xref: /freebsd/sys/dev/netmap/netmap_mem2.c (revision 59c8e88e72633afbc47a4ace0d2170d00d51f7dc)
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 #include <sys/types.h>
42 #include <sys/malloc.h>
43 #include <sys/kernel.h>		/* MALLOC_DEFINE */
44 #include <sys/proc.h>
45 #include <vm/vm.h>	/* vtophys */
46 #include <vm/pmap.h>	/* vtophys */
47 #include <sys/socket.h> /* sockaddrs */
48 #include <sys/selinfo.h>
49 #include <sys/sysctl.h>
50 #include <net/if.h>
51 #include <net/if_var.h>
52 #include <net/vnet.h>
53 #include <machine/bus.h>	/* bus_dmamap_* */
54 
55 /* M_NETMAP only used in here */
56 MALLOC_DECLARE(M_NETMAP);
57 MALLOC_DEFINE(M_NETMAP, "netmap", "Network memory map");
58 
59 #endif /* __FreeBSD__ */
60 
61 #ifdef _WIN32
62 #include <win_glue.h>
63 #endif
64 
65 #include <net/netmap.h>
66 #include <dev/netmap/netmap_kern.h>
67 #include <net/netmap_virt.h>
68 #include "netmap_mem2.h"
69 
70 #ifdef _WIN32_USE_SMALL_GENERIC_DEVICES_MEMORY
71 #define NETMAP_BUF_MAX_NUM  8*4096      /* if too big takes too much time to allocate */
72 #else
73 #define NETMAP_BUF_MAX_NUM 20*4096*2	/* large machine */
74 #endif
75 
76 #define NETMAP_POOL_MAX_NAMSZ	32
77 
78 
79 enum {
80 	NETMAP_IF_POOL   = 0,
81 	NETMAP_RING_POOL,
82 	NETMAP_BUF_POOL,
83 	NETMAP_POOLS_NR
84 };
85 
86 
87 struct netmap_obj_params {
88 	u_int size;
89 	u_int num;
90 
91 	u_int last_size;
92 	u_int last_num;
93 };
94 
95 struct netmap_obj_pool {
96 	char name[NETMAP_POOL_MAX_NAMSZ];	/* name of the allocator */
97 
98 	/* ---------------------------------------------------*/
99 	/* these are only meaningful if the pool is finalized */
100 	/* (see 'finalized' field in netmap_mem_d)            */
101 	size_t memtotal;	/* actual total memory space */
102 
103 	struct lut_entry *lut;  /* virt,phys addresses, objtotal entries */
104 	uint32_t *bitmap;       /* one bit per buffer, 1 means free */
105 	uint32_t *invalid_bitmap;/* one bit per buffer, 1 means invalid */
106 	uint32_t bitmap_slots;	/* number of uint32 entries in bitmap */
107 
108 	u_int objtotal;         /* actual total number of objects. */
109 	u_int numclusters;	/* actual number of clusters */
110 	u_int objfree;          /* number of free objects. */
111 
112 	int	alloc_done;	/* we have allocated the memory */
113 	/* ---------------------------------------------------*/
114 
115 	/* limits */
116 	u_int objminsize;	/* minimum object size */
117 	u_int objmaxsize;	/* maximum object size */
118 	u_int nummin;		/* minimum number of objects */
119 	u_int nummax;		/* maximum number of objects */
120 
121 	/* these are changed only by config */
122 	u_int _objtotal;	/* total number of objects */
123 	u_int _objsize;		/* object size */
124 	u_int _clustsize;       /* cluster size */
125 	u_int _clustentries;    /* objects per cluster */
126 	u_int _numclusters;	/* number of clusters */
127 
128 	/* requested values */
129 	u_int r_objtotal;
130 	u_int r_objsize;
131 };
132 
133 #define NMA_LOCK_T		NM_MTX_T
134 #define NMA_LOCK_INIT(n)	NM_MTX_INIT((n)->nm_mtx)
135 #define NMA_LOCK_DESTROY(n)	NM_MTX_DESTROY((n)->nm_mtx)
136 #define NMA_LOCK(n)		NM_MTX_LOCK((n)->nm_mtx)
137 #define NMA_SPINLOCK(n)         NM_MTX_SPINLOCK((n)->nm_mtx)
138 #define NMA_UNLOCK(n)		NM_MTX_UNLOCK((n)->nm_mtx)
139 
140 struct netmap_mem_ops {
141 	int (*nmd_get_lut)(struct netmap_mem_d *, struct netmap_lut*);
142 	int  (*nmd_get_info)(struct netmap_mem_d *, uint64_t *size,
143 			u_int *memflags, uint16_t *id);
144 
145 	vm_paddr_t (*nmd_ofstophys)(struct netmap_mem_d *, vm_ooffset_t);
146 	int (*nmd_config)(struct netmap_mem_d *);
147 	int (*nmd_finalize)(struct netmap_mem_d *, struct netmap_adapter *);
148 	void (*nmd_deref)(struct netmap_mem_d *, struct netmap_adapter *);
149 	ssize_t  (*nmd_if_offset)(struct netmap_mem_d *, const void *vaddr);
150 	void (*nmd_delete)(struct netmap_mem_d *);
151 
152 	struct netmap_if * (*nmd_if_new)(struct netmap_mem_d *,
153 			struct netmap_adapter *, struct netmap_priv_d *);
154 	void (*nmd_if_delete)(struct netmap_mem_d *,
155 			struct netmap_adapter *, struct netmap_if *);
156 	int  (*nmd_rings_create)(struct netmap_mem_d *,
157 			struct netmap_adapter *);
158 	void (*nmd_rings_delete)(struct netmap_mem_d *,
159 			struct netmap_adapter *);
160 };
161 
162 struct netmap_mem_d {
163 	NMA_LOCK_T nm_mtx;  /* protect the allocator */
164 	size_t nm_totalsize; /* shorthand */
165 
166 	u_int flags;
167 #define NETMAP_MEM_FINALIZED	0x1	/* preallocation done */
168 #define NETMAP_MEM_HIDDEN	0x8	/* being prepared */
169 #define NETMAP_MEM_NOMAP	0x10	/* do not map/unmap pdevs */
170 	int lasterr;		/* last error for curr config */
171 	int active;		/* active users */
172 	int refcount;
173 	/* the three allocators */
174 	struct netmap_obj_pool pools[NETMAP_POOLS_NR];
175 
176 	nm_memid_t nm_id;	/* allocator identifier */
177 	int nm_grp;	/* iommu group id */
178 
179 	/* list of all existing allocators, sorted by nm_id */
180 	struct netmap_mem_d *prev, *next;
181 
182 	struct netmap_mem_ops *ops;
183 
184 	struct netmap_obj_params params[NETMAP_POOLS_NR];
185 
186 #define NM_MEM_NAMESZ	16
187 	char name[NM_MEM_NAMESZ];
188 };
189 
190 int
191 netmap_mem_get_lut(struct netmap_mem_d *nmd, struct netmap_lut *lut)
192 {
193 	int rv;
194 
195 	NMA_LOCK(nmd);
196 	rv = nmd->ops->nmd_get_lut(nmd, lut);
197 	NMA_UNLOCK(nmd);
198 
199 	return rv;
200 }
201 
202 int
203 netmap_mem_get_info(struct netmap_mem_d *nmd, uint64_t *size,
204 		u_int *memflags, nm_memid_t *memid)
205 {
206 	int rv;
207 
208 	NMA_LOCK(nmd);
209 	rv = nmd->ops->nmd_get_info(nmd, size, memflags, memid);
210 	NMA_UNLOCK(nmd);
211 
212 	return rv;
213 }
214 
215 vm_paddr_t
216 netmap_mem_ofstophys(struct netmap_mem_d *nmd, vm_ooffset_t off)
217 {
218 	vm_paddr_t pa;
219 
220 #if defined(__FreeBSD__)
221 	/* This function is called by netmap_dev_pager_fault(), which holds a
222 	 * non-sleepable lock since FreeBSD 12. Since we cannot sleep, we
223 	 * spin on the trylock. */
224 	NMA_SPINLOCK(nmd);
225 #else
226 	NMA_LOCK(nmd);
227 #endif
228 	pa = nmd->ops->nmd_ofstophys(nmd, off);
229 	NMA_UNLOCK(nmd);
230 
231 	return pa;
232 }
233 
234 static int
235 netmap_mem_config(struct netmap_mem_d *nmd)
236 {
237 	if (nmd->active) {
238 		/* already in use. Not fatal, but we
239 		 * cannot change the configuration
240 		 */
241 		return 0;
242 	}
243 
244 	return nmd->ops->nmd_config(nmd);
245 }
246 
247 ssize_t
248 netmap_mem_if_offset(struct netmap_mem_d *nmd, const void *off)
249 {
250 	ssize_t rv;
251 
252 	NMA_LOCK(nmd);
253 	rv = nmd->ops->nmd_if_offset(nmd, off);
254 	NMA_UNLOCK(nmd);
255 
256 	return rv;
257 }
258 
259 static void
260 netmap_mem_delete(struct netmap_mem_d *nmd)
261 {
262 	nmd->ops->nmd_delete(nmd);
263 }
264 
265 struct netmap_if *
266 netmap_mem_if_new(struct netmap_adapter *na, struct netmap_priv_d *priv)
267 {
268 	struct netmap_if *nifp;
269 	struct netmap_mem_d *nmd = na->nm_mem;
270 
271 	NMA_LOCK(nmd);
272 	nifp = nmd->ops->nmd_if_new(nmd, na, priv);
273 	NMA_UNLOCK(nmd);
274 
275 	return nifp;
276 }
277 
278 void
279 netmap_mem_if_delete(struct netmap_adapter *na, struct netmap_if *nif)
280 {
281 	struct netmap_mem_d *nmd = na->nm_mem;
282 
283 	NMA_LOCK(nmd);
284 	nmd->ops->nmd_if_delete(nmd, na, nif);
285 	NMA_UNLOCK(nmd);
286 }
287 
288 int
289 netmap_mem_rings_create(struct netmap_adapter *na)
290 {
291 	int rv;
292 	struct netmap_mem_d *nmd = na->nm_mem;
293 
294 	NMA_LOCK(nmd);
295 	rv = nmd->ops->nmd_rings_create(nmd, na);
296 	NMA_UNLOCK(nmd);
297 
298 	return rv;
299 }
300 
301 void
302 netmap_mem_rings_delete(struct netmap_adapter *na)
303 {
304 	struct netmap_mem_d *nmd = na->nm_mem;
305 
306 	NMA_LOCK(nmd);
307 	nmd->ops->nmd_rings_delete(nmd, na);
308 	NMA_UNLOCK(nmd);
309 }
310 
311 static int netmap_mem_map(struct netmap_obj_pool *, struct netmap_adapter *);
312 static int netmap_mem_unmap(struct netmap_obj_pool *, struct netmap_adapter *);
313 static int nm_mem_check_group(struct netmap_mem_d *, bus_dma_tag_t);
314 static void nm_mem_release_id(struct netmap_mem_d *);
315 
316 nm_memid_t
317 netmap_mem_get_id(struct netmap_mem_d *nmd)
318 {
319 	return nmd->nm_id;
320 }
321 
322 #ifdef NM_DEBUG_MEM_PUTGET
323 #define NM_DBG_REFC(nmd, func, line)	\
324 	nm_prinf("%s:%d mem[%d:%d] -> %d", func, line, (nmd)->nm_id, (nmd)->nm_grp, (nmd)->refcount);
325 #else
326 #define NM_DBG_REFC(nmd, func, line)
327 #endif
328 
329 /* circular list of all existing allocators */
330 static struct netmap_mem_d *netmap_last_mem_d = &nm_mem;
331 static NM_MTX_T nm_mem_list_lock;
332 
333 struct netmap_mem_d *
334 __netmap_mem_get(struct netmap_mem_d *nmd, const char *func, int line)
335 {
336 	NM_MTX_LOCK(nm_mem_list_lock);
337 	nmd->refcount++;
338 	NM_DBG_REFC(nmd, func, line);
339 	NM_MTX_UNLOCK(nm_mem_list_lock);
340 	return nmd;
341 }
342 
343 void
344 __netmap_mem_put(struct netmap_mem_d *nmd, const char *func, int line)
345 {
346 	int last;
347 	NM_MTX_LOCK(nm_mem_list_lock);
348 	last = (--nmd->refcount == 0);
349 	if (last)
350 		nm_mem_release_id(nmd);
351 	NM_DBG_REFC(nmd, func, line);
352 	NM_MTX_UNLOCK(nm_mem_list_lock);
353 	if (last)
354 		netmap_mem_delete(nmd);
355 }
356 
357 int
358 netmap_mem_finalize(struct netmap_mem_d *nmd, struct netmap_adapter *na)
359 {
360 	int lasterr = 0;
361 	if (nm_mem_check_group(nmd, na->pdev) < 0) {
362 		return ENOMEM;
363 	}
364 
365 	NMA_LOCK(nmd);
366 
367 	if (netmap_mem_config(nmd))
368 		goto out;
369 
370 	nmd->active++;
371 
372 	nmd->lasterr = nmd->ops->nmd_finalize(nmd, na);
373 
374 	if (!nmd->lasterr && !(nmd->flags & NETMAP_MEM_NOMAP)) {
375 		nmd->lasterr = netmap_mem_map(&nmd->pools[NETMAP_BUF_POOL], na);
376 	}
377 
378 out:
379 	lasterr = nmd->lasterr;
380 	NMA_UNLOCK(nmd);
381 
382 	if (lasterr)
383 		netmap_mem_deref(nmd, na);
384 
385 	return lasterr;
386 }
387 
388 static int
389 nm_isset(uint32_t *bitmap, u_int i)
390 {
391 	return bitmap[ (i>>5) ] & ( 1U << (i & 31U) );
392 }
393 
394 
395 static int
396 netmap_init_obj_allocator_bitmap(struct netmap_obj_pool *p)
397 {
398 	u_int n, j;
399 
400 	if (p->bitmap == NULL) {
401 		/* Allocate the bitmap */
402 		n = (p->objtotal + 31) / 32;
403 		p->bitmap = nm_os_malloc(sizeof(p->bitmap[0]) * n);
404 		if (p->bitmap == NULL) {
405 			nm_prerr("Unable to create bitmap (%d entries) for allocator '%s'", (int)n,
406 			    p->name);
407 			return ENOMEM;
408 		}
409 		p->bitmap_slots = n;
410 	} else {
411 		memset(p->bitmap, 0, p->bitmap_slots * sizeof(p->bitmap[0]));
412 	}
413 
414 	p->objfree = 0;
415 	/*
416 	 * Set all the bits in the bitmap that have
417 	 * corresponding buffers to 1 to indicate they are
418 	 * free.
