xref: /freebsd/sys/dev/netmap/netmap_mem2.c (revision 535af610)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
3  *
4  * Copyright (C) 2012-2014 Matteo Landi
5  * Copyright (C) 2012-2016 Luigi Rizzo
6  * Copyright (C) 2012-2016 Giuseppe Lettieri
7  * All rights reserved.
8  *
9  * Redistribution and use in source and binary forms, with or without
10  * modification, are permitted provided that the following conditions
11  * are met:
12  *   1. Redistributions of source code must retain the above copyright
13  *      notice, this list of conditions and the following disclaimer.
14  *   2. Redistributions in binary form must reproduce the above copyright
15  *      notice, this list of conditions and the following disclaimer in the
16  *      documentation and/or other materials provided with the distribution.
17  *
18  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
19  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
20  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
21  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
22  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
23  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
24  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
25  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
26  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
27  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
28  * SUCH DAMAGE.
29  */
30 
31 #ifdef linux
32 #include "bsd_glue.h"
33 #endif /* linux */
34 
35 #ifdef __APPLE__
36 #include "osx_glue.h"
37 #endif /* __APPLE__ */
38 
39 #ifdef __FreeBSD__
40 #include <sys/cdefs.h> /* prerequisite */
41 __FBSDID("$FreeBSD$");
42 
43 #include <sys/types.h>
44 #include <sys/malloc.h>
45 #include <sys/kernel.h>		/* MALLOC_DEFINE */
46 #include <sys/proc.h>
47 #include <vm/vm.h>	/* vtophys */
48 #include <vm/pmap.h>	/* vtophys */
49 #include <sys/socket.h> /* sockaddrs */
50 #include <sys/selinfo.h>
51 #include <sys/sysctl.h>
52 #include <net/if.h>
53 #include <net/if_var.h>
54 #include <net/vnet.h>
55 #include <machine/bus.h>	/* bus_dmamap_* */
56 
57 /* M_NETMAP only used in here */
58 MALLOC_DECLARE(M_NETMAP);
59 MALLOC_DEFINE(M_NETMAP, "netmap", "Network memory map");
60 
61 #endif /* __FreeBSD__ */
62 
63 #ifdef _WIN32
64 #include <win_glue.h>
65 #endif
66 
67 #include <net/netmap.h>
68 #include <dev/netmap/netmap_kern.h>
69 #include <net/netmap_virt.h>
70 #include "netmap_mem2.h"
71 
72 #ifdef _WIN32_USE_SMALL_GENERIC_DEVICES_MEMORY
73 #define NETMAP_BUF_MAX_NUM  8*4096      /* if too big takes too much time to allocate */
74 #else
75 #define NETMAP_BUF_MAX_NUM 20*4096*2	/* large machine */
76 #endif
77 
78 #define NETMAP_POOL_MAX_NAMSZ	32
79 
80 
81 enum {
82 	NETMAP_IF_POOL   = 0,
83 	NETMAP_RING_POOL,
84 	NETMAP_BUF_POOL,
85 	NETMAP_POOLS_NR
86 };
87 
88 
89 struct netmap_obj_params {
90 	u_int size;
91 	u_int num;
92 
93 	u_int last_size;
94 	u_int last_num;
95 };
96 
97 struct netmap_obj_pool {
98 	char name[NETMAP_POOL_MAX_NAMSZ];	/* name of the allocator */
99 
100 	/* ---------------------------------------------------*/
101 	/* these are only meaningful if the pool is finalized */
102 	/* (see 'finalized' field in netmap_mem_d)            */
103 	size_t memtotal;	/* actual total memory space */
104 
105 	struct lut_entry *lut;  /* virt,phys addresses, objtotal entries */
106 	uint32_t *bitmap;       /* one bit per buffer, 1 means free */
107 	uint32_t *invalid_bitmap;/* one bit per buffer, 1 means invalid */
108 	uint32_t bitmap_slots;	/* number of uint32 entries in bitmap */
109 
110 	u_int objtotal;         /* actual total number of objects. */
111 	u_int numclusters;	/* actual number of clusters */
112 	u_int objfree;          /* number of free objects. */
113 
114 	int	alloc_done;	/* we have allocated the memory */
115 	/* ---------------------------------------------------*/
116 
117 	/* limits */
118 	u_int objminsize;	/* minimum object size */
119 	u_int objmaxsize;	/* maximum object size */
120 	u_int nummin;		/* minimum number of objects */
121 	u_int nummax;		/* maximum number of objects */
122 
123 	/* these are changed only by config */
124 	u_int _objtotal;	/* total number of objects */
125 	u_int _objsize;		/* object size */
126 	u_int _clustsize;       /* cluster size */
127 	u_int _clustentries;    /* objects per cluster */
128 	u_int _numclusters;	/* number of clusters */
129 
130 	/* requested values */
131 	u_int r_objtotal;
132 	u_int r_objsize;
133 };
134 
135 #define NMA_LOCK_T		NM_MTX_T
136 #define NMA_LOCK_INIT(n)	NM_MTX_INIT((n)->nm_mtx)
137 #define NMA_LOCK_DESTROY(n)	NM_MTX_DESTROY((n)->nm_mtx)
138 #define NMA_LOCK(n)		NM_MTX_LOCK((n)->nm_mtx)
139 #define NMA_SPINLOCK(n)         NM_MTX_SPINLOCK((n)->nm_mtx)
140 #define NMA_UNLOCK(n)		NM_MTX_UNLOCK((n)->nm_mtx)
141 
142 struct netmap_mem_ops {
143 	int (*nmd_get_lut)(struct netmap_mem_d *, struct netmap_lut*);
144 	int  (*nmd_get_info)(struct netmap_mem_d *, uint64_t *size,
145 			u_int *memflags, uint16_t *id);
146 
147 	vm_paddr_t (*nmd_ofstophys)(struct netmap_mem_d *, vm_ooffset_t);
148 	int (*nmd_config)(struct netmap_mem_d *);
149 	int (*nmd_finalize)(struct netmap_mem_d *, struct netmap_adapter *);
150 	void (*nmd_deref)(struct netmap_mem_d *, struct netmap_adapter *);
151 	ssize_t  (*nmd_if_offset)(struct netmap_mem_d *, const void *vaddr);
152 	void (*nmd_delete)(struct netmap_mem_d *);
153 
154 	struct netmap_if * (*nmd_if_new)(struct netmap_mem_d *,
155 			struct netmap_adapter *, struct netmap_priv_d *);
156 	void (*nmd_if_delete)(struct netmap_mem_d *,
157 			struct netmap_adapter *, struct netmap_if *);
158 	int  (*nmd_rings_create)(struct netmap_mem_d *,
159 			struct netmap_adapter *);
160 	void (*nmd_rings_delete)(struct netmap_mem_d *,
161 			struct netmap_adapter *);
162 };
163 
164 struct netmap_mem_d {
165 	NMA_LOCK_T nm_mtx;  /* protect the allocator */
166 	size_t nm_totalsize; /* shorthand */
167 
168 	u_int flags;
169 #define NETMAP_MEM_FINALIZED	0x1	/* preallocation done */
170 #define NETMAP_MEM_HIDDEN	0x8	/* being prepared */
171 #define NETMAP_MEM_NOMAP	0x10	/* do not map/unmap pdevs */
172 	int lasterr;		/* last error for curr config */
173 	int active;		/* active users */
174 	int refcount;
175 	/* the three allocators */
176 	struct netmap_obj_pool pools[NETMAP_POOLS_NR];
177 
178 	nm_memid_t nm_id;	/* allocator identifier */
179 	int nm_grp;	/* iommu group id */
180 
181 	/* list of all existing allocators, sorted by nm_id */
182 	struct netmap_mem_d *prev, *next;
183 
184 	struct netmap_mem_ops *ops;
185 
186 	struct netmap_obj_params params[NETMAP_POOLS_NR];
187 
188 #define NM_MEM_NAMESZ	16
189 	char name[NM_MEM_NAMESZ];
190 };
191 
192 int
193 netmap_mem_get_lut(struct netmap_mem_d *nmd, struct netmap_lut *lut)
194 {
195 	int rv;
196 
197 	NMA_LOCK(nmd);
198 	rv = nmd->ops->nmd_get_lut(nmd, lut);
199 	NMA_UNLOCK(nmd);
200 
201 	return rv;
202 }
203 
204 int
205 netmap_mem_get_info(struct netmap_mem_d *nmd, uint64_t *size,
206 		u_int *memflags, nm_memid_t *memid)
207 {
208 	int rv;
209 
210 	NMA_LOCK(nmd);
211 	rv = nmd->ops->nmd_get_info(nmd, size, memflags, memid);
212 	NMA_UNLOCK(nmd);
213 
214 	return rv;
215 }
216 
217 vm_paddr_t
218 netmap_mem_ofstophys(struct netmap_mem_d *nmd, vm_ooffset_t off)
219 {
220 	vm_paddr_t pa;
221 
222 #if defined(__FreeBSD__)
223 	/* This function is called by netmap_dev_pager_fault(), which holds a
224 	 * non-sleepable lock since FreeBSD 12. Since we cannot sleep, we
225 	 * spin on the trylock. */
226 	NMA_SPINLOCK(nmd);
227 #else
228 	NMA_LOCK(nmd);
229 #endif
230 	pa = nmd->ops->nmd_ofstophys(nmd, off);
231 	NMA_UNLOCK(nmd);
232 
233 	return pa;
234 }
235 
236 static int
237 netmap_mem_config(struct netmap_mem_d *nmd)
238 {
239 	if (nmd->active) {
240 		/* already in use. Not fatal, but we
241 		 * cannot change the configuration
242 		 */
243 		return 0;
244 	}
245 
246 	return nmd->ops->nmd_config(nmd);
247 }
248 
249 ssize_t
250 netmap_mem_if_offset(struct netmap_mem_d *nmd, const void *off)
251 {
252 	ssize_t rv;
253 
254 	NMA_LOCK(nmd);
255 	rv = nmd->ops->nmd_if_offset(nmd, off);
256 	NMA_UNLOCK(nmd);
257 
258 	return rv;
259 }
260 
261 static void
262 netmap_mem_delete(struct netmap_mem_d *nmd)
263 {
264 	nmd->ops->nmd_delete(nmd);
265 }
266 
267 struct netmap_if *
268 netmap_mem_if_new(struct netmap_adapter *na, struct netmap_priv_d *priv)
269 {
270 	struct netmap_if *nifp;
271 	struct netmap_mem_d *nmd = na->nm_mem;
272 
273 	NMA_LOCK(nmd);
274 	nifp = nmd->ops->nmd_if_new(nmd, na, priv);
275 	NMA_UNLOCK(nmd);
276 
277 	return nifp;
278 }
279 
280 void
281 netmap_mem_if_delete(struct netmap_adapter *na, struct netmap_if *nif)
282 {
283 	struct netmap_mem_d *nmd = na->nm_mem;
284 
285 	NMA_LOCK(nmd);
286 	nmd->ops->nmd_if_delete(nmd, na, nif);
287 	NMA_UNLOCK(nmd);
288 }
289 
290 int
291 netmap_mem_rings_create(struct netmap_adapter *na)
292 {
293 	int rv;
294 	struct netmap_mem_d *nmd = na->nm_mem;
295 
296 	NMA_LOCK(nmd);
297 	rv = nmd->ops->nmd_rings_create(nmd, na);
298 	NMA_UNLOCK(nmd);
299 
300 	return rv;
301 }
302 
303 void
304 netmap_mem_rings_delete(struct netmap_adapter *na)
305 {
306 	struct netmap_mem_d *nmd = na->nm_mem;
307 
308 	NMA_LOCK(nmd);
309 	nmd->ops->nmd_rings_delete(nmd, na);
310 	NMA_UNLOCK(nmd);
311 }
312 
313 static int netmap_mem_map(struct netmap_obj_pool *, struct netmap_adapter *);
314 static int netmap_mem_unmap(struct netmap_obj_pool *, struct netmap_adapter *);
315 static int nm_mem_check_group(struct netmap_mem_d *, bus_dma_tag_t);
316 static void nm_mem_release_id(struct netmap_mem_d *);
317 
318 nm_memid_t
319 netmap_mem_get_id(struct netmap_mem_d *nmd)
320 {
321 	return nmd->nm_id;
322 }
323 
324 #ifdef NM_DEBUG_MEM_PUTGET
325 #define NM_DBG_REFC(nmd, func, line)	\
326 	nm_prinf("%s:%d mem[%d:%d] -> %d", func, line, (nmd)->nm_id, (nmd)->nm_grp, (nmd)->refcount);
327 #else
328 #define NM_DBG_REFC(nmd, func, line)
329 #endif
330 
331 /* circular list of all existing allocators */
332 static struct netmap_mem_d *netmap_last_mem_d = &nm_mem;
333 static NM_MTX_T nm_mem_list_lock;
334 
335 struct netmap_mem_d *
336 __netmap_mem_get(struct netmap_mem_d *nmd, const char *func, int line)
337 {
338 	NM_MTX_LOCK(nm_mem_list_lock);
339 	nmd->refcount++;
340 	NM_DBG_REFC(nmd, func, line);
341 	NM_MTX_UNLOCK(nm_mem_list_lock);
342 	return nmd;
343 }
344 
345 void
346 __netmap_mem_put(struct netmap_mem_d *nmd, const char *func, int line)
347 {
348 	int last;
349 	NM_MTX_LOCK(nm_mem_list_lock);
350 	last = (--nmd->refcount == 0);
351 	if (last)
352 		nm_mem_release_id(nmd);
353 	NM_DBG_REFC(nmd, func, line);
354 	NM_MTX_UNLOCK(nm_mem_list_lock);
355 	if (last)
356 		netmap_mem_delete(nmd);
357 }
358 
359 int
360 netmap_mem_finalize(struct netmap_mem_d *nmd, struct netmap_adapter *na)
361 {
362 	int lasterr = 0;
363 	if (nm_mem_check_group(nmd, na->pdev) < 0) {
364 		return ENOMEM;
365 	}
366 
367 	NMA_LOCK(nmd);
368 
369 	if (netmap_mem_config(nmd))
370 		goto out;
371 
372 	nmd->active++;
373 
374 	nmd->lasterr = nmd->ops->nmd_finalize(nmd, na);
375 
376 	if (!nmd->lasterr && !(nmd->flags & NETMAP_MEM_NOMAP)) {
377 		nmd->lasterr = netmap_mem_map(&nmd->pools[NETMAP_BUF_POOL], na);
378 	}
379 
380 out:
381 	lasterr = nmd->lasterr;
382 	NMA_UNLOCK(nmd);
383 
384 	if (lasterr)
385 		netmap_mem_deref(nmd, na);
386 
387 	return lasterr;
388 }
389 
390 static int
391 nm_isset(uint32_t *bitmap, u_int i)
392 {
393 	return bitmap[ (i>>5) ] & ( 1U << (i & 31U) );
394 }
395 
396 
397 static int
398 netmap_init_obj_allocator_bitmap(struct netmap_obj_pool *p)
399 {
400 	u_int n, j;
401 
402 	if (p->bitmap == NULL) {
403 		/* Allocate the bitmap */
404 		n = (p->objtotal + 31) / 32;
405 		p->bitmap = nm_os_malloc(sizeof(p->bitmap[0]) * n);
406 		if (p->bitmap == NULL) {
407 			nm_prerr("Unable to create bitmap (%d entries) for allocator '%s'", (int)n,
408 			    p->name);
409 			return ENOMEM;
410 		}
411 		p->bitmap_slots = n;
412 	} else {
413 		memset(p->bitmap, 0, p->bitmap_slots * sizeof(p->bitmap[0]));
414 	}
415 
416 	p->objfree = 0;
417 	/*
418 	 * Set all the bits in the bitmap that have
419 	 * corresponding buffers to 1 to indicate they are
420 	 * free.
