netmap_mem2.c revision 270063
1/*
2 * Copyright (C) 2012-2014 Matteo Landi, Luigi Rizzo, Giuseppe Lettieri. All rights reserved.
3 *
4 * Redistribution and use in source and binary forms, with or without
5 * modification, are permitted provided that the following conditions
6 * are met:
7 *   1. Redistributions of source code must retain the above copyright
8 *      notice, this list of conditions and the following disclaimer.
9 *   2. Redistributions in binary form must reproduce the above copyright
10 *      notice, this list of conditions and the following disclaimer in the
11 *      documentation and/or other materials provided with the distribution.
12 *
13 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
14 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
15 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
16 * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
17 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
18 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
19 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
20 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
21 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
22 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
23 * SUCH DAMAGE.
24 */
25
26#ifdef linux
27#include "bsd_glue.h"
28#endif /* linux */
29
30#ifdef __APPLE__
31#include "osx_glue.h"
32#endif /* __APPLE__ */
33
34#ifdef __FreeBSD__
35#include <sys/cdefs.h> /* prerequisite */
36__FBSDID("$FreeBSD: head/sys/dev/netmap/netmap_mem2.c 270063 2014-08-16 15:00:01Z luigi $");
37
38#include <sys/types.h>
39#include <sys/malloc.h>
40#include <sys/proc.h>
41#include <vm/vm.h>	/* vtophys */
42#include <vm/pmap.h>	/* vtophys */
43#include <sys/socket.h> /* sockaddrs */
44#include <sys/selinfo.h>
45#include <sys/sysctl.h>
46#include <net/if.h>
47#include <net/if_var.h>
48#include <net/vnet.h>
49#include <machine/bus.h>	/* bus_dmamap_* */
50
51#endif /* __FreeBSD__ */
52
53#include <net/netmap.h>
54#include <dev/netmap/netmap_kern.h>
55#include "netmap_mem2.h"
56
57#define NETMAP_BUF_MAX_NUM	20*4096*2	/* large machine */
58
59#define NETMAP_POOL_MAX_NAMSZ	32
60
61
62enum {
63	NETMAP_IF_POOL   = 0,
64	NETMAP_RING_POOL,
65	NETMAP_BUF_POOL,
66	NETMAP_POOLS_NR
67};
68
69
70struct netmap_obj_params {
71	u_int size;
72	u_int num;
73};
74struct netmap_obj_pool {
75	char name[NETMAP_POOL_MAX_NAMSZ];	/* name of the allocator */
76
77	/* ---------------------------------------------------*/
78	/* these are only meaningful if the pool is finalized */
79	/* (see 'finalized' field in netmap_mem_d)            */
80	u_int objtotal;         /* actual total number of objects. */
81	u_int memtotal;		/* actual total memory space */
82	u_int numclusters;	/* actual number of clusters */
83
84	u_int objfree;          /* number of free objects. */
85
86	struct lut_entry *lut;  /* virt,phys addresses, objtotal entries */
87	uint32_t *bitmap;       /* one bit per buffer, 1 means free */
88	uint32_t bitmap_slots;	/* number of uint32 entries in bitmap */
89	/* ---------------------------------------------------*/
90
91	/* limits */
92	u_int objminsize;	/* minimum object size */
93	u_int objmaxsize;	/* maximum object size */
94	u_int nummin;		/* minimum number of objects */
95	u_int nummax;		/* maximum number of objects */
96
97	/* these are changed only by config */
98	u_int _objtotal;	/* total number of objects */
99	u_int _objsize;		/* object size */
100	u_int _clustsize;       /* cluster size */
101	u_int _clustentries;    /* objects per cluster */
102	u_int _numclusters;	/* number of clusters */
103
104	/* requested values */
105	u_int r_objtotal;
106	u_int r_objsize;
107};
108
109#ifdef linux
110// XXX a mtx would suffice here 20130415 lr
111#define NMA_LOCK_T		struct semaphore
112#else /* !linux */
113#define NMA_LOCK_T		struct mtx
114#endif /* linux */
115
116typedef int (*netmap_mem_config_t)(struct netmap_mem_d*);
117typedef int (*netmap_mem_finalize_t)(struct netmap_mem_d*);
118typedef void (*netmap_mem_deref_t)(struct netmap_mem_d*);
119
120typedef uint16_t nm_memid_t;
121
122struct netmap_mem_d {
123	NMA_LOCK_T nm_mtx;  /* protect the allocator */
124	u_int nm_totalsize; /* shorthand */
125
126	u_int flags;
127#define NETMAP_MEM_FINALIZED	0x1	/* preallocation done */
128	int lasterr;		/* last error for curr config */
129	int refcount;		/* existing priv structures */
130	/* the three allocators */
131	struct netmap_obj_pool pools[NETMAP_POOLS_NR];
132
133	netmap_mem_config_t   config;
134	netmap_mem_finalize_t finalize;
135	netmap_mem_deref_t    deref;
136
137	nm_memid_t nm_id;	/* allocator identifier */
138	int nm_grp;	/* iommu groupd id */
139
140	/* list of all existing allocators, sorted by nm_id */
141	struct netmap_mem_d *prev, *next;
142};
143
144/* accessor functions */
145struct lut_entry*
146netmap_mem_get_lut(struct netmap_mem_d *nmd)
147{
148	return nmd->pools[NETMAP_BUF_POOL].lut;
149}
150
151u_int
152netmap_mem_get_buftotal(struct netmap_mem_d *nmd)
153{
154	return nmd->pools[NETMAP_BUF_POOL].objtotal;
155}
156
157size_t
158netmap_mem_get_bufsize(struct netmap_mem_d *nmd)
159{
160	return nmd->pools[NETMAP_BUF_POOL]._objsize;
161}
162
163#ifdef linux
164#define NMA_LOCK_INIT(n)	sema_init(&(n)->nm_mtx, 1)
165#define NMA_LOCK_DESTROY(n)
166#define NMA_LOCK(n)		down(&(n)->nm_mtx)
167#define NMA_UNLOCK(n)		up(&(n)->nm_mtx)
168#else /* !linux */
169#define NMA_LOCK_INIT(n)	mtx_init(&(n)->nm_mtx, "netmap memory allocator lock", NULL, MTX_DEF)
170#define NMA_LOCK_DESTROY(n)	mtx_destroy(&(n)->nm_mtx)
171#define NMA_LOCK(n)		mtx_lock(&(n)->nm_mtx)
172#define NMA_UNLOCK(n)		mtx_unlock(&(n)->nm_mtx)
173#endif /* linux */
174
175
176struct netmap_obj_params netmap_params[NETMAP_POOLS_NR] = {
177	[NETMAP_IF_POOL] = {
178		.size = 1024,
179		.num  = 100,
180	},
181	[NETMAP_RING_POOL] = {
182		.size = 9*PAGE_SIZE,
183		.num  = 200,
184	},
185	[NETMAP_BUF_POOL] = {
186		.size = 2048,
187		.num  = NETMAP_BUF_MAX_NUM,
188	},
189};
190
191struct netmap_obj_params netmap_min_priv_params[NETMAP_POOLS_NR] = {
192	[NETMAP_IF_POOL] = {
193		.size = 1024,
194		.num  = 1,
195	},
196	[NETMAP_RING_POOL] = {
197		.size = 5*PAGE_SIZE,
198		.num  = 4,
199	},
200	[NETMAP_BUF_POOL] = {
201		.size = 2048,
202		.num  = 4098,
203	},
204};
205
206
207/*
208 * nm_mem is the memory allocator used for all physical interfaces
209 * running in netmap mode.
