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/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>
19 #include <asm/hypervisor.h>
20 #include <asm/iommu.h>
23 #include <asm/mdesc.h>
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"
30 static char version[] =
31 DRV_MODULE_NAME ".c:v" DRV_MODULE_VERSION " (" DRV_MODULE_RELDATE ")\n";
32 #define LDC_PACKET_SIZE 64
34 /* Packet header layout for unreliable and reliable mode frames.
35 * When in RAW mode, packets are simply straight 64-byte payloads
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
58 #define LDC_FRAG_MASK 0xc0
59 #define LDC_START 0x40
65 u8 u_data[LDC_PACKET_SIZE - 8];
69 u8 r_data[LDC_PACKET_SIZE - 8 - 8];
79 /* Ordered from largest major to lowest. */
80 static struct ldc_version ver_arr[] = {
81 { .major = 1, .minor = 0 },
84 #define LDC_DEFAULT_MTU (4 * LDC_PACKET_SIZE)
85 #define LDC_DEFAULT_NUM_ENTRIES (PAGE_SIZE / LDC_PACKET_SIZE)
90 int (*write)(struct ldc_channel *, const void *, unsigned int);
91 int (*read)(struct ldc_channel *, void *, unsigned int);
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;
98 int ldom_domaining_enabled;
101 /* Protects arena alloc/free. */
103 struct iommu_arena arena;
104 struct ldc_mtable_entry *page_table;
108 /* Protects all operations that depend upon channel state. */
117 struct ldc_packet *tx_base;
118 unsigned long tx_head;
119 unsigned long tx_tail;
120 unsigned long tx_num_entries;
123 unsigned long tx_acked;
125 struct ldc_packet *rx_base;
126 unsigned long rx_head;
127 unsigned long rx_tail;
128 unsigned long rx_num_entries;
134 unsigned long chan_state;
136 struct ldc_channel_config cfg;
139 const struct ldc_mode_ops *mops;
141 struct ldc_iommu iommu;
143 struct ldc_version ver;
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
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
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];
166 struct hlist_head mh_list;
168 struct hlist_node list;
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); \
176 static const char *state_to_str(u8 state)
179 case LDC_STATE_INVALID:
183 case LDC_STATE_BOUND:
185 case LDC_STATE_READY:
187 case LDC_STATE_CONNECTED:
194 static void ldc_set_state(struct ldc_channel *lp, u8 state)
196 ldcdbg(STATE, "STATE (%s) --> (%s)\n",
197 state_to_str(lp->state),
198 state_to_str(state));
203 static unsigned long __advance(unsigned long off, unsigned long num_entries)
205 off += LDC_PACKET_SIZE;
206 if (off == (num_entries * LDC_PACKET_SIZE))
212 static unsigned long rx_advance(struct ldc_channel *lp, unsigned long off)
214 return __advance(off, lp->rx_num_entries);
217 static unsigned long tx_advance(struct ldc_channel *lp, unsigned long off)
219 return __advance(off, lp->tx_num_entries);
222 static struct ldc_packet *handshake_get_tx_packet(struct ldc_channel *lp,
223 unsigned long *new_tail)
225 struct ldc_packet *p;
228 t = tx_advance(lp, lp->tx_tail);
229 if (t == lp->tx_head)
235 return p + (lp->tx_tail / LDC_PACKET_SIZE);
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
245 static unsigned long head_for_data(struct ldc_channel *lp)
247 if (lp->cfg.mode == LDC_MODE_STREAM)
252 static int tx_has_space_for(struct ldc_channel *lp, unsigned int size)
254 unsigned long limit, tail, new_tail, diff;
257 limit = head_for_data(lp);
259 new_tail = tx_advance(lp, tail);
260 if (new_tail == limit)
263 if (limit > new_tail)
264 diff = limit - new_tail;
267 ((lp->tx_num_entries * LDC_PACKET_SIZE) - new_tail));
268 diff /= LDC_PACKET_SIZE;
271 if (diff * mss < size)
277 static struct ldc_packet *data_get_tx_packet(struct ldc_channel *lp,
278 unsigned long *new_tail)
280 struct ldc_packet *p;
283 h = head_for_data(lp);
284 t = tx_advance(lp, lp->tx_tail);
291 return p + (lp->tx_tail / LDC_PACKET_SIZE);
294 static int set_tx_tail(struct ldc_channel *lp, unsigned long tail)
296 unsigned long orig_tail = lp->tx_tail;
300 while (limit-- > 0) {
303 err = sun4v_ldc_tx_set_qtail(lp->id, tail);
307 if (err != HV_EWOULDBLOCK) {
308 lp->tx_tail = orig_tail;
314 lp->tx_tail = orig_tail;
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.