419 	 */
420 	for (j = 0; j < p->objtotal; j++) {
421 		if (p->invalid_bitmap && nm_isset(p->invalid_bitmap, j)) {
422 			if (netmap_debug & NM_DEBUG_MEM)
423 				nm_prinf("skipping %s %d", p->name, j);
424 			continue;
425 		}
426 		p->bitmap[ (j>>5) ] |=  ( 1U << (j & 31U) );
427 		p->objfree++;
428 	}
429 
430 	if (netmap_verbose)
431 		nm_prinf("%s free %u", p->name, p->objfree);
432 	if (p->objfree == 0) {
433 		if (netmap_verbose)
434 			nm_prerr("%s: no objects available", p->name);
435 		return ENOMEM;
436 	}
437 
438 	return 0;
439 }
440 
441 static int
442 netmap_mem_init_bitmaps(struct netmap_mem_d *nmd)
443 {
444 	int i, error = 0;
445 
446 	for (i = 0; i < NETMAP_POOLS_NR; i++) {
447 		struct netmap_obj_pool *p = &nmd->pools[i];
448 
449 		error = netmap_init_obj_allocator_bitmap(p);
450 		if (error)
451 			return error;
452 	}
453 
454 	/*
455 	 * buffers 0 and 1 are reserved
456 	 */
457 	if (nmd->pools[NETMAP_BUF_POOL].objfree < 2) {
458 		nm_prerr("%s: not enough buffers", nmd->pools[NETMAP_BUF_POOL].name);
459 		return ENOMEM;
460 	}
461 
462 	nmd->pools[NETMAP_BUF_POOL].objfree -= 2;
463 	if (nmd->pools[NETMAP_BUF_POOL].bitmap) {
464 		/* XXX This check is a workaround that prevents a
465 		 * NULL pointer crash which currently happens only
466 		 * with ptnetmap guests.
467 		 * Removed shared-info --> is the bug still there? */
468 		nmd->pools[NETMAP_BUF_POOL].bitmap[0] = ~3U;
469 	}
470 	return 0;
471 }
472 
473 int
474 netmap_mem_deref(struct netmap_mem_d *nmd, struct netmap_adapter *na)
475 {
476 	int last_user = 0;
477 	NMA_LOCK(nmd);
478 	if (na->active_fds <= 0 && !(nmd->flags & NETMAP_MEM_NOMAP))
479 		netmap_mem_unmap(&nmd->pools[NETMAP_BUF_POOL], na);
480 	if (nmd->active == 1) {
481 		last_user = 1;
482 		/*
483 		 * Reset the allocator when it falls out of use so that any
484 		 * pool resources leaked by unclean application exits are
485 		 * reclaimed.
486 		 */
487 		netmap_mem_init_bitmaps(nmd);
488 	}
489 	nmd->ops->nmd_deref(nmd, na);
490 
491 	nmd->active--;
492 	if (last_user) {
493 		nmd->lasterr = 0;
494 	}
495 
496 	NMA_UNLOCK(nmd);
497 	return last_user;
498 }
499 
500 
501 /* accessor functions */
502 static int
503 netmap_mem2_get_lut(struct netmap_mem_d *nmd, struct netmap_lut *lut)
504 {
505 	lut->lut = nmd->pools[NETMAP_BUF_POOL].lut;
506 #ifdef __FreeBSD__
507 	lut->plut = lut->lut;
508 #endif
509 	lut->objtotal = nmd->pools[NETMAP_BUF_POOL].objtotal;
510 	lut->objsize = nmd->pools[NETMAP_BUF_POOL]._objsize;
511 
512 	return 0;
513 }
514 
515 static struct netmap_obj_params netmap_min_priv_params[NETMAP_POOLS_NR] = {
516 	[NETMAP_IF_POOL] = {
517 		.size = 1024,
518 		.num  = 2,
519 	},
520 	[NETMAP_RING_POOL] = {
521 		.size = 5*PAGE_SIZE,
522 		.num  = 4,
523 	},
524 	[NETMAP_BUF_POOL] = {
525 		.size = 2048,
526 		.num  = 4098,
527 	},
528 };
529 
530 
531 /*
532  * nm_mem is the memory allocator used for all physical interfaces
533  * running in netmap mode.
534  * Virtual (VALE) ports will have each its own allocator.
535  */
536 extern struct netmap_mem_ops netmap_mem_global_ops; /* forward */
537 struct netmap_mem_d nm_mem = {	/* Our memory allocator. */
538 	.pools = {
539 		[NETMAP_IF_POOL] = {
540 			.name 	= "netmap_if",
541 			.objminsize = sizeof(struct netmap_if),
542 			.objmaxsize = 4096,
543 			.nummin     = 10,	/* don't be stingy */
544 			.nummax	    = 10000,	/* XXX very large */
545 		},
546 		[NETMAP_RING_POOL] = {
547 			.name 	= "netmap_ring",
548 			.objminsize = sizeof(struct netmap_ring),
549 			.objmaxsize = 32*PAGE_SIZE,
550 			.nummin     = 2,
551 			.nummax	    = 1024,
552 		},
553 		[NETMAP_BUF_POOL] = {
554 			.name	= "netmap_buf",
555 			.objminsize = 64,
556 			.objmaxsize = 65536,
557 			.nummin     = 4,
558 			.nummax	    = 1000000, /* one million! */
559 		},
560 	},
561 
562 	.params = {
563 		[NETMAP_IF_POOL] = {
564 			.size = 1024,
565 			.num  = 100,
566 		},
567 		[NETMAP_RING_POOL] = {
568 			.size = 9*PAGE_SIZE,
569 			.num  = 200,
570 		},
571 		[NETMAP_BUF_POOL] = {
572 			.size = 2048,
573 			.num  = NETMAP_BUF_MAX_NUM,
574 		},
575 	},
576 
577 	.nm_id = 1,
578 	.nm_grp = -1,
579 
580 	.prev = &nm_mem,
581 	.next = &nm_mem,
582 
583 	.ops = &netmap_mem_global_ops,
584 
585 	.name = "1"
586 };
587 
588 static struct netmap_mem_d nm_mem_blueprint;
589 
590 /* blueprint for the private memory allocators */
591 /* XXX clang is not happy about using name as a print format */
592 static const struct netmap_mem_d nm_blueprint = {
593 	.pools = {
594 		[NETMAP_IF_POOL] = {
595 			.name 	= "%s_if",
596 			.objminsize = sizeof(struct netmap_if),
597 			.objmaxsize = 4096,
598 			.nummin     = 1,
599 			.nummax	    = 100,
600 		},
601 		[NETMAP_RING_POOL] = {
602 			.name 	= "%s_ring",
603 			.objminsize = sizeof(struct netmap_ring),
604 			.objmaxsize = 32*PAGE_SIZE,
605 			.nummin     = 2,
606 			.nummax	    = 1024,
607 		},
608 		[NETMAP_BUF_POOL] = {
609 			.name	= "%s_buf",
610 			.objminsize = 64,
611 			.objmaxsize = 65536,
612 			.nummin     = 4,
613 			.nummax	    = 1000000, /* one million! */
614 		},
615 	},
616 
617 	.nm_grp = -1,
618 
619 	.flags = NETMAP_MEM_PRIVATE,
620 
621 	.ops = &netmap_mem_global_ops,
622 };
623 
624 /* memory allocator related sysctls */
625 
626 #define STRINGIFY(x) #x
627 
628 
629 #define DECLARE_SYSCTLS(id, name) \
630 	SYSBEGIN(mem2_ ## name); \
631 	SYSCTL_INT(_dev_netmap, OID_AUTO, name##_size, \
632 	    CTLFLAG_RW, &nm_mem.params[id].size, 0, "Requested size of netmap " STRINGIFY(name) "s"); \
633 	SYSCTL_INT(_dev_netmap, OID_AUTO, name##_curr_size, \
634 	    CTLFLAG_RD, &nm_mem.pools[id]._objsize, 0, "Current size of netmap " STRINGIFY(name) "s"); \
635 	SYSCTL_INT(_dev_netmap, OID_AUTO, name##_num, \
636 	    CTLFLAG_RW, &nm_mem.params[id].num, 0, "Requested number of netmap " STRINGIFY(name) "s"); \
637 	SYSCTL_INT(_dev_netmap, OID_AUTO, name##_curr_num, \
638 	    CTLFLAG_RD, &nm_mem.pools[id].objtotal, 0, "Current number of netmap " STRINGIFY(name) "s"); \
639 	SYSCTL_INT(_dev_netmap, OID_AUTO, priv_##name##_size, \
640 	    CTLFLAG_RW, &netmap_min_priv_params[id].size, 0, \
641 	    "Default size of private netmap " STRINGIFY(name) "s"); \
642 	SYSCTL_INT(_dev_netmap, OID_AUTO, priv_##name##_num, \
643 	    CTLFLAG_RW, &netmap_min_priv_params[id].num, 0, \
644 	    "Default number of private netmap " STRINGIFY(name) "s");	\
645 	SYSEND
646 
647 SYSCTL_DECL(_dev_netmap);
648 DECLARE_SYSCTLS(NETMAP_IF_POOL, if);
649 DECLARE_SYSCTLS(NETMAP_RING_POOL, ring);
650 DECLARE_SYSCTLS(NETMAP_BUF_POOL, buf);
651 
652 /* call with nm_mem_list_lock held */
653 static int
654 nm_mem_assign_id_locked(struct netmap_mem_d *nmd, int grp_id)
655 {
656 	nm_memid_t id;
657 	struct netmap_mem_d *scan = netmap_last_mem_d;
658 	int error = ENOMEM;
659 
660 	do {
661 		/* we rely on unsigned wrap around */
662 		id = scan->nm_id + 1;
663 		if (id == 0) /* reserve 0 as error value */
664 			id = 1;
665 		scan = scan->next;
666 		if (id != scan->nm_id) {
667 			nmd->nm_id = id;
668 			nmd->nm_grp = grp_id;
669 			nmd->prev = scan->prev;
670 			nmd->next = scan;
671 			scan->prev->next = nmd;
672 			scan->prev = nmd;
673 			netmap_last_mem_d = nmd;
674 			nmd->refcount = 1;
675 			NM_DBG_REFC(nmd, __FUNCTION__, __LINE__);
676 			error = 0;
677 			break;
678 		}
679 	} while (scan != netmap_last_mem_d);
680 
681 	return error;
682 }
683 
684 /* call with nm_mem_list_lock *not* held */
685 static int
686 nm_mem_assign_id(struct netmap_mem_d *nmd, int grp_id)
687 {
688 	int ret;
689 
690 	NM_MTX_LOCK(nm_mem_list_lock);
691 	ret = nm_mem_assign_id_locked(nmd, grp_id);
692 	NM_MTX_UNLOCK(nm_mem_list_lock);
693 
694 	return ret;
695 }
696 
697 /* call with nm_mem_list_lock held */
698 static void
699 nm_mem_release_id(struct netmap_mem_d *nmd)
700 {
701 	nmd->prev->next = nmd->next;
702 	nmd->next->prev = nmd->prev;
703 
704 	if (netmap_last_mem_d == nmd)
705 		netmap_last_mem_d = nmd->prev;
706 
707 	nmd->prev = nmd->next = NULL;
708 }
709 
710 struct netmap_mem_d *
711 netmap_mem_find(nm_memid_t id)
712 {
713 	struct netmap_mem_d *nmd;
714 
715 	NM_MTX_LOCK(nm_mem_list_lock);
716 	nmd = netmap_last_mem_d;
717 	do {
718 		if (!(nmd->flags & NETMAP_MEM_HIDDEN) && nmd->nm_id == id) {
719 			nmd->refcount++;
720 			NM_DBG_REFC(nmd, __FUNCTION__, __LINE__);
721 			NM_MTX_UNLOCK(nm_mem_list_lock);
722 			return nmd;
723 		}
724 		nmd = nmd->next;
725 	} while (nmd != netmap_last_mem_d);
726 	NM_MTX_UNLOCK(nm_mem_list_lock);
727 	return NULL;
728 }
729 
730 static int
731 nm_mem_check_group(struct netmap_mem_d *nmd, bus_dma_tag_t dev)
732 {
733 	int err = 0, id;
734 
735 	/* Skip not hw adapters.
736 	 * Vale port can use particular allocator through vale-ctl -m option
737 	 */
738 	if (!dev)
739 		return 0;
740 	id = nm_iommu_group_id(dev);
741 	if (netmap_debug & NM_DEBUG_MEM)
742 		nm_prinf("iommu_group %d", id);
743 
744 	NMA_LOCK(nmd);
745 
746 	if (nmd->nm_grp != id) {
747 		if (netmap_verbose)
748 			nm_prerr("iommu group mismatch: %d vs %d",
749 					nmd->nm_grp, id);
750 		nmd->lasterr = err = ENOMEM;
751 	}
752 
753 	NMA_UNLOCK(nmd);
754 	return err;
755 }
756 
757 static struct lut_entry *
758 nm_alloc_lut(u_int nobj)
759 {
760 	size_t n = sizeof(struct lut_entry) * nobj;
761 	struct lut_entry *lut;
762 #ifdef linux
763 	lut = vmalloc(n);
764 #else
765 	lut = nm_os_malloc(n);
766 #endif
767 	return lut;
768 }
769 
770 static void
771 nm_free_lut(struct lut_entry *lut, u_int objtotal)
772 {
773 	bzero(lut, sizeof(struct lut_entry) * objtotal);
774 #ifdef linux
775 	vfree(lut);
776 #else
777 	nm_os_free(lut);
778 #endif
779 }
780 
781 #if defined(linux) || defined(_WIN32)
782 static struct plut_entry *
783 nm_alloc_plut(u_int nobj)
784 {
785 	size_t n = sizeof(struct plut_entry) * nobj;
786 	struct plut_entry *lut;
787 	lut = vmalloc(n);
788 	return lut;
789 }
790 
791 static void
792 nm_free_plut(struct plut_entry * lut)
793 {
794 	vfree(lut);
795 }
796 #endif /* linux or _WIN32 */
797 
798 
799 /*
800  * First, find the allocator that contains the requested offset,
801  * then locate the cluster through a lookup table.