210 * Virtual (VALE) ports will have each its own allocator.
211 */
212static int netmap_mem_global_config(struct netmap_mem_d *nmd);
213static int netmap_mem_global_finalize(struct netmap_mem_d *nmd);
214static void netmap_mem_global_deref(struct netmap_mem_d *nmd);
215struct netmap_mem_d nm_mem = {	/* Our memory allocator. */
216	.pools = {
217		[NETMAP_IF_POOL] = {
218			.name 	= "netmap_if",
219			.objminsize = sizeof(struct netmap_if),
220			.objmaxsize = 4096,
221			.nummin     = 10,	/* don't be stingy */
222			.nummax	    = 10000,	/* XXX very large */
223		},
224		[NETMAP_RING_POOL] = {
225			.name 	= "netmap_ring",
226			.objminsize = sizeof(struct netmap_ring),
227			.objmaxsize = 32*PAGE_SIZE,
228			.nummin     = 2,
229			.nummax	    = 1024,
230		},
231		[NETMAP_BUF_POOL] = {
232			.name	= "netmap_buf",
233			.objminsize = 64,
234			.objmaxsize = 65536,
235			.nummin     = 4,
236			.nummax	    = 1000000, /* one million! */
237		},
238	},
239	.config   = netmap_mem_global_config,
240	.finalize = netmap_mem_global_finalize,
241	.deref    = netmap_mem_global_deref,
242
243	.nm_id = 1,
244	.nm_grp = -1,
245
246	.prev = &nm_mem,
247	.next = &nm_mem,
248};
249
250
251struct netmap_mem_d *netmap_last_mem_d = &nm_mem;
252
253/* blueprint for the private memory allocators */
254static int netmap_mem_private_config(struct netmap_mem_d *nmd);
255static int netmap_mem_private_finalize(struct netmap_mem_d *nmd);
256static void netmap_mem_private_deref(struct netmap_mem_d *nmd);
257const struct netmap_mem_d nm_blueprint = {
258	.pools = {
259		[NETMAP_IF_POOL] = {
260			.name 	= "%s_if",
261			.objminsize = sizeof(struct netmap_if),
262			.objmaxsize = 4096,
263			.nummin     = 1,
264			.nummax	    = 100,
265		},
266		[NETMAP_RING_POOL] = {
267			.name 	= "%s_ring",
268			.objminsize = sizeof(struct netmap_ring),
269			.objmaxsize = 32*PAGE_SIZE,
270			.nummin     = 2,
271			.nummax	    = 1024,
272		},
273		[NETMAP_BUF_POOL] = {
274			.name	= "%s_buf",
275			.objminsize = 64,
276			.objmaxsize = 65536,
277			.nummin     = 4,
278			.nummax	    = 1000000, /* one million! */
279		},
280	},
281	.config   = netmap_mem_private_config,
282	.finalize = netmap_mem_private_finalize,
283	.deref    = netmap_mem_private_deref,
284
285	.flags = NETMAP_MEM_PRIVATE,
286};
287
288/* memory allocator related sysctls */
289
290#define STRINGIFY(x) #x
291
292
293#define DECLARE_SYSCTLS(id, name) \
294	SYSCTL_INT(_dev_netmap, OID_AUTO, name##_size, \
295	    CTLFLAG_RW, &netmap_params[id].size, 0, "Requested size of netmap " STRINGIFY(name) "s"); \
296	SYSCTL_INT(_dev_netmap, OID_AUTO, name##_curr_size, \
297	    CTLFLAG_RD, &nm_mem.pools[id]._objsize, 0, "Current size of netmap " STRINGIFY(name) "s"); \
298	SYSCTL_INT(_dev_netmap, OID_AUTO, name##_num, \
299	    CTLFLAG_RW, &netmap_params[id].num, 0, "Requested number of netmap " STRINGIFY(name) "s"); \
300	SYSCTL_INT(_dev_netmap, OID_AUTO, name##_curr_num, \
301	    CTLFLAG_RD, &nm_mem.pools[id].objtotal, 0, "Current number of netmap " STRINGIFY(name) "s"); \
302	SYSCTL_INT(_dev_netmap, OID_AUTO, priv_##name##_size, \
303	    CTLFLAG_RW, &netmap_min_priv_params[id].size, 0, \
304	    "Default size of private netmap " STRINGIFY(name) "s"); \
305	SYSCTL_INT(_dev_netmap, OID_AUTO, priv_##name##_num, \
306	    CTLFLAG_RW, &netmap_min_priv_params[id].num, 0, \
307	    "Default number of private netmap " STRINGIFY(name) "s")
308
309SYSCTL_DECL(_dev_netmap);
310DECLARE_SYSCTLS(NETMAP_IF_POOL, if);
311DECLARE_SYSCTLS(NETMAP_RING_POOL, ring);
312DECLARE_SYSCTLS(NETMAP_BUF_POOL, buf);
313
314static int
315nm_mem_assign_id(struct netmap_mem_d *nmd)
316{
317	nm_memid_t id;
318	struct netmap_mem_d *scan = netmap_last_mem_d;
319	int error = ENOMEM;
320
321	NMA_LOCK(&nm_mem);
322
323	do {
324		/* we rely on unsigned wrap around */
325		id = scan->nm_id + 1;
326		if (id == 0) /* reserve 0 as error value */
327			id = 1;
328		scan = scan->next;
329		if (id != scan->nm_id) {
330			nmd->nm_id = id;
331			nmd->prev = scan->prev;
332			nmd->next = scan;
333			scan->prev->next = nmd;
334			scan->prev = nmd;
335			netmap_last_mem_d = nmd;
336			error = 0;
337			break;
338		}
339	} while (scan != netmap_last_mem_d);
340
341	NMA_UNLOCK(&nm_mem);
342	return error;
343}
344
345static void
346nm_mem_release_id(struct netmap_mem_d *nmd)
347{
348	NMA_LOCK(&nm_mem);
349
350	nmd->prev->next = nmd->next;
351	nmd->next->prev = nmd->prev;
352
353	if (netmap_last_mem_d == nmd)
354		netmap_last_mem_d = nmd->prev;
355
356	nmd->prev = nmd->next = NULL;
357
358	NMA_UNLOCK(&nm_mem);
359}
360
361static int
362nm_mem_assign_group(struct netmap_mem_d *nmd, struct device *dev)
363{
364	int err = 0, id;
365	id = nm_iommu_group_id(dev);
366	if (netmap_verbose)
367		D("iommu_group %d", id);
368
369	NMA_LOCK(nmd);
370
371	if (nmd->nm_grp < 0)
372		nmd->nm_grp = id;
373
374	if (nmd->nm_grp != id)
375		nmd->lasterr = err = ENOMEM;
376
377	NMA_UNLOCK(nmd);
378	return err;
379}
380
381/*
382 * First, find the allocator that contains the requested offset,
383 * then locate the cluster through a lookup table.
