xref: /freebsd/sys/dev/netmap/netmap_mem2.c (revision d4ae33f0721c1b170fe37d97e395228ffcfb3f80)
1 /*
2  * Copyright (C) 2012-2013 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("$FreeBSD$");
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 #ifdef linux
58 #define NMA_LOCK_INIT(n)	sema_init(&(n)->nm_mtx, 1)
59 #define NMA_LOCK_DESTROY(n)
60 #define NMA_LOCK(n)		down(&(n)->nm_mtx)
61 #define NMA_UNLOCK(n)		up(&(n)->nm_mtx)
62 #else /* !linux */
63 #define NMA_LOCK_INIT(n)	mtx_init(&(n)->nm_mtx, "netmap memory allocator lock", NULL, MTX_DEF)
64 #define NMA_LOCK_DESTROY(n)	mtx_destroy(&(n)->nm_mtx)
65 #define NMA_LOCK(n)		mtx_lock(&(n)->nm_mtx)
66 #define NMA_UNLOCK(n)		mtx_unlock(&(n)->nm_mtx)
67 #endif /* linux */
68 
69 
70 struct netmap_obj_params netmap_params[NETMAP_POOLS_NR] = {
71 	[NETMAP_IF_POOL] = {
72 		.size = 1024,
73 		.num  = 100,
74 	},
75 	[NETMAP_RING_POOL] = {
76 		.size = 9*PAGE_SIZE,
77 		.num  = 200,
78 	},
79 	[NETMAP_BUF_POOL] = {
80 		.size = 2048,
81 		.num  = NETMAP_BUF_MAX_NUM,
82 	},
83 };
84 
85 
86 /*
87  * nm_mem is the memory allocator used for all physical interfaces
88  * running in netmap mode.
89  * Virtual (VALE) ports will have each its own allocator.
90  */
91 static int netmap_mem_global_config(struct netmap_mem_d *nmd);
92 static int netmap_mem_global_finalize(struct netmap_mem_d *nmd);
93 static void netmap_mem_global_deref(struct netmap_mem_d *nmd);
94 struct netmap_mem_d nm_mem = {	/* Our memory allocator. */
95 	.pools = {
96 		[NETMAP_IF_POOL] = {
97 			.name 	= "netmap_if",
98 			.objminsize = sizeof(struct netmap_if),
99 			.objmaxsize = 4096,
100 			.nummin     = 10,	/* don't be stingy */
101 			.nummax	    = 10000,	/* XXX very large */
102 		},
103 		[NETMAP_RING_POOL] = {
104 			.name 	= "netmap_ring",
105 			.objminsize = sizeof(struct netmap_ring),
106 			.objmaxsize = 32*PAGE_SIZE,
107 			.nummin     = 2,
108 			.nummax	    = 1024,
109 		},
110 		[NETMAP_BUF_POOL] = {
111 			.name	= "netmap_buf",
112 			.objminsize = 64,
113 			.objmaxsize = 65536,
114 			.nummin     = 4,
115 			.nummax	    = 1000000, /* one million! */
116 		},
117 	},
118 	.config   = netmap_mem_global_config,
119 	.finalize = netmap_mem_global_finalize,
120 	.deref    = netmap_mem_global_deref,
121 };
122 
123 
124 // XXX logically belongs to nm_mem
125 struct lut_entry *netmap_buffer_lut;	/* exported */
126 
127 /* blueprint for the private memory allocators */
128 static int netmap_mem_private_config(struct netmap_mem_d *nmd);
129 static int netmap_mem_private_finalize(struct netmap_mem_d *nmd);
130 static void netmap_mem_private_deref(struct netmap_mem_d *nmd);
131 const struct netmap_mem_d nm_blueprint = {
132 	.pools = {
133 		[NETMAP_IF_POOL] = {
134 			.name 	= "%s_if",
135 			.objminsize = sizeof(struct netmap_if),
136 			.objmaxsize = 4096,
137 			.nummin     = 1,
138 			.nummax	    = 10,
139 		},
140 		[NETMAP_RING_POOL] = {
141 			.name 	= "%s_ring",
142 			.objminsize = sizeof(struct netmap_ring),
143 			.objmaxsize = 32*PAGE_SIZE,
144 			.nummin     = 2,
145 			.nummax	    = 1024,
146 		},
147 		[NETMAP_BUF_POOL] = {
148 			.name	= "%s_buf",
149 			.objminsize = 64,
150 			.objmaxsize = 65536,
151 			.nummin     = 4,
152 			.nummax	    = 1000000, /* one million! */
153 		},
154 	},
155 	.config   = netmap_mem_private_config,
156 	.finalize = netmap_mem_private_finalize,
157 	.deref    = netmap_mem_private_deref,
158 
159 	.flags = NETMAP_MEM_PRIVATE,
160 };
161 
162 /* memory allocator related sysctls */
163 
164 #define STRINGIFY(x) #x
165 
166 
167 #define DECLARE_SYSCTLS(id, name) \
168 	SYSCTL_INT(_dev_netmap, OID_AUTO, name##_size, \
169 	    CTLFLAG_RW, &netmap_params[id].size, 0, "Requested size of netmap " STRINGIFY(name) "s"); \
170         SYSCTL_INT(_dev_netmap, OID_AUTO, name##_curr_size, \
171             CTLFLAG_RD, &nm_mem.pools[id]._objsize, 0, "Current size of netmap " STRINGIFY(name) "s"); \
172         SYSCTL_INT(_dev_netmap, OID_AUTO, name##_num, \
173             CTLFLAG_RW, &netmap_params[id].num, 0, "Requested number of netmap " STRINGIFY(name) "s"); \
174         SYSCTL_INT(_dev_netmap, OID_AUTO, name##_curr_num, \
175             CTLFLAG_RD, &nm_mem.pools[id].objtotal, 0, "Current number of netmap " STRINGIFY(name) "s")
176 
177 SYSCTL_DECL(_dev_netmap);
178 DECLARE_SYSCTLS(NETMAP_IF_POOL, if);
179 DECLARE_SYSCTLS(NETMAP_RING_POOL, ring);
180 DECLARE_SYSCTLS(NETMAP_BUF_POOL, buf);
181 
182 /*
183  * First, find the allocator that contains the requested offset,
184  * then locate the cluster through a lookup table.