802  */
803 static vm_paddr_t
804 netmap_mem2_ofstophys(struct netmap_mem_d* nmd, vm_ooffset_t offset)
805 {
806 	int i;
807 	vm_ooffset_t o = offset;
808 	vm_paddr_t pa;
809 	struct netmap_obj_pool *p;
810 
811 	p = nmd->pools;
812 
813 	for (i = 0; i < NETMAP_POOLS_NR; offset -= p[i].memtotal, i++) {
814 		if (offset >= p[i].memtotal)
815 			continue;
816 		// now lookup the cluster's address
817 #ifndef _WIN32
818 		pa = vtophys(p[i].lut[offset / p[i]._objsize].vaddr) +
819 			offset % p[i]._objsize;
820 #else
821 		pa = vtophys(p[i].lut[offset / p[i]._objsize].vaddr);
822 		pa.QuadPart += offset % p[i]._objsize;
823 #endif
824 		return pa;
825 	}
826 	/* this is only in case of errors */
827 	nm_prerr("invalid ofs 0x%x out of 0x%zx 0x%zx 0x%zx", (u_int)o,
828 		p[NETMAP_IF_POOL].memtotal,
829 		p[NETMAP_IF_POOL].memtotal
830 			+ p[NETMAP_RING_POOL].memtotal,
831 		p[NETMAP_IF_POOL].memtotal
832 			+ p[NETMAP_RING_POOL].memtotal
833 			+ p[NETMAP_BUF_POOL].memtotal);
834 #ifndef _WIN32
835 	return 0; /* bad address */
836 #else
837 	vm_paddr_t res;
838 	res.QuadPart = 0;
839 	return res;
840 #endif
841 }
842 
843 #ifdef _WIN32
844 
845 /*
846  * win32_build_virtual_memory_for_userspace
847  *
848  * This function get all the object making part of the pools and maps
849  * a contiguous virtual memory space for the userspace
850  * It works this way
851  * 1 - allocate a Memory Descriptor List wide as the sum
852  *		of the memory needed for the pools
853  * 2 - cycle all the objects in every pool and for every object do
854  *
855  *		2a - cycle all the objects in every pool, get the list
856  *				of the physical address descriptors
857  *		2b - calculate the offset in the array of pages descriptor in the
858  *				main MDL
859  *		2c - copy the descriptors of the object in the main MDL
860  *
861  * 3 - return the resulting MDL that needs to be mapped in userland
862  *
863  * In this way we will have an MDL that describes all the memory for the
864  * objects in a single object
865 */
866 
867 PMDL
868 win32_build_user_vm_map(struct netmap_mem_d* nmd)
869 {
870 	u_int memflags, ofs = 0;
871 	PMDL mainMdl, tempMdl;
872 	uint64_t memsize;
873 	int i, j;
874 
875 	if (netmap_mem_get_info(nmd, &memsize, &memflags, NULL)) {
876 		nm_prerr("memory not finalised yet");
877 		return NULL;
878 	}
879 
880 	mainMdl = IoAllocateMdl(NULL, memsize, FALSE, FALSE, NULL);
881 	if (mainMdl == NULL) {
882 		nm_prerr("failed to allocate mdl");
883 		return NULL;
884 	}
885 
886 	NMA_LOCK(nmd);
887 	for (i = 0; i < NETMAP_POOLS_NR; i++) {
888 		struct netmap_obj_pool *p = &nmd->pools[i];
889 		int clsz = p->_clustsize;
890 		int clobjs = p->_clustentries; /* objects per cluster */
891 		int mdl_len = sizeof(PFN_NUMBER) * BYTES_TO_PAGES(clsz);
892 		PPFN_NUMBER pSrc, pDst;
893 
894 		/* each pool has a different cluster size so we need to reallocate */
895 		tempMdl = IoAllocateMdl(p->lut[0].vaddr, clsz, FALSE, FALSE, NULL);
896 		if (tempMdl == NULL) {
897 			NMA_UNLOCK(nmd);
898 			nm_prerr("fail to allocate tempMdl");
899 			IoFreeMdl(mainMdl);
900 			return NULL;
901 		}
902 		pSrc = MmGetMdlPfnArray(tempMdl);
903 		/* create one entry per cluster, the lut[] has one entry per object */
904 		for (j = 0; j < p->numclusters; j++, ofs += clsz) {
905 			pDst = &MmGetMdlPfnArray(mainMdl)[BYTES_TO_PAGES(ofs)];
906 			MmInitializeMdl(tempMdl, p->lut[j*clobjs].vaddr, clsz);
907 			MmBuildMdlForNonPagedPool(tempMdl); /* compute physical page addresses */
908 			RtlCopyMemory(pDst, pSrc, mdl_len); /* copy the page descriptors */
909 			mainMdl->MdlFlags = tempMdl->MdlFlags; /* XXX what is in here ? */
910 		}
911 		IoFreeMdl(tempMdl);
912 	}
913 	NMA_UNLOCK(nmd);
914 	return mainMdl;
915 }
916 
917 #endif /* _WIN32 */
918 
919 /*
920  * helper function for OS-specific mmap routines (currently only windows).
921  * Given an nmd and a pool index, returns the cluster size and number of clusters.
922  * Returns 0 if memory is finalised and the pool is valid, otherwise 1.
923  * It should be called under NMA_LOCK(nmd) otherwise the underlying info can change.
924  */
925 
926 int
927 netmap_mem2_get_pool_info(struct netmap_mem_d* nmd, u_int pool, u_int *clustsize, u_int *numclusters)
928 {
929 	if (!nmd || !clustsize || !numclusters || pool >= NETMAP_POOLS_NR)
930 		return 1; /* invalid arguments */
931 	// NMA_LOCK_ASSERT(nmd);
932 	if (!(nmd->flags & NETMAP_MEM_FINALIZED)) {
933 		*clustsize = *numclusters = 0;
934 		return 1; /* not ready yet */
935 	}
936 	*clustsize = nmd->pools[pool]._clustsize;
937 	*numclusters = nmd->pools[pool].numclusters;
938 	return 0; /* success */
939 }
940 
941 static int
942 netmap_mem2_get_info(struct netmap_mem_d* nmd, uint64_t* size,
943 			u_int *memflags, nm_memid_t *id)
944 {
945 	int error = 0;
946 	error = netmap_mem_config(nmd);
947 	if (error)
948 		goto out;
949 	if (size) {
950 		if (nmd->flags & NETMAP_MEM_FINALIZED) {
951 			*size = nmd->nm_totalsize;
952 		} else {
953 			int i;
954 			*size = 0;
955 			for (i = 0; i < NETMAP_POOLS_NR; i++) {
956 				struct netmap_obj_pool *p = nmd->pools + i;
957 				*size += ((size_t)p->_numclusters * (size_t)p->_clustsize);
958 			}
959 		}
960 	}
961 	if (memflags)
962 		*memflags = nmd->flags;
963 	if (id)
964 		*id = nmd->nm_id;
965 out:
966 	return error;
967 }
968 
969 /*
970  * we store objects by kernel address, need to find the offset
971  * within the pool to export the value to userspace.
972  * Algorithm: scan until we find the cluster, then add the
973  * actual offset in the cluster
974  */
975 static ssize_t
976 netmap_obj_offset(struct netmap_obj_pool *p, const void *vaddr)
977 {
978 	int i, k = p->_clustentries, n = p->objtotal;
979 	ssize_t ofs = 0;
980 
981 	for (i = 0; i < n; i += k, ofs += p->_clustsize) {
982 		const char *base = p->lut[i].vaddr;
983 		ssize_t relofs = (const char *) vaddr - base;
984 
985 		if (relofs < 0 || relofs >= p->_clustsize)
986 			continue;
987 
988 		ofs = ofs + relofs;
989 		nm_prdis("%s: return offset %d (cluster %d) for pointer %p",
990 		    p->name, ofs, i, vaddr);
991 		return ofs;
992 	}
993 	nm_prerr("address %p is not contained inside any cluster (%s)",
994 	    vaddr, p->name);
995 	return 0; /* An error occurred */
996 }
997 
998 /* Helper functions which convert virtual addresses to offsets */
999 #define netmap_if_offset(n, v)					\
1000 	netmap_obj_offset(&(n)->pools[NETMAP_IF_POOL], (v))
1001 
1002 #define netmap_ring_offset(n, v)				\
1003     ((n)->pools[NETMAP_IF_POOL].memtotal + 			\
1004 	netmap_obj_offset(&(n)->pools[NETMAP_RING_POOL], (v)))
1005 
1006 static ssize_t
1007 netmap_mem2_if_offset(struct netmap_mem_d *nmd, const void *addr)
1008 {
1009 	return netmap_if_offset(nmd, addr);
1010 }
1011 
1012 /*
1013  * report the index, and use start position as a hint,
1014  * otherwise buffer allocation becomes terribly expensive.
1015  */
1016 static void *
1017 netmap_obj_malloc(struct netmap_obj_pool *p, u_int len, uint32_t *start, uint32_t *index)
1018 {
1019 	uint32_t i = 0;			/* index in the bitmap */
1020 	uint32_t mask, j = 0;		/* slot counter */
1021 	void *vaddr = NULL;
1022 
1023 	if (len > p->_objsize) {
1024 		nm_prerr("%s request size %d too large", p->name, len);
1025 		return NULL;
1026 	}
1027 
1028 	if (p->objfree == 0) {
1029 		nm_prerr("no more %s objects", p->name);
1030 		return NULL;
1031 	}
1032 	if (start)
1033 		i = *start;
1034 
1035 	/* termination is guaranteed by p->free, but better check bounds on i */
1036 	while (vaddr == NULL && i < p->bitmap_slots)  {
1037 		uint32_t cur = p->bitmap[i];
1038 		if (cur == 0) { /* bitmask is fully used */
1039 			i++;
1040 			continue;
1041 		}
1042 		/* locate a slot */
1043 		for (j = 0, mask = 1; (cur & mask) == 0; j++, mask <<= 1)
1044 			;
1045 
1046 		p->bitmap[i] &= ~mask; /* mark object as in use */
1047 		p->objfree--;
1048 
1049 		vaddr = p->lut[i * 32 + j].vaddr;
1050 		if (index)
1051 			*index = i * 32 + j;
1052 	}
1053 	nm_prdis("%s allocator: allocated object @ [%d][%d]: vaddr %p",p->name, i, j, vaddr);
1054 
1055 	if (start)
1056 		*start = i;
1057 	return vaddr;
1058 }
1059 
1060 
1061 /*
1062  * free by index, not by address.
1063  * XXX should we also cleanup the content ?
1064  */
1065 static int
1066 netmap_obj_free(struct netmap_obj_pool *p, uint32_t j)
1067 {
1068 	uint32_t *ptr, mask;
1069 
1070 	if (j >= p->objtotal) {
1071 		nm_prerr("invalid index %u, max %u", j, p->objtotal);
1072 		return 1;
1073 	}
1074 	ptr = &p->bitmap[j / 32];
1075 	mask = (1 << (j % 32));
1076 	if (*ptr & mask) {
1077 		nm_prerr("ouch, double free on buffer %d", j);
1078 		return 1;
1079 	} else {
1080 		*ptr |= mask;
1081 		p->objfree++;
1082 		return 0;
1083 	}
1084 }
1085 
1086 /*
1087  * free by address. This is slow but is only used for a few
1088  * objects (rings, nifp)
1089  */
1090 static void
1091 netmap_obj_free_va(struct netmap_obj_pool *p, void *vaddr)
1092 {
1093 	u_int i, j, n = p->numclusters;
1094 
1095 	for (i = 0, j = 0; i < n; i++, j += p->_clustentries) {
1096 		void *base = p->lut[i * p->_clustentries].vaddr;
1097 		ssize_t relofs = (ssize_t) vaddr - (ssize_t) base;
1098 
1099 		/* Given address, is out of the scope of the current cluster.*/
1100 		if (base == NULL || vaddr < base || relofs >= p->_clustsize)
1101 			continue;
1102 
1103 		j = j + relofs / p->_objsize;
1104 		/* KASSERT(j != 0, ("Cannot free object 0")); */
1105 		netmap_obj_free(p, j);
1106 		return;
1107 	}
1108 	nm_prerr("address %p is not contained inside any cluster (%s)",
1109 	    vaddr, p->name);
1110 }
1111 
1112 unsigned
1113 netmap_mem_bufsize(struct netmap_mem_d *nmd)
1114 {
1115 	return nmd->pools[NETMAP_BUF_POOL]._objsize;
1116 }
1117 
1118 #define netmap_if_malloc(n, len)	netmap_obj_malloc(&(n)->pools[NETMAP_IF_POOL], len, NULL, NULL)
1119 #define netmap_if_free(n, v)		netmap_obj_free_va(&(n)->pools[NETMAP_IF_POOL], (v))
1120 #define netmap_ring_malloc(n, len)	netmap_obj_malloc(&(n)->pools[NETMAP_RING_POOL], len, NULL, NULL)
1121 #define netmap_ring_free(n, v)		netmap_obj_free_va(&(n)->pools[NETMAP_RING_POOL], (v))
1122 #define netmap_buf_malloc(n, _pos, _index)			\
1123 	netmap_obj_malloc(&(n)->pools[NETMAP_BUF_POOL], netmap_mem_bufsize(n), _pos, _index)
1124 
1125 
1126 #if 0 /* currently unused */
1127 /* Return the index associated to the given packet buffer */
1128 #define netmap_buf_index(n, v)						\
1129     (netmap_obj_offset(&(n)->pools[NETMAP_BUF_POOL], (v)) / NETMAP_BDG_BUF_SIZE(n))
1130 #endif
1131 
1132 /*
1133  * allocate extra buffers in a linked list.
1134  * returns the actual number.