322 static int __set_rx_head(struct ldc_channel *lp, unsigned long head)
326 while (limit-- > 0) {
329 err = sun4v_ldc_rx_set_qhead(lp->id, head);
333 if (err != HV_EWOULDBLOCK)
342 static int send_tx_packet(struct ldc_channel *lp,
343 struct ldc_packet *p,
344 unsigned long new_tail)
346 BUG_ON(p != (lp->tx_base + (lp->tx_tail / LDC_PACKET_SIZE)));
348 return set_tx_tail(lp, new_tail);
351 static struct ldc_packet *handshake_compose_ctrl(struct ldc_channel *lp,
353 void *data, int dlen,
354 unsigned long *new_tail)
356 struct ldc_packet *p = handshake_get_tx_packet(lp, new_tail);
359 memset(p, 0, sizeof(*p));
364 memcpy(p->u.u_data, data, dlen);
369 static int start_handshake(struct ldc_channel *lp)
371 struct ldc_packet *p;
372 struct ldc_version *ver;
373 unsigned long new_tail;
377 ldcdbg(HS, "SEND VER INFO maj[%u] min[%u]\n",
378 ver->major, ver->minor);
380 p = handshake_compose_ctrl(lp, LDC_INFO, LDC_VERS,
381 ver, sizeof(*ver), &new_tail);
383 int err = send_tx_packet(lp, p, new_tail);
385 lp->flags &= ~LDC_FLAG_RESET;
391 static int send_version_nack(struct ldc_channel *lp,
392 u16 major, u16 minor)
394 struct ldc_packet *p;
395 struct ldc_version ver;
396 unsigned long new_tail;
401 p = handshake_compose_ctrl(lp, LDC_NACK, LDC_VERS,
402 &ver, sizeof(ver), &new_tail);
404 ldcdbg(HS, "SEND VER NACK maj[%u] min[%u]\n",
405 ver.major, ver.minor);
407 return send_tx_packet(lp, p, new_tail);
412 static int send_version_ack(struct ldc_channel *lp,
413 struct ldc_version *vp)
415 struct ldc_packet *p;
416 unsigned long new_tail;
418 p = handshake_compose_ctrl(lp, LDC_ACK, LDC_VERS,
419 vp, sizeof(*vp), &new_tail);
421 ldcdbg(HS, "SEND VER ACK maj[%u] min[%u]\n",
422 vp->major, vp->minor);
424 return send_tx_packet(lp, p, new_tail);
429 static int send_rts(struct ldc_channel *lp)
431 struct ldc_packet *p;
432 unsigned long new_tail;
434 p = handshake_compose_ctrl(lp, LDC_INFO, LDC_RTS, NULL, 0,
437 p->env = lp->cfg.mode;
441 ldcdbg(HS, "SEND RTS env[0x%x] seqid[0x%x]\n",
444 return send_tx_packet(lp, p, new_tail);
449 static int send_rtr(struct ldc_channel *lp)
451 struct ldc_packet *p;
452 unsigned long new_tail;
454 p = handshake_compose_ctrl(lp, LDC_INFO, LDC_RTR, NULL, 0,
457 p->env = lp->cfg.mode;
460 ldcdbg(HS, "SEND RTR env[0x%x] seqid[0x%x]\n",
463 return send_tx_packet(lp, p, new_tail);
468 static int send_rdx(struct ldc_channel *lp)
470 struct ldc_packet *p;
471 unsigned long new_tail;
473 p = handshake_compose_ctrl(lp, LDC_INFO, LDC_RDX, NULL, 0,
477 p->seqid = ++lp->snd_nxt;
478 p->u.r.ackid = lp->rcv_nxt;
480 ldcdbg(HS, "SEND RDX env[0x%x] seqid[0x%x] ackid[0x%x]\n",
481 p->env, p->seqid, p->u.r.ackid);
483 return send_tx_packet(lp, p, new_tail);
488 static int send_data_nack(struct ldc_channel *lp, struct ldc_packet *data_pkt)
490 struct ldc_packet *p;
491 unsigned long new_tail;
494 p = data_get_tx_packet(lp, &new_tail);
497 memset(p, 0, sizeof(*p));
498 p->type = data_pkt->type;
500 p->ctrl = data_pkt->ctrl & LDC_CTRL_MSK;
501 p->seqid = lp->snd_nxt + 1;
502 p->u.r.ackid = lp->rcv_nxt;
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);
507 err = send_tx_packet(lp, p, new_tail);
514 static int ldc_abort(struct ldc_channel *lp)
516 unsigned long hv_err;
518 ldcdbg(STATE, "ABORT\n");
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.
523 hv_err = sun4v_ldc_tx_qconf(lp->id, lp->tx_ra, lp->tx_num_entries);
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);
529 hv_err = sun4v_ldc_tx_get_state(lp->id,
534 printk(KERN_ERR PFX "ldc_abort: "
535 "sun4v_ldc_tx_get_state(%lx,...) failed, err=%lu\n",
538 hv_err = sun4v_ldc_rx_qconf(lp->id, lp->rx_ra, lp->rx_num_entries);
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);
544 /* Refetch the RX queue state as well, because we could be invoked
545 * here in the queue processing context.
547 hv_err = sun4v_ldc_rx_get_state(lp->id,
552 printk(KERN_ERR PFX "ldc_abort: "
553 "sun4v_ldc_rx_get_state(%lx,...) failed, err=%lu\n",
559 static struct ldc_version *find_by_major(u16 major)
561 struct ldc_version *ret = NULL;
564 for (i = 0; i < ARRAY_SIZE(ver_arr); i++) {
565 struct ldc_version *v = &ver_arr[i];
566 if (v->major <= major) {
574 static int process_ver_info(struct ldc_channel *lp, struct ldc_version *vp)
576 struct ldc_version *vap;
579 ldcdbg(HS, "GOT VERSION INFO major[%x] minor[%x]\n",
580 vp->major, vp->minor);
582 if (lp->hs_state == LDC_HS_GOTVERS) {
583 lp->hs_state = LDC_HS_OPEN;
584 memset(&lp->ver, 0, sizeof(lp->ver));
587 vap = find_by_major(vp->major);
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);
593 struct ldc_version ver = *vp;
594 if (ver.minor > vap->minor)
595 ver.minor = vap->minor;
596 err = send_version_ack(lp, &ver);
599 lp->hs_state = LDC_HS_GOTVERS;
603 return ldc_abort(lp);
608 static int process_ver_ack(struct ldc_channel *lp, struct ldc_version *vp)
610 ldcdbg(HS, "GOT VERSION ACK major[%x] minor[%x]\n",
611 vp->major, vp->minor);
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);
619 lp->hs_state = LDC_HS_GOTVERS;
