error.
1740 * Since we haven't updated the RX ring head to
1741 * consume any of the packets, signal the error
1742 * to the user and just leave the RX ring alone.
1744 * This seems the best behavior because this allows
1745 * a user of the LDC layer to start with a small
1746 * RX buffer for ldc_read() calls and use -EMSGSIZE
1747 * as a cue to enlarge it's read buffer.
1753 /* Ok, we are gonna eat this one. */
1754 new = rx_advance(lp, new);
1757 (lp->cfg.mode == LDC_MODE_UNRELIABLE ?
1758 p->u.u_data : p->u.r.r_data), pkt_len);
1762 if (p->env & LDC_STOP)
1766 if (new == lp->rx_tail) {
1767 err = rx_data_wait(lp, new);
1774 err = rx_set_head(lp, new);
1776 if (err && first_frag)
1777 lp->rcv_nxt = first_frag->seqid - 1;
1781 if (err > 0 && lp->cfg.mode != LDC_MODE_UNRELIABLE)
1788 static const struct ldc_mode_ops nonraw_ops = {
1789 .write = write_nonraw,
1790 .read = read_nonraw,
1793 static int write_stream(struct ldc_channel *lp, const void *buf,
1796 if (size > lp->cfg.mtu)
1798 return write_nonraw(lp, buf, size);
1801 static int read_stream(struct ldc_channel *lp, void *buf, unsigned int size)
1803 if (!lp->mssbuf_len) {
1804 int err = read_nonraw(lp, lp->mssbuf, lp->cfg.mtu);
1808 lp->mssbuf_len = err;
1812 if (size > lp->mssbuf_len)
1813 size = lp->mssbuf_len;
1814 memcpy(buf, lp->mssbuf + lp->mssbuf_off, size);
1816 lp->mssbuf_off += size;
1817 lp->mssbuf_len -= size;
1822 static const struct ldc_mode_ops stream_ops = {
1823 .write = write_stream,
1824 .read = read_stream,
1827 int ldc_write(struct ldc_channel *lp, const void *buf, unsigned int size)
1829 unsigned long flags;
1838 spin_lock_irqsave(&lp->lock, flags);
1840 if (lp->hs_state != LDC_HS_COMPLETE)
1843 err = lp->mops->write(lp, buf, size);
1845 spin_unlock_irqrestore(&lp->lock, flags);
1849 EXPORT_SYMBOL(ldc_write);
1851 int ldc_read(struct ldc_channel *lp, void *buf, unsigned int size)
1853 unsigned long flags;
1862 spin_lock_irqsave(&lp->lock, flags);
1864 if (lp->hs_state != LDC_HS_COMPLETE)
1867 err = lp->mops->read(lp, buf, size);
1869 spin_unlock_irqrestore(&lp->lock, flags);
1873 EXPORT_SYMBOL(ldc_read);
1875 static long arena_alloc(struct ldc_iommu *iommu, unsigned long npages)
1877 struct iommu_arena *arena = &iommu->arena;
1878 unsigned long n, i, start, end, limit;
1881 limit = arena->limit;
1882 start = arena->hint;
1886 n = find_next_zero_bit(arena->map, limit, start);
1888 if (unlikely(end >= limit)) {
1889 if (likely(pass < 1)) {
1895 /* Scanned the whole thing, give up. */
1900 for (i = n; i < end; i++) {
1901 if (test_bit(i, arena->map)) {
1907 for (i = n; i < end; i++)
1908 __set_bit(i, arena->map);
1915 #define COOKIE_PGSZ_CODE 0xf000000000000000ULL
1916 #define COOKIE_PGSZ_CODE_SHIFT 60ULL
1918 static u64 pagesize_code(void)
1920 switch (PAGE_SIZE) {
1922 case (8ULL * 1024ULL):
1924 case (64ULL * 1024ULL):
1926 case (512ULL * 1024ULL):
1928 case (4ULL * 1024ULL * 1024ULL):
1930 case (32ULL * 1024ULL * 1024ULL):
1932 case (256ULL * 1024ULL * 1024ULL):
1937 static u64 make_cookie(u64 index, u64 pgsz_code, u64 page_offset)
1939 return ((pgsz_code << COOKIE_PGSZ_CODE_SHIFT) |
1940 (index << PAGE_SHIFT) |
1944 static u64 cookie_to_index(u64 cookie, unsigned long *shift)
1946 u64 szcode = cookie >> COOKIE_PGSZ_CODE_SHIFT;
1948 cookie &= ~COOKIE_PGSZ_CODE;
1950 *shift = szcode * 3;
1952 return (cookie >> (13ULL + (szcode * 3ULL)));
1955 static struct ldc_mtable_entry *alloc_npages(struct ldc_iommu *iommu,
1956 unsigned long npages)
1960 entry = arena_alloc(iommu, npages);
1961 if (unlikely(entry < 0))
1964 return iommu->page_table + entry;
1967 static u64 perm_to_mte(unsigned int map_perm)
1971 mte_base = pagesize_code();