1135  */
1136 uint32_t
1137 netmap_extra_alloc(struct netmap_adapter *na, uint32_t *head, uint32_t n)
1138 {
1139 	struct netmap_mem_d *nmd = na->nm_mem;
1140 	uint32_t i, pos = 0; /* opaque, scan position in the bitmap */
1141 
1142 	NMA_LOCK(nmd);
1143 
1144 	*head = 0;	/* default, 'null' index ie empty list */
1145 	for (i = 0 ; i < n; i++) {
1146 		uint32_t cur = *head;	/* save current head */
1147 		uint32_t *p = netmap_buf_malloc(nmd, &pos, head);
1148 		if (p == NULL) {
1149 			nm_prerr("no more buffers after %d of %d", i, n);
1150 			*head = cur; /* restore */
1151 			break;
1152 		}
1153 		nm_prdis(5, "allocate buffer %d -> %d", *head, cur);
1154 		*p = cur; /* link to previous head */
1155 	}
1156 
1157 	NMA_UNLOCK(nmd);
1158 
1159 	return i;
1160 }
1161 
1162 static void
1163 netmap_extra_free(struct netmap_adapter *na, uint32_t head)
1164 {
1165 	struct lut_entry *lut = na->na_lut.lut;
1166 	struct netmap_mem_d *nmd = na->nm_mem;
1167 	struct netmap_obj_pool *p = &nmd->pools[NETMAP_BUF_POOL];
1168 	uint32_t i, cur, *buf;
1169 
1170 	nm_prdis("freeing the extra list");
1171 	for (i = 0; head >=2 && head < p->objtotal; i++) {
1172 		cur = head;
1173 		buf = lut[head].vaddr;
1174 		head = *buf;
1175 		*buf = 0;
1176 		if (netmap_obj_free(p, cur))
1177 			break;
1178 	}
1179 	if (head != 0)
1180 		nm_prerr("breaking with head %d", head);
1181 	if (netmap_debug & NM_DEBUG_MEM)
1182 		nm_prinf("freed %d buffers", i);
1183 }
1184 
1185 
1186 /* Return nonzero on error */
1187 static int
1188 netmap_new_bufs(struct netmap_mem_d *nmd, struct netmap_slot *slot, u_int n)
1189 {
1190 	struct netmap_obj_pool *p = &nmd->pools[NETMAP_BUF_POOL];
1191 	u_int i = 0;	/* slot counter */
1192 	uint32_t pos = 0;	/* slot in p->bitmap */
1193 	uint32_t index = 0;	/* buffer index */
1194 
1195 	for (i = 0; i < n; i++) {
1196 		void *vaddr = netmap_buf_malloc(nmd, &pos, &index);
1197 		if (vaddr == NULL) {
1198 			nm_prerr("no more buffers after %d of %d", i, n);
1199 			goto cleanup;
1200 		}
1201 		slot[i].buf_idx = index;
1202 		slot[i].len = p->_objsize;
1203 		slot[i].flags = 0;
1204 		slot[i].ptr = 0;
1205 	}
1206 
1207 	nm_prdis("%s: allocated %d buffers, %d available, first at %d", p->name, n, p->objfree, pos);
1208 	return (0);
1209 
1210 cleanup:
1211 	while (i > 0) {
1212 		i--;
1213 		netmap_obj_free(p, slot[i].buf_idx);
1214 	}
1215 	bzero(slot, n * sizeof(slot[0]));
1216 	return (ENOMEM);
1217 }
1218 
1219 static void
1220 netmap_mem_set_ring(struct netmap_mem_d *nmd, struct netmap_slot *slot, u_int n, uint32_t index)
1221 {
1222 	struct netmap_obj_pool *p = &nmd->pools[NETMAP_BUF_POOL];
1223 	u_int i;
1224 
1225 	for (i = 0; i < n; i++) {
1226 		slot[i].buf_idx = index;
1227 		slot[i].len = p->_objsize;
1228 		slot[i].flags = 0;
1229 	}
1230 }
1231 
1232 
1233 static void
1234 netmap_free_buf(struct netmap_mem_d *nmd, uint32_t i)
1235 {
1236 	struct netmap_obj_pool *p = &nmd->pools[NETMAP_BUF_POOL];
1237 
1238 	if (i < 2 || i >= p->objtotal) {
1239 		nm_prerr("Cannot free buf#%d: should be in [2, %d[", i, p->objtotal);
1240 		return;
1241 	}
1242 	netmap_obj_free(p, i);
1243 }
1244 
1245 
1246 static void
1247 netmap_free_bufs(struct netmap_mem_d *nmd, struct netmap_slot *slot, u_int n)
1248 {
1249 	u_int i;
1250 
1251 	for (i = 0; i < n; i++) {
1252 		if (slot[i].buf_idx > 1)
1253 			netmap_free_buf(nmd, slot[i].buf_idx);
1254 	}
1255 	nm_prdis("%s: released some buffers, available: %u",
1256 			p->name, p->objfree);
1257 }
1258 
1259 static void
1260 netmap_reset_obj_allocator(struct netmap_obj_pool *p)
1261 {
1262 
1263 	if (p == NULL)
1264 		return;
1265 	if (p->bitmap)
1266 		nm_os_free(p->bitmap);
1267 	p->bitmap = NULL;
1268 	if (p->invalid_bitmap)
1269 		nm_os_free(p->invalid_bitmap);
1270 	p->invalid_bitmap = NULL;
1271 	if (!p->alloc_done) {
1272 		/* allocation was done by somebody else.
1273 		 * Let them clean up after themselves.
1274 		 */
1275 		return;
1276 	}
1277 	if (p->lut) {
1278 		u_int i;
1279 
1280 		/*
1281 		 * Free each cluster allocated in
1282 		 * netmap_finalize_obj_allocator().  The cluster start
1283 		 * addresses are stored at multiples of p->_clusterentries
1284 		 * in the lut.
1285 		 */
1286 		for (i = 0; i < p->objtotal; i += p->_clustentries) {
1287 			contigfree(p->lut[i].vaddr, p->_clustsize, M_NETMAP);
1288 		}
1289 		nm_free_lut(p->lut, p->objtotal);
1290 	}
1291 	p->lut = NULL;
1292 	p->objtotal = 0;
1293 	p->memtotal = 0;
1294 	p->numclusters = 0;
1295 	p->objfree = 0;
1296 	p->alloc_done = 0;
1297 }
1298 
1299 /*
1300  * Free all resources related to an allocator.
1301  */
1302 static void
1303 netmap_destroy_obj_allocator(struct netmap_obj_pool *p)
1304 {
1305 	if (p == NULL)
1306 		return;
1307 	netmap_reset_obj_allocator(p);
1308 }
1309 
1310 /*
1311  * We receive a request for objtotal objects, of size objsize each.
1312  * Internally we may round up both numbers, as we allocate objects
1313  * in small clusters multiple of the page size.
1314  * We need to keep track of objtotal and clustentries,
1315  * as they are needed when freeing memory.
1316  *
1317  * XXX note -- userspace needs the buffers to be contiguous,
1318  *	so we cannot afford gaps at the end of a cluster.
1319  */
1320 
1321 
1322 /* call with NMA_LOCK held */
1323 static int
1324 netmap_config_obj_allocator(struct netmap_obj_pool *p, u_int objtotal, u_int objsize)
1325 {
1326 	int i;
1327 	u_int clustsize;	/* the cluster size, multiple of page size */
1328 	u_int clustentries;	/* how many objects per entry */
1329 
1330 	/* we store the current request, so we can
1331 	 * detect configuration changes later */
1332 	p->r_objtotal = objtotal;
1333 	p->r_objsize = objsize;
1334 
1335 #define MAX_CLUSTSIZE	(1<<22)		// 4 MB
1336 #define LINE_ROUND	NM_BUF_ALIGN	// 64
1337 	if (objsize >= MAX_CLUSTSIZE) {
1338 		/* we could do it but there is no point */
1339 		nm_prerr("unsupported allocation for %d bytes", objsize);
1340 		return EINVAL;
1341 	}
1342 	/* make sure objsize is a multiple of LINE_ROUND */
1343 	i = (objsize & (LINE_ROUND - 1));
1344 	if (i) {
1345 		nm_prinf("aligning object by %d bytes", LINE_ROUND - i);
1346 		objsize += LINE_ROUND - i;
1347 	}
1348 	if (objsize < p->objminsize || objsize > p->objmaxsize) {
1349 		nm_prerr("requested objsize %d out of range [%d, %d]",
1350 			objsize, p->objminsize, p->objmaxsize);
1351 		return EINVAL;
1352 	}
1353 	if (objtotal < p->nummin || objtotal > p->nummax) {
1354 		nm_prerr("requested objtotal %d out of range [%d, %d]",
1355 			objtotal, p->nummin, p->nummax);
1356 		return EINVAL;
1357 	}
1358 	/*
1359 	 * Compute number of objects using a brute-force approach:
1360 	 * given a max cluster size,
1361 	 * we try to fill it with objects keeping track of the
1362 	 * wasted space to the next page boundary.
1363 	 */
1364 	for (clustentries = 0, i = 1;; i++) {
1365 		u_int delta, used = i * objsize;
1366 		if (used > MAX_CLUSTSIZE)
1367 			break;
1368 		delta = used % PAGE_SIZE;
1369 		if (delta == 0) { // exact solution
1370 			clustentries = i;
1371 			break;
1372 		}
1373 	}
1374 	/* exact solution not found */
1375 	if (clustentries == 0) {
1376 		nm_prerr("unsupported allocation for %d bytes", objsize);
1377 		return EINVAL;
1378 	}
1379 	/* compute clustsize */
1380 	clustsize = clustentries * objsize;
1381 	if (netmap_debug & NM_DEBUG_MEM)
1382 		nm_prinf("objsize %d clustsize %d objects %d",
1383 			objsize, clustsize, clustentries);
1384 
1385 	/*
1386 	 * The number of clusters is n = ceil(objtotal/clustentries)
1387 	 * objtotal' = n * clustentries
1388 	 */
1389 	p->_clustentries = clustentries;
1390 	p->_clustsize = clustsize;
1391 	p->_numclusters = (objtotal + clustentries - 1) / clustentries;
1392 
1393 	/* actual values (may be larger than requested) */
1394 	p->_objsize = objsize;
1395 	p->_objtotal = p->_numclusters * clustentries;
1396 
1397 	return 0;
1398 }
1399 
1400 /* call with NMA_LOCK held */
1401 static int
1402 netmap_finalize_obj_allocator(struct netmap_obj_pool *p)
1403 {
1404 	int i; /* must be signed */
1405 	size_t n;
1406 
1407 	if (p->lut) {
1408 		/* if the lut is already there we assume that also all the
1409 		 * clusters have already been allocated, possibly by somebody
1410 		 * else (e.g., extmem). In the latter case, the alloc_done flag
1411 		 * will remain at zero, so that we will not attempt to
1412 		 * deallocate the clusters by ourselves in
1413 		 * netmap_reset_obj_allocator.
1414 		 */
1415 		return 0;
1416 	}
1417 
1418 	/* optimistically assume we have enough memory */
1419 	p->numclusters = p->_numclusters;
1420 	p->objtotal = p->_objtotal;
1421 	p->alloc_done = 1;
1422 
1423 	p->lut = nm_alloc_lut(p->objtotal);
1424 	if (p->lut == NULL) {
1425 		nm_prerr("Unable to create lookup table for '%s'", p->name);
1426 		goto clean;
1427 	}
1428 
1429 	/*
1430 	 * Allocate clusters, init pointers
1431 	 */
1432 
1433 	n = p->_clustsize;
1434 	for (i = 0; i < (int)p->objtotal;) {
1435 		int lim = i + p->_clustentries;
1436 		char *clust;
1437 
1438 		/*
1439 		 * XXX Note, we only need contigmalloc() for buffers attached
1440 		 * to native interfaces. In all other cases (nifp, netmap rings
1441 		 * and even buffers for VALE ports or emulated interfaces) we
1442 		 * can live with standard malloc, because the hardware will not
1443 		 * access the pages directly.
1444 		 */
1445 		clust = contigmalloc(n, M_NETMAP, M_NOWAIT | M_ZERO,
1446 		    (size_t)0, -1UL, PAGE_SIZE, 0);
1447 		if (clust == NULL) {
1448 			/*
1449 			 * If we get here, there is a severe memory shortage,
1450 			 * so halve the allocated memory to reclaim some.
1451 			 */
1452 			nm_prerr("Unable to create cluster at %d for '%s' allocator",
1453 			    i, p->name);
1454 			if (i < 2) /* nothing to halve */
1455 				goto out;
1456 			lim = i / 2;
1457 			for (i--; i >= lim; i--) {
1458 				if (i % p->_clustentries == 0 && p->lut[i].vaddr)
1459 					contigfree(p->lut[i].vaddr,
1460 						n, M_NETMAP);
1461 				p->lut[i].vaddr = NULL;
1462 			}
1463 		out:
1464 			p->objtotal = i;
1465 			/* we may have stopped in the middle of a cluster */
1466 			p->numclusters = (i + p->_clustentries - 1) / p->_clustentries;
1467 			break;
1468 		}
1469 		/*
1470 		 * Set lut state for all buffers in the current cluster.
1471 		 *
1472 		 * [i, lim) is the set of buffer indexes that cover the
1473 		 * current cluster.
1474 		 *
1475 		 * 'clust' is really the address of the current buffer in
1476 		 * the current cluster as we index through it with a stride
1477 		 * of p->_objsize.