622 return ldc_abort(lp);
626 static int process_ver_nack(struct ldc_channel *lp, struct ldc_version *vp)
628 struct ldc_version *vap;
629 struct ldc_packet *p;
630 unsigned long new_tail;
632 if (vp->major == 0 && vp->minor == 0)
633 return ldc_abort(lp);
635 vap = find_by_major(vp->major);
637 return ldc_abort(lp);
639 p = handshake_compose_ctrl(lp, LDC_INFO, LDC_VERS,
643 return ldc_abort(lp);
645 return send_tx_packet(lp, p, new_tail);
648 static int process_version(struct ldc_channel *lp,
649 struct ldc_packet *p)
651 struct ldc_version *vp;
653 vp = (struct ldc_version *) p->u.u_data;
657 return process_ver_info(lp, vp);
660 return process_ver_ack(lp, vp);
663 return process_ver_nack(lp, vp);
666 return ldc_abort(lp);
670 static int process_rts(struct ldc_channel *lp,
671 struct ldc_packet *p)
673 ldcdbg(HS, "GOT RTS stype[%x] seqid[%x] env[%x]\n",
674 p->stype, p->seqid, p->env);
676 if (p->stype != LDC_INFO ||
677 lp->hs_state != LDC_HS_GOTVERS ||
678 p->env != lp->cfg.mode)
679 return ldc_abort(lp);
681 lp->snd_nxt = p->seqid;
682 lp->rcv_nxt = p->seqid;
683 lp->hs_state = LDC_HS_SENTRTR;
685 return ldc_abort(lp);
690 static int process_rtr(struct ldc_channel *lp,
691 struct ldc_packet *p)
693 ldcdbg(HS, "GOT RTR stype[%x] seqid[%x] env[%x]\n",
694 p->stype, p->seqid, p->env);
696 if (p->stype != LDC_INFO ||
697 p->env != lp->cfg.mode)
698 return ldc_abort(lp);
700 lp->snd_nxt = p->seqid;
701 lp->hs_state = LDC_HS_COMPLETE;
702 ldc_set_state(lp, LDC_STATE_CONNECTED);
708 static int rx_seq_ok(struct ldc_channel *lp, u32 seqid)
710 return lp->rcv_nxt + 1 == seqid;
713 static int process_rdx(struct ldc_channel *lp,
714 struct ldc_packet *p)
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);
719 if (p->stype != LDC_INFO ||
720 !(rx_seq_ok(lp, p->seqid)))
721 return ldc_abort(lp);
723 lp->rcv_nxt = p->seqid;
725 lp->hs_state = LDC_HS_COMPLETE;
726 ldc_set_state(lp, LDC_STATE_CONNECTED);
731 static int process_control_frame(struct ldc_channel *lp,
732 struct ldc_packet *p)
736 return process_version(lp, p);
739 return process_rts(lp, p);
742 return process_rtr(lp, p);
745 return process_rdx(lp, p);
748 return ldc_abort(lp);
752 static int process_error_frame(struct ldc_channel *lp,
753 struct ldc_packet *p)
755 return ldc_abort(lp);
758 static int process_data_ack(struct ldc_channel *lp,
759 struct ldc_packet *ack)
761 unsigned long head = lp->tx_acked;
762 u32 ackid = ack->u.r.ackid;
765 struct ldc_packet *p = lp->tx_base + (head / LDC_PACKET_SIZE);
767 head = tx_advance(lp, head);
769 if (p->seqid == ackid) {
773 if (head == lp->tx_tail)
774 return ldc_abort(lp);
780 static void send_events(struct ldc_channel *lp, unsigned int event_mask)
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);
790 static irqreturn_t ldc_rx(int irq, void *dev_id)
792 struct ldc_channel *lp = dev_id;
793 unsigned long orig_state, flags;
794 unsigned int event_mask;
796 spin_lock_irqsave(&lp->lock, flags);
798 orig_state = lp->chan_state;
800 /* We should probably check for hypervisor errors here and
801 * reset the LDC channel if we get one.
803 sun4v_ldc_rx_get_state(lp->id,
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);
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);
818 event_mask |= LDC_EVENT_UP;
820 orig_state = lp->chan_state;
823 /* If we are in reset state, flush the RX queue and ignore
826 if (lp->flags & LDC_FLAG_RESET) {
827 (void) __set_rx_head(lp, lp->rx_tail);
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.
835 if (lp->hs_state == LDC_HS_COMPLETE) {
837 if (lp->chan_state != orig_state) {
838 unsigned int event = LDC_EVENT_RESET;
840 if (lp->chan_state == LDC_CHANNEL_UP)
841 event = LDC_EVENT_UP;
845 if (lp->rx_head != lp->rx_tail)
846 event_mask |= LDC_EVENT_DATA_READY;
851 if (lp->chan_state != orig_state)
854 while (lp->rx_head != lp->rx_tail) {
855 struct ldc_packet *p;
859 p = lp->rx_base + (lp->rx_head / LDC_PACKET_SIZE);
863 err = process_control_frame(lp, p);
869 event_mask |= LDC_EVENT_DATA_READY;
874 err = process_error_frame(lp, p);
886 new += LDC_PACKET_SIZE;
887 if (new == (lp->rx_num_entries * LDC_PACKET_SIZE))
891 err = __set_rx_head(lp, new);
893 (void) ldc_abort(lp);
896 if (lp->hs_state == LDC_HS_COMPLETE)
897 goto handshake_complete;
901 spin_unlock_irqrestore(&lp->lock, flags);
903 send_events(lp, event_mask);
908 static irqreturn_t ldc_tx(int irq, void *dev_id)
910 struct ldc_channel *lp = dev_id;
911 unsigned long flags, orig_state;
912 unsigned int event_mask = 0;
914 spin_lock_irqsave(&lp->lock, flags);
916 orig_state = lp->chan_state;
918 /* We should probably check for hypervisor errors here and
919 * reset the LDC channel if we get one.
921 sun4v_ldc_tx_get_state(lp->id,
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);
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);
934 event_mask |= LDC_EVENT_UP;
937 spin_unlock_irqrestore(&lp->lock, flags);
939 send_events(lp, event_mask);
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.