1973 if (map_perm & LDC_MAP_SHADOW) {
1974 if (map_perm & LDC_MAP_R)
1975 mte_base |= LDC_MTE_COPY_R;
1976 if (map_perm & LDC_MAP_W)
1977 mte_base |= LDC_MTE_COPY_W;
1979 if (map_perm & LDC_MAP_DIRECT) {
1980 if (map_perm & LDC_MAP_R)
1981 mte_base |= LDC_MTE_READ;
1982 if (map_perm & LDC_MAP_W)
1983 mte_base |= LDC_MTE_WRITE;
1984 if (map_perm & LDC_MAP_X)
1985 mte_base |= LDC_MTE_EXEC;
1987 if (map_perm & LDC_MAP_IO) {
1988 if (map_perm & LDC_MAP_R)
1989 mte_base |= LDC_MTE_IOMMU_R;
1990 if (map_perm & LDC_MAP_W)
1991 mte_base |= LDC_MTE_IOMMU_W;
1997 static int pages_in_region(unsigned long base, long len)
2002 unsigned long new = (base + PAGE_SIZE) & PAGE_MASK;
2004 len -= (new - base);
2012 struct cookie_state {
2013 struct ldc_mtable_entry *page_table;
2014 struct ldc_trans_cookie *cookies;
2021 static void fill_cookies(struct cookie_state *sp, unsigned long pa,
2022 unsigned long off, unsigned long len)
2025 unsigned long tlen, new = pa + PAGE_SIZE;
2028 sp->page_table[sp->pte_idx].mte = sp->mte_base | pa;
2032 tlen = PAGE_SIZE - off;
2036 this_cookie = make_cookie(sp->pte_idx,
2037 pagesize_code(), off);
2041 if (this_cookie == sp->prev_cookie) {
2042 sp->cookies[sp->nc - 1].cookie_size += tlen;
2044 sp->cookies[sp->nc].cookie_addr = this_cookie;
2045 sp->cookies[sp->nc].cookie_size = tlen;
2048 sp->prev_cookie = this_cookie + tlen;
2057 static int sg_count_one(struct scatterlist *sg)
2059 unsigned long base = page_to_pfn(sg_page(sg)) << PAGE_SHIFT;
2060 long len = sg->length;
2062 if ((sg->offset | len) & (8UL - 1))
2065 return pages_in_region(base + sg->offset, len);
2068 static int sg_count_pages(struct scatterlist *sg, int num_sg)
2074 for (i = 0; i < num_sg; i++) {
2075 int err = sg_count_one(sg + i);
2084 int ldc_map_sg(struct ldc_channel *lp,
2085 struct scatterlist *sg, int num_sg,
2086 struct ldc_trans_cookie *cookies, int ncookies,
2087 unsigned int map_perm)
2089 unsigned long i, npages, flags;
2090 struct ldc_mtable_entry *base;
2091 struct cookie_state state;
2092 struct ldc_iommu *iommu;
2095 if (map_perm & ~LDC_MAP_ALL)
2098 err = sg_count_pages(sg, num_sg);
2108 spin_lock_irqsave(&iommu->lock, flags);
2109 base = alloc_npages(iommu, npages);
2110 spin_unlock_irqrestore(&iommu->lock, flags);
2115 state.page_table = iommu->page_table;
2116 state.cookies = cookies;
2117 state.mte_base = perm_to_mte(map_perm);
2118 state.prev_cookie = ~(u64)0;
2119 state.pte_idx = (base - iommu->page_table);
2122 for (i = 0; i < num_sg; i++)
2123 fill_cookies(&state, page_to_pfn(sg_page(&sg[i])) << PAGE_SHIFT,
2124 sg[i].offset, sg[i].length);
2128 EXPORT_SYMBOL(ldc_map_sg);
2130 int ldc_map_single(struct ldc_channel *lp,
2131 void *buf, unsigned int len,
2132 struct ldc_trans_cookie *cookies, int ncookies,
2133 unsigned int map_perm)
2135 unsigned long npages, pa, flags;
2136 struct ldc_mtable_entry *base;
2137 struct cookie_state state;
2138 struct ldc_iommu *iommu;
2140 if ((map_perm & ~LDC_MAP_ALL) || (ncookies < 1))
2144 if ((pa | len) & (8UL - 1))
2147 npages = pages_in_region(pa, len);
2151 spin_lock_irqsave(&iommu->lock, flags);
2152 base = alloc_npages(iommu, npages);
2153 spin_unlock_irqrestore(&iommu->lock, flags);
2158 state.page_table = iommu->page_table;
2159 state.cookies = cookies;
2160 state.mte_base = perm_to_mte(map_perm);
2161 state.prev_cookie = ~(u64)0;
2162 state.pte_idx = (base - iommu->page_table);
2164 fill_cookies(&state, (pa & PAGE_MASK), (pa & ~PAGE_MASK), len);
2165 BUG_ON(state.nc != 1);
2169 EXPORT_SYMBOL(ldc_map_single);
2171 static void free_npages(unsigned long id, struct ldc_iommu *iommu,
2172 u64 cookie, u64 size)
2174 struct iommu_arena *arena = &iommu->arena;