384 */
385vm_paddr_t
386netmap_mem_ofstophys(struct netmap_mem_d* nmd, vm_ooffset_t offset)
387{
388	int i;
389	vm_ooffset_t o = offset;
390	vm_paddr_t pa;
391	struct netmap_obj_pool *p;
392
393	NMA_LOCK(nmd);
394	p = nmd->pools;
395
396	for (i = 0; i < NETMAP_POOLS_NR; offset -= p[i].memtotal, i++) {
397		if (offset >= p[i].memtotal)
398			continue;
399		// now lookup the cluster's address
400		pa = vtophys(p[i].lut[offset / p[i]._objsize].vaddr) +
401			offset % p[i]._objsize;
402		NMA_UNLOCK(nmd);
403		return pa;
404	}
405	/* this is only in case of errors */
406	D("invalid ofs 0x%x out of 0x%x 0x%x 0x%x", (u_int)o,
407		p[NETMAP_IF_POOL].memtotal,
408		p[NETMAP_IF_POOL].memtotal
409			+ p[NETMAP_RING_POOL].memtotal,
410		p[NETMAP_IF_POOL].memtotal
411			+ p[NETMAP_RING_POOL].memtotal
412			+ p[NETMAP_BUF_POOL].memtotal);
413	NMA_UNLOCK(nmd);
414	return 0;	// XXX bad address
415}
416
417int
418netmap_mem_get_info(struct netmap_mem_d* nmd, u_int* size, u_int *memflags,
419	nm_memid_t *id)
420{
421	int error = 0;
422	NMA_LOCK(nmd);
423	error = nmd->config(nmd);
424	if (error)
425		goto out;
426	if (size) {
427		if (nmd->flags & NETMAP_MEM_FINALIZED) {
428			*size = nmd->nm_totalsize;
429		} else {
430			int i;
431			*size = 0;
432			for (i = 0; i < NETMAP_POOLS_NR; i++) {
433				struct netmap_obj_pool *p = nmd->pools + i;
434				*size += (p->_numclusters * p->_clustsize);
435			}
436		}
437	}
438	if (memflags)
439		*memflags = nmd->flags;
440	if (id)
441		*id = nmd->nm_id;
442out:
443	NMA_UNLOCK(nmd);
444	return error;
445}
446
447/*
448 * we store objects by kernel address, need to find the offset
449 * within the pool to export the value to userspace.
450 * Algorithm: scan until we find the cluster, then add the
451 * actual offset in the cluster
452 */
453static ssize_t
454netmap_obj_offset(struct netmap_obj_pool *p, const void *vaddr)
455{
456	int i, k = p->_clustentries, n = p->objtotal;
457	ssize_t ofs = 0;
458
459	for (i = 0; i < n; i += k, ofs += p->_clustsize) {
460		const char *base = p->lut[i].vaddr;
461		ssize_t relofs = (const char *) vaddr - base;
462
463		if (relofs < 0 || relofs >= p->_clustsize)
464			continue;
465
466		ofs = ofs + relofs;
467		ND("%s: return offset %d (cluster %d) for pointer %p",
468		    p->name, ofs, i, vaddr);
469		return ofs;
470	}
471	D("address %p is not contained inside any cluster (%s)",
472	    vaddr, p->name);
473	return 0; /* An error occurred */
474}
475
476/* Helper functions which convert virtual addresses to offsets */
477#define netmap_if_offset(n, v)					\
478	netmap_obj_offset(&(n)->pools[NETMAP_IF_POOL], (v))
479
480#define netmap_ring_offset(n, v)				\
481    ((n)->pools[NETMAP_IF_POOL].memtotal + 			\
482	netmap_obj_offset(&(n)->pools[NETMAP_RING_POOL], (v)))
483
484#define netmap_buf_offset(n, v)					\
485    ((n)->pools[NETMAP_IF_POOL].memtotal +			\
486	(n)->pools[NETMAP_RING_POOL].memtotal +		\
487	netmap_obj_offset(&(n)->pools[NETMAP_BUF_POOL], (v)))
488
489
490ssize_t
491netmap_mem_if_offset(struct netmap_mem_d *nmd, const void *addr)
492{
493	ssize_t v;
494	NMA_LOCK(nmd);
495	v = netmap_if_offset(nmd, addr);
496	NMA_UNLOCK(nmd);
497	return v;
498}
499
500/*
501 * report the index, and use start position as a hint,
502 * otherwise buffer allocation becomes terribly expensive.
503 */
504static void *
505netmap_obj_malloc(struct netmap_obj_pool *p, u_int len, uint32_t *start, uint32_t *index)
506{
507	uint32_t i = 0;			/* index in the bitmap */
508	uint32_t mask, j;		/* slot counter */
509	void *vaddr = NULL;
510
511	if (len > p->_objsize) {
512		D("%s request size %d too large", p->name, len);
513		// XXX cannot reduce the size
514		return NULL;
515	}
516
517	if (p->objfree == 0) {
518		D("no more %s objects", p->name);
519		return NULL;
520	}
521	if (start)
522		i = *start;
523
524	/* termination is guaranteed by p->free, but better check bounds on i */
525	while (vaddr == NULL && i < p->bitmap_slots)  {
526		uint32_t cur = p->bitmap[i];
527		if (cur == 0) { /* bitmask is fully used */
528			i++;
529			continue;
530		}
531		/* locate a slot */
532		for (j = 0, mask = 1; (cur & mask) == 0; j++, mask <<= 1)
533			;
534
535		p->bitmap[i] &= ~mask; /* mark object as in use */
536		p->objfree--;
537
538		vaddr = p->lut[i * 32 + j].vaddr;
539		if (index)
540			*index = i * 32 + j;
541	}
542	ND("%s allocator: allocated object @ [%d][%d]: vaddr %p", i, j, vaddr);
543
544	if (start)
545		*start = i;
546	return vaddr;
547}
548
549
550/*
551 * free by index, not by address.
552 * XXX should we also cleanup the content ?