421 	 */
422 	for (j = 0; j < p->objtotal; j++) {
423 		if (p->invalid_bitmap && nm_isset(p->invalid_bitmap, j)) {
424 			if (netmap_debug & NM_DEBUG_MEM)
425 				nm_prinf("skipping %s %d", p->name, j);
426 			continue;
427 		}
428 		p->bitmap[ (j>>5) ] |=  ( 1U << (j & 31U) );
429 		p->objfree++;
430 	}
431 
432 	if (netmap_verbose)
433 		nm_prinf("%s free %u", p->name, p->objfree);
434 	if (p->objfree == 0) {
435 		if (netmap_verbose)
436 			nm_prerr("%s: no objects available", p->name);
437 		return ENOMEM;
438 	}
439 
440 	return 0;
441 }
442 
443 static int
444 netmap_mem_init_bitmaps(struct netmap_mem_d *nmd)
445 {
446 	int i, error = 0;
447 
448 	for (i = 0; i < NETMAP_POOLS_NR; i++) {
449 		struct netmap_obj_pool *p = &nmd->pools[i];
450 
451 		error = netmap_init_obj_allocator_bitmap(p);
452 		if (error)
453 			return error;
454 	}
455 
456 	/*
457 	 * buffers 0 and 1 are reserved
458 	 */
459 	if (nmd->pools[NETMAP_BUF_POOL].objfree < 2) {
460 		nm_prerr("%s: not enough buffers", nmd->pools[NETMAP_BUF_POOL].name);
461 		return ENOMEM;
462 	}
463 
464 	nmd->pools[NETMAP_BUF_POOL].objfree -= 2;
465 	if (nmd->pools[NETMAP_BUF_POOL].bitmap) {
466 		/* XXX This check is a workaround that prevents a
467 		 * NULL pointer crash which currently happens only
468 		 * with ptnetmap guests.
469 		 * Removed shared-info --> is the bug still there? */
470 		nmd->pools[NETMAP_BUF_POOL].bitmap[0] = ~3U;
471 	}
472 	return 0;
473 }
474 
475 int
476 netmap_mem_deref(struct netmap_mem_d *nmd, struct netmap_adapter *na)
477 {
478 	int last_user = 0;
479 	NMA_LOCK(nmd);
480 	if (na->active_fds <= 0 && !(nmd->flags & NETMAP_MEM_NOMAP))
481 		netmap_mem_unmap(&nmd->pools[NETMAP_BUF_POOL], na);
482 	if (nmd->active == 1) {
483 		last_user = 1;
484 		/*
485 		 * Reset the allocator when it falls out of use so that any
486 		 * pool resources leaked by unclean application exits are
487 		 * reclaimed.
488 		 */
489 		netmap_mem_init_bitmaps(nmd);
490 	}
491 	nmd->ops->nmd_deref(nmd, na);
492 
493 	nmd->active--;
494 	if (last_user) {
495 		nmd->lasterr = 0;
496 	}
497 
498 	NMA_UNLOCK(nmd);
499 	return last_user;
500 }
501 
502 
503 /* accessor functions */
504 static int
505 netmap_mem2_get_lut(struct netmap_mem_d *nmd, struct netmap_lut *lut)
506 {
507 	lut->lut = nmd->pools[NETMAP_BUF_POOL].lut;
508 #ifdef __FreeBSD__
509 	lut->plut = lut->lut;
510 #endif
511 	lut->objtotal = nmd->pools[NETMAP_BUF_POOL].objtotal;
512 	lut->objsize = nmd->pools[NETMAP_BUF_POOL]._objsize;
513 
514 	return 0;
515 }
516 
517 static struct netmap_obj_params netmap_min_priv_params[NETMAP_POOLS_NR] = {
518 	[NETMAP_IF_POOL] = {
519 		.size = 1024,
520 		.num  = 2,
521 	},
522 	[NETMAP_RING_POOL] = {
523 		.size = 5*PAGE_SIZE,
524 		.num  = 4,
525 	},
526 	[NETMAP_BUF_POOL] = {
527 		.size = 2048,
528 		.num  = 4098,
529 	},
530 };
531 
532 
533 /*
534  * nm_mem is the memory allocator used for all physical interfaces
535  * running in netmap mode.
536  * Virtual (VALE) ports will have each its own allocator.
537  */
538 extern struct netmap_mem_ops netmap_mem_global_ops; /* forward */
539 struct netmap_mem_d nm_mem = {	/* Our memory allocator. */
540 	.pools = {
541 		[NETMAP_IF_POOL] = {
542 			.name 	= "netmap_if",
543 			.objminsize = sizeof(struct netmap_if),
544 			.objmaxsize = 4096,
545 			.nummin     = 10,	/* don't be stingy */
546 			.nummax	    = 10000,	/* XXX very large */
547 		},
548 		[NETMAP_RING_POOL] = {
549 			.name 	= "netmap_ring",
550 			.objminsize = sizeof(struct netmap_ring),
551 			.objmaxsize = 32*PAGE_SIZE,
552 			.nummin     = 2,
553 			.nummax	    = 1024,
554 		},
555 		[NETMAP_BUF_POOL] = {
556 			.name	= "netmap_buf",
557 			.objminsize = 64,
558 			.objmaxsize = 65536,
559 			.nummin     = 4,
560 			.nummax	    = 1000000, /* one million! */
561 		},
562 	},
563 
564 	.params = {
565 		[NETMAP_IF_POOL] = {
566 			.size = 1024,
567 			.num  = 100,
568 		},
569 		[NETMAP_RING_POOL] = {
570 			.size = 9*PAGE_SIZE,
571 			.num  = 200,
572 		},
573 		[NETMAP_BUF_POOL] = {
574 			.size = 2048,
575 			.num  = NETMAP_BUF_MAX_NUM,
576 		},
577 	},
578 
579 	.nm_id = 1,
580 	.nm_grp = -1,
581 
582 	.prev = &nm_mem,
583 	.next = &nm_mem,
584 
585 	.ops = &netmap_mem_global_ops,
586 
587 	.name = "1"
588 };
589 
590 static struct netmap_mem_d nm_mem_blueprint;
591 
592 /* blueprint for the private memory allocators */
593 /* XXX clang is not happy about using name as a print format */
594 static const struct netmap_mem_d nm_blueprint = {
595 	.pools = {
596 		[NETMAP_IF_POOL] = {
597 			.name 	= "%s_if",
598 			.objminsize = sizeof(struct netmap_if),
599 			.objmaxsize = 4096,
600 			.nummin     = 1,
601 			.nummax	    = 100,
602 		},
603 		[NETMAP_RING_POOL] = {
604 			.name 	= "%s_ring",
605 			.objminsize = sizeof(struct netmap_ring),
606 			.objmaxsize = 32*PAGE_SIZE,
607 			.nummin     = 2,
608 			.nummax	    = 1024,
609 		},
610 		[NETMAP_BUF_POOL] = {
611 			.name	= "%s_buf",
612 			.objminsize = 64,
613 			.objmaxsize = 65536,
614 			.nummin     = 4,
615 			.nummax	    = 1000000, /* one million! */
616 		},
617 	},
618 
619 	.nm_grp = -1,
620 
621 	.flags = NETMAP_MEM_PRIVATE,
622 
623 	.ops = &netmap_mem_global_ops,
624 };
625 
626 /* memory allocator related sysctls */
627 
628 #define STRINGIFY(x) #x
629 
630 
631 #define DECLARE_SYSCTLS(id, name) \
632 	SYSBEGIN(mem2_ ## name); \
633 	SYSCTL_INT(_dev_netmap, OID_AUTO, name##_size, \
634 	    CTLFLAG_RW, &nm_mem.params[id].size, 0, "Requested size of netmap " STRINGIFY(name) "s"); \
635 	SYSCTL_INT(_dev_netmap, OID_AUTO, name##_curr_size, \
636 	    CTLFLAG_RD, &nm_mem.pools[id]._objsize, 0, "Current size of netmap " STRINGIFY(name) "s"); \
637 	SYSCTL_INT(_dev_netmap, OID_AUTO, name##_num, \
638 	    CTLFLAG_RW, &nm_mem.params[id].num, 0, "Requested number of netmap " STRINGIFY(name) "s"); \
639 	SYSCTL_INT(_dev_netmap, OID_AUTO, name##_curr_num, \
640 	    CTLFLAG_RD, &nm_mem.pools[id].objtotal, 0, "Current number of netmap " STRINGIFY(name) "s"); \
641 	SYSCTL_INT(_dev_netmap, OID_AUTO, priv_##name##_size, \
642 	    CTLFLAG_RW, &netmap_min_priv_params[id].size, 0, \
643 	    "Default size of private netmap " STRINGIFY(name) "s"); \
644 	SYSCTL_INT(_dev_netmap, OID_AUTO, priv_##name##_num, \
645 	    CTLFLAG_RW, &netmap_min_priv_params[id].num, 0, \
646 	    "Default number of private netmap " STRINGIFY(name) "s");	\
647 	SYSEND
648 
649 SYSCTL_DECL(_dev_netmap);
650 DECLARE_SYSCTLS(NETMAP_IF_POOL, if);
651 DECLARE_SYSCTLS(NETMAP_RING_POOL, ring);
652 DECLARE_SYSCTLS(NETMAP_BUF_POOL, buf);
653 
654 /* call with nm_mem_list_lock held */
655 static int
656 nm_mem_assign_id_locked(struct netmap_mem_d *nmd, int grp_id)
657 {
658 	nm_memid_t id;
659 	struct netmap_mem_d *scan = netmap_last_mem_d;
660 	int error = ENOMEM;
661 
662 	do {
663 		/* we rely on unsigned wrap around */
664 		id = scan->nm_id + 1;
665 		if (id == 0) /* reserve 0 as error value */
666 			id = 1;
667 		scan = scan->next;
668 		if (id != scan->nm_id) {
669 			nmd->nm_id = id;
670 			nmd->nm_grp = grp_id;
671 			nmd->prev = scan->prev;
672 			nmd->next = scan;
673 			scan->prev->next = nmd;
674 			scan->prev = nmd;
675 			netmap_last_mem_d = nmd;
676 			nmd->refcount = 1;
677 			NM_DBG_REFC(nmd, __FUNCTION__, __LINE__);
678 			error = 0;
679 			break;
680 		}
681 	} while (scan != netmap_last_mem_d);
682 
683 	return error;
684 }
685 
686 /* call with nm_mem_list_lock *not* held */
687 static int
688 nm_mem_assign_id(struct netmap_mem_d *nmd, int grp_id)
689 {
690 	int ret;
691 
692 	NM_MTX_LOCK(nm_mem_list_lock);
693 	ret = nm_mem_assign_id_locked(nmd, grp_id);
694 	NM_MTX_UNLOCK(nm_mem_list_lock);
695 
696 	return ret;
697 }
698 
699 /* call with nm_mem_list_lock held */
700 static void
701 nm_mem_release_id(struct netmap_mem_d *nmd)
702 {
703 	nmd->prev->next = nmd->next;
704 	nmd->next->prev = nmd->prev;
705 
706 	if (netmap_last_mem_d == nmd)
707 		netmap_last_mem_d = nmd->prev;
708 
709 	nmd->prev = nmd->next = NULL;
710 }
711 
712 struct netmap_mem_d *
713 netmap_mem_find(nm_memid_t id)
714 {
715 	struct netmap_mem_d *nmd;
716 
717 	NM_MTX_LOCK(nm_mem_list_lock);
718 	nmd = netmap_last_mem_d;
719 	do {
720 		if (!(nmd->flags & NETMAP_MEM_HIDDEN) && nmd->nm_id == id) {
721 			nmd->refcount++;
722 			NM_DBG_REFC(nmd, __FUNCTION__, __LINE__);
723 			NM_MTX_UNLOCK(nm_mem_list_lock);
724 			return nmd;
725 		}
726 		nmd = nmd->next;
727 	} while (nmd != netmap_last_mem_d);
728 	NM_MTX_UNLOCK(nm_mem_list_lock);
729 	return NULL;
730 }
731 
732 static int
733 nm_mem_check_group(struct netmap_mem_d *nmd, bus_dma_tag_t dev)
734 {
735 	int err = 0, id;
736 
737 	/* Skip not hw adapters.
738 	 * Vale port can use particular allocator through vale-ctl -m option
739 	 */
740 	if (!dev)
741 		return 0;
742 	id = nm_iommu_group_id(dev);
743 	if (netmap_debug & NM_DEBUG_MEM)
744 		nm_prinf("iommu_group %d", id);
745 
746 	NMA_LOCK(nmd);
747 
748 	if (nmd->nm_grp != id) {
749 		if (netmap_verbose)
750 			nm_prerr("iommu group mismatch: %d vs %d",
751 					nmd->nm_grp, id);
752 		nmd->lasterr = err = ENOMEM;
753 	}
754 
755 	NMA_UNLOCK(nmd);
756 	return err;
757 }
758 
759 static struct lut_entry *
760 nm_alloc_lut(u_int nobj)
761 {
762 	size_t n = sizeof(struct lut_entry) * nobj;
763 	struct lut_entry *lut;
764 #ifdef linux
765 	lut = vmalloc(n);
766 #else
767 	lut = nm_os_malloc(n);
768 #endif
769 	return lut;
770 }
771 
772 static void
773 nm_free_lut(struct lut_entry *lut, u_int objtotal)
774 {
775 	bzero(lut, sizeof(struct lut_entry) * objtotal);
776 #ifdef linux
777 	vfree(lut);
778 #else
779 	nm_os_free(lut);
780 #endif
781 }
782 
783 #if defined(linux) || defined(_WIN32)
784 static struct plut_entry *
785 nm_alloc_plut(u_int nobj)
786 {
787 	size_t n = sizeof(struct plut_entry) * nobj;
788 	struct plut_entry *lut;
789 	lut = vmalloc(n);
790 	return lut;
791 }
792 
793 static void
794 nm_free_plut(struct plut_entry * lut)
795 {
796 	vfree(lut);
797 }
798 #endif /* linux or _WIN32 */
799 
800 
801 /*
802  * First, find the allocator that contains the requested offset,
803  * then locate the cluster through a lookup table.