1478 		 */
1479 		for (; i < lim; i++, clust += p->_objsize) {
1480 			p->lut[i].vaddr = clust;
1481 #if !defined(linux) && !defined(_WIN32)
1482 			p->lut[i].paddr = vtophys(clust);
1483 #endif
1484 		}
1485 	}
1486 	p->memtotal = (size_t)p->numclusters * (size_t)p->_clustsize;
1487 	if (netmap_verbose)
1488 		nm_prinf("Pre-allocated %d clusters (%d/%zuKB) for '%s'",
1489 		    p->numclusters, p->_clustsize >> 10,
1490 		    p->memtotal >> 10, p->name);
1491 
1492 	return 0;
1493 
1494 clean:
1495 	netmap_reset_obj_allocator(p);
1496 	return ENOMEM;
1497 }
1498 
1499 /* call with lock held */
1500 static int
1501 netmap_mem_params_changed(struct netmap_obj_params* p)
1502 {
1503 	int i, rv = 0;
1504 
1505 	for (i = 0; i < NETMAP_POOLS_NR; i++) {
1506 		if (p[i].last_size != p[i].size || p[i].last_num != p[i].num) {
1507 			p[i].last_size = p[i].size;
1508 			p[i].last_num = p[i].num;
1509 			rv = 1;
1510 		}
1511 	}
1512 	return rv;
1513 }
1514 
1515 static void
1516 netmap_mem_reset_all(struct netmap_mem_d *nmd)
1517 {
1518 	int i;
1519 
1520 	if (netmap_debug & NM_DEBUG_MEM)
1521 		nm_prinf("resetting %p", nmd);
1522 	for (i = 0; i < NETMAP_POOLS_NR; i++) {
1523 		netmap_reset_obj_allocator(&nmd->pools[i]);
1524 	}
1525 	nmd->flags  &= ~NETMAP_MEM_FINALIZED;
1526 }
1527 
1528 static int
1529 netmap_mem_unmap(struct netmap_obj_pool *p, struct netmap_adapter *na)
1530 {
1531 	int i, lim = p->objtotal;
1532 	struct netmap_lut *lut;
1533 	if (na == NULL || na->pdev == NULL)
1534 		return 0;
1535 
1536 	lut = &na->na_lut;
1537 
1538 
1539 
1540 #if defined(__FreeBSD__)
1541 	/* On FreeBSD mapping and unmapping is performed by the txsync
1542 	 * and rxsync routine, packet by packet. */
1543 	(void)i;
1544 	(void)lim;
1545 	(void)lut;
1546 #elif defined(_WIN32)
1547 	(void)i;
1548 	(void)lim;
1549 	(void)lut;
1550 	nm_prerr("unsupported on Windows");
1551 #else /* linux */
1552 	nm_prdis("unmapping and freeing plut for %s", na->name);
1553 	if (lut->plut == NULL || na->pdev == NULL)
1554 		return 0;
1555 	for (i = 0; i < lim; i += p->_clustentries) {
1556 		if (lut->plut[i].paddr)
1557 			netmap_unload_map(na, (bus_dma_tag_t) na->pdev, &lut->plut[i].paddr, p->_clustsize);
1558 	}
1559 	nm_free_plut(lut->plut);
1560 	lut->plut = NULL;
1561 #endif /* linux */
1562 
1563 	return 0;
1564 }
1565 
1566 static int
1567 netmap_mem_map(struct netmap_obj_pool *p, struct netmap_adapter *na)
1568 {
1569 	int error = 0;
1570 	int i, lim = p->objtotal;
1571 	struct netmap_lut *lut = &na->na_lut;
1572 
1573 	if (na->pdev == NULL)
1574 		return 0;
1575 
1576 #if defined(__FreeBSD__)
1577 	/* On FreeBSD mapping and unmapping is performed by the txsync
1578 	 * and rxsync routine, packet by packet. */
1579 	(void)i;
1580 	(void)lim;
1581 	(void)lut;
1582 #elif defined(_WIN32)
1583 	(void)i;
1584 	(void)lim;
1585 	(void)lut;
1586 	nm_prerr("unsupported on Windows");
1587 #else /* linux */
1588 
1589 	if (lut->plut != NULL) {
1590 		nm_prdis("plut already allocated for %s", na->name);
1591 		return 0;
1592 	}
1593 
1594 	nm_prdis("allocating physical lut for %s", na->name);
1595 	lut->plut = nm_alloc_plut(lim);
1596 	if (lut->plut == NULL) {
1597 		nm_prerr("Failed to allocate physical lut for %s", na->name);
1598 		return ENOMEM;
1599 	}
1600 
1601 	for (i = 0; i < lim; i += p->_clustentries) {
1602 		lut->plut[i].paddr = 0;
1603 	}
1604 
1605 	for (i = 0; i < lim; i += p->_clustentries) {
1606 		int j;
1607 
1608 		if (p->lut[i].vaddr == NULL)
1609 			continue;
1610 
1611 		error = netmap_load_map(na, (bus_dma_tag_t) na->pdev, &lut->plut[i].paddr,
1612 				p->lut[i].vaddr, p->_clustsize);
1613 		if (error) {
1614 			nm_prerr("Failed to map cluster #%d from the %s pool", i, p->name);
1615 			break;
1616 		}
1617 
1618 		for (j = 1; j < p->_clustentries; j++) {
1619 			lut->plut[i + j].paddr = lut->plut[i + j - 1].paddr + p->_objsize;
1620 		}
1621 	}
1622 
1623 	if (error)
1624 		netmap_mem_unmap(p, na);
1625 
1626 #endif /* linux */
1627 
1628 	return error;
1629 }
1630 
1631 static int
1632 netmap_mem_finalize_all(struct netmap_mem_d *nmd)
1633 {
1634 	int i;
1635 	if (nmd->flags & NETMAP_MEM_FINALIZED)
1636 		return 0;
1637 	nmd->lasterr = 0;
1638 	nmd->nm_totalsize = 0;
1639 	for (i = 0; i < NETMAP_POOLS_NR; i++) {
1640 		nmd->lasterr = netmap_finalize_obj_allocator(&nmd->pools[i]);
1641 		if (nmd->lasterr)
1642 			goto error;
1643 		nmd->nm_totalsize += nmd->pools[i].memtotal;
1644 	}
1645 	nmd->nm_totalsize = (nmd->nm_totalsize + PAGE_SIZE - 1) & ~(PAGE_SIZE - 1);
1646 	nmd->lasterr = netmap_mem_init_bitmaps(nmd);
1647 	if (nmd->lasterr)
1648 		goto error;
1649 
1650 	nmd->flags |= NETMAP_MEM_FINALIZED;
1651 
1652 	if (netmap_verbose)
1653 		nm_prinf("interfaces %zd KB, rings %zd KB, buffers %zd MB",
1654 		    nmd->pools[NETMAP_IF_POOL].memtotal >> 10,
1655 		    nmd->pools[NETMAP_RING_POOL].memtotal >> 10,
1656 		    nmd->pools[NETMAP_BUF_POOL].memtotal >> 20);
1657 
1658 	if (netmap_verbose)
1659 		nm_prinf("Free buffers: %d", nmd->pools[NETMAP_BUF_POOL].objfree);
1660 
1661 
1662 	return 0;
1663 error:
1664 	netmap_mem_reset_all(nmd);
1665 	return nmd->lasterr;
1666 }
1667 
1668 /*
1669  * allocator for private memory
1670  */
1671 static void *
1672 _netmap_mem_private_new(size_t size, struct netmap_obj_params *p, int grp_id,
1673 		struct netmap_mem_ops *ops, uint64_t memtotal, int *perr)
1674 {
1675 	struct netmap_mem_d *d = NULL;
1676 	int i, err = 0;
1677 	int checksz = 0;
1678 
1679 	/* if memtotal is !=0 we check that the request fits the available
1680 	 * memory. Moreover, any surprlus memory is assigned to buffers.
1681 	 */
1682 	checksz = (memtotal > 0);
1683 
1684 	d = nm_os_malloc(size);
1685 	if (d == NULL) {
1686 		err = ENOMEM;
1687 		goto error;
1688 	}
1689 
1690 	*d = nm_blueprint;
1691 	d->ops = ops;
1692 
1693 	err = nm_mem_assign_id(d, grp_id);
1694 	if (err)
1695 		goto error_free;
1696 	snprintf(d->name, NM_MEM_NAMESZ, "%d", d->nm_id);
1697 
1698 	for (i = 0; i < NETMAP_POOLS_NR; i++) {
1699 		snprintf(d->pools[i].name, NETMAP_POOL_MAX_NAMSZ,
1700 				nm_blueprint.pools[i].name,
1701 				d->name);
1702 		if (checksz) {
1703 			uint64_t poolsz = (uint64_t)p[i].num * p[i].size;
1704 			if (memtotal < poolsz) {
1705 				nm_prerr("%s: request too large", d->pools[i].name);
1706 				err = ENOMEM;
1707 				goto error_rel_id;
1708 			}
1709 			memtotal -= poolsz;
1710 		}
1711 		d->params[i].num = p[i].num;
1712 		d->params[i].size = p[i].size;
1713 	}
1714 	if (checksz && memtotal > 0) {
1715 		uint64_t sz = d->params[NETMAP_BUF_POOL].size;
1716 		uint64_t n = (memtotal + sz - 1) / sz;
1717 
1718 		if (n) {
1719 			if (netmap_verbose) {
1720 				nm_prinf("%s: adding %llu more buffers",
1721 				    d->pools[NETMAP_BUF_POOL].name,
1722 				    (unsigned long long)n);
1723 			}
1724 			d->params[NETMAP_BUF_POOL].num += n;
1725 		}
1726 	}
1727 
1728 	NMA_LOCK_INIT(d);
1729 
1730 	err = netmap_mem_config(d);
1731 	if (err)
1732 		goto error_destroy_lock;
1733 
1734 	d->flags &= ~NETMAP_MEM_FINALIZED;
1735 
1736 	return d;
1737 
1738 error_destroy_lock:
1739 	NMA_LOCK_DESTROY(d);
1740 error_rel_id:
1741 	nm_mem_release_id(d);
1742 error_free:
1743 	nm_os_free(d);
1744 error:
1745 	if (perr)
1746 		*perr = err;
1747 	return NULL;
1748 }
1749 
1750 struct netmap_mem_d *
1751 netmap_mem_private_new(u_int txr, u_int txd, u_int rxr, u_int rxd,
1752 		u_int extra_bufs, u_int npipes, int *perr)
1753 {
1754 	struct netmap_mem_d *d = NULL;
1755 	struct netmap_obj_params p[NETMAP_POOLS_NR];
1756 	int i;
1757 	u_int v, maxd;
1758 	/* account for the fake host rings */
1759 	txr++;
1760 	rxr++;
1761 
1762 	/* copy the min values */
1763 	for (i = 0; i < NETMAP_POOLS_NR; i++) {
1764 		p[i] = netmap_min_priv_params[i];
1765 	}
1766 
1767 	/* possibly increase them to fit user request */
1768 	v = sizeof(struct netmap_if) + sizeof(ssize_t) * (txr + rxr);
1769 	if (p[NETMAP_IF_POOL].size < v)
1770 		p[NETMAP_IF_POOL].size = v;
1771 	v = 2 + 4 * npipes;
1772 	if (p[NETMAP_IF_POOL].num < v)
1773 		p[NETMAP_IF_POOL].num = v;
1774 	maxd = (txd > rxd) ? txd : rxd;
1775 	v = sizeof(struct netmap_ring) + sizeof(struct netmap_slot) * maxd;
1776 	if (p[NETMAP_RING_POOL].size < v)
1777 		p[NETMAP_RING_POOL].size = v;
1778 	/* each pipe endpoint needs two tx rings (1 normal + 1 host, fake)
1779 	 * and two rx rings (again, 1 normal and 1 fake host)
1780 	 */
1781 	v = txr + rxr + 8 * npipes;
1782 	if (p[NETMAP_RING_POOL].num < v)
1783 		p[NETMAP_RING_POOL].num = v;
1784 	/* for each pipe we only need the buffers for the 4 "real" rings.
1785 	 * On the other end, the pipe ring dimension may be different from
1786 	 * the parent port ring dimension. As a compromise, we allocate twice the
1787 	 * space actually needed if the pipe rings were the same size as the parent rings
1788 	 */
1789 	v = (4 * npipes + rxr) * rxd + (4 * npipes + txr) * txd + 2 + extra_bufs;
1790 		/* the +2 is for the tx and rx fake buffers (indices 0 and 1) */
1791 	if (p[NETMAP_BUF_POOL].num < v)
1792 		p[NETMAP_BUF_POOL].num = v;
1793 
1794 	if (netmap_verbose)
1795 		nm_prinf("req if %d*%d ring %d*%d buf %d*%d",
1796 			p[NETMAP_IF_POOL].num,
1797 			p[NETMAP_IF_POOL].size,
1798 			p[NETMAP_RING_POOL].num,
1799 			p[NETMAP_RING_POOL].size,
1800 			p[NETMAP_BUF_POOL].num,
1801 			p[NETMAP_BUF_POOL].size);
1802 
1803 	d = _netmap_mem_private_new(sizeof(*d), p, -1, &netmap_mem_global_ops, 0, perr);
1804 
1805 	return d;
1806 }
1807 
1808 /* Reference iommu allocator - find existing or create new,
1809  * for not hw addapeters fallback to global allocator.
1810  */
1811 struct netmap_mem_d *
1812 netmap_mem_get_iommu(struct netmap_adapter *na)
1813 {
1814 	int i, err, grp_id;
1815 	struct netmap_mem_d *nmd;
1816 
1817 	if (na == NULL || na->pdev == NULL)
1818 		return netmap_mem_get(&nm_mem);
1819 
1820 	grp_id = nm_iommu_group_id(na->pdev);
1821 
1822 	NM_MTX_LOCK(nm_mem_list_lock);
1823 	nmd = netmap_last_mem_d;
1824 	do {
1825 		if (!(nmd->flags & NETMAP_MEM_HIDDEN) && nmd->nm_grp == grp_id) {
1826 			nmd->refcount++;
1827 			NM_DBG_REFC(nmd, __FUNCTION__, __LINE__);
1828 			NM_MTX_UNLOCK(nm_mem_list_lock);
1829 			return nmd;
1830 		}
1831 		nmd = nmd->next;
1832 	} while (nmd != netmap_last_mem_d);
1833 
1834 	nmd = nm_os_malloc(sizeof(*nmd));
1835 	if (nmd == NULL)
1836 		goto error;
1837 
1838 	*nmd = nm_mem_blueprint;
1839 
1840 	err = nm_mem_assign_id_locked(nmd, grp_id);
1841 	if (err)
1842 		goto error_free;
1843 
1844 	snprintf(nmd->name, sizeof(nmd->name), "%d", nmd->nm_id);
1845 
1846 	for (i = 0; i < NETMAP_POOLS_NR; i++) {
1847 		snprintf(nmd->pools[i].name, NETMAP_POOL_MAX_NAMSZ, "%s-%s",
1848 			nm_mem_blueprint.pools[i].name, nmd->name);
1849 	}
1850 
1851 	NMA_LOCK_INIT(nmd);
1852 
1853 	NM_MTX_UNLOCK(nm_mem_list_lock);
1854 	return nmd;
1855 
1856 error_free:
1857 	nm_os_free(nmd);
1858 error:
1859 	NM_MTX_UNLOCK(nm_mem_list_lock);
1860 	return NULL;
1861 }
1862 
1863 /* call with lock held */
1864 static int
1865 netmap_mem2_config(struct netmap_mem_d *nmd)
1866 {
1867 	int i;
1868 
1869 	if (!netmap_mem_params_changed(nmd->params))
1870 		goto out;
1871 
1872 	nm_prdis("reconfiguring");
1873 
1874 	if (nmd->flags & NETMAP_MEM_FINALIZED) {
1875 		/* reset previous allocation */
1876 		for (i = 0; i < NETMAP_POOLS_NR; i++) {
1877 			netmap_reset_obj_allocator(&nmd->pools[i]);
1878 		}
1879 		nmd->flags &= ~NETMAP_MEM_FINALIZED;
1880 	}
1881 
1882 	for (i = 0; i < NETMAP_POOLS_NR; i++) {
1883 		nmd->lasterr = netmap_config_obj_allocator(&nmd->pools[i],
1884 				nmd->params[i].num, nmd->params[i].size);
1885 		if (nmd->lasterr)
1886 			goto out;
1887 	}
1888 
1889 out:
1890 
1891 	return nmd->lasterr;
1892 }
1893 
1894 static int
1895 netmap_mem2_finalize(struct netmap_mem_d *nmd, struct netmap_adapter *na)
1896 {
1897 	if (nmd->flags & NETMAP_MEM_FINALIZED)
1898 		goto out;
1899 
1900 	if (netmap_mem_finalize_all(nmd))
1901 		goto out;
1902 
1903 	nmd->lasterr = 0;
1904 
1905 out:
1906 	return nmd->lasterr;
1907 }
1908 
1909 static void
1910 netmap_mem2_delete(struct netmap_mem_d *nmd)
1911 {
1912 	int i;
1913 
1914 	for (i = 0; i < NETMAP_POOLS_NR; i++) {
1915 	    netmap_destroy_obj_allocator(&nmd->pools[i]);
1916 	}
1917 
1918 	NMA_LOCK_DESTROY(nmd);
1919 	if (nmd != &nm_mem)
1920 		nm_os_free(nmd);
1921 }
1922 
1923 #ifdef WITH_EXTMEM
1924 /* doubly linekd list of all existing external allocators */
1925 static struct netmap_mem_ext *netmap_mem_ext_list = NULL;
1926 NM_MTX_T nm_mem_ext_list_lock;
1927 #endif /* WITH_EXTMEM */
1928 
1929 int
1930 netmap_mem_init(void)
1931 {
1932 	nm_mem_blueprint = nm_mem;
1933 	NM_MTX_INIT(nm_mem_list_lock);
1934 	NMA_LOCK_INIT(&nm_mem);
1935 	netmap_mem_get(&nm_mem);
1936 #ifdef WITH_EXTMEM
1937 	NM_MTX_INIT(nm_mem_ext_list_lock);
1938 #endif /* WITH_EXTMEM */
1939 	return (0);
1940 }
1941 
1942 void
1943 netmap_mem_fini(void)
1944 {
1945 	netmap_mem_put(&nm_mem);
1946 }
1947 
1948 static int
1949 netmap_mem_ring_needed(struct netmap_kring *kring)
1950 {
1951 	return kring->ring == NULL &&
1952 		(kring->users > 0 ||
1953 		 (kring->nr_kflags & NKR_NEEDRING));
1954 }
1955 
1956 static int
1957 netmap_mem_ring_todelete(struct netmap_kring *kring)
1958 {
1959 	return kring->ring != NULL &&
1960 		kring->users == 0 &&
1961 		!(kring->nr_kflags & NKR_NEEDRING);
1962 }
1963 
1964 
1965 /* call with NMA_LOCK held *
1966  *
1967  * Allocate netmap rings and buffers for this card
1968  * The rings are contiguous, but have variable size.