185  */
186 vm_paddr_t
187 netmap_mem_ofstophys(struct netmap_mem_d* nmd, vm_ooffset_t offset)
188 {
189 	int i;
190 	vm_ooffset_t o = offset;
191 	vm_paddr_t pa;
192 	struct netmap_obj_pool *p;
193 
194 	NMA_LOCK(nmd);
195 	p = nmd->pools;
196 
197 	for (i = 0; i < NETMAP_POOLS_NR; offset -= p[i].memtotal, i++) {
198 		if (offset >= p[i].memtotal)
199 			continue;
200 		// now lookup the cluster's address
201 		pa = p[i].lut[offset / p[i]._objsize].paddr +
202 			offset % p[i]._objsize;
203 		NMA_UNLOCK(nmd);
204 		return pa;
205 	}
206 	/* this is only in case of errors */
207 	D("invalid ofs 0x%x out of 0x%x 0x%x 0x%x", (u_int)o,
208 		p[NETMAP_IF_POOL].memtotal,
209 		p[NETMAP_IF_POOL].memtotal
210 			+ p[NETMAP_RING_POOL].memtotal,
211 		p[NETMAP_IF_POOL].memtotal
212 			+ p[NETMAP_RING_POOL].memtotal
213 			+ p[NETMAP_BUF_POOL].memtotal);
214 	NMA_UNLOCK(nmd);
215 	return 0;	// XXX bad address
216 }
217 
218 int
219 netmap_mem_get_info(struct netmap_mem_d* nmd, u_int* size, u_int *memflags)
220 {
221 	int error = 0;
222 	NMA_LOCK(nmd);
223 	error = nmd->config(nmd);
224 	if (error)
225 		goto out;
226 	if (nmd->flags & NETMAP_MEM_FINALIZED) {
227 		*size = nmd->nm_totalsize;
228 	} else {
229 		int i;
230 		*size = 0;
231 		for (i = 0; i < NETMAP_POOLS_NR; i++) {
232 			struct netmap_obj_pool *p = nmd->pools + i;
233 			*size += (p->_numclusters * p->_clustsize);
234 		}
235 	}
236 	*memflags = nmd->flags;
237 out:
238 	NMA_UNLOCK(nmd);
239 	return error;
240 }
241 
242 /*
243  * we store objects by kernel address, need to find the offset
244  * within the pool to export the value to userspace.
245  * Algorithm: scan until we find the cluster, then add the
246  * actual offset in the cluster
247  */
248 static ssize_t
249 netmap_obj_offset(struct netmap_obj_pool *p, const void *vaddr)
250 {
251 	int i, k = p->_clustentries, n = p->objtotal;
252 	ssize_t ofs = 0;
253 
254 	for (i = 0; i < n; i += k, ofs += p->_clustsize) {
255 		const char *base = p->lut[i].vaddr;
256 		ssize_t relofs = (const char *) vaddr - base;
257 
258 		if (relofs < 0 || relofs >= p->_clustsize)
259 			continue;
260 
261 		ofs = ofs + relofs;
262 		ND("%s: return offset %d (cluster %d) for pointer %p",
263 		    p->name, ofs, i, vaddr);
264 		return ofs;
265 	}
266 	D("address %p is not contained inside any cluster (%s)",
267 	    vaddr, p->name);
268 	return 0; /* An error occurred */
269 }
270 
271 /* Helper functions which convert virtual addresses to offsets */
272 #define netmap_if_offset(n, v)					\
273 	netmap_obj_offset(&(n)->pools[NETMAP_IF_POOL], (v))
274 
275 #define netmap_ring_offset(n, v)				\
276     ((n)->pools[NETMAP_IF_POOL].memtotal + 			\
277 	netmap_obj_offset(&(n)->pools[NETMAP_RING_POOL], (v)))
278 
279 #define netmap_buf_offset(n, v)					\
280     ((n)->pools[NETMAP_IF_POOL].memtotal +			\
281 	(n)->pools[NETMAP_RING_POOL].memtotal +		\
282 	netmap_obj_offset(&(n)->pools[NETMAP_BUF_POOL], (v)))
283 
284 
285 ssize_t
286 netmap_mem_if_offset(struct netmap_mem_d *nmd, const void *addr)
287 {
288 	ssize_t v;
289 	NMA_LOCK(nmd);
290 	v = netmap_if_offset(nmd, addr);
291 	NMA_UNLOCK(nmd);
292 	return v;
293 }
294 
295 /*
296  * report the index, and use start position as a hint,
297  * otherwise buffer allocation becomes terribly expensive.