804  */
805 static vm_paddr_t
806 netmap_mem2_ofstophys(struct netmap_mem_d* nmd, vm_ooffset_t offset)
807 {
808 	int i;
809 	vm_ooffset_t o = offset;
810 	vm_paddr_t pa;
811 	struct netmap_obj_pool *p;
812 
813 	p = nmd->pools;
814 
815 	for (i = 0; i < NETMAP_POOLS_NR; offset -= p[i].memtotal, i++) {
816 		if (offset >= p[i].memtotal)
817 			continue;
818 		// now lookup the cluster's address
819 #ifndef _WIN32
820 		pa = vtophys(p[i].lut[offset / p[i]._objsize].vaddr) +
821 			offset % p[i]._objsize;
822 #else
823 		pa = vtophys(p[i].lut[offset / p[i]._objsize].vaddr);
824 		pa.QuadPart += offset % p[i]._objsize;
825 #endif
826 		return pa;
827 	}
828 	/* this is only in case of errors */
829 	nm_prerr("invalid ofs 0x%x out of 0x%zx 0x%zx 0x%zx", (u_int)o,
830 		p[NETMAP_IF_POOL].memtotal,
831 		p[NETMAP_IF_POOL].memtotal
832 			+ p[NETMAP_RING_POOL].memtotal,
833 		p[NETMAP_IF_POOL].memtotal
834 			+ p[NETMAP_RING_POOL].memtotal
835 			+ p[NETMAP_BUF_POOL].memtotal);
836 #ifndef _WIN32
837 	return 0; /* bad address */
838 #else
839 	vm_paddr_t res;
840 	res.QuadPart = 0;
841 	return res;
842 #endif
843 }
844 
845 #ifdef _WIN32
846 
847 /*
848  * win32_build_virtual_memory_for_userspace
849  *
850  * This function get all the object making part of the pools and maps
851  * a contiguous virtual memory space for the userspace
852  * It works this way
853  * 1 - allocate a Memory Descriptor List wide as the sum
854  *		of the memory needed for the pools
855  * 2 - cycle all the objects in every pool and for every object do
856  *
857  *		2a - cycle all the objects in every pool, get the list
858  *				of the physical address descriptors
859  *		2b - calculate the offset in the array of pages descriptor in the
860  *				main MDL
861  *		2c - copy the descriptors of the object in the main MDL
862  *
863  * 3 - return the resulting MDL that needs to be mapped in userland
864  *
865  * In this way we will have an MDL that describes all the memory for the
866  * objects in a single object
867 */
868 
869 PMDL
870 win32_build_user_vm_map(struct netmap_mem_d* nmd)
871 {
872 	u_int memflags, ofs = 0;
873 	PMDL mainMdl, tempMdl;
874 	uint64_t memsize;
875 	int i, j;
876 
877 	if (netmap_mem_get_info(nmd, &memsize, &memflags, NULL)) {
878 		nm_prerr("memory not finalised yet");
879 		return NULL;
880 	}
881 
882 	mainMdl = IoAllocateMdl(NULL, memsize, FALSE, FALSE, NULL);
883 	if (mainMdl == NULL) {
884 		nm_prerr("failed to allocate mdl");
885 		return NULL;
886 	}
887 
888 	NMA_LOCK(nmd);
889 	for (i = 0; i < NETMAP_POOLS_NR; i++) {
890 		struct netmap_obj_pool *p = &nmd->pools[i];
891 		int clsz = p->_clustsize;
892 		int clobjs = p->_clustentries; /* objects per cluster */
893 		int mdl_len = sizeof(PFN_NUMBER) * BYTES_TO_PAGES(clsz);
894 		PPFN_NUMBER pSrc, pDst;
895 
896 		/* each pool has a different cluster size so we need to reallocate */
897 		tempMdl = IoAllocateMdl(p->lut[0].vaddr, clsz, FALSE, FALSE, NULL);
898 		if (tempMdl == NULL) {
899 			NMA_UNLOCK(nmd);
900 			nm_prerr("fail to allocate tempMdl");
901 			IoFreeMdl(mainMdl);
902 			return NULL;
903 		}
904 		pSrc = MmGetMdlPfnArray(tempMdl);
905 		/* create one entry per cluster, the lut[] has one entry per object */
906 		for (j = 0; j < p->numclusters; j++, ofs += clsz) {
907 			pDst = &MmGetMdlPfnArray(mainMdl)[BYTES_TO_PAGES(ofs)];
908 			MmInitializeMdl(tempMdl, p->lut[j*clobjs].vaddr, clsz);
909 			MmBuildMdlForNonPagedPool(tempMdl); /* compute physical page addresses */
910 			RtlCopyMemory(pDst, pSrc, mdl_len); /* copy the page descriptors */
911 			mainMdl->MdlFlags = tempMdl->MdlFlags; /* XXX what is in here ? */
912 		}
913 		IoFreeMdl(tempMdl);
914 	}
915 	NMA_UNLOCK(nmd);
916 	return mainMdl;
917 }
918 
919 #endif /* _WIN32 */
920 
921 /*
922  * helper function for OS-specific mmap routines (currently only windows).
923  * Given an nmd and a pool index, returns the cluster size and number of clusters.
924  * Returns 0 if memory is finalised and the pool is valid, otherwise 1.
925  * It should be called under NMA_LOCK(nmd) otherwise the underlying info can change.
926  */
927 
928 int
929 netmap_mem2_get_pool_info(struct netmap_mem_d* nmd, u_int pool, u_int *clustsize, u_int *numclusters)
930 {
931 	if (!nmd || !clustsize || !numclusters || pool >= NETMAP_POOLS_NR)
932 		return 1; /* invalid arguments */
933 	// NMA_LOCK_ASSERT(nmd);
934 	if (!(nmd->flags & NETMAP_MEM_FINALIZED)) {
935 		*clustsize = *numclusters = 0;
936 		return 1; /* not ready yet */
937 	}
938 	*clustsize = nmd->pools[pool]._clustsize;
939 	*numclusters = nmd->pools[pool].numclusters;
940 	return 0; /* success */
941 }
942 
943 static int
944 netmap_mem2_get_info(struct netmap_mem_d* nmd, uint64_t* size,
945 			u_int *memflags, nm_memid_t *id)
946 {
947 	int error = 0;
948 	error = netmap_mem_config(nmd);
949 	if (error)
950 		goto out;
951 	if (size) {
952 		if (nmd->flags & NETMAP_MEM_FINALIZED) {
953 			*size = nmd->nm_totalsize;
954 		} else {
955 			int i;
956 			*size = 0;
957 			for (i = 0; i < NETMAP_POOLS_NR; i++) {
958 				struct netmap_obj_pool *p = nmd->pools + i;
959 				*size += ((size_t)p->_numclusters * (size_t)p->_clustsize);
960 			}
961 		}
962 	}
963 	if (memflags)
964 		*memflags = nmd->flags;
965 	if (id)
966 		*id = nmd->nm_id;
967 out:
968 	return error;
969 }
970 
971 /*
972  * we store objects by kernel address, need to find the offset
973  * within the pool to export the value to userspace.
974  * Algorithm: scan until we find the cluster, then add the
975  * actual offset in the cluster
976  */
977 static ssize_t
978 netmap_obj_offset(struct netmap_obj_pool *p, const void *vaddr)
979 {
980 	int i, k = p->_clustentries, n = p->objtotal;
981 	ssize_t ofs = 0;
982 
983 	for (i = 0; i < n; i += k, ofs += p->_clustsize) {
984 		const char *base = p->lut[i].vaddr;
985 		ssize_t relofs = (const char *) vaddr - base;
986 
987 		if (relofs < 0 || relofs >= p->_clustsize)
988 			continue;
989 
990 		ofs = ofs + relofs;
991 		nm_prdis("%s: return offset %d (cluster %d) for pointer %p",
992 		    p->name, ofs, i, vaddr);
993 		return ofs;
994 	}
995 	nm_prerr("address %p is not contained inside any cluster (%s)",
996 	    vaddr, p->name);
997 	return 0; /* An error occurred */
998 }
999 
1000 /* Helper functions which convert virtual addresses to offsets */
1001 #define netmap_if_offset(n, v)					\
1002 	netmap_obj_offset(&(n)->pools[NETMAP_IF_POOL], (v))
1003 
1004 #define netmap_ring_offset(n, v)				\
1005     ((n)->pools[NETMAP_IF_POOL].memtotal + 			\
1006 	netmap_obj_offset(&(n)->pools[NETMAP_RING_POOL], (v)))
1007 
1008 static ssize_t
1009 netmap_mem2_if_offset(struct netmap_mem_d *nmd, const void *addr)
1010 {
1011 	return netmap_if_offset(nmd, addr);
1012 }
1013 
1014 /*
1015  * report the index, and use start position as a hint,
1016  * otherwise buffer allocation becomes terribly expensive.
1017  */
1018 static void *
1019 netmap_obj_malloc(struct netmap_obj_pool *p, u_int len, uint32_t *start, uint32_t *index)
1020 {
1021 	uint32_t i = 0;			/* index in the bitmap */
1022 	uint32_t mask, j = 0;		/* slot counter */
1023 	void *vaddr = NULL;
1024 
1025 	if (len > p->_objsize) {
1026 		nm_prerr("%s request size %d too large", p->name, len);
1027 		return NULL;
1028 	}
1029 
1030 	if (p->objfree == 0) {
1031 		nm_prerr("no more %s objects", p->name);
1032 		return NULL;
1033 	}
1034 	if (start)
1035 		i = *start;
1036 
1037 	/* termination is guaranteed by p->free, but better check bounds on i */
1038 	while (vaddr == NULL && i < p->bitmap_slots)  {
1039 		uint32_t cur = p->bitmap[i];
1040 		if (cur == 0) { /* bitmask is fully used */
1041 			i++;
1042 			continue;
1043 		}
1044 		/* locate a slot */
1045 		for (j = 0, mask = 1; (cur & mask) == 0; j++, mask <<= 1)
1046 			;
1047 
1048 		p->bitmap[i] &= ~mask; /* mark object as in use */
1049 		p->objfree--;
1050 
1051 		vaddr = p->lut[i * 32 + j].vaddr;
1052 		if (index)
1053 			*index = i * 32 + j;
1054 	}
1055 	nm_prdis("%s allocator: allocated object @ [%d][%d]: vaddr %p",p->name, i, j, vaddr);
1056 
1057 	if (start)
1058 		*start = i;
1059 	return vaddr;
1060 }
1061 
1062 
1063 /*
1064  * free by index, not by address.
1065  * XXX should we also cleanup the content ?
1066  */
1067 static int
1068 netmap_obj_free(struct netmap_obj_pool *p, uint32_t j)
1069 {
1070 	uint32_t *ptr, mask;
1071 
1072 	if (j >= p->objtotal) {
1073 		nm_prerr("invalid index %u, max %u", j, p->objtotal);
1074 		return 1;
1075 	}
1076 	ptr = &p->bitmap[j / 32];
1077 	mask = (1 << (j % 32));
1078 	if (*ptr & mask) {
1079 		nm_prerr("ouch, double free on buffer %d", j);
1080 		return 1;
1081 	} else {
1082 		*ptr |= mask;
1083 		p->objfree++;
1084 		return 0;
1085 	}
1086 }
1087 
1088 /*
1089  * free by address. This is slow but is only used for a few
1090  * objects (rings, nifp)
1091  */
1092 static void
1093 netmap_obj_free_va(struct netmap_obj_pool *p, void *vaddr)
1094 {
1095 	u_int i, j, n = p->numclusters;
1096 
1097 	for (i = 0, j = 0; i < n; i++, j += p->_clustentries) {
1098 		void *base = p->lut[i * p->_clustentries].vaddr;
1099 		ssize_t relofs = (ssize_t) vaddr - (ssize_t) base;
1100 
1101 		/* Given address, is out of the scope of the current cluster.*/
1102 		if (base == NULL || vaddr < base || relofs >= p->_clustsize)
1103 			continue;
1104 
1105 		j = j + relofs / p->_objsize;
1106 		/* KASSERT(j != 0, ("Cannot free object 0")); */
1107 		netmap_obj_free(p, j);
1108 		return;
1109 	}
1110 	nm_prerr("address %p is not contained inside any cluster (%s)",
1111 	    vaddr, p->name);
1112 }
1113 
1114 unsigned
1115 netmap_mem_bufsize(struct netmap_mem_d *nmd)
1116 {
1117 	return nmd->pools[NETMAP_BUF_POOL]._objsize;
1118 }
1119 
1120 #define netmap_if_malloc(n, len)	netmap_obj_malloc(&(n)->pools[NETMAP_IF_POOL], len, NULL, NULL)
1121 #define netmap_if_free(n, v)		netmap_obj_free_va(&(n)->pools[NETMAP_IF_POOL], (v))
1122 #define netmap_ring_malloc(n, len)	netmap_obj_malloc(&(n)->pools[NETMAP_RING_POOL], len, NULL, NULL)
1123 #define netmap_ring_free(n, v)		netmap_obj_free_va(&(n)->pools[NETMAP_RING_POOL], (v))
1124 #define netmap_buf_malloc(n, _pos, _index)			\
1125 	netmap_obj_malloc(&(n)->pools[NETMAP_BUF_POOL], netmap_mem_bufsize(n), _pos, _index)
1126 
1127 
1128 #if 0 /* currently unused */
1129 /* Return the index associated to the given packet buffer */
1130 #define netmap_buf_index(n, v)						\
1131     (netmap_obj_offset(&(n)->pools[NETMAP_BUF_POOL], (v)) / NETMAP_BDG_BUF_SIZE(n))
1132 #endif
1133 
1134 /*
1135  * allocate extra buffers in a linked list.
1136  * returns the actual number.