951 static HLIST_HEAD(ldc_channel_list);
953 static int __ldc_channel_exists(unsigned long id)
955 struct ldc_channel *lp;
957 hlist_for_each_entry(lp, &ldc_channel_list, list) {
964 static int alloc_queue(const char *name, unsigned long num_entries,
965 struct ldc_packet **base, unsigned long *ra)
967 unsigned long size, order;
970 size = num_entries * LDC_PACKET_SIZE;
971 order = get_order(size);
973 q = (void *) __get_free_pages(GFP_KERNEL, order);
975 printk(KERN_ERR PFX "Alloc of %s queue failed with "
976 "size=%lu order=%lu\n", name, size, order);
980 memset(q, 0, PAGE_SIZE << order);
988 static void free_queue(unsigned long num_entries, struct ldc_packet *q)
990 unsigned long size, order;
995 size = num_entries * LDC_PACKET_SIZE;
996 order = get_order(size);
998 free_pages((unsigned long)q, order);
1001 /* XXX Make this configurable... XXX */
1002 #define LDC_IOTABLE_SIZE (8 * 1024)
1004 static int ldc_iommu_init(struct ldc_channel *lp)
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;
1012 num_tsb_entries = LDC_IOTABLE_SIZE;
1013 tsbsize = num_tsb_entries * sizeof(struct ldc_mtable_entry);
1015 spin_lock_init(&iommu->lock);
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);
1025 iommu->arena.limit = num_tsb_entries;
1027 order = get_order(tsbsize);
1029 table = (struct ldc_mtable_entry *)
1030 __get_free_pages(GFP_KERNEL, order);
1033 printk(KERN_ERR PFX "Alloc of MTE table failed, "
1034 "size=%lu order=%lu\n", tsbsize, order);
1038 memset(table, 0, PAGE_SIZE << order);
1040 iommu->page_table = table;
1042 hv_err = sun4v_ldc_set_map_table(lp->id, __pa(table),
1046 goto out_free_table;
1051 free_pages((unsigned long) table, order);
1052 iommu->page_table = NULL;
1055 kfree(iommu->arena.map);
1056 iommu->arena.map = NULL;
1061 static void ldc_iommu_release(struct ldc_channel *lp)
1063 struct ldc_iommu *iommu = &lp->iommu;
1064 unsigned long num_tsb_entries, tsbsize, order;
1066 (void) sun4v_ldc_set_map_table(lp->id, 0, 0);
1068 num_tsb_entries = iommu->arena.limit;
1069 tsbsize = num_tsb_entries * sizeof(struct ldc_mtable_entry);
1070 order = get_order(tsbsize);
1072 free_pages((unsigned long) iommu->page_table, order);
1073 iommu->page_table = NULL;
1075 kfree(iommu->arena.map);
1076 iommu->arena.map = NULL;
1079 struct ldc_channel *ldc_alloc(unsigned long id,
1080 const struct ldc_channel_config *cfgp,
1084 struct ldc_channel *lp;
1085 const struct ldc_mode_ops *mops;
1086 unsigned long dummy1, dummy2, hv_err;
1091 if (!ldom_domaining_enabled)
1100 switch (cfgp->mode) {
1103 mss = LDC_PACKET_SIZE;
1106 case LDC_MODE_UNRELIABLE:
1108 mss = LDC_PACKET_SIZE - 8;
1111 case LDC_MODE_STREAM:
1113 mss = LDC_PACKET_SIZE - 8 - 8;
1120 if (!cfgp->event || !event_arg || !cfgp->rx_irq || !cfgp->tx_irq)
1123 hv_err = sun4v_ldc_tx_qinfo(id, &dummy1, &dummy2);
1125 if (hv_err == HV_ECHANNEL)
1129 if (__ldc_channel_exists(id))
1134 lp = kzalloc(sizeof(*lp), GFP_KERNEL);
1139 spin_lock_init(&lp->lock);
1143 err = ldc_iommu_init(lp);
1152 lp->cfg.mtu = LDC_DEFAULT_MTU;
1154 if (lp->cfg.mode == LDC_MODE_STREAM) {
1155 mssbuf = kzalloc(lp->cfg.mtu, GFP_KERNEL);
1158 goto out_free_iommu;
1160 lp->mssbuf = mssbuf;
1163 lp->event_arg = event_arg;
1165 /* XXX allow setting via ldc_channel_config to override defaults
1166 * XXX or use some formula based upon mtu
1168 lp->tx_num_entries = LDC_DEFAULT_NUM_ENTRIES;
1169 lp->rx_num_entries = LDC_DEFAULT_NUM_ENTRIES;
1171 err = alloc_queue("TX", lp->tx_num_entries,
1172 &lp->tx_base, &lp->tx_ra);
1174 goto out_free_mssbuf;
1176 err = alloc_queue("RX", lp->rx_num_entries,
1177 &lp->rx_base, &lp->rx_ra);
1181 lp->flags |= LDC_FLAG_ALLOCED_QUEUES;
1183 lp->hs_state = LDC_HS_CLOSED;
1184 ldc_set_state(lp, LDC_STATE_INIT);
1186 INIT_HLIST_NODE(&lp->list);
1187 hlist_add_head(&lp->list, &ldc_channel_list);
1189 INIT_HLIST_HEAD(&lp->mh_list);
1191 snprintf(lp->rx_irq_name, LDC_IRQ_NAME_MAX, "%s RX", name);
1192 snprintf(lp->tx_irq_name, LDC_IRQ_NAME_MAX, "%s TX", name);
1194 err = request_irq(lp->cfg.rx_irq, ldc_rx, 0,
1195 lp->rx_irq_name, lp);
1199 err = request_irq(lp->cfg.tx_irq, ldc_tx, 0,
1200 lp->tx_irq_name, lp);
1202 free_irq(lp->cfg.rx_irq, lp);
1209 free_queue(lp->tx_num_entries, lp->tx_base);
1215 ldc_iommu_release(lp);
1221 return ERR_PTR(err);
1223 EXPORT_SYMBOL(ldc_alloc);
1225 void ldc_unbind(struct ldc_channel *lp)
1227 if (lp->flags & LDC_FLAG_REGISTERED_IRQS) {
1228 free_irq(lp->cfg.rx_irq, lp);
1229 free_irq(lp->cfg.tx_irq, lp);
1230 lp->flags &= ~LDC_FLAG_REGISTERED_IRQS;
1233 if (lp->flags & LDC_FLAG_REGISTERED_QUEUES) {
1234 sun4v_ldc_tx_qconf(lp->id, 0, 0);
1235 sun4v_ldc_rx_qconf(lp->id, 0, 0);