298  */
299 static void *
300 netmap_obj_malloc(struct netmap_obj_pool *p, u_int len, uint32_t *start, uint32_t *index)
301 {
302 	uint32_t i = 0;			/* index in the bitmap */
303 	uint32_t mask, j;		/* slot counter */
304 	void *vaddr = NULL;
305 
306 	if (len > p->_objsize) {
307 		D("%s request size %d too large", p->name, len);
308 		// XXX cannot reduce the size
309 		return NULL;
310 	}
311 
312 	if (p->objfree == 0) {
313 		D("%s allocator: run out of memory", p->name);
314 		return NULL;
315 	}
316 	if (start)
317 		i = *start;
318 
319 	/* termination is guaranteed by p->free, but better check bounds on i */
320 	while (vaddr == NULL && i < p->bitmap_slots)  {
321 		uint32_t cur = p->bitmap[i];
322 		if (cur == 0) { /* bitmask is fully used */
323 			i++;
324 			continue;
325 		}
326 		/* locate a slot */
327 		for (j = 0, mask = 1; (cur & mask) == 0; j++, mask <<= 1)
328 			;
329 
330 		p->bitmap[i] &= ~mask; /* mark object as in use */
331 		p->objfree--;
332 
333 		vaddr = p->lut[i * 32 + j].vaddr;
334 		if (index)
335 			*index = i * 32 + j;
336 	}
337 	ND("%s allocator: allocated object @ [%d][%d]: vaddr %p", i, j, vaddr);
338 
339 	if (start)
340 		*start = i;
341 	return vaddr;
342 }
343 
344 
345 /*
346  * free by index, not by address. This is slow, but is only used
347  * for a small number of objects (rings, nifp)
348  */
349 static void
350 netmap_obj_free(struct netmap_obj_pool *p, uint32_t j)
351 {
352 	if (j >= p->objtotal) {
353 		D("invalid index %u, max %u", j, p->objtotal);
354 		return;
355 	}
356 	p->bitmap[j / 32] |= (1 << (j % 32));
357 	p->objfree++;
358 	return;
359 }
360 
361 static void
362 netmap_obj_free_va(struct netmap_obj_pool *p, void *vaddr)
363 {
364 	u_int i, j, n = p->numclusters;
365 
366 	for (i = 0, j = 0; i < n; i++, j += p->_clustentries) {
367 		void *base = p->lut[i * p->_clustentries].vaddr;
368 		ssize_t relofs = (ssize_t) vaddr - (ssize_t) base;
369 
370 		/* Given address, is out of the scope of the current cluster.*/
371 		if (vaddr < base || relofs >= p->_clustsize)
372 			continue;
373 
374 		j = j + relofs / p->_objsize;
375 		/* KASSERT(j != 0, ("Cannot free object 0")); */
376 		netmap_obj_free(p, j);
377 		return;
378 	}
379 	D("address %p is not contained inside any cluster (%s)",
380 	    vaddr, p->name);
381 }
382 
383 #define netmap_if_malloc(n, len)	netmap_obj_malloc(&(n)->pools[NETMAP_IF_POOL], len, NULL, NULL)
384 #define netmap_if_free(n, v)		netmap_obj_free_va(&(n)->pools[NETMAP_IF_POOL], (v))
385 #define netmap_ring_malloc(n, len)	netmap_obj_malloc(&(n)->pools[NETMAP_RING_POOL], len, NULL, NULL)
386 #define netmap_ring_free(n, v)		netmap_obj_free_va(&(n)->pools[NETMAP_RING_POOL], (v))
387 #define netmap_buf_malloc(n, _pos, _index)			\
388 	netmap_obj_malloc(&(n)->pools[NETMAP_BUF_POOL], NETMAP_BDG_BUF_SIZE(n), _pos, _index)
389 
390 
391 /* Return the index associated to the given packet buffer */
392 #define netmap_buf_index(n, v)						\
393     (netmap_obj_offset(&(n)->pools[NETMAP_BUF_POOL], (v)) / NETMAP_BDG_BUF_SIZE(n))
394 
395 
396 /* Return nonzero on error */
397 static int
398 netmap_new_bufs(struct netmap_mem_d *nmd, struct netmap_if *nifp,
399                 struct netmap_slot *slot, u_int n)
400 {
401 	struct netmap_obj_pool *p = &nmd->pools[NETMAP_BUF_POOL];
402 	u_int i = 0;	/* slot counter */
403 	uint32_t pos = 0;	/* slot in p->bitmap */
404 	uint32_t index = 0;	/* buffer index */
405 
406 	(void)nifp;	/* UNUSED */
407 	for (i = 0; i < n; i++) {
408 		void *vaddr = netmap_buf_malloc(nmd, &pos, &index);
409 		if (vaddr == NULL) {
410 			D("unable to locate empty packet buffer");
411 			goto cleanup;
412 		}
413 		slot[i].buf_idx = index;
414 		slot[i].len = p->_objsize;
415 		/* XXX setting flags=NS_BUF_CHANGED forces a pointer reload
416 		 * in the NIC ring. This is a hack that hides missing
417 		 * initializations in the drivers, and should go away.
418 		 */
419 		// slot[i].flags = NS_BUF_CHANGED;
420 	}
421 
422 	ND("allocated %d buffers, %d available, first at %d", n, p->objfree, pos);
423 	return (0);
424 
425 cleanup:
426 	while (i > 0) {
427 		i--;
428 		netmap_obj_free(p, slot[i].buf_idx);
429 	}
430 	bzero(slot, n * sizeof(slot[0]));
431 	return (ENOMEM);
432 }
433 
434 
435 static void
436 netmap_free_buf(struct netmap_mem_d *nmd, struct netmap_if *nifp, uint32_t i)
437 {
438 	struct netmap_obj_pool *p = &nmd->pools[NETMAP_BUF_POOL];
439 
440 	(void)nifp;
441 	if (i < 2 || i >= p->objtotal) {
442 		D("Cannot free buf#%d: should be in [2, %d[", i, p->objtotal);
443 		return;
444 	}
445 	netmap_obj_free(p, i);
446 }
447 
448 static void
449 netmap_reset_obj_allocator(struct netmap_obj_pool *p)
450 {
451 
452 	if (p == NULL)
453 		return;
454 	if (p->bitmap)
455 		free(p->bitmap, M_NETMAP);
456 	p->bitmap = NULL;
457 	if (p->lut) {
458 		u_int i;
459 		size_t sz = p->_clustsize;
460 
461 		for (i = 0; i < p->objtotal; i += p->_clustentries) {
462 			if (p->lut[i].vaddr)
463 				contigfree(p->lut[i].vaddr, sz, M_NETMAP);
464 		}
465 		bzero(p->lut, sizeof(struct lut_entry) * p->objtotal);
466 #ifdef linux
467 		vfree(p->lut);
468 #else
469 		free(p->lut, M_NETMAP);
470 #endif
471 	}
472 	p->lut = NULL;
473 	p->objtotal = 0;
474 	p->memtotal = 0;
475 	p->numclusters = 0;
476 	p->objfree = 0;
477 }
478 
479 /*
480  * Free all resources related to an allocator.