1137  */
1138 uint32_t
1139 netmap_extra_alloc(struct netmap_adapter *na, uint32_t *head, uint32_t n)
1140 {
1141 	struct netmap_mem_d *nmd = na->nm_mem;
1142 	uint32_t i, pos = 0; /* opaque, scan position in the bitmap */
1143 
1144 	NMA_LOCK(nmd);
1145 
1146 	*head = 0;	/* default, 'null' index ie empty list */
1147 	for (i = 0 ; i < n; i++) {
1148 		uint32_t cur = *head;	/* save current head */
1149 		uint32_t *p = netmap_buf_malloc(nmd, &pos, head);
1150 		if (p == NULL) {
1151 			nm_prerr("no more buffers after %d of %d", i, n);
1152 			*head = cur; /* restore */
1153 			break;
1154 		}
1155 		nm_prdis(5, "allocate buffer %d -> %d", *head, cur);
1156 		*p = cur; /* link to previous head */
1157 	}
1158 
1159 	NMA_UNLOCK(nmd);
1160 
1161 	return i;
1162 }
1163 
1164 static void
1165 netmap_extra_free(struct netmap_adapter *na, uint32_t head)
1166 {
1167 	struct lut_entry *lut = na->na_lut.lut;
1168 	struct netmap_mem_d *nmd = na->nm_mem;
1169 	struct netmap_obj_pool *p = &nmd->pools[NETMAP_BUF_POOL];
1170 	uint32_t i, cur, *buf;
1171 
1172 	nm_prdis("freeing the extra list");
1173 	for (i = 0; head >=2 && head < p->objtotal; i++) {
1174 		cur = head;
1175 		buf = lut[head].vaddr;
1176 		head = *buf;
1177 		*buf = 0;
1178 		if (netmap_obj_free(p, cur))
1179 			break;
1180 	}
1181 	if (head != 0)
1182 		nm_prerr("breaking with head %d", head);
1183 	if (netmap_debug & NM_DEBUG_MEM)
1184 		nm_prinf("freed %d buffers", i);
1185 }
1186 
1187 
1188 /* Return nonzero on error */
1189 static int
1190 netmap_new_bufs(struct netmap_mem_d *nmd, struct netmap_slot *slot, u_int n)
1191 {
1192 	struct netmap_obj_pool *p = &nmd->pools[NETMAP_BUF_POOL];
1193 	u_int i = 0;	/* slot counter */
1194 	uint32_t pos = 0;	/* slot in p->bitmap */
1195 	uint32_t index = 0;	/* buffer index */
1196 
1197 	for (i = 0; i < n; i++) {
1198 		void *vaddr = netmap_buf_malloc(nmd, &pos, &index);
1199 		if (vaddr == NULL) {
1200 			nm_prerr("no more buffers after %d of %d", i, n);
1201 			goto cleanup;
1202 		}
1203 		slot[i].buf_idx = index;
1204 		slot[i].len = p->_objsize;
1205 		slot[i].flags = 0;
1206 		slot[i].ptr = 0;
1207 	}
1208 
1209 	nm_prdis("%s: allocated %d buffers, %d available, first at %d", p->name, n, p->objfree, pos);
1210 	return (0);
1211 
1212 cleanup:
1213 	while (i > 0) {
1214 		i--;
1215 		netmap_obj_free(p, slot[i].buf_idx);
1216 	}
1217 	bzero(slot, n * sizeof(slot[0]));
1218 	return (ENOMEM);
1219 }
1220 
1221 static void
1222 netmap_mem_set_ring(struct netmap_mem_d *nmd, struct netmap_slot *slot, u_int n, uint32_t index)
1223 {
1224 	struct netmap_obj_pool *p = &nmd->pools[NETMAP_BUF_POOL];
1225 	u_int i;
1226 
1227 	for (i = 0; i < n; i++) {
1228 		slot[i].buf_idx = index;
1229 		slot[i].len = p->_objsize;
1230 		slot[i].flags = 0;
1231 	}
1232 }
1233 
1234 
1235 static void
1236 netmap_free_buf(struct netmap_mem_d *nmd, uint32_t i)
1237 {
1238 	struct netmap_obj_pool *p = &nmd->pools[NETMAP_BUF_POOL];
1239 
1240 	if (i < 2 || i >= p->objtotal) {
1241 		nm_prerr("Cannot free buf#%d: should be in [2, %d[", i, p->objtotal);
1242 		return;
1243 	}
1244 	netmap_obj_free(p, i);
1245 }
1246 
1247 
1248 static void
1249 netmap_free_bufs(struct netmap_mem_d *nmd, struct netmap_slot *slot, u_int n)
1250 {
1251 	u_int i;
1252 
1253 	for (i = 0; i < n; i++) {
1254 		if (slot[i].buf_idx > 1)
1255 			netmap_free_buf(nmd, slot[i].buf_idx);
1256 	}
1257 	nm_prdis("%s: released some buffers, available: %u",
1258 			p->name, p->objfree);
1259 }
1260 
1261 static void
1262 netmap_reset_obj_allocator(struct netmap_obj_pool *p)
1263 {
1264 
1265 	if (p == NULL)
1266 		return;
1267 	if (p->bitmap)
1268 		nm_os_free(p->bitmap);
1269 	p->bitmap = NULL;
1270 	if (p->invalid_bitmap)
1271 		nm_os_free(p->invalid_bitmap);
1272 	p->invalid_bitmap = NULL;
1273 	if (!p->alloc_done) {
1274 		/* allocation was done by somebody else.
1275 		 * Let them clean up after themselves.
1276 		 */
1277 		return;
1278 	}
1279 	if (p->lut) {
1280 		u_int i;
1281 
1282 		/*
1283 		 * Free each cluster allocated in
1284 		 * netmap_finalize_obj_allocator().  The cluster start
1285 		 * addresses are stored at multiples of p->_clusterentries
1286 		 * in the lut.
1287 		 */
1288 		for (i = 0; i < p->objtotal; i += p->_clustentries) {
1289 			contigfree(p->lut[i].vaddr, p->_clustsize, M_NETMAP);
1290 		}
1291 		nm_free_lut(p->lut, p->objtotal);
1292 	}
1293 	p->lut = NULL;
1294 	p->objtotal = 0;
1295 	p->memtotal = 0;
1296 	p->numclusters = 0;
1297 	p->objfree = 0;
1298 	p->alloc_done = 0;
1299 }
1300 
1301 /*
1302  * Free all resources related to an allocator.
1303  */
1304 static void
1305 netmap_destroy_obj_allocator(struct netmap_obj_pool *p)
1306 {
1307 	if (p == NULL)
1308 		return;
1309 	netmap_reset_obj_allocator(p);
1310 }
1311 
1312 /*
1313  * We receive a request for objtotal objects, of size objsize each.
1314  * Internally we may round up both numbers, as we allocate objects
1315  * in small clusters multiple of the page size.
1316  * We need to keep track of objtotal and clustentries,
1317  * as they are needed when freeing memory.
1318  *
1319  * XXX note -- userspace needs the buffers to be contiguous,
1320  *	so we cannot afford gaps at the end of a cluster.
1321  */
1322 
1323 
1324 /* call with NMA_LOCK held */
1325 static int
1326 netmap_config_obj_allocator(struct netmap_obj_pool *p, u_int objtotal, u_int objsize)
1327 {
1328 	int i;
1329 	u_int clustsize;	/* the cluster size, multiple of page size */
1330 	u_int clustentries;	/* how many objects per entry */
1331 
1332 	/* we store the current request, so we can
1333 	 * detect configuration changes later */
1334 	p->r_objtotal = objtotal;
1335 	p->r_objsize = objsize;
1336 
1337 #define MAX_CLUSTSIZE	(1<<22)		// 4 MB
1338 #define LINE_ROUND	NM_BUF_ALIGN	// 64
1339 	if (objsize >= MAX_CLUSTSIZE) {
1340 		/* we could do it but there is no point */
1341 		nm_prerr("unsupported allocation for %d bytes", objsize);
1342 		return EINVAL;
1343 	}
1344 	/* make sure objsize is a multiple of LINE_ROUND */
1345 	i = (objsize & (LINE_ROUND - 1));
1346 	if (i) {
1347 		nm_prinf("aligning object by %d bytes", LINE_ROUND - i);
1348 		objsize += LINE_ROUND - i;
1349 	}
1350 	if (objsize < p->objminsize || objsize > p->objmaxsize) {
1351 		nm_prerr("requested objsize %d out of range [%d, %d]",
1352 			objsize, p->objminsize, p->objmaxsize);
1353 		return EINVAL;
1354 	}
1355 	if (objtotal < p->nummin || objtotal > p->nummax) {
1356 		nm_prerr("requested objtotal %d out of range [%d, %d]",
1357 			objtotal, p->nummin, p->nummax);
1358 		return EINVAL;
1359 	}
1360 	/*
1361 	 * Compute number of objects using a brute-force approach:
1362 	 * given a max cluster size,
1363 	 * we try to fill it with objects keeping track of the
1364 	 * wasted space to the next page boundary.
1365 	 */
1366 	for (clustentries = 0, i = 1;; i++) {
1367 		u_int delta, used = i * objsize;
1368 		if (used > MAX_CLUSTSIZE)
1369 			break;
1370 		delta = used % PAGE_SIZE;
1371 		if (delta == 0) { // exact solution
1372 			clustentries = i;
1373 			break;
1374 		}
1375 	}
1376 	/* exact solution not found */
1377 	if (clustentries == 0) {
1378 		nm_prerr("unsupported allocation for %d bytes", objsize);
1379 		return EINVAL;
1380 	}
1381 	/* compute clustsize */
1382 	clustsize = clustentries * objsize;
1383 	if (netmap_debug & NM_DEBUG_MEM)
1384 		nm_prinf("objsize %d clustsize %d objects %d",
1385 			objsize, clustsize, clustentries);
1386 
1387 	/*
1388 	 * The number of clusters is n = ceil(objtotal/clustentries)
1389 	 * objtotal' = n * clustentries
1390 	 */
1391 	p->_clustentries = clustentries;
1392 	p->_clustsize = clustsize;
1393 	p->_numclusters = (objtotal + clustentries - 1) / clustentries;
1394 
1395 	/* actual values (may be larger than requested) */
1396 	p->_objsize = objsize;
1397 	p->_objtotal = p->_numclusters * clustentries;
1398 
1399 	return 0;
1400 }
1401 
1402 /* call with NMA_LOCK held */
1403 static int
1404 netmap_finalize_obj_allocator(struct netmap_obj_pool *p)
1405 {
1406 	int i; /* must be signed */
1407 	size_t n;
1408 
1409 	if (p->lut) {
1410 		/* if the lut is already there we assume that also all the
1411 		 * clusters have already been allocated, possibly by somebody
1412 		 * else (e.g., extmem). In the latter case, the alloc_done flag
1413 		 * will remain at zero, so that we will not attempt to
1414 		 * deallocate the clusters by ourselves in
1415 		 * netmap_reset_obj_allocator.
1416 		 */
1417 		return 0;
1418 	}
1419 
1420 	/* optimistically assume we have enough memory */
1421 	p->numclusters = p->_numclusters;
1422 	p->objtotal = p->_objtotal;
1423 	p->alloc_done = 1;
1424 
1425 	p->lut = nm_alloc_lut(p->objtotal);
1426 	if (p->lut == NULL) {
1427 		nm_prerr("Unable to create lookup table for '%s'", p->name);
1428 		goto clean;
1429 	}
1430 
1431 	/*
1432 	 * Allocate clusters, init pointers
1433 	 */
1434 
1435 	n = p->_clustsize;
1436 	for (i = 0; i < (int)p->objtotal;) {
1437 		int lim = i + p->_clustentries;
1438 		char *clust;
1439 
1440 		/*
1441 		 * XXX Note, we only need contigmalloc() for buffers attached
1442 		 * to native interfaces. In all other cases (nifp, netmap rings
1443 		 * and even buffers for VALE ports or emulated interfaces) we
1444 		 * can live with standard malloc, because the hardware will not
1445 		 * access the pages directly.
1446 		 */
1447 		clust = contigmalloc(n, M_NETMAP, M_NOWAIT | M_ZERO,
1448 		    (size_t)0, -1UL, PAGE_SIZE, 0);
1449 		if (clust == NULL) {
1450 			/*
1451 			 * If we get here, there is a severe memory shortage,
1452 			 * so halve the allocated memory to reclaim some.
1453 			 */
1454 			nm_prerr("Unable to create cluster at %d for '%s' allocator",
1455 			    i, p->name);
1456 			if (i < 2) /* nothing to halve */
1457 				goto out;
1458 			lim = i / 2;
1459 			for (i--; i >= lim; i--) {
1460 				if (i % p->_clustentries == 0 && p->lut[i].vaddr)
1461 					contigfree(p->lut[i].vaddr,
1462 						n, M_NETMAP);
1463 				p->lut[i].vaddr = NULL;
1464 			}
1465 		out:
1466 			p->objtotal = i;
1467 			/* we may have stopped in the middle of a cluster */
1468 			p->numclusters = (i + p->_clustentries - 1) / p->_clustentries;
1469 			break;
1470 		}
1471 		/*
1472 		 * Set lut state for all buffers in the current cluster.
1473 		 *
1474 		 * [i, lim) is the set of buffer indexes that cover the
1475 		 * current cluster.
1476 		 *
1477 		 * 'clust' is really the address of the current buffer in
1478 		 * the current cluster as we index through it with a stride
1479 		 * of p->_objsize.