1969  * The kring array must follow the layout described
1970  * in netmap_krings_create().
1971  */
1972 static int
1973 netmap_mem2_rings_create(struct netmap_mem_d *nmd, struct netmap_adapter *na)
1974 {
1975 	enum txrx t;
1976 
1977 	for_rx_tx(t) {
1978 		u_int i;
1979 
1980 		for (i = 0; i < netmap_all_rings(na, t); i++) {
1981 			struct netmap_kring *kring = NMR(na, t)[i];
1982 			struct netmap_ring *ring = kring->ring;
1983 			u_int len, ndesc;
1984 
1985 			if (!netmap_mem_ring_needed(kring)) {
1986 				/* unneeded, or already created by somebody else */
1987 				if (netmap_debug & NM_DEBUG_MEM)
1988 					nm_prinf("NOT creating ring %s (ring %p, users %d neekring %d)",
1989 						kring->name, ring, kring->users, kring->nr_kflags & NKR_NEEDRING);
1990 				continue;
1991 			}
1992 			if (netmap_debug & NM_DEBUG_MEM)
1993 				nm_prinf("creating %s", kring->name);
1994 			ndesc = kring->nkr_num_slots;
1995 			len = sizeof(struct netmap_ring) +
1996 				  ndesc * sizeof(struct netmap_slot);
1997 			ring = netmap_ring_malloc(nmd, len);
1998 			if (ring == NULL) {
1999 				nm_prerr("Cannot allocate %s_ring", nm_txrx2str(t));
2000 				goto cleanup;
2001 			}
2002 			nm_prdis("txring at %p", ring);
2003 			kring->ring = ring;
2004 			*(uint32_t *)(uintptr_t)&ring->num_slots = ndesc;
2005 			*(int64_t *)(uintptr_t)&ring->buf_ofs =
2006 			    (nmd->pools[NETMAP_IF_POOL].memtotal +
2007 				nmd->pools[NETMAP_RING_POOL].memtotal) -
2008 				netmap_ring_offset(nmd, ring);
2009 
2010 			/* copy values from kring */
2011 			ring->head = kring->rhead;
2012 			ring->cur = kring->rcur;
2013 			ring->tail = kring->rtail;
2014 			*(uint32_t *)(uintptr_t)&ring->nr_buf_size =
2015 				netmap_mem_bufsize(nmd);
2016 			nm_prdis("%s h %d c %d t %d", kring->name,
2017 				ring->head, ring->cur, ring->tail);
2018 			nm_prdis("initializing slots for %s_ring", nm_txrx2str(t));
2019 			if (!(kring->nr_kflags & NKR_FAKERING)) {
2020 				/* this is a real ring */
2021 				if (netmap_debug & NM_DEBUG_MEM)
2022 					nm_prinf("allocating buffers for %s", kring->name);
2023 				if (netmap_new_bufs(nmd, ring->slot, ndesc)) {
2024 					nm_prerr("Cannot allocate buffers for %s_ring", nm_txrx2str(t));
2025 					goto cleanup;
2026 				}
2027 			} else {
2028 				/* this is a fake ring, set all indices to 0 */
2029 				if (netmap_debug & NM_DEBUG_MEM)
2030 					nm_prinf("NOT allocating buffers for %s", kring->name);
2031 				netmap_mem_set_ring(nmd, ring->slot, ndesc, 0);
2032 			}
2033 		        /* ring info */
2034 		        *(uint16_t *)(uintptr_t)&ring->ringid = kring->ring_id;
2035 		        *(uint16_t *)(uintptr_t)&ring->dir = kring->tx;
2036 		}
2037 	}
2038 
2039 	return 0;
2040 
2041 cleanup:
2042 	/* we cannot actually cleanup here, since we don't own kring->users
2043 	 * and kring->nr_klags & NKR_NEEDRING. The caller must decrement
2044 	 * the first or zero-out the second, then call netmap_free_rings()
2045 	 * to do the cleanup
2046 	 */
2047 
2048 	return ENOMEM;
2049 }
2050 
2051 static void
2052 netmap_mem2_rings_delete(struct netmap_mem_d *nmd, struct netmap_adapter *na)
2053 {
2054 	enum txrx t;
2055 
2056 	for_rx_tx(t) {
2057 		u_int i;
2058 		for (i = 0; i < netmap_all_rings(na, t); i++) {
2059 			struct netmap_kring *kring = NMR(na, t)[i];
2060 			struct netmap_ring *ring = kring->ring;
2061 
2062 			if (!netmap_mem_ring_todelete(kring)) {
2063 				if (netmap_debug & NM_DEBUG_MEM)
2064 					nm_prinf("NOT deleting ring %s (ring %p, users %d neekring %d)",
2065 						kring->name, ring, kring->users, kring->nr_kflags & NKR_NEEDRING);
2066 				continue;
2067 			}
2068 			if (netmap_debug & NM_DEBUG_MEM)
2069 				nm_prinf("deleting ring %s", kring->name);
2070 			if (!(kring->nr_kflags & NKR_FAKERING)) {
2071 				nm_prdis("freeing bufs for %s", kring->name);
2072 				netmap_free_bufs(nmd, ring->slot, kring->nkr_num_slots);
2073 			} else {
2074 				nm_prdis("NOT freeing bufs for %s", kring->name);
2075 			}
2076 			netmap_ring_free(nmd, ring);
2077 			kring->ring = NULL;
2078 		}
2079 	}
2080 }
2081 
2082 /* call with NMA_LOCK held */
2083 /*
2084  * Allocate the per-fd structure netmap_if.
2085  *
2086  * We assume that the configuration stored in na
2087  * (number of tx/rx rings and descs) does not change while
2088  * the interface is in netmap mode.
2089  */
2090 static struct netmap_if *
2091 netmap_mem2_if_new(struct netmap_mem_d *nmd,
2092 		struct netmap_adapter *na, struct netmap_priv_d *priv)
2093 {
2094 	struct netmap_if *nifp;
2095 	ssize_t base; /* handy for relative offsets between rings and nifp */
2096 	u_int i, len, n[NR_TXRX], ntot;
2097 	enum txrx t;
2098 
2099 	ntot = 0;
2100 	for_rx_tx(t) {
2101 		/* account for the (eventually fake) host rings */
2102 		n[t] = netmap_all_rings(na, t);
2103 		ntot += n[t];
2104 	}
2105 	/*
2106 	 * the descriptor is followed inline by an array of offsets
2107 	 * to the tx and rx rings in the shared memory region.
2108 	 */
2109 
2110 	len = sizeof(struct netmap_if) + (ntot * sizeof(ssize_t));
2111 	nifp = netmap_if_malloc(nmd, len);
2112 	if (nifp == NULL) {
2113 		return NULL;
2114 	}
2115 
2116 	/* initialize base fields -- override const */
2117 	*(u_int *)(uintptr_t)&nifp->ni_tx_rings = na->num_tx_rings;
2118 	*(u_int *)(uintptr_t)&nifp->ni_rx_rings = na->num_rx_rings;
2119 	*(u_int *)(uintptr_t)&nifp->ni_host_tx_rings =
2120 		(na->num_host_tx_rings ? na->num_host_tx_rings : 1);
2121 	*(u_int *)(uintptr_t)&nifp->ni_host_rx_rings =
2122 		(na->num_host_rx_rings ? na->num_host_rx_rings : 1);
2123 	strlcpy(nifp->ni_name, na->name, sizeof(nifp->ni_name));
2124 
2125 	/*
2126 	 * fill the slots for the rx and tx rings. They contain the offset
2127 	 * between the ring and nifp, so the information is usable in
2128 	 * userspace to reach the ring from the nifp.
2129 	 */
2130 	base = netmap_if_offset(nmd, nifp);
2131 	for (i = 0; i < n[NR_TX]; i++) {
2132 		/* XXX instead of ofs == 0 maybe use the offset of an error
2133 		 * ring, like we do for buffers? */
2134 		ssize_t ofs = 0;
2135 
2136 		if (na->tx_rings[i]->ring != NULL && i >= priv->np_qfirst[NR_TX]
2137 				&& i < priv->np_qlast[NR_TX]) {
2138 			ofs = netmap_ring_offset(nmd,
2139 						 na->tx_rings[i]->ring) - base;
2140 		}
2141 		*(ssize_t *)(uintptr_t)&nifp->ring_ofs[i] = ofs;
2142 	}
2143 	for (i = 0; i < n[NR_RX]; i++) {
2144 		/* XXX instead of ofs == 0 maybe use the offset of an error
2145 		 * ring, like we do for buffers? */
2146 		ssize_t ofs = 0;
2147 
2148 		if (na->rx_rings[i]->ring != NULL && i >= priv->np_qfirst[NR_RX]
2149 				&& i < priv->np_qlast[NR_RX]) {
2150 			ofs = netmap_ring_offset(nmd,
2151 						 na->rx_rings[i]->ring) - base;
2152 		}
2153 		*(ssize_t *)(uintptr_t)&nifp->ring_ofs[i+n[NR_TX]] = ofs;
2154 	}
2155 
2156 	return (nifp);
2157 }
2158 
2159 static void
2160 netmap_mem2_if_delete(struct netmap_mem_d *nmd,
2161 		struct netmap_adapter *na, struct netmap_if *nifp)
2162 {
2163 	if (nifp == NULL)
2164 		/* nothing to do */
2165 		return;
2166 	if (nifp->ni_bufs_head)
2167 		netmap_extra_free(na, nifp->ni_bufs_head);
2168 	netmap_if_free(nmd, nifp);
2169 }
2170 
2171 static void
2172 netmap_mem2_deref(struct netmap_mem_d *nmd, struct netmap_adapter *na)
2173 {
2174 
2175 	if (netmap_debug & NM_DEBUG_MEM)
2176 		nm_prinf("active = %d", nmd->active);
2177 
2178 }
2179 
2180 struct netmap_mem_ops netmap_mem_global_ops = {
2181 	.nmd_get_lut = netmap_mem2_get_lut,
2182 	.nmd_get_info = netmap_mem2_get_info,
2183 	.nmd_ofstophys = netmap_mem2_ofstophys,
2184 	.nmd_config = netmap_mem2_config,
2185 	.nmd_finalize = netmap_mem2_finalize,
2186 	.nmd_deref = netmap_mem2_deref,
2187 	.nmd_delete = netmap_mem2_delete,
2188 	.nmd_if_offset = netmap_mem2_if_offset,
2189 	.nmd_if_new = netmap_mem2_if_new,
2190 	.nmd_if_delete = netmap_mem2_if_delete,
2191 	.nmd_rings_create = netmap_mem2_rings_create,
2192 	.nmd_rings_delete = netmap_mem2_rings_delete
2193 };
2194 
2195 int
2196 netmap_mem_pools_info_get(struct nmreq_pools_info *req,
2197 				struct netmap_mem_d *nmd)
2198 {
2199 	int ret;
2200 
2201 	ret = netmap_mem_get_info(nmd, &req->nr_memsize, NULL,
2202 					&req->nr_mem_id);
2203 	if (ret) {
2204 		return ret;
2205 	}
2206 
2207 	NMA_LOCK(nmd);
2208 	req->nr_if_pool_offset = 0;
2209 	req->nr_if_pool_objtotal = nmd->pools[NETMAP_IF_POOL].objtotal;
2210 	req->nr_if_pool_objsize = nmd->pools[NETMAP_IF_POOL]._objsize;
2211 
2212 	req->nr_ring_pool_offset = nmd->pools[NETMAP_IF_POOL].memtotal;
2213 	req->nr_ring_pool_objtotal = nmd->pools[NETMAP_RING_POOL].objtotal;
2214 	req->nr_ring_pool_objsize = nmd->pools[NETMAP_RING_POOL]._objsize;
2215 
2216 	req->nr_buf_pool_offset = nmd->pools[NETMAP_IF_POOL].memtotal +
2217 			     nmd->pools[NETMAP_RING_POOL].memtotal;
2218 	req->nr_buf_pool_objtotal = nmd->pools[NETMAP_BUF_POOL].objtotal;
2219 	req->nr_buf_pool_objsize = nmd->pools[NETMAP_BUF_POOL]._objsize;
2220 	NMA_UNLOCK(nmd);
2221 
2222 	return 0;
2223 }
2224 
2225 #ifdef WITH_EXTMEM
2226 struct netmap_mem_ext {
2227 	struct netmap_mem_d up;
2228 
2229 	struct nm_os_extmem *os;
2230 	struct netmap_mem_ext *next, *prev;
2231 };
2232 
2233 /* call with nm_mem_list_lock held */
2234 static void
2235 netmap_mem_ext_register(struct netmap_mem_ext *e)
2236 {
2237 	NM_MTX_LOCK(nm_mem_ext_list_lock);
2238 	if (netmap_mem_ext_list)
2239 		netmap_mem_ext_list->prev = e;
2240 	e->next = netmap_mem_ext_list;
2241 	netmap_mem_ext_list = e;
2242 	e->prev = NULL;
2243 	NM_MTX_UNLOCK(nm_mem_ext_list_lock);
2244 }
2245 
2246 /* call with nm_mem_list_lock held */
2247 static void
2248 netmap_mem_ext_unregister(struct netmap_mem_ext *e)
2249 {
2250 	if (e->prev)
2251 		e->prev->next = e->next;
2252 	else
2253 		netmap_mem_ext_list = e->next;
2254 	if (e->next)
2255 		e->next->prev = e->prev;
2256 	e->prev = e->next = NULL;
2257 }
2258 
2259 static struct netmap_mem_ext *
2260 netmap_mem_ext_search(struct nm_os_extmem *os)
2261 {
2262 	struct netmap_mem_ext *e;
2263 
2264 	NM_MTX_LOCK(nm_mem_ext_list_lock);
2265 	for (e = netmap_mem_ext_list; e; e = e->next) {
2266 		if (nm_os_extmem_isequal(e->os, os)) {
2267 			netmap_mem_get(&e->up);
2268 			break;
2269 		}
2270 	}
2271 	NM_MTX_UNLOCK(nm_mem_ext_list_lock);
2272 	return e;
2273 }
2274 
2275 
2276 static void
2277 netmap_mem_ext_delete(struct netmap_mem_d *d)
2278 {
2279 	int i;
2280 	struct netmap_mem_ext *e =
2281 		(struct netmap_mem_ext *)d;
2282 
2283 	netmap_mem_ext_unregister(e);
2284 
2285 	for (i = 0; i < NETMAP_POOLS_NR; i++) {
2286 		struct netmap_obj_pool *p = &d->pools[i];
2287 
2288 		if (p->lut) {
2289 			nm_free_lut(p->lut, p->objtotal);
2290 			p->lut = NULL;
2291 		}
2292 	}
2293 	if (e->os)
2294 		nm_os_extmem_delete(e->os);
2295 	netmap_mem2_delete(d);
2296 }
2297 
2298 static int
2299 netmap_mem_ext_config(struct netmap_mem_d *nmd)
2300 {
2301 	return 0;
2302 }
2303 
2304 struct netmap_mem_ops netmap_mem_ext_ops = {
2305 	.nmd_get_lut = netmap_mem2_get_lut,
2306 	.nmd_get_info = netmap_mem2_get_info,
2307 	.nmd_ofstophys = netmap_mem2_ofstophys,
2308 	.nmd_config = netmap_mem_ext_config,
2309 	.nmd_finalize = netmap_mem2_finalize,
2310 	.nmd_deref = netmap_mem2_deref,
2311 	.nmd_delete = netmap_mem_ext_delete,