553 */
554static int
555netmap_obj_free(struct netmap_obj_pool *p, uint32_t j)
556{
557	uint32_t *ptr, mask;
558
559	if (j >= p->objtotal) {
560		D("invalid index %u, max %u", j, p->objtotal);
561		return 1;
562	}
563	ptr = &p->bitmap[j / 32];
564	mask = (1 << (j % 32));
565	if (*ptr & mask) {
566		D("ouch, double free on buffer %d", j);
567		return 1;
568	} else {
569		*ptr |= mask;
570		p->objfree++;
571		return 0;
572	}
573}
574
575/*
576 * free by address. This is slow but is only used for a few
577 * objects (rings, nifp)
578 */
579static void
580netmap_obj_free_va(struct netmap_obj_pool *p, void *vaddr)
581{
582	u_int i, j, n = p->numclusters;
583
584	for (i = 0, j = 0; i < n; i++, j += p->_clustentries) {
585		void *base = p->lut[i * p->_clustentries].vaddr;
586		ssize_t relofs = (ssize_t) vaddr - (ssize_t) base;
587
588		/* Given address, is out of the scope of the current cluster.*/
589		if (vaddr < base || relofs >= p->_clustsize)
590			continue;
591
592		j = j + relofs / p->_objsize;
593		/* KASSERT(j != 0, ("Cannot free object 0")); */
594		netmap_obj_free(p, j);
595		return;
596	}
597	D("address %p is not contained inside any cluster (%s)",
598	    vaddr, p->name);
599}
600
601#define netmap_mem_bufsize(n)	\
602	((n)->pools[NETMAP_BUF_POOL]._objsize)
603
604#define netmap_if_malloc(n, len)	netmap_obj_malloc(&(n)->pools[NETMAP_IF_POOL], len, NULL, NULL)
605#define netmap_if_free(n, v)		netmap_obj_free_va(&(n)->pools[NETMAP_IF_POOL], (v))
606#define netmap_ring_malloc(n, len)	netmap_obj_malloc(&(n)->pools[NETMAP_RING_POOL], len, NULL, NULL)
607#define netmap_ring_free(n, v)		netmap_obj_free_va(&(n)->pools[NETMAP_RING_POOL], (v))
608#define netmap_buf_malloc(n, _pos, _index)			\
609	netmap_obj_malloc(&(n)->pools[NETMAP_BUF_POOL], netmap_mem_bufsize(n), _pos, _index)
610
611
612#if 0 // XXX unused
613/* Return the index associated to the given packet buffer */
614#define netmap_buf_index(n, v)						\
615    (netmap_obj_offset(&(n)->pools[NETMAP_BUF_POOL], (v)) / NETMAP_BDG_BUF_SIZE(n))
616#endif
617
618/*
619 * allocate extra buffers in a linked list.
620 * returns the actual number.
621 */
622uint32_t
623netmap_extra_alloc(struct netmap_adapter *na, uint32_t *head, uint32_t n)
624{
625	struct netmap_mem_d *nmd = na->nm_mem;
626	uint32_t i, pos = 0; /* opaque, scan position in the bitmap */
627
628	NMA_LOCK(nmd);
629
630	*head = 0;	/* default, 'null' index ie empty list */
631	for (i = 0 ; i < n; i++) {
632		uint32_t cur = *head;	/* save current head */
633		uint32_t *p = netmap_buf_malloc(nmd, &pos, head);
634		if (p == NULL) {
635			D("no more buffers after %d of %d", i, n);
636			*head = cur; /* restore */
637			break;
638		}
639		RD(5, "allocate buffer %d -> %d", *head, cur);
640		*p = cur; /* link to previous head */
641	}
642
643	NMA_UNLOCK(nmd);
644
645	return i;
646}
647
648static void
649netmap_extra_free(struct netmap_adapter *na, uint32_t head)
650{
651        struct lut_entry *lut = na->na_lut;
652	struct netmap_mem_d *nmd = na->nm_mem;
653	struct netmap_obj_pool *p = &nmd->pools[NETMAP_BUF_POOL];
654	uint32_t i, cur, *buf;
655
656	D("freeing the extra list");
657	for (i = 0; head >=2 && head < p->objtotal; i++) {
658		cur = head;
659		buf = lut[head].vaddr;
660		head = *buf;
661		*buf = 0;
662		if (netmap_obj_free(p, cur))
663			break;
664	}
665	if (head != 0)
666		D("breaking with head %d", head);
667	D("freed %d buffers", i);
668}
669
670
671/* Return nonzero on error */
672static int
673netmap_new_bufs(struct netmap_mem_d *nmd, struct netmap_slot *slot, u_int n)
674{
675	struct netmap_obj_pool *p = &nmd->pools[NETMAP_BUF_POOL];
676	u_int i = 0;	/* slot counter */
677	uint32_t pos = 0;	/* slot in p->bitmap */
678	uint32_t index = 0;	/* buffer index */
679
680	for (i = 0; i < n; i++) {
681		void *vaddr = netmap_buf_malloc(nmd, &pos, &index);
682		if (vaddr == NULL) {
683			D("no more buffers after %d of %d", i, n);
684			goto cleanup;
685		}
686		slot[i].buf_idx = index;
687		slot[i].len = p->_objsize;
688		slot[i].flags = 0;
689	}
690
691	ND("allocated %d buffers, %d available, first at %d", n, p->objfree, pos);
692	return (0);
693
694cleanup:
695	while (i > 0) {
696		i--;
697		netmap_obj_free(p, slot[i].buf_idx);
698	}
699	bzero(slot, n * sizeof(slot[0]));
700	return (ENOMEM);
701}
702
703static void
704netmap_mem_set_ring(struct netmap_mem_d *nmd, struct netmap_slot *slot, u_int n, uint32_t index)
705{
706	struct netmap_obj_pool *p = &nmd->pools[NETMAP_BUF_POOL];
707	u_int i;
708
709	for (i = 0; i < n; i++) {
710		slot[i].buf_idx = index;
711		slot[i].len = p->_objsize;
712		slot[i].flags = 0;
713	}
714}
715
716
717static void
718netmap_free_buf(struct netmap_mem_d *nmd, uint32_t i)
719{
720	struct netmap_obj_pool *p = &nmd->pools[NETMAP_BUF_POOL];
721
722	if (i < 2 || i >= p->objtotal) {
723		D("Cannot free buf#%d: should be in [2, %d[", i, p->objtotal);
724		return;
725	}
726	netmap_obj_free(p, i);
727}
728
729
730static void
731netmap_free_bufs(struct netmap_mem_d *nmd, struct netmap_slot *slot, u_int n)
732{
733	u_int i;
734
735	for (i = 0; i < n; i++) {
736		if (slot[i].buf_idx > 2)
737			netmap_free_buf(nmd, slot[i].buf_idx);
738	}
739}
740
741static void
742netmap_reset_obj_allocator(struct netmap_obj_pool *p)
743{
744
745	if (p == NULL)
746		return;
747	if (p->bitmap)
748		free(p->bitmap, M_NETMAP);
749	p->bitmap = NULL;
750	if (p->lut) {
751		u_int i;
752		size_t sz = p->_clustsize;
753
754		for (i = 0; i < p->objtotal; i += p->_clustentries) {
755			if (p->lut[i].vaddr)
756				contigfree(p->lut[i].vaddr, sz, M_NETMAP);
757		}
758		bzero(p->lut, sizeof(struct lut_entry) * p->objtotal);
759#ifdef linux
760		vfree(p->lut);
761#else
762		free(p->lut, M_NETMAP);
763#endif
764	}
765	p->lut = NULL;
766	p->objtotal = 0;
767	p->memtotal = 0;
768	p->numclusters = 0;
769	p->objfree = 0;
770}
771
772/*
773 * Free all resources related to an allocator.
774 */
775static void
776netmap_destroy_obj_allocator(struct netmap_obj_pool *p)
777{
778	if (p == NULL)
779		return;
780	netmap_reset_obj_allocator(p);
781}
782
783/*
784 * We receive a request for objtotal objects, of size objsize each.