481  */
482 static void
483 netmap_destroy_obj_allocator(struct netmap_obj_pool *p)
484 {
485 	if (p == NULL)
486 		return;
487 	netmap_reset_obj_allocator(p);
488 }
489 
490 /*
491  * We receive a request for objtotal objects, of size objsize each.
492  * Internally we may round up both numbers, as we allocate objects
493  * in small clusters multiple of the page size.
494  * We need to keep track of objtotal and clustentries,
495  * as they are needed when freeing memory.
496  *
497  * XXX note -- userspace needs the buffers to be contiguous,
498  *	so we cannot afford gaps at the end of a cluster.
499  */
500 
501 
502 /* call with NMA_LOCK held */
503 static int
504 netmap_config_obj_allocator(struct netmap_obj_pool *p, u_int objtotal, u_int objsize)
505 {
506 	int i;
507 	u_int clustsize;	/* the cluster size, multiple of page size */
508 	u_int clustentries;	/* how many objects per entry */
509 
510 	/* we store the current request, so we can
511 	 * detect configuration changes later */
512 	p->r_objtotal = objtotal;
513 	p->r_objsize = objsize;
514 
515 #define MAX_CLUSTSIZE	(1<<17)
516 #define LINE_ROUND	64
517 	if (objsize >= MAX_CLUSTSIZE) {
518 		/* we could do it but there is no point */
519 		D("unsupported allocation for %d bytes", objsize);
520 		return EINVAL;
521 	}
522 	/* make sure objsize is a multiple of LINE_ROUND */
523 	i = (objsize & (LINE_ROUND - 1));
524 	if (i) {
525 		D("XXX aligning object by %d bytes", LINE_ROUND - i);
526 		objsize += LINE_ROUND - i;
527 	}
528 	if (objsize < p->objminsize || objsize > p->objmaxsize) {
529 		D("requested objsize %d out of range [%d, %d]",
530 			objsize, p->objminsize, p->objmaxsize);
531 		return EINVAL;
532 	}
533 	if (objtotal < p->nummin || objtotal > p->nummax) {
534 		D("requested objtotal %d out of range [%d, %d]",
535 			objtotal, p->nummin, p->nummax);
536 		return EINVAL;
537 	}
538 	/*
539 	 * Compute number of objects using a brute-force approach:
540 	 * given a max cluster size,
541 	 * we try to fill it with objects keeping track of the
542 	 * wasted space to the next page boundary.
543 	 */
544 	for (clustentries = 0, i = 1;; i++) {
545 		u_int delta, used = i * objsize;
546 		if (used > MAX_CLUSTSIZE)
547 			break;
548 		delta = used % PAGE_SIZE;
549 		if (delta == 0) { // exact solution
550 			clustentries = i;
551 			break;
552 		}
553 		if (delta > ( (clustentries*objsize) % PAGE_SIZE) )
554 			clustentries = i;
555 	}
556 	// D("XXX --- ouch, delta %d (bad for buffers)", delta);
557 	/* compute clustsize and round to the next page */
558 	clustsize = clustentries * objsize;
559 	i =  (clustsize & (PAGE_SIZE - 1));
560 	if (i)
561 		clustsize += PAGE_SIZE - i;
562 	if (netmap_verbose)
563 		D("objsize %d clustsize %d objects %d",
564 			objsize, clustsize, clustentries);
565 
566 	/*
567 	 * The number of clusters is n = ceil(objtotal/clustentries)
568 	 * objtotal' = n * clustentries
569 	 */
570 	p->_clustentries = clustentries;
571 	p->_clustsize = clustsize;
572 	p->_numclusters = (objtotal + clustentries - 1) / clustentries;
573 
574 	/* actual values (may be larger than requested) */
575 	p->_objsize = objsize;
576 	p->_objtotal = p->_numclusters * clustentries;
577 
578 	return 0;
579 }
580 
581 
582 /* call with NMA_LOCK held */
583 static int
584 netmap_finalize_obj_allocator(struct netmap_obj_pool *p)
585 {
586 	int i; /* must be signed */
587 	size_t n;
588 
589 	/* optimistically assume we have enough memory */
590 	p->numclusters = p->_numclusters;
591 	p->objtotal = p->_objtotal;
592 
593 	n = sizeof(struct lut_entry) * p->objtotal;
594 #ifdef linux
595 	p->lut = vmalloc(n);
596 #else
597 	p->lut = malloc(n, M_NETMAP, M_NOWAIT | M_ZERO);
598 #endif
599 	if (p->lut == NULL) {
600 		D("Unable to create lookup table (%d bytes) for '%s'", (int)n, p->name);
601 		goto clean;
602 	}
603 
604 	/* Allocate the bitmap */
605 	n = (p->objtotal + 31) / 32;
606 	p->bitmap = malloc(sizeof(uint32_t) * n, M_NETMAP, M_NOWAIT | M_ZERO);
607 	if (p->bitmap == NULL) {
608 		D("Unable to create bitmap (%d entries) for allocator '%s'", (int)n,
609 		    p->name);
610 		goto clean;
611 	}
612 	p->bitmap_slots = n;
613 
614 	/*
615 	 * Allocate clusters, init pointers and bitmap
616 	 */
617 
618 	n = p->_clustsize;
619 	for (i = 0; i < (int)p->objtotal;) {
620 		int lim = i + p->_clustentries;
621 		char *clust;
622 
623 		clust = contigmalloc(n, M_NETMAP, M_NOWAIT | M_ZERO,
624 		    (size_t)0, -1UL, PAGE_SIZE, 0);
625 		if (clust == NULL) {
626 			/*
627 			 * If we get here, there is a severe memory shortage,
628 			 * so halve the allocated memory to reclaim some.