1236 lp->flags &= ~LDC_FLAG_REGISTERED_QUEUES;
1238 if (lp->flags & LDC_FLAG_ALLOCED_QUEUES) {
1239 free_queue(lp->tx_num_entries, lp->tx_base);
1240 free_queue(lp->rx_num_entries, lp->rx_base);
1241 lp->flags &= ~LDC_FLAG_ALLOCED_QUEUES;
1244 ldc_set_state(lp, LDC_STATE_INIT);
1246 EXPORT_SYMBOL(ldc_unbind);
1248 void ldc_free(struct ldc_channel *lp)
1251 hlist_del(&lp->list);
1253 ldc_iommu_release(lp);
1257 EXPORT_SYMBOL(ldc_free);
1259 /* Bind the channel. This registers the LDC queues with
1260 * the hypervisor and puts the channel into a pseudo-listening
1261 * state. This does not initiate a handshake, ldc_connect() does
1264 int ldc_bind(struct ldc_channel *lp)
1266 unsigned long hv_err, flags;
1269 if (lp->state != LDC_STATE_INIT)
1272 spin_lock_irqsave(&lp->lock, flags);
1274 enable_irq(lp->cfg.rx_irq);
1275 enable_irq(lp->cfg.tx_irq);
1277 lp->flags |= LDC_FLAG_REGISTERED_IRQS;
1280 hv_err = sun4v_ldc_tx_qconf(lp->id, 0, 0);
1284 hv_err = sun4v_ldc_tx_qconf(lp->id, lp->tx_ra, lp->tx_num_entries);
1288 hv_err = sun4v_ldc_rx_qconf(lp->id, 0, 0);
1292 hv_err = sun4v_ldc_rx_qconf(lp->id, lp->rx_ra, lp->rx_num_entries);
1296 lp->flags |= LDC_FLAG_REGISTERED_QUEUES;
1298 hv_err = sun4v_ldc_tx_get_state(lp->id,
1306 lp->tx_acked = lp->tx_head;
1308 lp->hs_state = LDC_HS_OPEN;
1309 ldc_set_state(lp, LDC_STATE_BOUND);
1311 spin_unlock_irqrestore(&lp->lock, flags);
1316 lp->flags &= ~LDC_FLAG_REGISTERED_QUEUES;
1317 sun4v_ldc_rx_qconf(lp->id, 0, 0);
1320 sun4v_ldc_tx_qconf(lp->id, 0, 0);
1323 lp->flags &= ~LDC_FLAG_REGISTERED_IRQS;
1324 free_irq(lp->cfg.tx_irq, lp);
1325 free_irq(lp->cfg.rx_irq, lp);
1327 spin_unlock_irqrestore(&lp->lock, flags);
1331 EXPORT_SYMBOL(ldc_bind);
1333 int ldc_connect(struct ldc_channel *lp)
1335 unsigned long flags;
1338 if (lp->cfg.mode == LDC_MODE_RAW)
1341 spin_lock_irqsave(&lp->lock, flags);
1343 if (!(lp->flags & LDC_FLAG_ALLOCED_QUEUES) ||
1344 !(lp->flags & LDC_FLAG_REGISTERED_QUEUES) ||
1345 lp->hs_state != LDC_HS_OPEN)
1346 err = ((lp->hs_state > LDC_HS_OPEN) ? 0 : -EINVAL);
1348 err = start_handshake(lp);
1350 spin_unlock_irqrestore(&lp->lock, flags);
1354 EXPORT_SYMBOL(ldc_connect);
1356 int ldc_disconnect(struct ldc_channel *lp)
1358 unsigned long hv_err, flags;
1361 if (lp->cfg.mode == LDC_MODE_RAW)
1364 if (!(lp->flags & LDC_FLAG_ALLOCED_QUEUES) ||
1365 !(lp->flags & LDC_FLAG_REGISTERED_QUEUES))
1368 spin_lock_irqsave(&lp->lock, flags);
1371 hv_err = sun4v_ldc_tx_qconf(lp->id, 0, 0);
1375 hv_err = sun4v_ldc_tx_qconf(lp->id, lp->tx_ra, lp->tx_num_entries);
1379 hv_err = sun4v_ldc_rx_qconf(lp->id, 0, 0);
1383 hv_err = sun4v_ldc_rx_qconf(lp->id, lp->rx_ra, lp->rx_num_entries);
1387 ldc_set_state(lp, LDC_STATE_BOUND);
1388 lp->hs_state = LDC_HS_OPEN;
1389 lp->flags |= LDC_FLAG_RESET;
1391 spin_unlock_irqrestore(&lp->lock, flags);
1396 sun4v_ldc_tx_qconf(lp->id, 0, 0);
1397 sun4v_ldc_rx_qconf(lp->id, 0, 0);
1398 free_irq(lp->cfg.tx_irq, lp);
1399 free_irq(lp->cfg.rx_irq, lp);
1400 lp->flags &= ~(LDC_FLAG_REGISTERED_IRQS |
1401 LDC_FLAG_REGISTERED_QUEUES);
1402 ldc_set_state(lp, LDC_STATE_INIT);
1404 spin_unlock_irqrestore(&lp->lock, flags);
1408 EXPORT_SYMBOL(ldc_disconnect);
1410 int ldc_state(struct ldc_channel *lp)
1414 EXPORT_SYMBOL(ldc_state);
1416 static int write_raw(struct ldc_channel *lp, const void *buf, unsigned int size)
1418 struct ldc_packet *p;
1419 unsigned long new_tail;
1422 if (size > LDC_PACKET_SIZE)
1425 p = data_get_tx_packet(lp, &new_tail);
1429 memcpy(p, buf, size);
1431 err = send_tx_packet(lp, p, new_tail);
1438 static int read_raw(struct ldc_channel *lp, void *buf, unsigned int size)
1440 struct ldc_packet *p;
1441 unsigned long hv_err, new;
1444 if (size < LDC_PACKET_SIZE)
1447 hv_err = sun4v_ldc_rx_get_state(lp->id,
1452 return ldc_abort(lp);
1454 if (lp->chan_state == LDC_CHANNEL_DOWN ||
1455 lp->chan_state == LDC_CHANNEL_RESETTING)
1458 if (lp->rx_head == lp->rx_tail)
1461 p = lp->rx_base + (lp->rx_head / LDC_PACKET_SIZE);
1462 memcpy(buf, p, LDC_PACKET_SIZE);
1464 new = rx_advance(lp, lp->rx_head);
1467 err = __set_rx_head(lp, new);
1471 err = LDC_PACKET_SIZE;
1476 static const struct ldc_mode_ops raw_ops = {
1481 static int write_nonraw(struct ldc_channel *lp, const void *buf,
1484 unsigned long hv_err, tail;
1485 unsigned int copied;
1489 hv_err = sun4v_ldc_tx_get_state(lp->id, &lp->tx_head, &lp->tx_tail,
1491 if (unlikely(hv_err))
1494 if (unlikely(lp->chan_state != LDC_CHANNEL_UP))
1495 return ldc_abort(lp);
1497 if (!tx_has_space_for(lp, size))
1503 while (copied < size) {
1504 struct ldc_packet *p = lp->tx_base + (tail / LDC_PACKET_SIZE);
1505 u8 *data = ((lp->cfg.mode == LDC_MODE_UNRELIABLE) ?