785 * Internally we may round up both numbers, as we allocate objects
786 * in small clusters multiple of the page size.
787 * We need to keep track of objtotal and clustentries,
788 * as they are needed when freeing memory.
789 *
790 * XXX note -- userspace needs the buffers to be contiguous,
791 *	so we cannot afford gaps at the end of a cluster.
792 */
793
794
795/* call with NMA_LOCK held */
796static int
797netmap_config_obj_allocator(struct netmap_obj_pool *p, u_int objtotal, u_int objsize)
798{
799	int i;
800	u_int clustsize;	/* the cluster size, multiple of page size */
801	u_int clustentries;	/* how many objects per entry */
802
803	/* we store the current request, so we can
804	 * detect configuration changes later */
805	p->r_objtotal = objtotal;
806	p->r_objsize = objsize;
807
808#define MAX_CLUSTSIZE	(1<<22)		// 4 MB
809#define LINE_ROUND	NM_CACHE_ALIGN	// 64
810	if (objsize >= MAX_CLUSTSIZE) {
811		/* we could do it but there is no point */
812		D("unsupported allocation for %d bytes", objsize);
813		return EINVAL;
814	}
815	/* make sure objsize is a multiple of LINE_ROUND */
816	i = (objsize & (LINE_ROUND - 1));
817	if (i) {
818		D("XXX aligning object by %d bytes", LINE_ROUND - i);
819		objsize += LINE_ROUND - i;
820	}
821	if (objsize < p->objminsize || objsize > p->objmaxsize) {
822		D("requested objsize %d out of range [%d, %d]",
823			objsize, p->objminsize, p->objmaxsize);
824		return EINVAL;
825	}
826	if (objtotal < p->nummin || objtotal > p->nummax) {
827		D("requested objtotal %d out of range [%d, %d]",
828			objtotal, p->nummin, p->nummax);
829		return EINVAL;
830	}
831	/*
832	 * Compute number of objects using a brute-force approach:
833	 * given a max cluster size,
834	 * we try to fill it with objects keeping track of the
835	 * wasted space to the next page boundary.
836	 */
837	for (clustentries = 0, i = 1;; i++) {
838		u_int delta, used = i * objsize;
839		if (used > MAX_CLUSTSIZE)
840			break;
841		delta = used % PAGE_SIZE;
842		if (delta == 0) { // exact solution
843			clustentries = i;
844			break;
845		}
846	}
847	/* exact solution not found */
848	if (clustentries == 0) {
849		D("unsupported allocation for %d bytes", objsize);
850		return EINVAL;
851	}
852	/* compute clustsize */
853	clustsize = clustentries * objsize;
854	if (netmap_verbose)
855		D("objsize %d clustsize %d objects %d",
856			objsize, clustsize, clustentries);
857
858	/*
859	 * The number of clusters is n = ceil(objtotal/clustentries)
860	 * objtotal' = n * clustentries
861	 */
862	p->_clustentries = clustentries;
863	p->_clustsize = clustsize;
864	p->_numclusters = (objtotal + clustentries - 1) / clustentries;
865
866	/* actual values (may be larger than requested) */
867	p->_objsize = objsize;
868	p->_objtotal = p->_numclusters * clustentries;
869
870	return 0;
871}
872
873
874/* call with NMA_LOCK held */
875static int
876netmap_finalize_obj_allocator(struct netmap_obj_pool *p)
877{
878	int i; /* must be signed */
879	size_t n;
880
881	/* optimistically assume we have enough memory */
882	p->numclusters = p->_numclusters;
883	p->objtotal = p->_objtotal;
884
885	n = sizeof(struct lut_entry) * p->objtotal;
886#ifdef linux
887	p->lut = vmalloc(n);
888#else
889	p->lut = malloc(n, M_NETMAP, M_NOWAIT | M_ZERO);
890#endif
891	if (p->lut == NULL) {
892		D("Unable to create lookup table (%d bytes) for '%s'", (int)n, p->name);
893		goto clean;
894	}
895
896	/* Allocate the bitmap */
897	n = (p->objtotal + 31) / 32;
898	p->bitmap = malloc(sizeof(uint32_t) * n, M_NETMAP, M_NOWAIT | M_ZERO);
899	if (p->bitmap == NULL) {
900		D("Unable to create bitmap (%d entries) for allocator '%s'", (int)n,
901		    p->name);
902		goto clean;
903	}
904	p->bitmap_slots = n;
905
906	/*
907	 * Allocate clusters, init pointers and bitmap
908	 */
909
910	n = p->_clustsize;
911	for (i = 0; i < (int)p->objtotal;) {
912		int lim = i + p->_clustentries;
913		char *clust;
914
915		clust = contigmalloc(n, M_NETMAP, M_NOWAIT | M_ZERO,
916		    (size_t)0, -1UL, PAGE_SIZE, 0);
917		if (clust == NULL) {
918			/*
919			 * If we get here, there is a severe memory shortage,
920			 * so halve the allocated memory to reclaim some.
921			 */
922			D("Unable to create cluster at %d for '%s' allocator",
923			    i, p->name);
924			if (i < 2) /* nothing to halve */
925				goto out;
926			lim = i / 2;
927			for (i--; i >= lim; i--) {
928				p->bitmap[ (i>>5) ] &=  ~( 1 << (i & 31) );
929				if (i % p->_clustentries == 0 && p->lut[i].vaddr)
930					contigfree(p->lut[i].vaddr,
931						n, M_NETMAP);
932			}
933		out:
934			p->objtotal = i;
935			/* we may have stopped in the middle of a cluster */
936			p->numclusters = (i + p->_clustentries - 1) / p->_clustentries;
937			break;
938		}
939		for (; i < lim; i++, clust += p->_objsize) {
940			p->bitmap[ (i>>5) ] |=  ( 1 << (i & 31) );
941			p->lut[i].vaddr = clust;
942			p->lut[i].paddr = vtophys(clust);
943		}
944	}
945	p->objfree = p->objtotal;
946	p->memtotal = p->numclusters * p->_clustsize;
947	if (p->objfree == 0)
948		goto clean;
949	if (netmap_verbose)
950		D("Pre-allocated %d clusters (%d/%dKB) for '%s'",
951		    p->numclusters, p->_clustsize >> 10,
952		    p->memtotal >> 10, p->name);
953
954	return 0;
955
956clean:
957	netmap_reset_obj_allocator(p);