1480 		 */
1481 		for (; i < lim; i++, clust += p->_objsize) {
1482 			p->lut[i].vaddr = clust;
1483 #if !defined(linux) && !defined(_WIN32)
1484 			p->lut[i].paddr = vtophys(clust);
1485 #endif
1486 		}
1487 	}
1488 	p->memtotal = (size_t)p->numclusters * (size_t)p->_clustsize;
1489 	if (netmap_verbose)
1490 		nm_prinf("Pre-allocated %d clusters (%d/%zuKB) for '%s'",
1491 		    p->numclusters, p->_clustsize >> 10,
1492 		    p->memtotal >> 10, p->name);
1493 
1494 	return 0;
1495 
1496 clean:
1497 	netmap_reset_obj_allocator(p);
1498 	return ENOMEM;
1499 }
1500 
1501 /* call with lock held */
1502 static int
1503 netmap_mem_params_changed(struct netmap_obj_params* p)
1504 {
1505 	int i, rv = 0;
1506 
1507 	for (i = 0; i < NETMAP_POOLS_NR; i++) {
1508 		if (p[i].last_size != p[i].size || p[i].last_num != p[i].num) {
1509 			p[i].last_size = p[i].size;
1510 			p[i].last_num = p[i].num;
1511 			rv = 1;
1512 		}
1513 	}
1514 	return rv;
1515 }
1516 
1517 static void
1518 netmap_mem_reset_all(struct netmap_mem_d *nmd)
1519 {
1520 	int i;
1521 
1522 	if (netmap_debug & NM_DEBUG_MEM)
1523 		nm_prinf("resetting %p", nmd);
1524 	for (i = 0; i < NETMAP_POOLS_NR; i++) {
1525 		netmap_reset_obj_allocator(&nmd->pools[i]);
1526 	}
1527 	nmd->flags  &= ~NETMAP_MEM_FINALIZED;
1528 }
1529 
1530 static int
1531 netmap_mem_unmap(struct netmap_obj_pool *p, struct netmap_adapter *na)
1532 {
1533 	int i, lim = p->objtotal;
1534 	struct netmap_lut *lut;
1535 	if (na == NULL || na->pdev == NULL)
1536 		return 0;
1537 
1538 	lut = &na->na_lut;
1539 
1540 
1541 
1542 #if defined(__FreeBSD__)
1543 	/* On FreeBSD mapping and unmapping is performed by the txsync
1544 	 * and rxsync routine, packet by packet. */
1545 	(void)i;
1546 	(void)lim;
1547 	(void)lut;
1548 #elif defined(_WIN32)
1549 	(void)i;
1550 	(void)lim;
1551 	(void)lut;
1552 	nm_prerr("unsupported on Windows");
1553 #else /* linux */
1554 	nm_prdis("unmapping and freeing plut for %s", na->name);
1555 	if (lut->plut == NULL || na->pdev == NULL)
1556 		return 0;
1557 	for (i = 0; i < lim; i += p->_clustentries) {
1558 		if (lut->plut[i].paddr)
1559 			netmap_unload_map(na, (bus_dma_tag_t) na->pdev, &lut->plut[i].paddr, p->_clustsize);
1560 	}
1561 	nm_free_plut(lut->plut);
1562 	lut->plut = NULL;
1563 #endif /* linux */
1564 
1565 	return 0;
1566 }
1567 
1568 static int
1569 netmap_mem_map(struct netmap_obj_pool *p, struct netmap_adapter *na)
1570 {
1571 	int error = 0;
1572 	int i, lim = p->objtotal;
1573 	struct netmap_lut *lut = &na->na_lut;
1574 
1575 	if (na->pdev == NULL)
1576 		return 0;
1577 
1578 #if defined(__FreeBSD__)
1579 	/* On FreeBSD mapping and unmapping is performed by the txsync
1580 	 * and rxsync routine, packet by packet. */
1581 	(void)i;
1582 	(void)lim;
1583 	(void)lut;
1584 #elif defined(_WIN32)
1585 	(void)i;
1586 	(void)lim;
1587 	(void)lut;
1588 	nm_prerr("unsupported on Windows");
1589 #else /* linux */
1590 
1591 	if (lut->plut != NULL) {
1592 		nm_prdis("plut already allocated for %s", na->name);
1593 		return 0;
1594 	}
1595 
1596 	nm_prdis("allocating physical lut for %s", na->name);
1597 	lut->plut = nm_alloc_plut(lim);
1598 	if (lut->plut == NULL) {
1599 		nm_prerr("Failed to allocate physical lut for %s", na->name);
1600 		return ENOMEM;
1601 	}
1602 
1603 	for (i = 0; i < lim; i += p->_clustentries) {
1604 		lut->plut[i].paddr = 0;
1605 	}
1606 
1607 	for (i = 0; i < lim; i += p->_clustentries) {
1608 		int j;
1609 
1610 		if (p->lut[i].vaddr == NULL)
1611 			continue;
1612 
1613 		error = netmap_load_map(na, (bus_dma_tag_t) na->pdev, &lut->plut[i].paddr,
1614 				p->lut[i].vaddr, p->_clustsize);
1615 		if (error) {
1616 			nm_prerr("Failed to map cluster #%d from the %s pool", i, p->name);
1617 			break;
1618 		}
1619 
1620 		for (j = 1; j < p->_clustentries; j++) {
1621 			lut->plut[i + j].paddr = lut->plut[i + j - 1].paddr + p->_objsize;
1622 		}
1623 	}
1624 
1625 	if (error)
1626 		netmap_mem_unmap(p, na);
1627 
1628 #endif /* linux */
1629 
1630 	return error;
1631 }
1632 
1633 static int
1634 netmap_mem_finalize_all(struct netmap_mem_d *nmd)
1635 {
1636 	int i;
1637 	if (nmd->flags & NETMAP_MEM_FINALIZED)
1638 		return 0;
1639 	nmd->lasterr = 0;
1640 	nmd->nm_totalsize = 0;
1641 	for (i = 0; i < NETMAP_POOLS_NR; i++) {
1642 		nmd->lasterr = netmap_finalize_obj_allocator(&nmd->pools[i]);
1643 		if (nmd->lasterr)
1644 			goto error;
1645 		nmd->nm_totalsize += nmd->pools[i].memtotal;
1646 	}
1647 	nmd->nm_totalsize = (nmd->nm_totalsize + PAGE_SIZE - 1) & ~(PAGE_SIZE - 1);
1648 	nmd->lasterr = netmap_mem_init_bitmaps(nmd);
1649 	if (nmd->lasterr)
1650 		goto error;
1651 
1652 	nmd->flags |= NETMAP_MEM_FINALIZED;
1653 
1654 	if (netmap_verbose)
1655 		nm_prinf("interfaces %zd KB, rings %zd KB, buffers %zd MB",
1656 		    nmd->pools[NETMAP_IF_POOL].memtotal >> 10,
1657 		    nmd->pools[NETMAP_RING_POOL].memtotal >> 10,
1658 		    nmd->pools[NETMAP_BUF_POOL].memtotal >> 20);
1659 
1660 	if (netmap_verbose)
1661 		nm_prinf("Free buffers: %d", nmd->pools[NETMAP_BUF_POOL].objfree);
1662 
1663 
1664 	return 0;
1665 error:
1666 	netmap_mem_reset_all(nmd);
1667 	return nmd->lasterr;
1668 }
1669 
1670 /*
1671  * allocator for private memory
1672  */
1673 static void *
1674 _netmap_mem_private_new(size_t size, struct netmap_obj_params *p, int grp_id,
1675 		struct netmap_mem_ops *ops, uint64_t memtotal, int *perr)
1676 {
1677 	struct netmap_mem_d *d = NULL;
1678 	int i, err = 0;
1679 	int checksz = 0;
1680 
1681 	/* if memtotal is !=0 we check that the request fits the available
1682 	 * memory. Moreover, any surprlus memory is assigned to buffers.
1683 	 */
1684 	checksz = (memtotal > 0);
1685 
1686 	d = nm_os_malloc(size);
1687 	if (d == NULL) {
1688 		err = ENOMEM;
1689 		goto error;
1690 	}
1691 
1692 	*d = nm_blueprint;
1693 	d->ops = ops;
1694 
1695 	err = nm_mem_assign_id(d, grp_id);
1696 	if (err)
1697 		goto error_free;
1698 	snprintf(d->name, NM_MEM_NAMESZ, "%d", d->nm_id);
1699 
1700 	for (i = 0; i < NETMAP_POOLS_NR; i++) {
1701 		snprintf(d->pools[i].name, NETMAP_POOL_MAX_NAMSZ,
1702 				nm_blueprint.pools[i].name,
1703 				d->name);
1704 		if (checksz) {
1705 			uint64_t poolsz = (uint64_t)p[i].num * p[i].size;
1706 			if (memtotal < poolsz) {
1707 				nm_prerr("%s: request too large", d->pools[i].name);
1708 				err = ENOMEM;
1709 				goto error_rel_id;
1710 			}
1711 			memtotal -= poolsz;
1712 		}
1713 		d->params[i].num = p[i].num;
1714 		d->params[i].size = p[i].size;
1715 	}
1716 	if (checksz && memtotal > 0) {
1717 		uint64_t sz = d->params[NETMAP_BUF_POOL].size;
1718 		uint64_t n = (memtotal + sz - 1) / sz;
1719 
1720 		if (n) {
1721 			if (netmap_verbose) {
1722 				nm_prinf("%s: adding %llu more buffers",
1723 				    d->pools[NETMAP_BUF_POOL].name,
1724 				    (unsigned long long)n);
1725 			}
1726 			d->params[NETMAP_BUF_POOL].num += n;
1727 		}
1728 	}
1729 
1730 	NMA_LOCK_INIT(d);
1731 
1732 	err = netmap_mem_config(d);
1733 	if (err)
1734 		goto error_destroy_lock;
1735 
1736 	d->flags &= ~NETMAP_MEM_FINALIZED;
1737 
1738 	return d;
1739 
1740 error_destroy_lock:
1741 	NMA_LOCK_DESTROY(d);
1742 error_rel_id:
1743 	nm_mem_release_id(d);
1744 error_free:
1745 	nm_os_free(d);
1746 error:
1747 	if (perr)
1748 		*perr = err;
1749 	return NULL;
1750 }
1751 
1752 struct netmap_mem_d *
1753 netmap_mem_private_new(u_int txr, u_int txd, u_int rxr, u_int rxd,
1754 		u_int extra_bufs, u_int npipes, int *perr)
1755 {
1756 	struct netmap_mem_d *d = NULL;
1757 	struct netmap_obj_params p[NETMAP_POOLS_NR];
1758 	int i;
1759 	u_int v, maxd;
1760 	/* account for the fake host rings */
1761 	txr++;
1762 	rxr++;
1763 
1764 	/* copy the min values */
1765 	for (i = 0; i < NETMAP_POOLS_NR; i++) {
1766 		p[i] = netmap_min_priv_params[i];
1767 	}
1768 
1769 	/* possibly increase them to fit user request */
1770 	v = sizeof(struct netmap_if) + sizeof(ssize_t) * (txr + rxr);
1771 	if (p[NETMAP_IF_POOL].size < v)
1772 		p[NETMAP_IF_POOL].size = v;
1773 	v = 2 + 4 * npipes;
1774 	if (p[NETMAP_IF_POOL].num < v)
1775 		p[NETMAP_IF_POOL].num = v;
1776 	maxd = (txd > rxd) ? txd : rxd;
1777 	v = sizeof(struct netmap_ring) + sizeof(struct netmap_slot) * maxd;
1778 	if (p[NETMAP_RING_POOL].size < v)
1779 		p[NETMAP_RING_POOL].size = v;
1780 	/* each pipe endpoint needs two tx rings (1 normal + 1 host, fake)
1781 	 * and two rx rings (again, 1 normal and 1 fake host)
1782 	 */
1783 	v = txr + rxr + 8 * npipes;
1784 	if (p[NETMAP_RING_POOL].num < v)
1785 		p[NETMAP_RING_POOL].num = v;
1786 	/* for each pipe we only need the buffers for the 4 "real" rings.
1787 	 * On the other end, the pipe ring dimension may be different from
1788 	 * the parent port ring dimension. As a compromise, we allocate twice the
1789 	 * space actually needed if the pipe rings were the same size as the parent rings
1790 	 */
1791 	v = (4 * npipes + rxr) * rxd + (4 * npipes + txr) * txd + 2 + extra_bufs;
1792 		/* the +2 is for the tx and rx fake buffers (indices 0 and 1) */
1793 	if (p[NETMAP_BUF_POOL].num < v)
1794 		p[NETMAP_BUF_POOL].num = v;
1795 
1796 	if (netmap_verbose)
1797 		nm_prinf("req if %d*%d ring %d*%d buf %d*%d",
1798 			p[NETMAP_IF_POOL].num,
1799 			p[NETMAP_IF_POOL].size,
1800 			p[NETMAP_RING_POOL].num,
1801 			p[NETMAP_RING_POOL].size,
1802 			p[NETMAP_BUF_POOL].num,
1803 			p[NETMAP_BUF_POOL].size);
1804 
1805 	d = _netmap_mem_private_new(sizeof(*d), p, -1, &netmap_mem_global_ops, 0, perr);
1806 
1807 	return d;
1808 }
1809 
1810 /* Reference iommu allocator - find existing or create new,
1811  * for not hw addapeters fallback to global allocator.
1812  */
1813 struct netmap_mem_d *
1814 netmap_mem_get_iommu(struct netmap_adapter *na)
1815 {
1816 	int i, err, grp_id;
1817 	struct netmap_mem_d *nmd;
1818 
1819 	if (na == NULL || na->pdev == NULL)
1820 		return netmap_mem_get(&nm_mem);
1821 
1822 	grp_id = nm_iommu_group_id(na->pdev);
1823 
1824 	NM_MTX_LOCK(nm_mem_list_lock);
1825 	nmd = netmap_last_mem_d;
1826 	do {
1827 		if (!(nmd->flags & NETMAP_MEM_HIDDEN) && nmd->nm_grp == grp_id) {
1828 			nmd->refcount++;
1829 			NM_DBG_REFC(nmd, __FUNCTION__, __LINE__);
1830 			NM_MTX_UNLOCK(nm_mem_list_lock);
1831 			return nmd;
1832 		}
1833 		nmd = nmd->next;
1834 	} while (nmd != netmap_last_mem_d);
1835 
1836 	nmd = nm_os_malloc(sizeof(*nmd));
1837 	if (nmd == NULL)
1838 		goto error;
1839 
1840 	*nmd = nm_mem_blueprint;
1841 
1842 	err = nm_mem_assign_id_locked(nmd, grp_id);
1843 	if (err)
1844 		goto error_free;
1845 
1846 	snprintf(nmd->name, sizeof(nmd->name), "%d", nmd->nm_id);
1847 
1848 	for (i = 0; i < NETMAP_POOLS_NR; i++) {
1849 		snprintf(nmd->pools[i].name, NETMAP_POOL_MAX_NAMSZ, "%s-%s",
1850 			nm_mem_blueprint.pools[i].name, nmd->name);
1851 	}
1852 
1853 	NMA_LOCK_INIT(nmd);
1854 
1855 	NM_MTX_UNLOCK(nm_mem_list_lock);
1856 	return nmd;
1857 
1858 error_free:
1859 	nm_os_free(nmd);
1860 error:
1861 	NM_MTX_UNLOCK(nm_mem_list_lock);
1862 	return NULL;
1863 }
1864 
1865 /* call with lock held */
1866 static int
1867 netmap_mem2_config(struct netmap_mem_d *nmd)
1868 {
1869 	int i;
1870 
1871 	if (!netmap_mem_params_changed(nmd->params))
1872 		goto out;
1873 
1874 	nm_prdis("reconfiguring");
1875 
1876 	if (nmd->flags & NETMAP_MEM_FINALIZED) {
1877 		/* reset previous allocation */
1878 		for (i = 0; i < NETMAP_POOLS_NR; i++) {
1879 			netmap_reset_obj_allocator(&nmd->pools[i]);
1880 		}
1881 		nmd->flags &= ~NETMAP_MEM_FINALIZED;
1882 	}
1883 
1884 	for (i = 0; i < NETMAP_POOLS_NR; i++) {
1885 		nmd->lasterr = netmap_config_obj_allocator(&nmd->pools[i],
1886 				nmd->params[i].num, nmd->params[i].size);
1887 		if (nmd->lasterr)
1888 			goto out;
1889 	}
1890 
1891 out:
1892 
1893 	return nmd->lasterr;
1894 }
1895 
1896 static int
1897 netmap_mem2_finalize(struct netmap_mem_d *nmd, struct netmap_adapter *na)
1898 {
1899 	if (nmd->flags & NETMAP_MEM_FINALIZED)
1900 		goto out;
1901 
1902 	if (netmap_mem_finalize_all(nmd))
1903 		goto out;
1904 
1905 	nmd->lasterr = 0;
1906 
1907 out:
1908 	return nmd->lasterr;
1909 }
1910 
1911 static void
1912 netmap_mem2_delete(struct netmap_mem_d *nmd)
1913 {
1914 	int i;
1915 
1916 	for (i = 0; i < NETMAP_POOLS_NR; i++) {
1917 	    netmap_destroy_obj_allocator(&nmd->pools[i]);
1918 	}
1919 
1920 	NMA_LOCK_DESTROY(nmd);
1921 	if (nmd != &nm_mem)
1922 		nm_os_free(nmd);
1923 }
1924 
1925 #ifdef WITH_EXTMEM
1926 /* doubly linekd list of all existing external allocators */
1927 static struct netmap_mem_ext *netmap_mem_ext_list = NULL;
1928 NM_MTX_T nm_mem_ext_list_lock;
1929 #endif /* WITH_EXTMEM */
1930 
1931 int
1932 netmap_mem_init(void)
1933 {
1934 	nm_mem_blueprint = nm_mem;
1935 	NM_MTX_INIT(nm_mem_list_lock);
1936 	NMA_LOCK_INIT(&nm_mem);
1937 	netmap_mem_get(&nm_mem);
1938 #ifdef WITH_EXTMEM
1939 	NM_MTX_INIT(nm_mem_ext_list_lock);
1940 #endif /* WITH_EXTMEM */
1941 	return (0);
1942 }
1943 
1944 void
1945 netmap_mem_fini(void)
1946 {
1947 	netmap_mem_put(&nm_mem);
1948 }
1949 
1950 static int
1951 netmap_mem_ring_needed(struct netmap_kring *kring)
1952 {
1953 	return kring->ring == NULL &&
1954 		(kring->users > 0 ||
1955 		 (kring->nr_kflags & NKR_NEEDRING));
1956 }
1957 
1958 static int
1959 netmap_mem_ring_todelete(struct netmap_kring *kring)
1960 {
1961 	return kring->ring != NULL &&
1962 		kring->users == 0 &&
1963 		!(kring->nr_kflags & NKR_NEEDRING);
1964 }
1965 
1966 
1967 /* call with NMA_LOCK held *
1968  *
1969  * Allocate netmap rings and buffers for this card
1970  * The rings are contiguous, but have variable size.