629 			 */
630 			D("Unable to create cluster at %d for '%s' allocator",
631 			    i, p->name);
632 			if (i < 2) /* nothing to halve */
633 				goto out;
634 			lim = i / 2;
635 			for (i--; i >= lim; i--) {
636 				p->bitmap[ (i>>5) ] &=  ~( 1 << (i & 31) );
637 				if (i % p->_clustentries == 0 && p->lut[i].vaddr)
638 					contigfree(p->lut[i].vaddr,
639 						n, M_NETMAP);
640 			}
641 		out:
642 			p->objtotal = i;
643 			/* we may have stopped in the middle of a cluster */
644 			p->numclusters = (i + p->_clustentries - 1) / p->_clustentries;
645 			break;
646 		}
647 		for (; i < lim; i++, clust += p->_objsize) {
648 			p->bitmap[ (i>>5) ] |=  ( 1 << (i & 31) );
649 			p->lut[i].vaddr = clust;
650 			p->lut[i].paddr = vtophys(clust);
651 		}
652 	}
653 	p->objfree = p->objtotal;
654 	p->memtotal = p->numclusters * p->_clustsize;
655 	if (p->objfree == 0)
656 		goto clean;
657 	if (netmap_verbose)
658 		D("Pre-allocated %d clusters (%d/%dKB) for '%s'",
659 		    p->numclusters, p->_clustsize >> 10,
660 		    p->memtotal >> 10, p->name);
661 
662 	return 0;
663 
664 clean:
665 	netmap_reset_obj_allocator(p);
666 	return ENOMEM;
667 }
668 
669 /* call with lock held */
670 static int
671 netmap_memory_config_changed(struct netmap_mem_d *nmd)
672 {
673 	int i;
674 
675 	for (i = 0; i < NETMAP_POOLS_NR; i++) {
676 		if (nmd->pools[i].r_objsize != netmap_params[i].size ||
677 		    nmd->pools[i].r_objtotal != netmap_params[i].num)
678 		    return 1;
679 	}
680 	return 0;
681 }
682 
683 static void
684 netmap_mem_reset_all(struct netmap_mem_d *nmd)
685 {
686 	int i;
687 	D("resetting %p", nmd);
688 	for (i = 0; i < NETMAP_POOLS_NR; i++) {
689 		netmap_reset_obj_allocator(&nmd->pools[i]);
690 	}
691 	nmd->flags  &= ~NETMAP_MEM_FINALIZED;
692 }
693 
694 static int
695 netmap_mem_finalize_all(struct netmap_mem_d *nmd)
696 {
697 	int i;
698 	if (nmd->flags & NETMAP_MEM_FINALIZED)
699 		return 0;
700 	nmd->lasterr = 0;
701 	nmd->nm_totalsize = 0;
702 	for (i = 0; i < NETMAP_POOLS_NR; i++) {
703 		nmd->lasterr = netmap_finalize_obj_allocator(&nmd->pools[i]);
704 		if (nmd->lasterr)
705 			goto error;
706 		nmd->nm_totalsize += nmd->pools[i].memtotal;
707 	}
708 	/* buffers 0 and 1 are reserved */
709 	nmd->pools[NETMAP_BUF_POOL].objfree -= 2;
710 	nmd->pools[NETMAP_BUF_POOL].bitmap[0] = ~3;
711 	nmd->flags |= NETMAP_MEM_FINALIZED;
712 
713 	D("Have %d KB for interfaces, %d KB for rings and %d MB for buffers",
714 	    nmd->pools[NETMAP_IF_POOL].memtotal >> 10,
715 	    nmd->pools[NETMAP_RING_POOL].memtotal >> 10,
716 	    nmd->pools[NETMAP_BUF_POOL].memtotal >> 20);
717 
718 	D("Free buffers: %d", nmd->pools[NETMAP_BUF_POOL].objfree);
719 
720 
721 	return 0;
722 error:
723 	netmap_mem_reset_all(nmd);
724 	return nmd->lasterr;
725 }
726 
727 
728 
729 void
730 netmap_mem_private_delete(struct netmap_mem_d *nmd)
731 {
732 	if (nmd == NULL)
733 		return;
734 	D("deleting %p", nmd);
735 	if (nmd->refcount > 0)
736 		D("bug: deleting mem allocator with refcount=%d!", nmd->refcount);
737 	D("done deleting %p", nmd);
738 	NMA_LOCK_DESTROY(nmd);
739 	free(nmd, M_DEVBUF);
740 }
741 
742 static int
743 netmap_mem_private_config(struct netmap_mem_d *nmd)
744 {
745 	/* nothing to do, we are configured on creation
746  	 * and configuration never changes thereafter
747  	 */
748 	return 0;
749 }
750 
751 static int
752 netmap_mem_private_finalize(struct netmap_mem_d *nmd)
753 {
754 	int err;
755 	NMA_LOCK(nmd);
756 	nmd->refcount++;
757 	err = netmap_mem_finalize_all(nmd);
758 	NMA_UNLOCK(nmd);
759 	return err;
760 
761 }
762 
763 static void netmap_mem_private_deref(struct netmap_mem_d *nmd)
764 {
765 	NMA_LOCK(nmd);
766 	if (--nmd->refcount <= 0)
767 		netmap_mem_reset_all(nmd);
768 	NMA_UNLOCK(nmd);
769 }
770 
771 struct netmap_mem_d *
772 netmap_mem_private_new(const char *name, u_int txr, u_int txd, u_int rxr, u_int rxd)
773 {
774 	struct netmap_mem_d *d = NULL;
775 	struct netmap_obj_params p[NETMAP_POOLS_NR];
776 	int i;
777 	u_int maxd;
778 
779 	d = malloc(sizeof(struct netmap_mem_d),
780 			M_DEVBUF, M_NOWAIT | M_ZERO);