958	return ENOMEM;
959}
960
961/* call with lock held */
962static int
963netmap_memory_config_changed(struct netmap_mem_d *nmd)
964{
965	int i;
966
967	for (i = 0; i < NETMAP_POOLS_NR; i++) {
968		if (nmd->pools[i].r_objsize != netmap_params[i].size ||
969		    nmd->pools[i].r_objtotal != netmap_params[i].num)
970		    return 1;
971	}
972	return 0;
973}
974
975static void
976netmap_mem_reset_all(struct netmap_mem_d *nmd)
977{
978	int i;
979
980	if (netmap_verbose)
981		D("resetting %p", nmd);
982	for (i = 0; i < NETMAP_POOLS_NR; i++) {
983		netmap_reset_obj_allocator(&nmd->pools[i]);
984	}
985	nmd->flags  &= ~NETMAP_MEM_FINALIZED;
986}
987
988static int
989netmap_mem_unmap(struct netmap_obj_pool *p, struct netmap_adapter *na)
990{
991	int i, lim = p->_objtotal;
992
993	if (na->pdev == NULL)
994		return 0;
995
996#ifdef __FreeBSD__
997	(void)i;
998	(void)lim;
999	D("unsupported on FreeBSD");
1000#else /* linux */
1001	for (i = 2; i < lim; i++) {
1002		netmap_unload_map(na, (bus_dma_tag_t) na->pdev, &p->lut[i].paddr);
1003	}
1004#endif /* linux */
1005
1006	return 0;
1007}
1008
1009static int
1010netmap_mem_map(struct netmap_obj_pool *p, struct netmap_adapter *na)
1011{
1012#ifdef __FreeBSD__
1013	D("unsupported on FreeBSD");
1014#else /* linux */
1015	int i, lim = p->_objtotal;
1016
1017	if (na->pdev == NULL)
1018		return 0;
1019
1020	for (i = 2; i < lim; i++) {
1021		netmap_load_map(na, (bus_dma_tag_t) na->pdev, &p->lut[i].paddr,
1022				p->lut[i].vaddr);
1023	}
1024#endif /* linux */
1025
1026	return 0;
1027}
1028
1029static int
1030netmap_mem_finalize_all(struct netmap_mem_d *nmd)
1031{
1032	int i;
1033	if (nmd->flags & NETMAP_MEM_FINALIZED)
1034		return 0;
1035	nmd->lasterr = 0;
1036	nmd->nm_totalsize = 0;
1037	for (i = 0; i < NETMAP_POOLS_NR; i++) {
1038		nmd->lasterr = netmap_finalize_obj_allocator(&nmd->pools[i]);
1039		if (nmd->lasterr)
1040			goto error;
1041		nmd->nm_totalsize += nmd->pools[i].memtotal;
1042	}
1043	/* buffers 0 and 1 are reserved */
1044	nmd->pools[NETMAP_BUF_POOL].objfree -= 2;
1045	nmd->pools[NETMAP_BUF_POOL].bitmap[0] = ~3;
1046	nmd->flags |= NETMAP_MEM_FINALIZED;
1047
1048	if (netmap_verbose)
1049		D("interfaces %d KB, rings %d KB, buffers %d MB",
1050		    nmd->pools[NETMAP_IF_POOL].memtotal >> 10,
1051		    nmd->pools[NETMAP_RING_POOL].memtotal >> 10,
1052		    nmd->pools[NETMAP_BUF_POOL].memtotal >> 20);
1053
1054	if (netmap_verbose)
1055		D("Free buffers: %d", nmd->pools[NETMAP_BUF_POOL].objfree);
1056
1057
1058	return 0;
1059error:
1060	netmap_mem_reset_all(nmd);
1061	return nmd->lasterr;
1062}
1063
1064
1065
1066void
1067netmap_mem_private_delete(struct netmap_mem_d *nmd)
1068{
1069	if (nmd == NULL)
1070		return;
1071	if (netmap_verbose)
1072		D("deleting %p", nmd);
1073	if (nmd->refcount > 0)
1074		D("bug: deleting mem allocator with refcount=%d!", nmd->refcount);
1075	nm_mem_release_id(nmd);
1076	if (netmap_verbose)
1077		D("done deleting %p", nmd);
1078	NMA_LOCK_DESTROY(nmd);
1079	free(nmd, M_DEVBUF);
1080}
1081
1082static int
1083netmap_mem_private_config(struct netmap_mem_d *nmd)
1084{
1085	/* nothing to do, we are configured on creation
1086 	 * and configuration never changes thereafter
1087 	 */
1088	return 0;
1089}
1090
1091static int
1092netmap_mem_private_finalize(struct netmap_mem_d *nmd)
1093{
1094	int err;
1095	NMA_LOCK(nmd);
1096	nmd->refcount++;
1097	err = netmap_mem_finalize_all(nmd);
1098	NMA_UNLOCK(nmd);
1099	return err;
1100
1101}
1102
1103static void
1104netmap_mem_private_deref(struct netmap_mem_d *nmd)
1105{
1106	NMA_LOCK(nmd);
1107	if (--nmd->refcount <= 0)
1108		netmap_mem_reset_all(nmd);
1109	NMA_UNLOCK(nmd);
1110}
1111
1112
1113/*
1114 * allocator for private memory
1115 */
1116struct netmap_mem_d *
1117netmap_mem_private_new(const char *name, u_int txr, u_int txd,
1118	u_int rxr, u_int rxd, u_int extra_bufs, u_int npipes, int *perr)
1119{
1120	struct netmap_mem_d *d = NULL;
1121	struct netmap_obj_params p[NETMAP_POOLS_NR];
1122	int i, err;
1123	u_int v, maxd;
1124
1125	d = malloc(sizeof(struct netmap_mem_d),
1126			M_DEVBUF, M_NOWAIT | M_ZERO);
1127	if (d == NULL) {
1128		err = ENOMEM;
1129		goto error;
1130	}
1131
1132	*d = nm_blueprint;
1133
1134	err = nm_mem_assign_id(d);
1135	if (err)
1136		goto error;
1137
1138	/* account for the fake host rings */
1139	txr++;
1140	rxr++;
1141
1142	/* copy the min values */
1143	for (i = 0; i < NETMAP_POOLS_NR; i++) {
1144		p[i] = netmap_min_priv_params[i];
1145	}
1146
1147	/* possibly increase them to fit user request */
1148	v = sizeof(struct netmap_if) + sizeof(ssize_t) * (txr + rxr);
1149	if (p[NETMAP_IF_POOL].size < v)
1150		p[NETMAP_IF_POOL].size = v;
1151	v = 2 + 4 * npipes;
1152	if (p[NETMAP_IF_POOL].num < v)
1153		p[NETMAP_IF_POOL].num = v;
1154	maxd = (txd > rxd) ? txd : rxd;
1155	v = sizeof(struct netmap_ring) + sizeof(struct netmap_slot) * maxd;
1156	if (p[NETMAP_RING_POOL].size < v)
1157		p[NETMAP_RING_POOL].size = v;
1158	/* each pipe endpoint needs two tx rings (1 normal + 1 host, fake)
1159         * and two rx rings (again, 1 normal and 1 fake host)
1160         */
1161	v = txr + rxr + 8 * npipes;
1162	if (p[NETMAP_RING_POOL].num < v)
1163		p[NETMAP_RING_POOL].num = v;
1164	/* for each pipe we only need the buffers for the 4 "real" rings.