1971  * The kring array must follow the layout described
1972  * in netmap_krings_create().
1973  */
1974 static int
1975 netmap_mem2_rings_create(struct netmap_mem_d *nmd, struct netmap_adapter *na)
1976 {
1977 	enum txrx t;
1978 
1979 	for_rx_tx(t) {
1980 		u_int i;
1981 
1982 		for (i = 0; i < netmap_all_rings(na, t); i++) {
1983 			struct netmap_kring *kring = NMR(na, t)[i];
1984 			struct netmap_ring *ring = kring->ring;
1985 			u_int len, ndesc;
1986 
1987 			if (!netmap_mem_ring_needed(kring)) {
1988 				/* unneeded, or already created by somebody else */
1989 				if (netmap_debug & NM_DEBUG_MEM)
1990 					nm_prinf("NOT creating ring %s (ring %p, users %d neekring %d)",
1991 						kring->name, ring, kring->users, kring->nr_kflags & NKR_NEEDRING);
1992 				continue;
1993 			}
1994 			if (netmap_debug & NM_DEBUG_MEM)
1995 				nm_prinf("creating %s", kring->name);
1996 			ndesc = kring->nkr_num_slots;
1997 			len = sizeof(struct netmap_ring) +
1998 				  ndesc * sizeof(struct netmap_slot);
1999 			ring = netmap_ring_malloc(nmd, len);
2000 			if (ring == NULL) {
2001 				nm_prerr("Cannot allocate %s_ring", nm_txrx2str(t));
2002 				goto cleanup;
2003 			}
2004 			nm_prdis("txring at %p", ring);
2005 			kring->ring = ring;
2006 			*(uint32_t *)(uintptr_t)&ring->num_slots = ndesc;
2007 			*(int64_t *)(uintptr_t)&ring->buf_ofs =
2008 			    (nmd->pools[NETMAP_IF_POOL].memtotal +
2009 				nmd->pools[NETMAP_RING_POOL].memtotal) -
2010 				netmap_ring_offset(nmd, ring);
2011 
2012 			/* copy values from kring */
2013 			ring->head = kring->rhead;
2014 			ring->cur = kring->rcur;
2015 			ring->tail = kring->rtail;
2016 			*(uint32_t *)(uintptr_t)&ring->nr_buf_size =
2017 				netmap_mem_bufsize(nmd);
2018 			nm_prdis("%s h %d c %d t %d", kring->name,
2019 				ring->head, ring->cur, ring->tail);
2020 			nm_prdis("initializing slots for %s_ring", nm_txrx2str(t));
2021 			if (!(kring->nr_kflags & NKR_FAKERING)) {
2022 				/* this is a real ring */
2023 				if (netmap_debug & NM_DEBUG_MEM)
2024 					nm_prinf("allocating buffers for %s", kring->name);
2025 				if (netmap_new_bufs(nmd, ring->slot, ndesc)) {
2026 					nm_prerr("Cannot allocate buffers for %s_ring", nm_txrx2str(t));
2027 					goto cleanup;
2028 				}
2029 			} else {
2030 				/* this is a fake ring, set all indices to 0 */
2031 				if (netmap_debug & NM_DEBUG_MEM)
2032 					nm_prinf("NOT allocating buffers for %s", kring->name);
2033 				netmap_mem_set_ring(nmd, ring->slot, ndesc, 0);
2034 			}
2035 		        /* ring info */
2036 		        *(uint16_t *)(uintptr_t)&ring->ringid = kring->ring_id;
2037 		        *(uint16_t *)(uintptr_t)&ring->dir = kring->tx;
2038 		}
2039 	}
2040 
2041 	return 0;
2042 
2043 cleanup:
2044 	/* we cannot actually cleanup here, since we don't own kring->users
2045 	 * and kring->nr_klags & NKR_NEEDRING. The caller must decrement
2046 	 * the first or zero-out the second, then call netmap_free_rings()
2047 	 * to do the cleanup
2048 	 */
2049 
2050 	return ENOMEM;
2051 }
2052 
2053 static void
2054 netmap_mem2_rings_delete(struct netmap_mem_d *nmd, struct netmap_adapter *na)
2055 {
2056 	enum txrx t;
2057 
2058 	for_rx_tx(t) {
2059 		u_int i;
2060 		for (i = 0; i < netmap_all_rings(na, t); i++) {
2061 			struct netmap_kring *kring = NMR(na, t)[i];
2062 			struct netmap_ring *ring = kring->ring;
2063 
2064 			if (!netmap_mem_ring_todelete(kring)) {
2065 				if (netmap_debug & NM_DEBUG_MEM)
2066 					nm_prinf("NOT deleting ring %s (ring %p, users %d neekring %d)",
2067 						kring->name, ring, kring->users, kring->nr_kflags & NKR_NEEDRING);
2068 				continue;
2069 			}
2070 			if (netmap_debug & NM_DEBUG_MEM)
2071 				nm_prinf("deleting ring %s", kring->name);
2072 			if (!(kring->nr_kflags & NKR_FAKERING)) {
2073 				nm_prdis("freeing bufs for %s", kring->name);
2074 				netmap_free_bufs(nmd, ring->slot, kring->nkr_num_slots);
2075 			} else {
2076 				nm_prdis("NOT freeing bufs for %s", kring->name);
2077 			}
2078 			netmap_ring_free(nmd, ring);
2079 			kring->ring = NULL;
2080 		}
2081 	}
2082 }
2083 
2084 /* call with NMA_LOCK held */
2085 /*
2086  * Allocate the per-fd structure netmap_if.
2087  *
2088  * We assume that the configuration stored in na
2089  * (number of tx/rx rings and descs) does not change while
2090  * the interface is in netmap mode.
2091  */
2092 static struct netmap_if *
2093 netmap_mem2_if_new(struct netmap_mem_d *nmd,
2094 		struct netmap_adapter *na, struct netmap_priv_d *priv)
2095 {
2096 	struct netmap_if *nifp;
2097 	ssize_t base; /* handy for relative offsets between rings and nifp */
2098 	u_int i, len, n[NR_TXRX], ntot;
2099 	enum txrx t;
2100 
2101 	ntot = 0;
2102 	for_rx_tx(t) {
2103 		/* account for the (eventually fake) host rings */
2104 		n[t] = netmap_all_rings(na, t);
2105 		ntot += n[t];
2106 	}
2107 	/*
2108 	 * the descriptor is followed inline by an array of offsets
2109 	 * to the tx and rx rings in the shared memory region.
2110 	 */
2111 
2112 	len = sizeof(struct netmap_if) + (ntot * sizeof(ssize_t));
2113 	nifp = netmap_if_malloc(nmd, len);
2114 	if (nifp == NULL) {
2115 		return NULL;
2116 	}
2117 
2118 	/* initialize base fields -- override const */
2119 	*(u_int *)(uintptr_t)&nifp->ni_tx_rings = na->num_tx_rings;
2120 	*(u_int *)(uintptr_t)&nifp->ni_rx_rings = na->num_rx_rings;
2121 	*(u_int *)(uintptr_t)&nifp->ni_host_tx_rings =
2122 		(na->num_host_tx_rings ? na->num_host_tx_rings : 1);
2123 	*(u_int *)(uintptr_t)&nifp->ni_host_rx_rings =
2124 		(na->num_host_rx_rings ? na->num_host_rx_rings : 1);
2125 	strlcpy(nifp->ni_name, na->name, sizeof(nifp->ni_name));
2126 
2127 	/*
2128 	 * fill the slots for the rx and tx rings. They contain the offset
2129 	 * between the ring and nifp, so the information is usable in
2130 	 * userspace to reach the ring from the nifp.
2131 	 */
2132 	base = netmap_if_offset(nmd, nifp);
2133 	for (i = 0; i < n[NR_TX]; i++) {
2134 		/* XXX instead of ofs == 0 maybe use the offset of an error
2135 		 * ring, like we do for buffers? */
2136 		ssize_t ofs = 0;
2137 
2138 		if (na->tx_rings[i]->ring != NULL && i >= priv->np_qfirst[NR_TX]
2139 				&& i < priv->np_qlast[NR_TX]) {
2140 			ofs = netmap_ring_offset(nmd,
2141 						 na->tx_rings[i]->ring) - base;
2142 		}
2143 		*(ssize_t *)(uintptr_t)&nifp->ring_ofs[i] = ofs;
2144 	}
2145 	for (i = 0; i < n[NR_RX]; i++) {
2146 		/* XXX instead of ofs == 0 maybe use the offset of an error
2147 		 * ring, like we do for buffers? */
2148 		ssize_t ofs = 0;
2149 
2150 		if (na->rx_rings[i]->ring != NULL && i >= priv->np_qfirst[NR_RX]
2151 				&& i < priv->np_qlast[NR_RX]) {
2152 			ofs = netmap_ring_offset(nmd,
2153 						 na->rx_rings[i]->ring) - base;
2154 		}
2155 		*(ssize_t *)(uintptr_t)&nifp->ring_ofs[i+n[NR_TX]] = ofs;
2156 	}
2157 
2158 	return (nifp);
2159 }
2160 
2161 static void
2162 netmap_mem2_if_delete(struct netmap_mem_d *nmd,
2163 		struct netmap_adapter *na, struct netmap_if *nifp)
2164 {
2165 	if (nifp == NULL)
2166 		/* nothing to do */
2167 		return;
2168 	if (nifp->ni_bufs_head)
2169 		netmap_extra_free(na, nifp->ni_bufs_head);
2170 	netmap_if_free(nmd, nifp);
2171 }
2172 
2173 static void
2174 netmap_mem2_deref(struct netmap_mem_d *nmd, struct netmap_adapter *na)
2175 {
2176 
2177 	if (netmap_debug & NM_DEBUG_MEM)
2178 		nm_prinf("active = %d", nmd->active);
2179 
2180 }
2181 
2182 struct netmap_mem_ops netmap_mem_global_ops = {
2183 	.nmd_get_lut = netmap_mem2_get_lut,
2184 	.nmd_get_info = netmap_mem2_get_info,
2185 	.nmd_ofstophys = netmap_mem2_ofstophys,
2186 	.nmd_config = netmap_mem2_config,
2187 	.nmd_finalize = netmap_mem2_finalize,
2188 	.nmd_deref = netmap_mem2_deref,
2189 	.nmd_delete = netmap_mem2_delete,
2190 	.nmd_if_offset = netmap_mem2_if_offset,
2191 	.nmd_if_new = netmap_mem2_if_new,
2192 	.nmd_if_delete = netmap_mem2_if_delete,
2193 	.nmd_rings_create = netmap_mem2_rings_create,
2194 	.nmd_rings_delete = netmap_mem2_rings_delete
2195 };
2196 
2197 int
2198 netmap_mem_pools_info_get(struct nmreq_pools_info *req,
2199 				struct netmap_mem_d *nmd)
2200 {
2201 	int ret;
2202 
2203 	ret = netmap_mem_get_info(nmd, &req->nr_memsize, NULL,
2204 					&req->nr_mem_id);
2205 	if (ret) {
2206 		return ret;
2207 	}
2208 
2209 	NMA_LOCK(nmd);
2210 	req->nr_if_pool_offset = 0;
2211 	req->nr_if_pool_objtotal = nmd->pools[NETMAP_IF_POOL].objtotal;
2212 	req->nr_if_pool_objsize = nmd->pools[NETMAP_IF_POOL]._objsize;
2213 
2214 	req->nr_ring_pool_offset = nmd->pools[NETMAP_IF_POOL].memtotal;
2215 	req->nr_ring_pool_objtotal = nmd->pools[NETMAP_RING_POOL].objtotal;
2216 	req->nr_ring_pool_objsize = nmd->pools[NETMAP_RING_POOL]._objsize;
2217 
2218 	req->nr_buf_pool_offset = nmd->pools[NETMAP_IF_POOL].memtotal +
2219 			     nmd->pools[NETMAP_RING_POOL].memtotal;
2220 	req->nr_buf_pool_objtotal = nmd->pools[NETMAP_BUF_POOL].objtotal;
2221 	req->nr_buf_pool_objsize = nmd->pools[NETMAP_BUF_POOL]._objsize;
2222 	NMA_UNLOCK(nmd);
2223 
2224 	return 0;
2225 }
2226 
2227 #ifdef WITH_EXTMEM
2228 struct netmap_mem_ext {
2229 	struct netmap_mem_d up;
2230 
2231 	struct nm_os_extmem *os;
2232 	struct netmap_mem_ext *next, *prev;
2233 };
2234 
2235 /* call with nm_mem_list_lock held */
2236 static void
2237 netmap_mem_ext_register(struct netmap_mem_ext *e)
2238 {
2239 	NM_MTX_LOCK(nm_mem_ext_list_lock);
2240 	if (netmap_mem_ext_list)
2241 		netmap_mem_ext_list->prev = e;
2242 	e->next = netmap_mem_ext_list;
2243 	netmap_mem_ext_list = e;
2244 	e->prev = NULL;
2245 	NM_MTX_UNLOCK(nm_mem_ext_list_lock);
2246 }
2247 
2248 /* call with nm_mem_list_lock held */
2249 static void
2250 netmap_mem_ext_unregister(struct netmap_mem_ext *e)
2251 {
2252 	if (e->prev)
2253 		e->prev->next = e->next;
2254 	else
2255 		netmap_mem_ext_list = e->next;
2256 	if (e->next)
2257 		e->next->prev = e->prev;
2258 	e->prev = e->next = NULL;
2259 }
2260 
2261 static struct netmap_mem_ext *
2262 netmap_mem_ext_search(struct nm_os_extmem *os)
2263 {
2264 	struct netmap_mem_ext *e;
2265 
2266 	NM_MTX_LOCK(nm_mem_ext_list_lock);
2267 	for (e = netmap_mem_ext_list; e; e = e->next) {
2268 		if (nm_os_extmem_isequal(e->os, os)) {
2269 			netmap_mem_get(&e->up);
2270 			break;
2271 		}
2272 	}
2273 	NM_MTX_UNLOCK(nm_mem_ext_list_lock);
2274 	return e;
2275 }
2276 
2277 
2278 static void
2279 netmap_mem_ext_delete(struct netmap_mem_d *d)
2280 {
2281 	int i;
2282 	struct netmap_mem_ext *e =
2283 		(struct netmap_mem_ext *)d;
2284 
2285 	netmap_mem_ext_unregister(e);
2286 
2287 	for (i = 0; i < NETMAP_POOLS_NR; i++) {
2288 		struct netmap_obj_pool *p = &d->pools[i];
2289 
2290 		if (p->lut) {
2291 			nm_free_lut(p->lut, p->objtotal);
2292 			p->lut = NULL;
2293 		}
2294 	}
2295 	if (e->os)
2296 		nm_os_extmem_delete(e->os);
2297 	netmap_mem2_delete(d);
2298 }
2299 
2300 static int
2301 netmap_mem_ext_config(struct netmap_mem_d *nmd)
2302 {
2303 	return 0;
2304 }
2305 
2306 struct netmap_mem_ops netmap_mem_ext_ops = {
2307 	.nmd_get_lut = netmap_mem2_get_lut,
2308 	.nmd_get_info = netmap_mem2_get_info,
2309 	.nmd_ofstophys = netmap_mem2_ofstophys,
2310 	.nmd_config = netmap_mem_ext_config,
2311 	.nmd_finalize = netmap_mem2_finalize,
2312 	.nmd_deref = netmap_mem2_deref,
2313 	.nmd_delete = netmap_mem_ext_delete,