781 	if (d == NULL)
782 		return NULL;
783 
784 	*d = nm_blueprint;
785 
786 	/* XXX the rest of the code assumes the stack rings are alwasy present */
787 	txr++;
788 	rxr++;
789 	p[NETMAP_IF_POOL].size = sizeof(struct netmap_if) +
790 		sizeof(ssize_t) * (txr + rxr);
791 	p[NETMAP_IF_POOL].num = 2;
792 	maxd = (txd > rxd) ? txd : rxd;
793 	p[NETMAP_RING_POOL].size = sizeof(struct netmap_ring) +
794 		sizeof(struct netmap_slot) * maxd;
795 	p[NETMAP_RING_POOL].num = txr + rxr;
796 	p[NETMAP_BUF_POOL].size = 2048; /* XXX find a way to let the user choose this */
797 	p[NETMAP_BUF_POOL].num = rxr * (rxd + 2) + txr * (txd + 2);
798 
799 	D("req if %d*%d ring %d*%d buf %d*%d",
800 			p[NETMAP_IF_POOL].num,
801 			p[NETMAP_IF_POOL].size,
802 			p[NETMAP_RING_POOL].num,
803 			p[NETMAP_RING_POOL].size,
804 			p[NETMAP_BUF_POOL].num,
805 			p[NETMAP_BUF_POOL].size);
806 
807 	for (i = 0; i < NETMAP_POOLS_NR; i++) {
808 		snprintf(d->pools[i].name, NETMAP_POOL_MAX_NAMSZ,
809 				nm_blueprint.pools[i].name,
810 				name);
811 		if (netmap_config_obj_allocator(&d->pools[i],
812 				p[i].num, p[i].size))
813 			goto error;
814 	}
815 
816 	d->flags &= ~NETMAP_MEM_FINALIZED;
817 
818 	NMA_LOCK_INIT(d);
819 
820 	return d;
821 error:
822 	netmap_mem_private_delete(d);
823 	return NULL;
824 }
825 
826 
827 /* call with lock held */
828 static int
829 netmap_mem_global_config(struct netmap_mem_d *nmd)
830 {
831 	int i;
832 
833 	if (nmd->refcount)
834 		/* already in use, we cannot change the configuration */
835 		goto out;
836 
837 	if (!netmap_memory_config_changed(nmd))
838 		goto out;
839 
840 	D("reconfiguring");
841 
842 	if (nmd->flags & NETMAP_MEM_FINALIZED) {
843 		/* reset previous allocation */
844 		for (i = 0; i < NETMAP_POOLS_NR; i++) {
845 			netmap_reset_obj_allocator(&nmd->pools[i]);
846 		}
847 		nmd->flags &= ~NETMAP_MEM_FINALIZED;
848         }
849 
850 	for (i = 0; i < NETMAP_POOLS_NR; i++) {
851 		nmd->lasterr = netmap_config_obj_allocator(&nmd->pools[i],
852 				netmap_params[i].num, netmap_params[i].size);
853 		if (nmd->lasterr)
854 			goto out;
855 	}
856 
857 out:
858 
859 	return nmd->lasterr;
860 }
861 
862 static int
863 netmap_mem_global_finalize(struct netmap_mem_d *nmd)
864 {
865 	int err;
866 
867 	NMA_LOCK(nmd);
868 
869 
870 	/* update configuration if changed */
871 	if (netmap_mem_global_config(nmd))
872 		goto out;
873 
874 	nmd->refcount++;
875 
876 	if (nmd->flags & NETMAP_MEM_FINALIZED) {
877 		/* may happen if config is not changed */
878 		ND("nothing to do");
879 		goto out;
880 	}
881 
882 	if (netmap_mem_finalize_all(nmd))
883 		goto out;
884 
885 	/* backward compatibility */
886 	netmap_buf_size = nmd->pools[NETMAP_BUF_POOL]._objsize;
887 	netmap_total_buffers = nmd->pools[NETMAP_BUF_POOL].objtotal;
888 
889 	netmap_buffer_lut = nmd->pools[NETMAP_BUF_POOL].lut;
890 	netmap_buffer_base = nmd->pools[NETMAP_BUF_POOL].lut[0].vaddr;
891 
892 	nmd->lasterr = 0;
893 
894 out:
895 	if (nmd->lasterr)
896 		nmd->refcount--;
897 	err = nmd->lasterr;
898 
899 	NMA_UNLOCK(nmd);
900 
901 	return err;
902 
903 }
904 
905 int
906 netmap_mem_init(void)
907 {
908 	NMA_LOCK_INIT(&nm_mem);
909 	return (0);
910 }
911 
912 void
913 netmap_mem_fini(void)
914 {
915 	int i;
916 
917 	for (i = 0; i < NETMAP_POOLS_NR; i++) {
918 	    netmap_destroy_obj_allocator(&nm_mem.pools[i]);
919 	}
920 	NMA_LOCK_DESTROY(&nm_mem);
921 }
922 
923 static void
924 netmap_free_rings(struct netmap_adapter *na)
925 {
926 	u_int i;
927 	if (!na->tx_rings)
928 		return;
929 	for (i = 0; i < na->num_tx_rings + 1; i++) {
930 		if (na->tx_rings[i].ring) {
931 			netmap_ring_free(na->nm_mem, na->tx_rings[i].ring);
932 			na->tx_rings[i].ring = NULL;
933 		}
934 	}
935 	for (i = 0; i < na->num_rx_rings + 1; i++) {
936 		if (na->rx_rings[i].ring) {
937 			netmap_ring_free(na->nm_mem, na->rx_rings[i].ring);
938 			na->rx_rings[i].ring = NULL;
939 		}
940 	}
941 	free(na->tx_rings, M_DEVBUF);
942 	na->tx_rings = na->rx_rings = NULL;
943 }
944 
945 
946 
947 /* call with NMA_LOCK held */
948 /*
949  * Allocate the per-fd structure netmap_if.