1511 p->stype = LDC_INFO;
1514 data_len = size - copied;
1515 if (data_len > lp->mss)
1518 BUG_ON(data_len > LDC_LEN);
1520 p->env = (data_len |
1521 (copied == 0 ? LDC_START : 0) |
1522 (data_len == size - copied ? LDC_STOP : 0));
1526 ldcdbg(DATA, "SENT DATA [%02x:%02x:%02x:%02x:%08x]\n",
1533 memcpy(data, buf, data_len);
1537 tail = tx_advance(lp, tail);
1540 err = set_tx_tail(lp, tail);
1549 static int rx_bad_seq(struct ldc_channel *lp, struct ldc_packet *p,
1550 struct ldc_packet *first_frag)
1555 lp->rcv_nxt = first_frag->seqid - 1;
1557 err = send_data_nack(lp, p);
1561 err = __set_rx_head(lp, lp->rx_tail);
1563 return ldc_abort(lp);
1568 static int data_ack_nack(struct ldc_channel *lp, struct ldc_packet *p)
1570 if (p->stype & LDC_ACK) {
1571 int err = process_data_ack(lp, p);
1575 if (p->stype & LDC_NACK)
1576 return ldc_abort(lp);
1581 static int rx_data_wait(struct ldc_channel *lp, unsigned long cur_head)
1583 unsigned long dummy;
1586 ldcdbg(DATA, "DATA WAIT cur_head[%lx] rx_head[%lx] rx_tail[%lx]\n",
1587 cur_head, lp->rx_head, lp->rx_tail);
1588 while (limit-- > 0) {
1589 unsigned long hv_err;
1591 hv_err = sun4v_ldc_rx_get_state(lp->id,
1596 return ldc_abort(lp);
1598 if (lp->chan_state == LDC_CHANNEL_DOWN ||
1599 lp->chan_state == LDC_CHANNEL_RESETTING)
1602 if (cur_head != lp->rx_tail) {
1603 ldcdbg(DATA, "DATA WAIT DONE "
1604 "head[%lx] tail[%lx] chan_state[%lx]\n",
1605 dummy, lp->rx_tail, lp->chan_state);
1614 static int rx_set_head(struct ldc_channel *lp, unsigned long head)
1616 int err = __set_rx_head(lp, head);
1619 return ldc_abort(lp);
1625 static void send_data_ack(struct ldc_channel *lp)
1627 unsigned long new_tail;
1628 struct ldc_packet *p;
1630 p = data_get_tx_packet(lp, &new_tail);
1634 memset(p, 0, sizeof(*p));
1638 p->seqid = lp->snd_nxt + 1;
1639 p->u.r.ackid = lp->rcv_nxt;
1641 err = send_tx_packet(lp, p, new_tail);
1647 static int read_nonraw(struct ldc_channel *lp, void *buf, unsigned int size)
1649 struct ldc_packet *first_frag;
1650 unsigned long hv_err, new;
1653 hv_err = sun4v_ldc_rx_get_state(lp->id,
1658 return ldc_abort(lp);
1660 if (lp->chan_state == LDC_CHANNEL_DOWN ||
1661 lp->chan_state == LDC_CHANNEL_RESETTING)
1664 if (lp->rx_head == lp->rx_tail)
1671 struct ldc_packet *p;
1674 BUG_ON(new == lp->rx_tail);
1675 p = lp->rx_base + (new / LDC_PACKET_SIZE);
1677 ldcdbg(RX, "RX read pkt[%02x:%02x:%02x:%02x:%08x:%08x] "
1687 if (unlikely(!rx_seq_ok(lp, p->seqid))) {
1688 err = rx_bad_seq(lp, p, first_frag);
1693 if (p->type & LDC_CTRL) {
1694 err = process_control_frame(lp, p);
1700 lp->rcv_nxt = p->seqid;
1702 if (!(p->type & LDC_DATA)) {
1703 new = rx_advance(lp, new);
1706 if (p->stype & (LDC_ACK | LDC_NACK)) {
1707 err = data_ack_nack(lp, p);
1711 if (!(p->stype & LDC_INFO)) {
1712 new = rx_advance(lp, new);
1713 err = rx_set_head(lp, new);
1719 pkt_len = p->env & LDC_LEN;
1721 /* Every initial packet starts with the START bit set.
1723 * Singleton packets will have both START+STOP set.
1725 * Fragments will have START set in the first frame, STOP
1726 * set in the last frame, and neither bit set in middle
1727 * frames of the packet.
1729 * Therefore if we are at the beginning of a packet and
1730 * we don't see START, or we are in the middle of a fragmented
1731 * packet and do see START, we are unsynchronized and should
1732 * flush the RX queue.
1734 if ((first_frag == NULL && !(p->env & LDC_START)) ||
1735 (first_frag != NULL && (p->env & LDC_START))) {
1737 new = rx_advance(lp, new);
1739 err = rx_set_head(lp, new);
1749 if (pkt_len > size - copied) {
1750 /* User didn't give us a big enough buffer,
1751 * what to do? This is a pretty serious error.
1753 * Since we haven't updated the RX ring head to
1754 * consume any of the packets, signal the error
1755 * to the user and just leave the RX ring alone.
1757 * This seems the best behavior because this allows
1758 * a user of the LDC layer to start with a small
1759 * RX buffer for ldc_read() calls and use -EMSGSIZE
1760 * as a cue to enlarge it's read buffer.
1766 /* Ok, we are gonna eat this one. */
1767 new = rx_advance(lp, new);
1770 (lp->cfg.mode == LDC_MODE_UNRELIABLE ?