2312 	.nmd_if_offset = netmap_mem2_if_offset,
2313 	.nmd_if_new = netmap_mem2_if_new,
2314 	.nmd_if_delete = netmap_mem2_if_delete,
2315 	.nmd_rings_create = netmap_mem2_rings_create,
2316 	.nmd_rings_delete = netmap_mem2_rings_delete
2317 };
2318 
2319 struct netmap_mem_d *
2320 netmap_mem_ext_create(uint64_t usrptr, struct nmreq_pools_info *pi, int *perror)
2321 {
2322 	int error = 0;
2323 	int i, j;
2324 	struct netmap_mem_ext *nme;
2325 	char *clust;
2326 	size_t off;
2327 	struct nm_os_extmem *os = NULL;
2328 	int nr_pages;
2329 
2330 	// XXX sanity checks
2331 	if (pi->nr_if_pool_objtotal == 0)
2332 		pi->nr_if_pool_objtotal = netmap_min_priv_params[NETMAP_IF_POOL].num;
2333 	if (pi->nr_if_pool_objsize == 0)
2334 		pi->nr_if_pool_objsize = netmap_min_priv_params[NETMAP_IF_POOL].size;
2335 	if (pi->nr_ring_pool_objtotal == 0)
2336 		pi->nr_ring_pool_objtotal = netmap_min_priv_params[NETMAP_RING_POOL].num;
2337 	if (pi->nr_ring_pool_objsize == 0)
2338 		pi->nr_ring_pool_objsize = netmap_min_priv_params[NETMAP_RING_POOL].size;
2339 	if (pi->nr_buf_pool_objtotal == 0)
2340 		pi->nr_buf_pool_objtotal = netmap_min_priv_params[NETMAP_BUF_POOL].num;
2341 	if (pi->nr_buf_pool_objsize == 0)
2342 		pi->nr_buf_pool_objsize = netmap_min_priv_params[NETMAP_BUF_POOL].size;
2343 	if (netmap_verbose & NM_DEBUG_MEM)
2344 		nm_prinf("if %d %d ring %d %d buf %d %d",
2345 			pi->nr_if_pool_objtotal, pi->nr_if_pool_objsize,
2346 			pi->nr_ring_pool_objtotal, pi->nr_ring_pool_objsize,
2347 			pi->nr_buf_pool_objtotal, pi->nr_buf_pool_objsize);
2348 
2349 	os = nm_os_extmem_create(usrptr, pi, &error);
2350 	if (os == NULL) {
2351 		nm_prerr("os extmem creation failed");
2352 		goto out;
2353 	}
2354 
2355 	nme = netmap_mem_ext_search(os);
2356 	if (nme) {
2357 		nm_os_extmem_delete(os);
2358 		return &nme->up;
2359 	}
2360 	if (netmap_verbose & NM_DEBUG_MEM)
2361 		nm_prinf("not found, creating new");
2362 
2363 	nme = _netmap_mem_private_new(sizeof(*nme),
2364 
2365 			(struct netmap_obj_params[]){
2366 				{ pi->nr_if_pool_objsize, pi->nr_if_pool_objtotal },
2367 				{ pi->nr_ring_pool_objsize, pi->nr_ring_pool_objtotal },
2368 				{ pi->nr_buf_pool_objsize, pi->nr_buf_pool_objtotal }},
2369 			-1,
2370 			&netmap_mem_ext_ops,
2371 			pi->nr_memsize,
2372 			&error);
2373 	if (nme == NULL)
2374 		goto out_unmap;
2375 
2376 	nr_pages = nm_os_extmem_nr_pages(os);
2377 
2378 	/* from now on pages will be released by nme destructor;
2379 	 * we let res = 0 to prevent release in out_unmap below
2380 	 */
2381 	nme->os = os;
2382 	os = NULL; /* pass ownership */
2383 
2384 	clust = nm_os_extmem_nextpage(nme->os);
2385 	off = 0;
2386 	for (i = 0; i < NETMAP_POOLS_NR; i++) {
2387 		struct netmap_obj_pool *p = &nme->up.pools[i];
2388 		struct netmap_obj_params *o = &nme->up.params[i];
2389 
2390 		p->_objsize = o->size;
2391 		p->_clustsize = o->size;
2392 		p->_clustentries = 1;
2393 
2394 		p->lut = nm_alloc_lut(o->num);
2395 		if (p->lut == NULL) {
2396 			error = ENOMEM;
2397 			goto out_delete;
2398 		}
2399 
2400 		p->bitmap_slots = (o->num + sizeof(uint32_t) - 1) / sizeof(uint32_t);
2401 		p->invalid_bitmap = nm_os_malloc(sizeof(uint32_t) * p->bitmap_slots);
2402 		if (p->invalid_bitmap == NULL) {
2403 			error = ENOMEM;
2404 			goto out_delete;
2405 		}
2406 
2407 		if (nr_pages == 0) {
2408 			p->objtotal = 0;
2409 			p->memtotal = 0;
2410 			p->objfree = 0;
2411 			continue;
2412 		}
2413 
2414 		for (j = 0; j < o->num && nr_pages > 0; j++) {
2415 			size_t noff;
2416 
2417 			p->lut[j].vaddr = clust + off;
2418 #if !defined(linux) && !defined(_WIN32)
2419 			p->lut[j].paddr = vtophys(p->lut[j].vaddr);
2420 #endif
2421 			nm_prdis("%s %d at %p", p->name, j, p->lut[j].vaddr);
2422 			noff = off + p->_objsize;
2423 			if (noff < PAGE_SIZE) {
2424 				off = noff;
2425 				continue;
2426 			}
2427 			nm_prdis("too big, recomputing offset...");
2428 			while (noff >= PAGE_SIZE) {
2429 				char *old_clust = clust;
2430 				noff -= PAGE_SIZE;
2431 				clust = nm_os_extmem_nextpage(nme->os);
2432 				nr_pages--;
2433 				nm_prdis("noff %zu page %p nr_pages %d", noff,
2434 						page_to_virt(*pages), nr_pages);
2435 				if (noff > 0 && !nm_isset(p->invalid_bitmap, j) &&
2436 					(nr_pages == 0 ||
2437 					 old_clust + PAGE_SIZE != clust))
2438 				{
2439 					/* out of space or non contiguous,
2440 					 * drop this object
2441 					 * */
2442 					p->invalid_bitmap[ (j>>5) ] |= 1U << (j & 31U);
2443 					nm_prdis("non contiguous at off %zu, drop", noff);
2444 				}
2445 				if (nr_pages == 0)
2446 					break;
2447 			}
2448 			off = noff;
2449 		}
2450 		p->objtotal = j;
2451 		p->numclusters = p->objtotal;
2452 		p->memtotal = j * (size_t)p->_objsize;
2453 		nm_prdis("%d memtotal %zu", j, p->memtotal);
2454 	}
2455 
2456 	netmap_mem_ext_register(nme);
2457 
2458 	return &nme->up;
2459 
2460 out_delete:
2461 	netmap_mem_put(&nme->up);
2462 out_unmap:
2463 	if (os)
2464 		nm_os_extmem_delete(os);
2465 out:
2466 	if (perror)
2467 		*perror = error;
2468 	return NULL;
2469 
2470 }
2471 #endif /* WITH_EXTMEM */
2472 
2473 
2474 #ifdef WITH_PTNETMAP
2475 struct mem_pt_if {
2476 	struct mem_pt_if *next;
2477 	if_t ifp;
2478 	unsigned int nifp_offset;
2479 };
2480 
2481 /* Netmap allocator for ptnetmap guests. */
2482 struct netmap_mem_ptg {
2483 	struct netmap_mem_d up;
2484 
2485 	vm_paddr_t nm_paddr;            /* physical address in the guest */
2486 	void *nm_addr;                  /* virtual address in the guest */
2487 	struct netmap_lut buf_lut;      /* lookup table for BUF pool in the guest */
2488 	nm_memid_t host_mem_id;         /* allocator identifier in the host */
2489 	struct ptnetmap_memdev *ptn_dev;/* ptnetmap memdev */
2490 	struct mem_pt_if *pt_ifs;	/* list of interfaces in passthrough */
2491 };
2492 
2493 /* Link a passthrough interface to a passthrough netmap allocator. */
2494 static int
2495 netmap_mem_pt_guest_ifp_add(struct netmap_mem_d *nmd, if_t ifp,
2496 			    unsigned int nifp_offset)
2497 {
2498 	struct netmap_mem_ptg *ptnmd = (struct netmap_mem_ptg *)nmd;
2499 	struct mem_pt_if *ptif = nm_os_malloc(sizeof(*ptif));
2500 
2501 	if (!ptif) {
2502 		return ENOMEM;
2503 	}
2504 
2505 	NMA_LOCK(nmd);
2506 
2507 	ptif->ifp = ifp;
2508 	ptif->nifp_offset = nifp_offset;
2509 
2510 	if (ptnmd->pt_ifs) {
2511 		ptif->next = ptnmd->pt_ifs;
2512 	}
2513 	ptnmd->pt_ifs = ptif;
2514 
2515 	NMA_UNLOCK(nmd);
2516 
2517 	nm_prinf("ifp=%s,nifp_offset=%u",
2518 		if_name(ptif->ifp), ptif->nifp_offset);
2519 
2520 	return 0;
2521 }
2522 
2523 /* Called with NMA_LOCK(nmd) held. */
2524 static struct mem_pt_if *
2525 netmap_mem_pt_guest_ifp_lookup(struct netmap_mem_d *nmd, if_t ifp)
2526 {
2527 	struct netmap_mem_ptg *ptnmd = (struct netmap_mem_ptg *)nmd;
2528 	struct mem_pt_if *curr;
2529 
2530 	for (curr = ptnmd->pt_ifs; curr; curr = curr->next) {
2531 		if (curr->ifp == ifp) {
2532 			return curr;
2533 		}
2534 	}
2535 
2536 	return NULL;
2537 }
2538 
2539 /* Unlink a passthrough interface from a passthrough netmap allocator. */
2540 int
2541 netmap_mem_pt_guest_ifp_del(struct netmap_mem_d *nmd, if_t ifp)
2542 {
2543 	struct netmap_mem_ptg *ptnmd = (struct netmap_mem_ptg *)nmd;
2544 	struct mem_pt_if *prev = NULL;
2545 	struct mem_pt_if *curr;
2546 	int ret = -1;
2547 
2548 	NMA_LOCK(nmd);
2549 
2550 	for (curr = ptnmd->pt_ifs; curr; curr = curr->next) {
2551 		if (curr->ifp == ifp) {
2552 			if (prev) {
2553 				prev->next = curr->next;
2554 			} else {
2555 				ptnmd->pt_ifs = curr->next;
2556 			}
2557 			nm_prinf("removed (ifp=%s,nifp_offset=%u)",
2558 			  if_name(curr->ifp), curr->nifp_offset);
2559 			nm_os_free(curr);
2560 			ret = 0;
2561 			break;
2562 		}
2563 		prev = curr;
2564 	}
2565 
2566 	NMA_UNLOCK(nmd);
2567 
2568 	return ret;
2569 }
2570 
2571 static int
2572 netmap_mem_pt_guest_get_lut(struct netmap_mem_d *nmd, struct netmap_lut *lut)
2573 {
2574 	struct netmap_mem_ptg *ptnmd = (struct netmap_mem_ptg *)nmd;
2575 
2576 	if (!(nmd->flags & NETMAP_MEM_FINALIZED)) {
2577 		return EINVAL;
2578 	}
2579 
2580 	*lut = ptnmd->buf_lut;
2581 	return 0;
2582 }
2583 
2584 static int
2585 netmap_mem_pt_guest_get_info(struct netmap_mem_d *nmd, uint64_t *size,
2586 			     u_int *memflags, uint16_t *id)
2587 {
2588 	int error = 0;
2589 
2590 	error = nmd->ops->nmd_config(nmd);
2591 	if (error)
2592 		goto out;
2593 
2594 	if (size)
2595 		*size = nmd->nm_totalsize;
2596 	if (memflags)
2597 		*memflags = nmd->flags;
2598 	if (id)
2599 		*id = nmd->nm_id;
2600 
2601 out:
2602 
2603 	return error;
2604 }
2605 
2606 static vm_paddr_t
2607 netmap_mem_pt_guest_ofstophys(struct netmap_mem_d *nmd, vm_ooffset_t off)
2608 {
2609 	struct netmap_mem_ptg *ptnmd = (struct netmap_mem_ptg *)nmd;
2610 	vm_paddr_t paddr;
2611 	/* if the offset is valid, just return csb->base_addr + off */
2612 	paddr = (vm_paddr_t)(ptnmd->nm_paddr + off);
2613 	nm_prdis("off %lx padr %lx", off, (unsigned long)paddr);
2614 	return paddr;
2615 }
2616 
2617 static int
2618 netmap_mem_pt_guest_config(struct netmap_mem_d *nmd)
2619 {
2620 	/* nothing to do, we are configured on creation
2621 	 * and configuration never changes thereafter
2622 	 */
2623 	return 0;
2624 }
2625 
2626 static int
2627 netmap_mem_pt_guest_finalize(struct netmap_mem_d *nmd, struct netmap_adapter *na)
2628 {
2629 	struct netmap_mem_ptg *ptnmd = (struct netmap_mem_ptg *)nmd;
2630 	uint64_t mem_size;
2631 	uint32_t bufsize;
2632 	uint32_t nbuffers;
2633 	uint32_t poolofs;
2634 	vm_paddr_t paddr;
2635 	char *vaddr;
2636 	int i;
2637 	int error = 0;
2638 
2639 	if (nmd->flags & NETMAP_MEM_FINALIZED)
2640 		goto out;
2641 
2642 	if (ptnmd->ptn_dev == NULL) {
2643 		nm_prerr("ptnetmap memdev not attached");
2644 		error = ENOMEM;
2645 		goto out;
2646 	}
2647 	/* Map memory through ptnetmap-memdev BAR. */
2648 	error = nm_os_pt_memdev_iomap(ptnmd->ptn_dev, &ptnmd->nm_paddr,
2649 				      &ptnmd->nm_addr, &mem_size);
2650 	if (error)
2651 		goto out;
2652 
2653 	/* Initialize the lut using the information contained in the
2654 	 * ptnetmap memory device. */
2655 	bufsize = nm_os_pt_memdev_ioread(ptnmd->ptn_dev,
2656 					 PTNET_MDEV_IO_BUF_POOL_OBJSZ);
2657 	nbuffers = nm_os_pt_memdev_ioread(ptnmd->ptn_dev,
2658 					 PTNET_MDEV_IO_BUF_POOL_OBJNUM);
2659 
2660 	/* allocate the lut */
2661 	if (ptnmd->buf_lut.lut == NULL) {
2662 		nm_prinf("allocating lut");
2663 		ptnmd->buf_lut.lut = nm_alloc_lut(nbuffers);
2664 		if (ptnmd->buf_lut.lut == NULL) {
2665 			nm_prerr("lut allocation failed");
2666 			return ENOMEM;
2667 		}
2668 	}
2669 
2670 	/* we have physically contiguous memory mapped through PCI BAR */
2671 	poolofs = nm_os_pt_memdev_ioread(ptnmd->ptn_dev,
2672 					 PTNET_MDEV_IO_BUF_POOL_OFS);
2673 	vaddr = (char *)(ptnmd->nm_addr) + poolofs;
2674 	paddr = ptnmd->nm_paddr + poolofs;
2675 
2676 	for (i = 0; i < nbuffers; i++) {
2677 		ptnmd->buf_lut.lut[i].vaddr = vaddr;
2678 		vaddr += bufsize;
2679 		paddr += bufsize;
2680 	}
2681 
2682 	ptnmd->buf_lut.objtotal = nbuffers;
2683 	ptnmd->buf_lut.objsize = bufsize;
2684 	nmd->nm_totalsize = mem_size;
2685 
2686 	/* Initialize these fields as are needed by
2687 	 * netmap_mem_bufsize().