1165         * On the other end, the pipe ring dimension may be different from
1166         * the parent port ring dimension. As a compromise, we allocate twice the
1167         * space actually needed if the pipe rings were the same size as the parent rings
1168         */
1169	v = (4 * npipes + rxr) * rxd + (4 * npipes + txr) * txd + 2 + extra_bufs;
1170		/* the +2 is for the tx and rx fake buffers (indices 0 and 1) */
1171	if (p[NETMAP_BUF_POOL].num < v)
1172		p[NETMAP_BUF_POOL].num = v;
1173
1174	if (netmap_verbose)
1175		D("req if %d*%d ring %d*%d buf %d*%d",
1176			p[NETMAP_IF_POOL].num,
1177			p[NETMAP_IF_POOL].size,
1178			p[NETMAP_RING_POOL].num,
1179			p[NETMAP_RING_POOL].size,
1180			p[NETMAP_BUF_POOL].num,
1181			p[NETMAP_BUF_POOL].size);
1182
1183	for (i = 0; i < NETMAP_POOLS_NR; i++) {
1184		snprintf(d->pools[i].name, NETMAP_POOL_MAX_NAMSZ,
1185				nm_blueprint.pools[i].name,
1186				name);
1187		err = netmap_config_obj_allocator(&d->pools[i],
1188				p[i].num, p[i].size);
1189		if (err)
1190			goto error;
1191	}
1192
1193	d->flags &= ~NETMAP_MEM_FINALIZED;
1194
1195	NMA_LOCK_INIT(d);
1196
1197	return d;
1198error:
1199	netmap_mem_private_delete(d);
1200	if (perr)
1201		*perr = err;
1202	return NULL;
1203}
1204
1205
1206/* call with lock held */
1207static int
1208netmap_mem_global_config(struct netmap_mem_d *nmd)
1209{
1210	int i;
1211
1212	if (nmd->refcount)
1213		/* already in use, we cannot change the configuration */
1214		goto out;
1215
1216	if (!netmap_memory_config_changed(nmd))
1217		goto out;
1218
1219	D("reconfiguring");
1220
1221	if (nmd->flags & NETMAP_MEM_FINALIZED) {
1222		/* reset previous allocation */
1223		for (i = 0; i < NETMAP_POOLS_NR; i++) {
1224			netmap_reset_obj_allocator(&nmd->pools[i]);
1225		}
1226		nmd->flags &= ~NETMAP_MEM_FINALIZED;
1227	}
1228
1229	for (i = 0; i < NETMAP_POOLS_NR; i++) {
1230		nmd->lasterr = netmap_config_obj_allocator(&nmd->pools[i],
1231				netmap_params[i].num, netmap_params[i].size);
1232		if (nmd->lasterr)
1233			goto out;
1234	}
1235
1236out:
1237
1238	return nmd->lasterr;
1239}
1240
1241static int
1242netmap_mem_global_finalize(struct netmap_mem_d *nmd)
1243{
1244	int err;
1245
1246	NMA_LOCK(nmd);
1247
1248
1249	/* update configuration if changed */
1250	if (netmap_mem_global_config(nmd))
1251		goto out;
1252
1253	nmd->refcount++;
1254
1255	if (nmd->flags & NETMAP_MEM_FINALIZED) {
1256		/* may happen if config is not changed */
1257		ND("nothing to do");
1258		goto out;
1259	}
1260
1261	if (netmap_mem_finalize_all(nmd))
1262		goto out;
1263
1264	nmd->lasterr = 0;
1265
1266out:
1267	if (nmd->lasterr)
1268		nmd->refcount--;
1269	err = nmd->lasterr;
1270
1271	NMA_UNLOCK(nmd);
1272
1273	return err;
1274
1275}
1276
1277int
1278netmap_mem_init(void)
1279{
1280	NMA_LOCK_INIT(&nm_mem);
1281	return (0);
1282}
1283
1284void
1285netmap_mem_fini(void)
1286{
1287	int i;
1288
1289	for (i = 0; i < NETMAP_POOLS_NR; i++) {
1290	    netmap_destroy_obj_allocator(&nm_mem.pools[i]);
1291	}
1292	NMA_LOCK_DESTROY(&nm_mem);
1293}
1294
1295static void
1296netmap_free_rings(struct netmap_adapter *na)
1297{
1298	struct netmap_kring *kring;
1299	struct netmap_ring *ring;
1300	if (!na->tx_rings)
1301		return;
1302	for (kring = na->tx_rings; kring != na->rx_rings; kring++) {
1303		ring = kring->ring;
1304		if (ring == NULL)
1305			continue;
1306		netmap_free_bufs(na->nm_mem, ring->slot, kring->nkr_num_slots);
1307		netmap_ring_free(na->nm_mem, ring);
1308		kring->ring = NULL;
1309	}
1310	for (/* cont'd from above */; kring != na->tailroom; kring++) {
1311		ring = kring->ring;
1312		if (ring == NULL)
1313			continue;
1314		netmap_free_bufs(na->nm_mem, ring->slot, kring->nkr_num_slots);
1315		netmap_ring_free(na->nm_mem, ring);
1316		kring->ring = NULL;
1317	}
1318}
1319
1320/* call with NMA_LOCK held *
1321 *
1322 * Allocate netmap rings and buffers for this card
1323 * The rings are contiguous, but have variable size.
1324 * The kring array must follow the layout described
1325 * in netmap_krings_create().
1326 */
1327int
1328netmap_mem_rings_create(struct netmap_adapter *na)
1329{
1330	struct netmap_ring *ring;
1331	u_int len, ndesc;
1332	struct netmap_kring *kring;
1333	u_int i;
1334
1335	NMA_LOCK(na->nm_mem);
1336
1337        /* transmit rings */
1338	for (i =0, kring = na->tx_rings; kring != na->rx_rings; kring++, i++) {
1339		if (kring->ring) {
1340			ND("%s %ld already created", kring->name, kring - na->tx_rings);
1341			continue; /* already created by somebody else */
1342		}
1343		ndesc = kring->nkr_num_slots;
1344		len = sizeof(struct netmap_ring) +
1345			  ndesc * sizeof(struct netmap_slot);
1346		ring = netmap_ring_malloc(na->nm_mem, len);
1347		if (ring == NULL) {
1348			D("Cannot allocate tx_ring");
1349			goto cleanup;
1350		}
1351		ND("txring at %p", ring);
1352		kring->ring = ring;
1353		*(uint32_t *)(uintptr_t)&ring->num_slots = ndesc;
1354		*(int64_t *)(uintptr_t)&ring->buf_ofs =
1355		    (na->nm_mem->pools[NETMAP_IF_POOL].memtotal +
1356			na->nm_mem->pools[NETMAP_RING_POOL].memtotal) -
1357			netmap_ring_offset(na->nm_mem, ring);
1358
1359		/* copy values from kring */
1360		ring->head = kring->rhead;
1361		ring->cur = kring->rcur;
1362		ring->tail = kring->rtail;
1363		*(uint16_t *)(uintptr_t)&ring->nr_buf_size =
1364			netmap_mem_bufsize(na->nm_mem);
1365		ND("%s h %d c %d t %d", kring->name,
1366			ring->head, ring->cur, ring->tail);
1367		ND("initializing slots for txring");