1771 p->u.u_data : p->u.r.r_data), pkt_len);
1775 if (p->env & LDC_STOP)
1779 if (new == lp->rx_tail) {
1780 err = rx_data_wait(lp, new);
1787 err = rx_set_head(lp, new);
1789 if (err && first_frag)
1790 lp->rcv_nxt = first_frag->seqid - 1;
1794 if (err > 0 && lp->cfg.mode != LDC_MODE_UNRELIABLE)
1801 static const struct ldc_mode_ops nonraw_ops = {
1802 .write = write_nonraw,
1803 .read = read_nonraw,
1806 static int write_stream(struct ldc_channel *lp, const void *buf,
1809 if (size > lp->cfg.mtu)
1811 return write_nonraw(lp, buf, size);
1814 static int read_stream(struct ldc_channel *lp, void *buf, unsigned int size)
1816 if (!lp->mssbuf_len) {
1817 int err = read_nonraw(lp, lp->mssbuf, lp->cfg.mtu);
1821 lp->mssbuf_len = err;
1825 if (size > lp->mssbuf_len)
1826 size = lp->mssbuf_len;
1827 memcpy(buf, lp->mssbuf + lp->mssbuf_off, size);
1829 lp->mssbuf_off += size;
1830 lp->mssbuf_len -= size;
1835 static const struct ldc_mode_ops stream_ops = {
1836 .write = write_stream,
1837 .read = read_stream,
1840 int ldc_write(struct ldc_channel *lp, const void *buf, unsigned int size)
1842 unsigned long flags;
1851 spin_lock_irqsave(&lp->lock, flags);
1853 if (lp->hs_state != LDC_HS_COMPLETE)
1856 err = lp->mops->write(lp, buf, size);
1858 spin_unlock_irqrestore(&lp->lock, flags);
1862 EXPORT_SYMBOL(ldc_write);
1864 int ldc_read(struct ldc_channel *lp, void *buf, unsigned int size)
1866 unsigned long flags;
1875 spin_lock_irqsave(&lp->lock, flags);
1877 if (lp->hs_state != LDC_HS_COMPLETE)
1880 err = lp->mops->read(lp, buf, size);
1882 spin_unlock_irqrestore(&lp->lock, flags);
1886 EXPORT_SYMBOL(ldc_read);
1888 static long arena_alloc(struct ldc_iommu *iommu, unsigned long npages)
1890 struct iommu_arena *arena = &iommu->arena;
1891 unsigned long n, start, end, limit;
1894 limit = arena->limit;
1895 start = arena->hint;
1899 n = bitmap_find_next_zero_area(arena->map, limit, start, npages, 0);
1901 if (unlikely(end >= limit)) {
1902 if (likely(pass < 1)) {
1908 /* Scanned the whole thing, give up. */
1912 bitmap_set(arena->map, n, npages);
1919 #define COOKIE_PGSZ_CODE 0xf000000000000000ULL
1920 #define COOKIE_PGSZ_CODE_SHIFT 60ULL
1922 static u64 pagesize_code(void)
1924 switch (PAGE_SIZE) {
1926 case (8ULL * 1024ULL):
1928 case (64ULL * 1024ULL):
1930 case (512ULL * 1024ULL):
1932 case (4ULL * 1024ULL * 1024ULL):
1934 case (32ULL * 1024ULL * 1024ULL):
1936 case (256ULL * 1024ULL * 1024ULL):
1941 static u64 make_cookie(u64 index, u64 pgsz_code, u64 page_offset)
1943 return ((pgsz_code << COOKIE_PGSZ_CODE_SHIFT) |
1944 (index << PAGE_SHIFT) |
1948 static u64 cookie_to_index(u64 cookie, unsigned long *shift)
1950 u64 szcode = cookie >> COOKIE_PGSZ_CODE_SHIFT;
1952 cookie &= ~COOKIE_PGSZ_CODE;
1954 *shift = szcode * 3;
1956 return (cookie >> (13ULL + (szcode * 3ULL)));
1959 static struct ldc_mtable_entry *alloc_npages(struct ldc_iommu *iommu,
1960 unsigned long npages)
1964 entry = arena_alloc(iommu, npages);
1965 if (unlikely(entry < 0))
1968 return iommu->page_table + entry;
1971 static u64 perm_to_mte(unsigned int map_perm)
1975 mte_base = pagesize_code();
1977 if (map_perm & LDC_MAP_SHADOW) {
1978 if (map_perm & LDC_MAP_R)
1979 mte_base |= LDC_MTE_COPY_R;
1980 if (map_perm & LDC_MAP_W)
1981 mte_base |= LDC_MTE_COPY_W;
1983 if (map_perm & LDC_MAP_DIRECT) {
1984 if (map_perm & LDC_MAP_R)
1985 mte_base |= LDC_MTE_READ;
1986 if (map_perm & LDC_MAP_W)
1987 mte_base |= LDC_MTE_WRITE;
1988 if (map_perm & LDC_MAP_X)
1989 mte_base |= LDC_MTE_EXEC;
1991 if (map_perm & LDC_MAP_IO) {
1992 if (map_perm & LDC_MAP_R)
1993 mte_base |= LDC_MTE_IOMMU_R;
1994 if (map_perm & LDC_MAP_W)
1995 mte_base |= LDC_MTE_IOMMU_W;
2001 static int pages_in_region(unsigned long base, long len)
2006 unsigned long new = (base + PAGE_SIZE) & PAGE_MASK;
2008 len -= (new - base);
2016 struct cookie_state {
2017 struct ldc_mtable_entry *page_table;
2018 struct ldc_trans_cookie *cookies;
2025 static void fill_cookies(struct cookie_state *sp, unsigned long pa,
2026 unsigned long off, unsigned long len)
2029 unsigned long tlen, new = pa + PAGE_SIZE;
2032 sp->page_table[sp->pte_idx].mte = sp->mte_base | pa;
2036 tlen = PAGE_SIZE - off;
2040 this_cookie = make_cookie(sp->pte_idx,
2041 pagesize_code(), off);
2045 if (this_cookie == sp->prev_cookie) {
2046 sp->cookies[sp->nc - 1].cookie_size += tlen;
2048 sp->cookies[sp->nc].cookie_addr = this_cookie;
2049 sp->cookies[sp->nc].cookie_size = tlen;
2052 sp->prev_cookie = this_cookie + tlen;
2061 static int sg_count_one(struct scatterlist *sg)
2063 unsigned long base = page_to_pfn(sg_page(sg)) << PAGE_SHIFT;
2064 long len = sg->length;
2066 if ((sg->offset | len) & (8UL - 1))
2069 return pages_in_region(base + sg->offset, len);
2072 static int sg_count_pages(struct scatterlist *sg, int num_sg)
2078 for (i = 0; i < num_sg; i++) {
2079 int err = sg_count_one(sg + i);
2088 int ldc_map_sg(struct ldc_channel *lp,
2089 struct scatterlist *sg, int num_sg,
2090 struct ldc_trans_cookie *cookies, int ncookies,
2091 unsigned int map_perm)
2093 unsigned long i, npages, flags;
2094 struct ldc_mtable_entry *base;
2095 struct cookie_state state;
2096 struct ldc_iommu *iommu;
2099 if (map_perm & ~LDC_MAP_ALL)
2102 err = sg_count_pages(sg, num_sg);