2688 	 * XXX please improve this, why do we need this
2689 	 * replication? maybe we nmd->pools[] should no be
2690 	 * there for the guest allocator? */
2691 	nmd->pools[NETMAP_BUF_POOL]._objsize = bufsize;
2692 	nmd->pools[NETMAP_BUF_POOL]._objtotal = nbuffers;
2693 
2694 	nmd->flags |= NETMAP_MEM_FINALIZED;
2695 out:
2696 	return error;
2697 }
2698 
2699 static void
2700 netmap_mem_pt_guest_deref(struct netmap_mem_d *nmd, struct netmap_adapter *na)
2701 {
2702 	struct netmap_mem_ptg *ptnmd = (struct netmap_mem_ptg *)nmd;
2703 
2704 	if (nmd->active == 1 &&
2705 		(nmd->flags & NETMAP_MEM_FINALIZED)) {
2706 	    nmd->flags  &= ~NETMAP_MEM_FINALIZED;
2707 	    /* unmap ptnetmap-memdev memory */
2708 	    if (ptnmd->ptn_dev) {
2709 		nm_os_pt_memdev_iounmap(ptnmd->ptn_dev);
2710 	    }
2711 	    ptnmd->nm_addr = NULL;
2712 	    ptnmd->nm_paddr = 0;
2713 	}
2714 }
2715 
2716 static ssize_t
2717 netmap_mem_pt_guest_if_offset(struct netmap_mem_d *nmd, const void *vaddr)
2718 {
2719 	struct netmap_mem_ptg *ptnmd = (struct netmap_mem_ptg *)nmd;
2720 
2721 	return (const char *)(vaddr) - (char *)(ptnmd->nm_addr);
2722 }
2723 
2724 static void
2725 netmap_mem_pt_guest_delete(struct netmap_mem_d *nmd)
2726 {
2727 	if (nmd == NULL)
2728 		return;
2729 	if (netmap_verbose)
2730 		nm_prinf("deleting %p", nmd);
2731 	if (nmd->active > 0)
2732 		nm_prerr("bug: deleting mem allocator with active=%d!", nmd->active);
2733 	if (netmap_verbose)
2734 		nm_prinf("done deleting %p", nmd);
2735 	NMA_LOCK_DESTROY(nmd);
2736 	nm_os_free(nmd);
2737 }
2738 
2739 static struct netmap_if *
2740 netmap_mem_pt_guest_if_new(struct netmap_mem_d *nmd,
2741 		struct netmap_adapter *na, struct netmap_priv_d *priv)
2742 {
2743 	struct netmap_mem_ptg *ptnmd = (struct netmap_mem_ptg *)nmd;
2744 	struct mem_pt_if *ptif;
2745 	struct netmap_if *nifp = NULL;
2746 
2747 	ptif = netmap_mem_pt_guest_ifp_lookup(nmd, na->ifp);
2748 	if (ptif == NULL) {
2749 		nm_prerr("interface %s is not in passthrough", na->name);
2750 		goto out;
2751 	}
2752 
2753 	nifp = (struct netmap_if *)((char *)(ptnmd->nm_addr) +
2754 				    ptif->nifp_offset);
2755 out:
2756 	return nifp;
2757 }
2758 
2759 static void
2760 netmap_mem_pt_guest_if_delete(struct netmap_mem_d * nmd,
2761 		struct netmap_adapter *na, struct netmap_if *nifp)
2762 {
2763 	struct mem_pt_if *ptif;
2764 
2765 	ptif = netmap_mem_pt_guest_ifp_lookup(nmd, na->ifp);
2766 	if (ptif == NULL) {
2767 		nm_prerr("interface %s is not in passthrough", na->name);
2768 	}
2769 }
2770 
2771 static int
2772 netmap_mem_pt_guest_rings_create(struct netmap_mem_d *nmd,
2773 		struct netmap_adapter *na)
2774 {
2775 	struct netmap_mem_ptg *ptnmd = (struct netmap_mem_ptg *)nmd;
2776 	struct mem_pt_if *ptif;
2777 	struct netmap_if *nifp;
2778 	int i, error = -1;
2779 
2780 	ptif = netmap_mem_pt_guest_ifp_lookup(nmd, na->ifp);
2781 	if (ptif == NULL) {
2782 		nm_prerr("interface %s is not in passthrough", na->name);
2783 		goto out;
2784 	}
2785 
2786 
2787 	/* point each kring to the corresponding backend ring */
2788 	nifp = (struct netmap_if *)((char *)ptnmd->nm_addr + ptif->nifp_offset);
2789 	for (i = 0; i < netmap_all_rings(na, NR_TX); i++) {
2790 		struct netmap_kring *kring = na->tx_rings[i];
2791 		if (kring->ring)
2792 			continue;
2793 		kring->ring = (struct netmap_ring *)
2794 			((char *)nifp + nifp->ring_ofs[i]);
2795 	}
2796 	for (i = 0; i < netmap_all_rings(na, NR_RX); i++) {
2797 		struct netmap_kring *kring = na->rx_rings[i];
2798 		if (kring->ring)
2799 			continue;
2800 		kring->ring = (struct netmap_ring *)
2801 			((char *)nifp +
2802 			 nifp->ring_ofs[netmap_all_rings(na, NR_TX) + i]);
2803 	}
2804 
2805 	error = 0;
2806 out:
2807 	return error;
2808 }
2809 
2810 static void
2811 netmap_mem_pt_guest_rings_delete(struct netmap_mem_d *nmd, struct netmap_adapter *na)
2812 {
2813 #if 0
2814 	enum txrx t;
2815 
2816 	for_rx_tx(t) {
2817 		u_int i;
2818 		for (i = 0; i < nma_get_nrings(na, t) + 1; i++) {
2819 			struct netmap_kring *kring = &NMR(na, t)[i];
2820 
2821 			kring->ring = NULL;
2822 		}
2823 	}
2824 #endif
2825 	(void)nmd;
2826 	(void)na;
2827 }
2828 
2829 static struct netmap_mem_ops netmap_mem_pt_guest_ops = {
2830 	.nmd_get_lut = netmap_mem_pt_guest_get_lut,
2831 	.nmd_get_info = netmap_mem_pt_guest_get_info,
2832 	.nmd_ofstophys = netmap_mem_pt_guest_ofstophys,
2833 	.nmd_config = netmap_mem_pt_guest_config,
2834 	.nmd_finalize = netmap_mem_pt_guest_finalize,
2835 	.nmd_deref = netmap_mem_pt_guest_deref,
2836 	.nmd_if_offset = netmap_mem_pt_guest_if_offset,
2837 	.nmd_delete = netmap_mem_pt_guest_delete,
2838 	.nmd_if_new = netmap_mem_pt_guest_if_new,
2839 	.nmd_if_delete = netmap_mem_pt_guest_if_delete,
2840 	.nmd_rings_create = netmap_mem_pt_guest_rings_create,
2841 	.nmd_rings_delete = netmap_mem_pt_guest_rings_delete
2842 };
2843 
2844 /* Called with nm_mem_list_lock held. */
2845 static struct netmap_mem_d *
2846 netmap_mem_pt_guest_find_memid(nm_memid_t mem_id)
2847 {
2848 	struct netmap_mem_d *mem = NULL;
2849 	struct netmap_mem_d *scan = netmap_last_mem_d;
2850 
2851 	do {
2852 		/* find ptnetmap allocator through host ID */
2853 		if (scan->ops->nmd_deref == netmap_mem_pt_guest_deref &&
2854 			((struct netmap_mem_ptg *)(scan))->host_mem_id == mem_id) {
2855 			mem = scan;
2856 			mem->refcount++;
2857 			NM_DBG_REFC(mem, __FUNCTION__, __LINE__);
2858 			break;
2859 		}
2860 		scan = scan->next;
2861 	} while (scan != netmap_last_mem_d);
2862 
2863 	return mem;
2864 }
2865 
2866 /* Called with nm_mem_list_lock held. */
2867 static struct netmap_mem_d *
2868 netmap_mem_pt_guest_create(nm_memid_t mem_id)
2869 {
2870 	struct netmap_mem_ptg *ptnmd;
2871 	int err = 0;
2872 
2873 	ptnmd = nm_os_malloc(sizeof(struct netmap_mem_ptg));
2874 	if (ptnmd == NULL) {
2875 		err = ENOMEM;
2876 		goto error;
2877 	}
2878 
2879 	ptnmd->up.ops = &netmap_mem_pt_guest_ops;
2880 	ptnmd->host_mem_id = mem_id;
2881 	ptnmd->pt_ifs = NULL;
2882 
2883 	/* Assign new id in the guest (We have the lock) */
2884 	err = nm_mem_assign_id_locked(&ptnmd->up, -1);
2885 	if (err)
2886 		goto error;
2887 
2888 	ptnmd->up.flags &= ~NETMAP_MEM_FINALIZED;
2889 	ptnmd->up.flags |= NETMAP_MEM_IO;
2890 
2891 	NMA_LOCK_INIT(&ptnmd->up);
2892 
2893 	snprintf(ptnmd->up.name, NM_MEM_NAMESZ, "%d", ptnmd->up.nm_id);
2894 
2895 
2896 	return &ptnmd->up;
2897 error:
2898 	netmap_mem_pt_guest_delete(&ptnmd->up);
2899 	return NULL;
2900 }
2901 
2902 /*
2903  * find host id in guest allocators and create guest allocator
2904  * if it is not there
2905  */
2906 static struct netmap_mem_d *
2907 netmap_mem_pt_guest_get(nm_memid_t mem_id)
2908 {
2909 	struct netmap_mem_d *nmd;
2910 
2911 	NM_MTX_LOCK(nm_mem_list_lock);
2912 	nmd = netmap_mem_pt_guest_find_memid(mem_id);
2913 	if (nmd == NULL) {
2914 		nmd = netmap_mem_pt_guest_create(mem_id);
2915 	}
2916 	NM_MTX_UNLOCK(nm_mem_list_lock);
2917 
2918 	return nmd;
2919 }
2920 
2921 /*
2922  * The guest allocator can be created by ptnetmap_memdev (during the device
2923  * attach) or by ptnetmap device (ptnet), during the netmap_attach.
2924  *
2925  * The order is not important (we have different order in LINUX and FreeBSD).
2926  * The first one, creates the device, and the second one simply attaches it.
2927  */
2928 
2929 /* Called when ptnetmap_memdev is attaching, to attach a new allocator in
2930  * the guest */
2931 struct netmap_mem_d *
2932 netmap_mem_pt_guest_attach(struct ptnetmap_memdev *ptn_dev, nm_memid_t mem_id)
2933 {
2934 	struct netmap_mem_d *nmd;
2935 	struct netmap_mem_ptg *ptnmd;
2936 
2937 	nmd = netmap_mem_pt_guest_get(mem_id);
2938 
2939 	/* assign this device to the guest allocator */
2940 	if (nmd) {
2941 		ptnmd = (struct netmap_mem_ptg *)nmd;
2942 		ptnmd->ptn_dev = ptn_dev;
2943 	}
2944 
2945 	return nmd;
2946 }
2947 
2948 /* Called when ptnet device is attaching */
2949 struct netmap_mem_d *
2950 netmap_mem_pt_guest_new(if_t ifp,
2951 			unsigned int nifp_offset,
2952 			unsigned int memid)
2953 {
2954 	struct netmap_mem_d *nmd;
2955 
2956 	if (ifp == NULL) {
2957 		return NULL;
2958 	}
2959 
2960 	nmd = netmap_mem_pt_guest_get((nm_memid_t)memid);
2961 
2962 	if (nmd) {
2963 		netmap_mem_pt_guest_ifp_add(nmd, ifp, nifp_offset);
2964 	}
2965 
2966 	return nmd;
2967 }
2968 
2969 #endif /* WITH_PTNETMAP */
2970