950  * If this is the first instance, also allocate the krings, rings etc.
951  *
952  * We assume that the configuration stored in na
953  * (number of tx/rx rings and descs) does not change while
954  * the interface is in netmap mode.
955  */
956 extern int nma_is_vp(struct netmap_adapter *na);
957 struct netmap_if *
958 netmap_mem_if_new(const char *ifname, struct netmap_adapter *na)
959 {
960 	struct netmap_if *nifp;
961 	struct netmap_ring *ring;
962 	ssize_t base; /* handy for relative offsets between rings and nifp */
963 	u_int i, len, ndesc, ntx, nrx;
964 	struct netmap_kring *kring;
965 	uint32_t *tx_leases = NULL, *rx_leases = NULL;
966 
967 	/*
968 	 * verify whether virtual port need the stack ring
969 	 */
970 	ntx = na->num_tx_rings + 1; /* shorthand, include stack ring */
971 	nrx = na->num_rx_rings + 1; /* shorthand, include stack ring */
972 	/*
973 	 * the descriptor is followed inline by an array of offsets
974 	 * to the tx and rx rings in the shared memory region.
975 	 * For virtual rx rings we also allocate an array of
976 	 * pointers to assign to nkr_leases.
977 	 */
978 
979 	NMA_LOCK(na->nm_mem);
980 
981 	len = sizeof(struct netmap_if) + (nrx + ntx) * sizeof(ssize_t);
982 	nifp = netmap_if_malloc(na->nm_mem, len);
983 	if (nifp == NULL) {
984 		NMA_UNLOCK(na->nm_mem);
985 		return NULL;
986 	}
987 
988 	/* initialize base fields -- override const */
989 	*(u_int *)(uintptr_t)&nifp->ni_tx_rings = na->num_tx_rings;
990 	*(u_int *)(uintptr_t)&nifp->ni_rx_rings = na->num_rx_rings;
991 	strncpy(nifp->ni_name, ifname, (size_t)IFNAMSIZ);
992 
993 	if (na->refcount) { /* already setup, we are done */
994 		goto final;
995 	}
996 
997 	len = (ntx + nrx) * sizeof(struct netmap_kring);
998 	/*
999 	 * Leases are attached to TX rings on NIC/host ports,
1000 	 * and to RX rings on VALE ports.
1001 	 */
1002 	if (nma_is_vp(na)) {
1003 		len += sizeof(uint32_t) * na->num_rx_desc * na->num_rx_rings;
1004 	} else {
1005 		len += sizeof(uint32_t) * na->num_tx_desc * ntx;
1006 	}
1007 
1008 	na->tx_rings = malloc((size_t)len, M_DEVBUF, M_NOWAIT | M_ZERO);
1009 	if (na->tx_rings == NULL) {
1010 		D("Cannot allocate krings for %s", ifname);
1011 		goto cleanup;
1012 	}
1013 	na->rx_rings = na->tx_rings + ntx;
1014 
1015 	if (nma_is_vp(na)) {
1016 		rx_leases = (uint32_t *)(na->rx_rings + nrx);
1017 	} else {
1018 		tx_leases = (uint32_t *)(na->rx_rings + nrx);
1019 	}
1020 
1021 	/*
1022 	 * First instance, allocate netmap rings and buffers for this card
1023 	 * The rings are contiguous, but have variable size.
1024 	 */
1025 	for (i = 0; i < ntx; i++) { /* Transmit rings */
1026 		kring = &na->tx_rings[i];
1027 		ndesc = na->num_tx_desc;
1028 		bzero(kring, sizeof(*kring));
1029 		len = sizeof(struct netmap_ring) +
1030 			  ndesc * sizeof(struct netmap_slot);
1031 		ring = netmap_ring_malloc(na->nm_mem, len);
1032 		if (ring == NULL) {
1033 			D("Cannot allocate tx_ring[%d] for %s", i, ifname);
1034 			goto cleanup;
1035 		}
1036 		ND("txring[%d] at %p ofs %d", i, ring);
1037 		kring->na = na;
1038 		kring->ring = ring;
1039 		if (tx_leases) {
1040 			kring->nkr_leases = tx_leases;
1041 			tx_leases += ndesc;
1042 		}
1043 		*(uint32_t *)(uintptr_t)&ring->num_slots = kring->nkr_num_slots = ndesc;
1044 		*(ssize_t *)(uintptr_t)&ring->buf_ofs =
1045 		    (na->nm_mem->pools[NETMAP_IF_POOL].memtotal +
1046 			na->nm_mem->pools[NETMAP_RING_POOL].memtotal) -
1047 			netmap_ring_offset(na->nm_mem, ring);
1048 
1049 		/*
1050 		 * IMPORTANT:
1051 		 * Always keep one slot empty, so we can detect new
1052 		 * transmissions comparing cur and nr_hwcur (they are
1053 		 * the same only if there are no new transmissions).