2112 spin_lock_irqsave(&iommu->lock, flags);
2113 base = alloc_npages(iommu, npages);
2114 spin_unlock_irqrestore(&iommu->lock, flags);
2119 state.page_table = iommu->page_table;
2120 state.cookies = cookies;
2121 state.mte_base = perm_to_mte(map_perm);
2122 state.prev_cookie = ~(u64)0;
2123 state.pte_idx = (base - iommu->page_table);
2126 for (i = 0; i < num_sg; i++)
2127 fill_cookies(&state, page_to_pfn(sg_page(&sg[i])) << PAGE_SHIFT,
2128 sg[i].offset, sg[i].length);
2132 EXPORT_SYMBOL(ldc_map_sg);
2134 int ldc_map_single(struct ldc_channel *lp,
2135 void *buf, unsigned int len,
2136 struct ldc_trans_cookie *cookies, int ncookies,
2137 unsigned int map_perm)
2139 unsigned long npages, pa, flags;
2140 struct ldc_mtable_entry *base;
2141 struct cookie_state state;
2142 struct ldc_iommu *iommu;
2144 if ((map_perm & ~LDC_MAP_ALL) || (ncookies < 1))
2148 if ((pa | len) & (8UL - 1))
2151 npages = pages_in_region(pa, len);
2155 spin_lock_irqsave(&iommu->lock, flags);
2156 base = alloc_npages(iommu, npages);
2157 spin_unlock_irqrestore(&iommu->lock, flags);
2162 state.page_table = iommu->page_table;
2163 state.cookies = cookies;
2164 state.mte_base = perm_to_mte(map_perm);
2165 state.prev_cookie = ~(u64)0;
2166 state.pte_idx = (base - iommu->page_table);
2168 fill_cookies(&state, (pa & PAGE_MASK), (pa & ~PAGE_MASK), len);
2169 BUG_ON(state.nc > ncookies);
2173 EXPORT_SYMBOL(ldc_map_single);
2175 static void free_npages(unsigned long id, struct ldc_iommu *iommu,
2176 u64 cookie, u64 size)
2178 struct iommu_arena *arena = &iommu->arena;
2179 unsigned long i, shift, index, npages;
2180 struct ldc_mtable_entry *base;
2182 npages = PAGE_ALIGN(((cookie & ~PAGE_MASK) + size)) >> PAGE_SHIFT;
2183 index = cookie_to_index(cookie, &shift);
2184 base = iommu->page_table + index;
2186 BUG_ON(index > arena->limit ||
2187 (index + npages) > arena->limit);
2189 for (i = 0; i < npages; i++) {
2191 sun4v_ldc_revoke(id, cookie + (i << shift),
2194 __clear_bit(index + i, arena->map);
2198 void ldc_unmap(struct ldc_channel *lp, struct ldc_trans_cookie *cookies,
2201 struct ldc_iommu *iommu = &lp->iommu;
2202 unsigned long flags;
2205 spin_lock_irqsave(&iommu->lock, flags);
2206 for (i = 0; i < ncookies; i++) {
2207 u64 addr = cookies[i].cookie_addr;
2208 u64 size = cookies[i].cookie_size;
2210 free_npages(lp->id, iommu, addr, size);
2212 spin_unlock_irqrestore(&iommu->lock, flags);
2214 EXPORT_SYMBOL(ldc_unmap);
2216 int ldc_copy(struct ldc_channel *lp, int copy_dir,
2217 void *buf, unsigned int len, unsigned long offset,
2218 struct ldc_trans_cookie *cookies, int ncookies)
2220 unsigned int orig_len;
2224 if (copy_dir != LDC_COPY_IN && copy_dir != LDC_COPY_OUT) {
2225 printk(KERN_ERR PFX "ldc_copy: ID[%lu] Bad copy_dir[%d]\n",
2231 if ((ra | len | offset) & (8UL - 1)) {
2232 printk(KERN_ERR PFX "ldc_copy: ID[%lu] Unaligned buffer "
2233 "ra[%lx] len[%x] offset[%lx]\n",
2234 lp->id, ra, len, offset);
2238 if (lp->hs_state != LDC_HS_COMPLETE ||
2239 (lp->flags & LDC_FLAG_RESET)) {
2240 printk(KERN_ERR PFX "ldc_copy: ID[%lu] Link down hs_state[%x] "
2241 "flags[%x]\n", lp->id, lp->hs_state, lp->flags);
2246 for (i = 0; i < ncookies; i++) {
2247 unsigned long cookie_raddr = cookies[i].cookie_addr;
2248 unsigned long this_len = cookies[i].cookie_size;
2249 unsigned long actual_len;
2251 if (unlikely(offset)) {
2252 unsigned long this_off = offset;
2254 if (this_off > this_len)
2255 this_off = this_len;
2258 this_len -= this_off;
2261 cookie_raddr += this_off;
2268 unsigned long hv_err;
2270 hv_err = sun4v_ldc_copy(lp->id, copy_dir,
2272 this_len, &actual_len);
2273 if (unlikely(hv_err)) {
2274 printk(KERN_ERR PFX "ldc_copy: ID[%lu] "
2277 if (lp->hs_state != LDC_HS_COMPLETE ||
2278 (lp->flags & LDC_FLAG_RESET))
2284 cookie_raddr += actual_len;
2287 if (actual_len == this_len)
2290 this_len -= actual_len;
2297 /* It is caller policy what to do about short copies.
2298 * For example, a networking driver can declare the
2299 * packet a runt and drop it.
2302 return orig_len - len;
2304 EXPORT_SYMBOL(ldc_copy);
2306 void *ldc_alloc_exp_dring(struct ldc_channel *lp, unsigned int len,
2307 struct ldc_trans_cookie *cookies, int *ncookies,
2308 unsigned int map_perm)
2313 if (len & (8UL - 1))
2314 return ERR_PTR(-EINVAL);
2316 buf = kzalloc(len, GFP_KERNEL);
2318 return ERR_PTR(-ENOMEM);
2320 err = ldc_map_single(lp, buf, len, cookies, *ncookies, map_perm);
2323 return ERR_PTR(err);
2329 EXPORT_SYMBOL(ldc_alloc_exp_dring);
2331 void ldc_free_exp_dring(struct ldc_channel *lp, void *buf, unsigned int len,
2332 struct ldc_trans_cookie *cookies, int ncookies)
2334 ldc_unmap(lp, cookies, ncookies);
2337 EXPORT_SYMBOL(ldc_free_exp_dring);
2339 static int __init ldc_init(void)
2341 unsigned long major, minor;
2342 struct mdesc_handle *hp;
2351 mp = mdesc_node_by_name(hp, MDESC_NODE_NULL, "platform");
2353 if (mp == MDESC_NODE_NULL)
2356 v = mdesc_get_property(hp, mp, "domaining-enabled", NULL);
2362 if (sun4v_hvapi_register(HV_GRP_LDOM, major, &minor)) {
2363 printk(KERN_INFO PFX "Could not register LDOM hvapi.\n");
2367 printk(KERN_INFO "%s", version);
2370 printk(KERN_INFO PFX "Domaining disabled.\n");
2373 ldom_domaining_enabled = 1;
2381 core_initcall(ldc_init);