2175 unsigned long i, shift, index, npages;
2176 struct ldc_mtable_entry *base;
2178 npages = PAGE_ALIGN(((cookie & ~PAGE_MASK) + size)) >> PAGE_SHIFT;
2179 index = cookie_to_index(cookie, &shift);
2180 base = iommu->page_table + index;
2182 BUG_ON(index > arena->limit ||
2183 (index + npages) > arena->limit);
2185 for (i = 0; i < npages; i++) {
2187 sun4v_ldc_revoke(id, cookie + (i << shift),
2190 __clear_bit(index + i, arena->map);
2194 void ldc_unmap(struct ldc_channel *lp, struct ldc_trans_cookie *cookies,
2197 struct ldc_iommu *iommu = &lp->iommu;
2198 unsigned long flags;
2201 spin_lock_irqsave(&iommu->lock, flags);
2202 for (i = 0; i < ncookies; i++) {
2203 u64 addr = cookies[i].cookie_addr;
2204 u64 size = cookies[i].cookie_size;
2206 free_npages(lp->id, iommu, addr, size);
2208 spin_unlock_irqrestore(&iommu->lock, flags);
2210 EXPORT_SYMBOL(ldc_unmap);
2212 int ldc_copy(struct ldc_channel *lp, int copy_dir,
2213 void *buf, unsigned int len, unsigned long offset,
2214 struct ldc_trans_cookie *cookies, int ncookies)
2216 unsigned int orig_len;
2220 if (copy_dir != LDC_COPY_IN && copy_dir != LDC_COPY_OUT) {
2221 printk(KERN_ERR PFX "ldc_copy: ID[%lu] Bad copy_dir[%d]\n",
2227 if ((ra | len | offset) & (8UL - 1)) {
2228 printk(KERN_ERR PFX "ldc_copy: ID[%lu] Unaligned buffer "
2229 "ra[%lx] len[%x] offset[%lx]\n",
2230 lp->id, ra, len, offset);
2234 if (lp->hs_state != LDC_HS_COMPLETE ||
2235 (lp->flags & LDC_FLAG_RESET)) {
2236 printk(KERN_ERR PFX "ldc_copy: ID[%lu] Link down hs_state[%x] "
2237 "flags[%x]\n", lp->id, lp->hs_state, lp->flags);
2242 for (i = 0; i < ncookies; i++) {
2243 unsigned long cookie_raddr = cookies[i].cookie_addr;
2244 unsigned long this_len = cookies[i].cookie_size;
2245 unsigned long actual_len;
2247 if (unlikely(offset)) {
2248 unsigned long this_off = offset;
2250 if (this_off > this_len)
2251 this_off = this_len;
2254 this_len -= this_off;
2257 cookie_raddr += this_off;
2264 unsigned long hv_err;
2266 hv_err = sun4v_ldc_copy(lp->id, copy_dir,
2268 this_len, &actual_len);
2269 if (unlikely(hv_err)) {
2270 printk(KERN_ERR PFX "ldc_copy: ID[%lu] "
2273 if (lp->hs_state != LDC_HS_COMPLETE ||
2274 (lp->flags & LDC_FLAG_RESET))
2280 cookie_raddr += actual_len;
2283 if (actual_len == this_len)
2286 this_len -= actual_len;
2293 /* It is caller policy what to do about short copies.
2294 * For example, a networking driver can declare the
2295 * packet a runt and drop it.
2298 return orig_len - len;
2300 EXPORT_SYMBOL(ldc_copy);
2302 void *ldc_alloc_exp_dring(struct ldc_channel *lp, unsigned int len,
2303 struct ldc_trans_cookie *cookies, int *ncookies,
2304 unsigned int map_perm)
2309 if (len & (8UL - 1))
2310 return ERR_PTR(-EINVAL);
2312 buf = kzalloc(len, GFP_KERNEL);
2314 return ERR_PTR(-ENOMEM);
2316 err = ldc_map_single(lp, buf, len, cookies, *ncookies, map_perm);
2319 return ERR_PTR(err);
2325 EXPORT_SYMBOL(ldc_alloc_exp_dring);
2327 void ldc_free_exp_dring(struct ldc_channel *lp, void *buf, unsigned int len,
2328 struct ldc_trans_cookie *cookies, int ncookies)
2330 ldc_unmap(lp, cookies, ncookies);
2333 EXPORT_SYMBOL(ldc_free_exp_dring);
2335 static int __init ldc_init(void)
2337 unsigned long major, minor;
2338 struct mdesc_handle *hp;
2347 mp = mdesc_node_by_name(hp, MDESC_NODE_NULL, "platform");
2349 if (mp == MDESC_NODE_NULL)
2352 v = mdesc_get_property(hp, mp, "domaining-enabled", NULL);
2358 if (sun4v_hvapi_register(HV_GRP_LDOM, major, &minor)) {
2359 printk(KERN_INFO PFX "Could not register LDOM hvapi.\n");
2363 printk(KERN_INFO "%s", version);
2366 printk(KERN_INFO PFX "Domaining disabled.\n");
2369 ldom_domaining_enabled = 1;
2377 core_initcall(ldc_init);