1054 		 */
1055 		ring->avail = kring->nr_hwavail = ndesc - 1;
1056 		ring->cur = kring->nr_hwcur = 0;
1057 		*(uint16_t *)(uintptr_t)&ring->nr_buf_size =
1058 			NETMAP_BDG_BUF_SIZE(na->nm_mem);
1059 		ND("initializing slots for txring[%d]", i);
1060 		if (netmap_new_bufs(na->nm_mem, nifp, ring->slot, ndesc)) {
1061 			D("Cannot allocate buffers for tx_ring[%d] for %s", i, ifname);
1062 			goto cleanup;
1063 		}
1064 	}
1065 
1066 	for (i = 0; i < nrx; i++) { /* Receive rings */
1067 		kring = &na->rx_rings[i];
1068 		ndesc = na->num_rx_desc;
1069 		bzero(kring, sizeof(*kring));
1070 		len = sizeof(struct netmap_ring) +
1071 			  ndesc * sizeof(struct netmap_slot);
1072 		ring = netmap_ring_malloc(na->nm_mem, len);
1073 		if (ring == NULL) {
1074 			D("Cannot allocate rx_ring[%d] for %s", i, ifname);
1075 			goto cleanup;
1076 		}
1077 		ND("rxring[%d] at %p ofs %d", i, ring);
1078 
1079 		kring->na = na;
1080 		kring->ring = ring;
1081 		if (rx_leases && i < na->num_rx_rings) {
1082 			kring->nkr_leases = rx_leases;
1083 			rx_leases += ndesc;
1084 		}
1085 		*(uint32_t *)(uintptr_t)&ring->num_slots = kring->nkr_num_slots = ndesc;
1086 		*(ssize_t *)(uintptr_t)&ring->buf_ofs =
1087 		    (na->nm_mem->pools[NETMAP_IF_POOL].memtotal +
1088 		        na->nm_mem->pools[NETMAP_RING_POOL].memtotal) -
1089 			netmap_ring_offset(na->nm_mem, ring);
1090 
1091 		ring->cur = kring->nr_hwcur = 0;
1092 		ring->avail = kring->nr_hwavail = 0; /* empty */
1093 		*(int *)(uintptr_t)&ring->nr_buf_size =
1094 			NETMAP_BDG_BUF_SIZE(na->nm_mem);
1095 		ND("initializing slots for rxring[%d]", i);
1096 		if (netmap_new_bufs(na->nm_mem, nifp, ring->slot, ndesc)) {
1097 			D("Cannot allocate buffers for rx_ring[%d] for %s", i, ifname);
1098 			goto cleanup;
1099 		}
1100 	}
1101 #ifdef linux
1102 	// XXX initialize the selrecord structs.
1103 	for (i = 0; i < ntx; i++)
1104 		init_waitqueue_head(&na->tx_rings[i].si);
1105 	for (i = 0; i < nrx; i++)
1106 		init_waitqueue_head(&na->rx_rings[i].si);
1107 	init_waitqueue_head(&na->tx_si);
1108 	init_waitqueue_head(&na->rx_si);
1109 #endif
1110 final:
1111 	/*
1112 	 * fill the slots for the rx and tx rings. They contain the offset
1113 	 * between the ring and nifp, so the information is usable in
1114 	 * userspace to reach the ring from the nifp.
1115 	 */
1116 	base = netmap_if_offset(na->nm_mem, nifp);
1117 	for (i = 0; i < ntx; i++) {
1118 		*(ssize_t *)(uintptr_t)&nifp->ring_ofs[i] =
1119 			netmap_ring_offset(na->nm_mem, na->tx_rings[i].ring) - base;
1120 	}
1121 	for (i = 0; i < nrx; i++) {
1122 		*(ssize_t *)(uintptr_t)&nifp->ring_ofs[i+ntx] =
1123 			netmap_ring_offset(na->nm_mem, na->rx_rings[i].ring) - base;
1124 	}
1125 
1126 	NMA_UNLOCK(na->nm_mem);
1127 
1128 	return (nifp);
1129 cleanup:
1130 	netmap_free_rings(na);
1131 	netmap_if_free(na->nm_mem, nifp);
1132 
1133 	NMA_UNLOCK(na->nm_mem);
1134 
1135 	return NULL;
1136 }
1137 
1138 void
1139 netmap_mem_if_delete(struct netmap_adapter *na, struct netmap_if *nifp)
1140 {
1141 	if (nifp == NULL)
1142 		/* nothing to do */
1143 		return;
1144 	NMA_LOCK(na->nm_mem);
1145 
1146 	if (na->refcount <= 0) {
1147 		/* last instance, release bufs and rings */
1148 		u_int i, j, lim;
1149 		struct netmap_ring *ring;
1150 
1151 		for (i = 0; i < na->num_tx_rings + 1; i++) {
1152 			ring = na->tx_rings[i].ring;
1153 			lim = na->tx_rings[i].nkr_num_slots;
1154 			for (j = 0; j < lim; j++)
1155 				netmap_free_buf(na->nm_mem, nifp, ring->slot[j].buf_idx);
1156 		}
1157 		for (i = 0; i < na->num_rx_rings + 1; i++) {
1158 			ring = na->rx_rings[i].ring;
1159 			lim = na->rx_rings[i].nkr_num_slots;
1160 			for (j = 0; j < lim; j++)
1161 				netmap_free_buf(na->nm_mem, nifp, ring->slot[j].buf_idx);
1162 		}
1163 		netmap_free_rings(na);
1164 	}
1165 	netmap_if_free(na->nm_mem, nifp);
1166 
1167 	NMA_UNLOCK(na->nm_mem);
1168 }
1169 
1170 static void
1171 netmap_mem_global_deref(struct netmap_mem_d *nmd)
1172 {
1173 	NMA_LOCK(nmd);
1174 
1175 	nmd->refcount--;
1176 	if (netmap_verbose)
1177 		D("refcount = %d", nmd->refcount);
1178 
1179 	NMA_UNLOCK(nmd);
1180 }
1181 
1182 int netmap_mem_finalize(struct netmap_mem_d *nmd)
1183 {
1184 	return nmd->finalize(nmd);
1185 }
1186 
1187 void netmap_mem_deref(struct netmap_mem_d *nmd)
1188 {
1189 	return nmd->deref(nmd);
1190 }
1191