2314 	.nmd_if_offset = netmap_mem2_if_offset,
2315 	.nmd_if_new = netmap_mem2_if_new,
2316 	.nmd_if_delete = netmap_mem2_if_delete,
2317 	.nmd_rings_create = netmap_mem2_rings_create,
2318 	.nmd_rings_delete = netmap_mem2_rings_delete
2319 };
2320 
2321 struct netmap_mem_d *
2322 netmap_mem_ext_create(uint64_t usrptr, struct nmreq_pools_info *pi, int *perror)
2323 {
2324 	int error = 0;
2325 	int i, j;
2326 	struct netmap_mem_ext *nme;
2327 	char *clust;
2328 	size_t off;
2329 	struct nm_os_extmem *os = NULL;
2330 	int nr_pages;
2331 
2332 	// XXX sanity checks
2333 	if (pi->nr_if_pool_objtotal == 0)
2334 		pi->nr_if_pool_objtotal = netmap_min_priv_params[NETMAP_IF_POOL].num;
2335 	if (pi->nr_if_pool_objsize == 0)
2336 		pi->nr_if_pool_objsize = netmap_min_priv_params[NETMAP_IF_POOL].size;
2337 	if (pi->nr_ring_pool_objtotal == 0)
2338 		pi->nr_ring_pool_objtotal = netmap_min_priv_params[NETMAP_RING_POOL].num;
2339 	if (pi->nr_ring_pool_objsize == 0)
2340 		pi->nr_ring_pool_objsize = netmap_min_priv_params[NETMAP_RING_POOL].size;
2341 	if (pi->nr_buf_pool_objtotal == 0)
2342 		pi->nr_buf_pool_objtotal = netmap_min_priv_params[NETMAP_BUF_POOL].num;
2343 	if (pi->nr_buf_pool_objsize == 0)
2344 		pi->nr_buf_pool_objsize = netmap_min_priv_params[NETMAP_BUF_POOL].size;
2345 	if (netmap_verbose & NM_DEBUG_MEM)
2346 		nm_prinf("if %d %d ring %d %d buf %d %d",
2347 			pi->nr_if_pool_objtotal, pi->nr_if_pool_objsize,
2348 			pi->nr_ring_pool_objtotal, pi->nr_ring_pool_objsize,
2349 			pi->nr_buf_pool_objtotal, pi->nr_buf_pool_objsize);
2350 
2351 	os = nm_os_extmem_create(usrptr, pi, &error);
2352 	if (os == NULL) {
2353 		nm_prerr("os extmem creation failed");
2354 		goto out;
2355 	}
2356 
2357 	nme = netmap_mem_ext_search(os);
2358 	if (nme) {
2359 		nm_os_extmem_delete(os);
2360 		return &nme->up;
2361 	}
2362 	if (netmap_verbose & NM_DEBUG_MEM)
2363 		nm_prinf("not found, creating new");
2364 
2365 	nme = _netmap_mem_private_new(sizeof(*nme),
2366 
2367 			(struct netmap_obj_params[]){
2368 				{ pi->nr_if_pool_objsize, pi->nr_if_pool_objtotal },
2369 				{ pi->nr_ring_pool_objsize, pi->nr_ring_pool_objtotal },
2370 				{ pi->nr_buf_pool_objsize, pi->nr_buf_pool_objtotal }},
2371 			-1,
2372 			&netmap_mem_ext_ops,
2373 			pi->nr_memsize,
2374 			&error);
2375 	if (nme == NULL)
2376 		goto out_unmap;
2377 
2378 	nr_pages = nm_os_extmem_nr_pages(os);
2379 
2380 	/* from now on pages will be released by nme destructor;
2381 	 * we let res = 0 to prevent release in out_unmap below
2382 	 */
2383 	nme->os = os;
2384 	os = NULL; /* pass ownership */
2385 
2386 	clust = nm_os_extmem_nextpage(nme->os);
2387 	off = 0;
2388 	for (i = 0; i < NETMAP_POOLS_NR; i++) {
2389 		struct netmap_obj_pool *p = &nme->up.pools[i];
2390 		struct netmap_obj_params *o = &nme->up.params[i];
2391 
2392 		p->_objsize = o->size;
2393 		p->_clustsize = o->size;
2394 		p->_clustentries = 1;
2395 
2396 		p->lut = nm_alloc_lut(o->num);
2397 		if (p->lut == NULL) {
2398 			error = ENOMEM;
2399 			goto out_delete;
2400 		}
2401 
2402 		p->bitmap_slots = (o->num + sizeof(uint32_t) - 1) / sizeof(uint32_t);
2403 		p->invalid_bitmap = nm_os_malloc(sizeof(uint32_t) * p->bitmap_slots);
2404 		if (p->invalid_bitmap == NULL) {
2405 			error = ENOMEM;
2406 			goto out_delete;
2407 		}
2408 
2409 		if (nr_pages == 0) {
2410 			p->objtotal = 0;
2411 			p->memtotal = 0;
2412 			p->objfree = 0;
2413 			continue;
2414 		}
2415 
2416 		for (j = 0; j < o->num && nr_pages > 0; j++) {
2417 			size_t noff;
2418 
2419 			p->lut[j].vaddr = clust + off;
2420 #if !defined(linux) && !defined(_WIN32)
2421 			p->lut[j].paddr = vtophys(p->lut[j].vaddr);
2422 #endif
2423 			nm_prdis("%s %d at %p", p->name, j, p->lut[j].vaddr);
2424 			noff = off + p->_objsize;
2425 			if (noff < PAGE_SIZE) {
2426 				off = noff;
2427 				continue;
2428 			}
2429 			nm_prdis("too big, recomputing offset...");
2430 			while (noff >= PAGE_SIZE) {
2431 				char *old_clust = clust;
2432 				noff -= PAGE_SIZE;
2433 				clust = nm_os_extmem_nextpage(nme->os);
2434 				nr_pages--;
2435 				nm_prdis("noff %zu page %p nr_pages %d", noff,
2436 						page_to_virt(*pages), nr_pages);
2437 				if (noff > 0 && !nm_isset(p->invalid_bitmap, j) &&
2438 					(nr_pages == 0 ||
2439 					 old_clust + PAGE_SIZE != clust))
2440 				{
2441 					/* out of space or non contiguous,
2442 					 * drop this object
2443 					 * */
2444 					p->invalid_bitmap[ (j>>5) ] |= 1U << (j & 31U);
2445 					nm_prdis("non contiguous at off %zu, drop", noff);
2446 				}
2447 				if (nr_pages == 0)
2448 					break;
2449 			}
2450 			off = noff;
2451 		}
2452 		p->objtotal = j;
2453 		p->numclusters = p->objtotal;
2454 		p->memtotal = j * (size_t)p->_objsize;
2455 		nm_prdis("%d memtotal %zu", j, p->memtotal);
2456 	}
2457 
2458 	netmap_mem_ext_register(nme);
2459 
2460 	return &nme->up;
2461 
2462 out_delete:
2463 	netmap_mem_put(&nme->up);
2464 out_unmap:
2465 	if (os)
2466 		nm_os_extmem_delete(os);
2467 out:
2468 	if (perror)
2469 		*perror = error;
2470 	return NULL;
2471 
2472 }
2473 #endif /* WITH_EXTMEM */
2474 
2475 
2476 #ifdef WITH_PTNETMAP
2477 struct mem_pt_if {
2478 	struct mem_pt_if *next;
2479 	if_t ifp;
2480 	unsigned int nifp_offset;
2481 };
2482 
2483 /* Netmap allocator for ptnetmap guests. */
2484 struct netmap_mem_ptg {
2485 	struct netmap_mem_d up;
2486 
2487 	vm_paddr_t nm_paddr;            /* physical address in the guest */
2488 	void *nm_addr;                  /* virtual address in the guest */
2489 	struct netmap_lut buf_lut;      /* lookup table for BUF pool in the guest */
2490 	nm_memid_t host_mem_id;         /* allocator identifier in the host */
2491 	struct ptnetmap_memdev *ptn_dev;/* ptnetmap memdev */
2492 	struct mem_pt_if *pt_ifs;	/* list of interfaces in passthrough */
2493 };
2494 
2495 /* Link a passthrough interface to a passthrough netmap allocator. */
2496 static int
2497 netmap_mem_pt_guest_ifp_add(struct netmap_mem_d *nmd, if_t ifp,
2498 			    unsigned int nifp_offset)
2499 {
2500 	struct netmap_mem_ptg *ptnmd = (struct netmap_mem_ptg *)nmd;
2501 	struct mem_pt_if *ptif = nm_os_malloc(sizeof(*ptif));
2502 
2503 	if (!ptif) {
2504 		return ENOMEM;
2505 	}
2506 
2507 	NMA_LOCK(nmd);
2508 
2509 	ptif->ifp = ifp;
2510 	ptif->nifp_offset = nifp_offset;
2511 
2512 	if (ptnmd->pt_ifs) {
2513 		ptif->next = ptnmd->pt_ifs;
2514 	}
2515 	ptnmd->pt_ifs = ptif;
2516 
2517 	NMA_UNLOCK(nmd);
2518 
2519 	nm_prinf("ifp=%s,nifp_offset=%u",
2520 		if_name(ptif->ifp), ptif->nifp_offset);
2521 
2522 	return 0;
2523 }
2524 
2525 /* Called with NMA_LOCK(nmd) held. */
2526 static struct mem_pt_if *
2527 netmap_mem_pt_guest_ifp_lookup(struct netmap_mem_d *nmd, if_t ifp)
2528 {
2529 	struct netmap_mem_ptg *ptnmd = (struct netmap_mem_ptg *)nmd;
2530 	struct mem_pt_if *curr;
2531 
2532 	for (curr = ptnmd->pt_ifs; curr; curr = curr->next) {
2533 		if (curr->ifp == ifp) {
2534 			return curr;
2535 		}
2536 	}
2537 
2538 	return NULL;
2539 }
2540 
2541 /* Unlink a passthrough interface from a passthrough netmap allocator. */
2542 int
2543 netmap_mem_pt_guest_ifp_del(struct netmap_mem_d *nmd, if_t ifp)
2544 {
2545 	struct netmap_mem_ptg *ptnmd = (struct netmap_mem_ptg *)nmd;
2546 	struct mem_pt_if *prev = NULL;
2547 	struct mem_pt_if *curr;
2548 	int ret = -1;
2549 
2550 	NMA_LOCK(nmd);
2551 
2552 	for (curr = ptnmd->pt_ifs; curr; curr = curr->next) {
2553 		if (curr->ifp == ifp) {
2554 			if (prev) {
2555 				prev->next = curr->next;
2556 			} else {
2557 				ptnmd->pt_ifs = curr->next;
2558 			}
2559 			nm_prinf("removed (ifp=%s,nifp_offset=%u)",
2560 			  if_name(curr->ifp), curr->nifp_offset);
2561 			nm_os_free(curr);
2562 			ret = 0;
2563 			break;
2564 		}
2565 		prev = curr;
2566 	}
2567 
2568 	NMA_UNLOCK(nmd);
2569 
2570 	return ret;
2571 }
2572 
2573 static int
2574 netmap_mem_pt_guest_get_lut(struct netmap_mem_d *nmd, struct netmap_lut *lut)
2575 {
2576 	struct netmap_mem_ptg *ptnmd = (struct netmap_mem_ptg *)nmd;
2577 
2578 	if (!(nmd->flags & NETMAP_MEM_FINALIZED)) {
2579 		return EINVAL;
2580 	}
2581 
2582 	*lut = ptnmd->buf_lut;
2583 	return 0;
2584 }
2585 
2586 static int
2587 netmap_mem_pt_guest_get_info(struct netmap_mem_d *nmd, uint64_t *size,
2588 			     u_int *memflags, uint16_t *id)
2589 {
2590 	int error = 0;
2591 
2592 	error = nmd->ops->nmd_config(nmd);
2593 	if (error)
2594 		goto out;
2595 
2596 	if (size)
2597 		*size = nmd->nm_totalsize;
2598 	if (memflags)
2599 		*memflags = nmd->flags;
2600 	if (id)
2601 		*id = nmd->nm_id;
2602 
2603 out:
2604 
2605 	return error;
2606 }
2607 
2608 static vm_paddr_t
2609 netmap_mem_pt_guest_ofstophys(struct netmap_mem_d *nmd, vm_ooffset_t off)
2610 {
2611 	struct netmap_mem_ptg *ptnmd = (struct netmap_mem_ptg *)nmd;
2612 	vm_paddr_t paddr;
2613 	/* if the offset is valid, just return csb->base_addr + off */
2614 	paddr = (vm_paddr_t)(ptnmd->nm_paddr + off);
2615 	nm_prdis("off %lx padr %lx", off, (unsigned long)paddr);
2616 	return paddr;
2617 }
2618 
2619 static int
2620 netmap_mem_pt_guest_config(struct netmap_mem_d *nmd)
2621 {
2622 	/* nothing to do, we are configured on creation
2623 	 * and configuration never changes thereafter
2624 	 */
2625 	return 0;
2626 }
2627 
2628 static int
2629 netmap_mem_pt_guest_finalize(struct netmap_mem_d *nmd, struct netmap_adapter *na)
2630 {
2631 	struct netmap_mem_ptg *ptnmd = (struct netmap_mem_ptg *)nmd;
2632 	uint64_t mem_size;
2633 	uint32_t bufsize;
2634 	uint32_t nbuffers;
2635 	uint32_t poolofs;
2636 	vm_paddr_t paddr;
2637 	char *vaddr;
2638 	int i;
2639 	int error = 0;
2640 
2641 	if (nmd->flags & NETMAP_MEM_FINALIZED)
2642 		goto out;
2643 
2644 	if (ptnmd->ptn_dev == NULL) {
2645 		nm_prerr("ptnetmap memdev not attached");
2646 		error = ENOMEM;
2647 		goto out;
2648 	}
2649 	/* Map memory through ptnetmap-memdev BAR. */
2650 	error = nm_os_pt_memdev_iomap(ptnmd->ptn_dev, &ptnmd->nm_paddr,
2651 				      &ptnmd->nm_addr, &mem_size);
2652 	if (error)
2653 		goto out;
2654 
2655 	/* Initialize the lut using the information contained in the
2656 	 * ptnetmap memory device. */
2657 	bufsize = nm_os_pt_memdev_ioread(ptnmd->ptn_dev,
2658 					 PTNET_MDEV_IO_BUF_POOL_OBJSZ);
2659 	nbuffers = nm_os_pt_memdev_ioread(ptnmd->ptn_dev,
2660 					 PTNET_MDEV_IO_BUF_POOL_OBJNUM);
2661 
2662 	/* allocate the lut */
2663 	if (ptnmd->buf_lut.lut == NULL) {
2664 		nm_prinf("allocating lut");
2665 		ptnmd->buf_lut.lut = nm_alloc_lut(nbuffers);
2666 		if (ptnmd->buf_lut.lut == NULL) {
2667 			nm_prerr("lut allocation failed");
2668 			return ENOMEM;
2669 		}
2670 	}
2671 
2672 	/* we have physically contiguous memory mapped through PCI BAR */
2673 	poolofs = nm_os_pt_memdev_ioread(ptnmd->ptn_dev,
2674 					 PTNET_MDEV_IO_BUF_POOL_OFS);
2675 	vaddr = (char *)(ptnmd->nm_addr) + poolofs;
2676 	paddr = ptnmd->nm_paddr + poolofs;
2677 
2678 	for (i = 0; i < nbuffers; i++) {
2679 		ptnmd->buf_lut.lut[i].vaddr = vaddr;
2680 		vaddr += bufsize;
2681 		paddr += bufsize;
2682 	}
2683 
2684 	ptnmd->buf_lut.objtotal = nbuffers;
2685 	ptnmd->buf_lut.objsize = bufsize;
2686 	nmd->nm_totalsize = mem_size;
2687 
2688 	/* Initialize these fields as are needed by
2689 	 * netmap_mem_bufsize().