1368		if (i != na->num_tx_rings || (na->na_flags & NAF_HOST_RINGS)) {
1369			/* this is a real ring */
1370			if (netmap_new_bufs(na->nm_mem, ring->slot, ndesc)) {
1371				D("Cannot allocate buffers for tx_ring");
1372				goto cleanup;
1373			}
1374		} else {
1375			/* this is a fake tx ring, set all indices to 0 */
1376			netmap_mem_set_ring(na->nm_mem, ring->slot, ndesc, 0);
1377		}
1378	}
1379
1380	/* receive rings */
1381	for ( i = 0 /* kring cont'd from above */ ; kring != na->tailroom; kring++, i++) {
1382		if (kring->ring) {
1383			ND("%s %ld already created", kring->name, kring - na->rx_rings);
1384			continue; /* already created by somebody else */
1385		}
1386		ndesc = kring->nkr_num_slots;
1387		len = sizeof(struct netmap_ring) +
1388			  ndesc * sizeof(struct netmap_slot);
1389		ring = netmap_ring_malloc(na->nm_mem, len);
1390		if (ring == NULL) {
1391			D("Cannot allocate rx_ring");
1392			goto cleanup;
1393		}
1394		ND("rxring at %p", ring);
1395		kring->ring = ring;
1396		*(uint32_t *)(uintptr_t)&ring->num_slots = ndesc;
1397		*(int64_t *)(uintptr_t)&ring->buf_ofs =
1398		    (na->nm_mem->pools[NETMAP_IF_POOL].memtotal +
1399		        na->nm_mem->pools[NETMAP_RING_POOL].memtotal) -
1400			netmap_ring_offset(na->nm_mem, ring);
1401
1402		/* copy values from kring */
1403		ring->head = kring->rhead;
1404		ring->cur = kring->rcur;
1405		ring->tail = kring->rtail;
1406		*(int *)(uintptr_t)&ring->nr_buf_size =
1407			netmap_mem_bufsize(na->nm_mem);
1408		ND("%s h %d c %d t %d", kring->name,
1409			ring->head, ring->cur, ring->tail);
1410		ND("initializing slots for rxring %p", ring);
1411		if (i != na->num_rx_rings || (na->na_flags & NAF_HOST_RINGS)) {
1412			/* this is a real ring */
1413			if (netmap_new_bufs(na->nm_mem, ring->slot, ndesc)) {
1414				D("Cannot allocate buffers for rx_ring");
1415				goto cleanup;
1416			}
1417		} else {
1418			/* this is a fake rx ring, set all indices to 1 */
1419			netmap_mem_set_ring(na->nm_mem, ring->slot, ndesc, 1);
1420		}
1421	}
1422
1423	NMA_UNLOCK(na->nm_mem);
1424
1425	return 0;
1426
1427cleanup:
1428	netmap_free_rings(na);
1429
1430	NMA_UNLOCK(na->nm_mem);
1431
1432	return ENOMEM;
1433}
1434
1435void
1436netmap_mem_rings_delete(struct netmap_adapter *na)
1437{
1438	/* last instance, release bufs and rings */
1439	NMA_LOCK(na->nm_mem);
1440
1441	netmap_free_rings(na);
1442
1443	NMA_UNLOCK(na->nm_mem);
1444}
1445
1446
1447/* call with NMA_LOCK held */
1448/*
1449 * Allocate the per-fd structure netmap_if.
1450 *
1451 * We assume that the configuration stored in na
1452 * (number of tx/rx rings and descs) does not change while
1453 * the interface is in netmap mode.
1454 */
1455struct netmap_if *
1456netmap_mem_if_new(struct netmap_adapter *na)
1457{
1458	struct netmap_if *nifp;
1459	ssize_t base; /* handy for relative offsets between rings and nifp */
1460	u_int i, len, ntx, nrx;
1461
1462	/* account for the (eventually fake) host rings */
1463	ntx = na->num_tx_rings + 1;
1464	nrx = na->num_rx_rings + 1;
1465	/*
1466	 * the descriptor is followed inline by an array of offsets
1467	 * to the tx and rx rings in the shared memory region.
1468	 */
1469
1470	NMA_LOCK(na->nm_mem);
1471
1472	len = sizeof(struct netmap_if) + (nrx + ntx) * sizeof(ssize_t);
1473	nifp = netmap_if_malloc(na->nm_mem, len);
1474	if (nifp == NULL) {
1475		NMA_UNLOCK(na->nm_mem);
1476		return NULL;
1477	}
1478
1479	/* initialize base fields -- override const */
1480	*(u_int *)(uintptr_t)&nifp->ni_tx_rings = na->num_tx_rings;
1481	*(u_int *)(uintptr_t)&nifp->ni_rx_rings = na->num_rx_rings;
1482	strncpy(nifp->ni_name, na->name, (size_t)IFNAMSIZ);
1483
1484	/*
1485	 * fill the slots for the rx and tx rings. They contain the offset
1486	 * between the ring and nifp, so the information is usable in
1487	 * userspace to reach the ring from the nifp.
1488	 */
1489	base = netmap_if_offset(na->nm_mem, nifp);
1490	for (i = 0; i < ntx; i++) {
1491		*(ssize_t *)(uintptr_t)&nifp->ring_ofs[i] =
1492			netmap_ring_offset(na->nm_mem, na->tx_rings[i].ring) - base;
1493	}
1494	for (i = 0; i < nrx; i++) {
1495		*(ssize_t *)(uintptr_t)&nifp->ring_ofs[i+ntx] =
1496			netmap_ring_offset(na->nm_mem, na->rx_rings[i].ring) - base;
1497	}
1498
1499	NMA_UNLOCK(na->nm_mem);
1500
1501	return (nifp);
1502}
1503
1504void
1505netmap_mem_if_delete(struct netmap_adapter *na, struct netmap_if *nifp)
1506{
1507	if (nifp == NULL)
1508		/* nothing to do */
1509		return;
1510	NMA_LOCK(na->nm_mem);
1511	if (nifp->ni_bufs_head)
1512		netmap_extra_free(na, nifp->ni_bufs_head);
1513	netmap_if_free(na->nm_mem, nifp);
1514
1515	NMA_UNLOCK(na->nm_mem);
1516}
1517
1518static void
1519netmap_mem_global_deref(struct netmap_mem_d *nmd)
1520{
1521	NMA_LOCK(nmd);
1522
1523	nmd->refcount--;
1524	if (!nmd->refcount)
1525		nmd->nm_grp = -1;
1526	if (netmap_verbose)
1527		D("refcount = %d", nmd->refcount);
1528
1529	NMA_UNLOCK(nmd);
1530}
1531
1532int
1533netmap_mem_finalize(struct netmap_mem_d *nmd, struct netmap_adapter *na)
1534{
1535	if (nm_mem_assign_group(nmd, na->pdev) < 0) {
1536		return ENOMEM;
1537	} else {
1538		nmd->finalize(nmd);
1539	}
1540
1541	if (!nmd->lasterr && na->pdev)
1542		netmap_mem_map(&nmd->pools[NETMAP_BUF_POOL], na);
1543
1544	return nmd->lasterr;
1545}
1546
1547void
1548netmap_mem_deref(struct netmap_mem_d *nmd, struct netmap_adapter *na)
1549{
1550	NMA_LOCK(nmd);
1551	netmap_mem_unmap(&nmd->pools[NETMAP_BUF_POOL], na);
1552	NMA_UNLOCK(nmd);
1553	return nmd->deref(nmd);
1554}
1555