2690 	 * XXX please improve this, why do we need this
2691 	 * replication? maybe we nmd->pools[] should no be
2692 	 * there for the guest allocator? */
2693 	nmd->pools[NETMAP_BUF_POOL]._objsize = bufsize;
2694 	nmd->pools[NETMAP_BUF_POOL]._objtotal = nbuffers;
2695 
2696 	nmd->flags |= NETMAP_MEM_FINALIZED;
2697 out:
2698 	return error;
2699 }
2700 
2701 static void
2702 netmap_mem_pt_guest_deref(struct netmap_mem_d *nmd, struct netmap_adapter *na)
2703 {
2704 	struct netmap_mem_ptg *ptnmd = (struct netmap_mem_ptg *)nmd;
2705 
2706 	if (nmd->active == 1 &&
2707 		(nmd->flags & NETMAP_MEM_FINALIZED)) {
2708 	    nmd->flags  &= ~NETMAP_MEM_FINALIZED;
2709 	    /* unmap ptnetmap-memdev memory */
2710 	    if (ptnmd->ptn_dev) {
2711 		nm_os_pt_memdev_iounmap(ptnmd->ptn_dev);
2712 	    }
2713 	    ptnmd->nm_addr = NULL;
2714 	    ptnmd->nm_paddr = 0;
2715 	}
2716 }
2717 
2718 static ssize_t
2719 netmap_mem_pt_guest_if_offset(struct netmap_mem_d *nmd, const void *vaddr)
2720 {
2721 	struct netmap_mem_ptg *ptnmd = (struct netmap_mem_ptg *)nmd;
2722 
2723 	return (const char *)(vaddr) - (char *)(ptnmd->nm_addr);
2724 }
2725 
2726 static void
2727 netmap_mem_pt_guest_delete(struct netmap_mem_d *nmd)
2728 {
2729 	if (nmd == NULL)
2730 		return;
2731 	if (netmap_verbose)
2732 		nm_prinf("deleting %p", nmd);
2733 	if (nmd->active > 0)
2734 		nm_prerr("bug: deleting mem allocator with active=%d!", nmd->active);
2735 	if (netmap_verbose)
2736 		nm_prinf("done deleting %p", nmd);
2737 	NMA_LOCK_DESTROY(nmd);
2738 	nm_os_free(nmd);
2739 }
2740 
2741 static struct netmap_if *
2742 netmap_mem_pt_guest_if_new(struct netmap_mem_d *nmd,
2743 		struct netmap_adapter *na, struct netmap_priv_d *priv)
2744 {
2745 	struct netmap_mem_ptg *ptnmd = (struct netmap_mem_ptg *)nmd;
2746 	struct mem_pt_if *ptif;
2747 	struct netmap_if *nifp = NULL;
2748 
2749 	ptif = netmap_mem_pt_guest_ifp_lookup(nmd, na->ifp);
2750 	if (ptif == NULL) {
2751 		nm_prerr("interface %s is not in passthrough", na->name);
2752 		goto out;
2753 	}
2754 
2755 	nifp = (struct netmap_if *)((char *)(ptnmd->nm_addr) +
2756 				    ptif->nifp_offset);
2757 out:
2758 	return nifp;
2759 }
2760 
2761 static void
2762 netmap_mem_pt_guest_if_delete(struct netmap_mem_d * nmd,
2763 		struct netmap_adapter *na, struct netmap_if *nifp)
2764 {
2765 	struct mem_pt_if *ptif;
2766 
2767 	ptif = netmap_mem_pt_guest_ifp_lookup(nmd, na->ifp);
2768 	if (ptif == NULL) {
2769 		nm_prerr("interface %s is not in passthrough", na->name);
2770 	}
2771 }
2772 
2773 static int
2774 netmap_mem_pt_guest_rings_create(struct netmap_mem_d *nmd,
2775 		struct netmap_adapter *na)
2776 {
2777 	struct netmap_mem_ptg *ptnmd = (struct netmap_mem_ptg *)nmd;
2778 	struct mem_pt_if *ptif;
2779 	struct netmap_if *nifp;
2780 	int i, error = -1;
2781 
2782 	ptif = netmap_mem_pt_guest_ifp_lookup(nmd, na->ifp);
2783 	if (ptif == NULL) {
2784 		nm_prerr("interface %s is not in passthrough", na->name);
2785 		goto out;
2786 	}
2787 
2788 
2789 	/* point each kring to the corresponding backend ring */
2790 	nifp = (struct netmap_if *)((char *)ptnmd->nm_addr + ptif->nifp_offset);
2791 	for (i = 0; i < netmap_all_rings(na, NR_TX); i++) {
2792 		struct netmap_kring *kring = na->tx_rings[i];
2793 		if (kring->ring)
2794 			continue;
2795 		kring->ring = (struct netmap_ring *)
2796 			((char *)nifp + nifp->ring_ofs[i]);
2797 	}
2798 	for (i = 0; i < netmap_all_rings(na, NR_RX); i++) {
2799 		struct netmap_kring *kring = na->rx_rings[i];
2800 		if (kring->ring)
2801 			continue;
2802 		kring->ring = (struct netmap_ring *)
2803 			((char *)nifp +
2804 			 nifp->ring_ofs[netmap_all_rings(na, NR_TX) + i]);
2805 	}
2806 
2807 	error = 0;
2808 out:
2809 	return error;
2810 }
2811 
2812 static void
2813 netmap_mem_pt_guest_rings_delete(struct netmap_mem_d *nmd, struct netmap_adapter *na)
2814 {
2815 #if 0
2816 	enum txrx t;
2817 
2818 	for_rx_tx(t) {
2819 		u_int i;
2820 		for (i = 0; i < nma_get_nrings(na, t) + 1; i++) {
2821 			struct netmap_kring *kring = &NMR(na, t)[i];
2822 
2823 			kring->ring = NULL;
2824 		}
2825 	}
2826 #endif
2827 	(void)nmd;
2828 	(void)na;
2829 }
2830 
2831 static struct netmap_mem_ops netmap_mem_pt_guest_ops = {
2832 	.nmd_get_lut = netmap_mem_pt_guest_get_lut,
2833 	.nmd_get_info = netmap_mem_pt_guest_get_info,
2834 	.nmd_ofstophys = netmap_mem_pt_guest_ofstophys,
2835 	.nmd_config = netmap_mem_pt_guest_config,
2836 	.nmd_finalize = netmap_mem_pt_guest_finalize,
2837 	.nmd_deref = netmap_mem_pt_guest_deref,
2838 	.nmd_if_offset = netmap_mem_pt_guest_if_offset,
2839 	.nmd_delete = netmap_mem_pt_guest_delete,
2840 	.nmd_if_new = netmap_mem_pt_guest_if_new,
2841 	.nmd_if_delete = netmap_mem_pt_guest_if_delete,
2842 	.nmd_rings_create = netmap_mem_pt_guest_rings_create,
2843 	.nmd_rings_delete = netmap_mem_pt_guest_rings_delete
2844 };
2845 
2846 /* Called with nm_mem_list_lock held. */
2847 static struct netmap_mem_d *
2848 netmap_mem_pt_guest_find_memid(nm_memid_t mem_id)
2849 {
2850 	struct netmap_mem_d *mem = NULL;
2851 	struct netmap_mem_d *scan = netmap_last_mem_d;
2852 
2853 	do {
2854 		/* find ptnetmap allocator through host ID */
2855 		if (scan->ops->nmd_deref == netmap_mem_pt_guest_deref &&
2856 			((struct netmap_mem_ptg *)(scan))->host_mem_id == mem_id) {
2857 			mem = scan;
2858 			mem->refcount++;
2859 			NM_DBG_REFC(mem, __FUNCTION__, __LINE__);
2860 			break;
2861 		}
2862 		scan = scan->next;
2863 	} while (scan != netmap_last_mem_d);
2864 
2865 	return mem;
2866 }
2867 
2868 /* Called with nm_mem_list_lock held. */
2869 static struct netmap_mem_d *
2870 netmap_mem_pt_guest_create(nm_memid_t mem_id)
2871 {
2872 	struct netmap_mem_ptg *ptnmd;
2873 	int err = 0;
2874 
2875 	ptnmd = nm_os_malloc(sizeof(struct netmap_mem_ptg));
2876 	if (ptnmd == NULL) {
2877 		err = ENOMEM;
2878 		goto error;
2879 	}
2880 
2881 	ptnmd->up.ops = &netmap_mem_pt_guest_ops;
2882 	ptnmd->host_mem_id = mem_id;
2883 	ptnmd->pt_ifs = NULL;
2884 
2885 	/* Assign new id in the guest (We have the lock) */
2886 	err = nm_mem_assign_id_locked(&ptnmd->up, -1);
2887 	if (err)
2888 		goto error;
2889 
2890 	ptnmd->up.flags &= ~NETMAP_MEM_FINALIZED;
2891 	ptnmd->up.flags |= NETMAP_MEM_IO;
2892 
2893 	NMA_LOCK_INIT(&ptnmd->up);
2894 
2895 	snprintf(ptnmd->up.name, NM_MEM_NAMESZ, "%d", ptnmd->up.nm_id);
2896 
2897 
2898 	return &ptnmd->up;
2899 error:
2900 	netmap_mem_pt_guest_delete(&ptnmd->up);
2901 	return NULL;
2902 }
2903 
2904 /*
2905  * find host id in guest allocators and create guest allocator
2906  * if it is not there
2907  */
2908 static struct netmap_mem_d *
2909 netmap_mem_pt_guest_get(nm_memid_t mem_id)
2910 {
2911 	struct netmap_mem_d *nmd;
2912 
2913 	NM_MTX_LOCK(nm_mem_list_lock);
2914 	nmd = netmap_mem_pt_guest_find_memid(mem_id);
2915 	if (nmd == NULL) {
2916 		nmd = netmap_mem_pt_guest_create(mem_id);
2917 	}
2918 	NM_MTX_UNLOCK(nm_mem_list_lock);
2919 
2920 	return nmd;
2921 }
2922 
2923 /*
2924  * The guest allocator can be created by ptnetmap_memdev (during the device
2925  * attach) or by ptnetmap device (ptnet), during the netmap_attach.
2926  *
2927  * The order is not important (we have different order in LINUX and FreeBSD).
2928  * The first one, creates the device, and the second one simply attaches it.
2929  */
2930 
2931 /* Called when ptnetmap_memdev is attaching, to attach a new allocator in
2932  * the guest */
2933 struct netmap_mem_d *
2934 netmap_mem_pt_guest_attach(struct ptnetmap_memdev *ptn_dev, nm_memid_t mem_id)
2935 {
2936 	struct netmap_mem_d *nmd;
2937 	struct netmap_mem_ptg *ptnmd;
2938 
2939 	nmd = netmap_mem_pt_guest_get(mem_id);
2940 
2941 	/* assign this device to the guest allocator */
2942 	if (nmd) {
2943 		ptnmd = (struct netmap_mem_ptg *)nmd;
2944 		ptnmd->ptn_dev = ptn_dev;
2945 	}
2946 
2947 	return nmd;
2948 }
2949 
2950 /* Called when ptnet device is attaching */
2951 struct netmap_mem_d *
2952 netmap_mem_pt_guest_new(if_t ifp,
2953 			unsigned int nifp_offset,
2954 			unsigned int memid)
2955 {
2956 	struct netmap_mem_d *nmd;
2957 
2958 	if (ifp == NULL) {
2959 		return NULL;
2960 	}
2961 
2962 	nmd = netmap_mem_pt_guest_get((nm_memid_t)memid);
2963 
2964 	if (nmd) {
2965 		netmap_mem_pt_guest_ifp_add(nmd, ifp, nifp_offset);
2966 	}
2967 
2968 	return nmd;
2969 }
2970 
2971 #endif /* WITH_PTNETMAP */
2972