1 /* 2 * Copyright (C) 2012-2014 Matteo Landi 3 * Copyright (C) 2012-2016 Luigi Rizzo 4 * Copyright (C) 2012-2016 Giuseppe Lettieri 5 * All rights reserved. 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 1. Redistributions of source code must retain the above copyright 11 * notice, this list of conditions and the following disclaimer. 12 * 2. Redistributions in binary form must reproduce the above copyright 13 * notice, this list of conditions and the following disclaimer in the 14 * documentation and/or other materials provided with the distribution. 15 * 16 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 17 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 18 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 19 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 20 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 21 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 22 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 23 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 24 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 25 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 26 * SUCH DAMAGE. 27 */ 28 29 #ifdef linux 30 #include "bsd_glue.h" 31 #endif /* linux */ 32 33 #ifdef __APPLE__ 34 #include "osx_glue.h" 35 #endif /* __APPLE__ */ 36 37 #ifdef __FreeBSD__ 38 #include <sys/cdefs.h> /* prerequisite */ 39 __FBSDID("$FreeBSD$"); 40 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 u_int objtotal; /* actual total number of objects. */ 102 u_int memtotal; /* actual total memory space */ 103 u_int numclusters; /* actual number of clusters */ 104 105 u_int objfree; /* number of free objects. */ 106 107 struct lut_entry *lut; /* virt,phys addresses, objtotal entries */ 108 uint32_t *bitmap; /* one bit per buffer, 1 means free */ 109 uint32_t bitmap_slots; /* number of uint32 entries in bitmap */ 110 /* ---------------------------------------------------*/ 111 112 /* limits */ 113 u_int objminsize; /* minimum object size */ 114 u_int objmaxsize; /* maximum object size */ 115 u_int nummin; /* minimum number of objects */ 116 u_int nummax; /* maximum number of objects */ 117 118 /* these are changed only by config */ 119 u_int _objtotal; /* total number of objects */ 120 u_int _objsize; /* object size */ 121 u_int _clustsize; /* cluster size */ 122 u_int _clustentries; /* objects per cluster */ 123 u_int _numclusters; /* number of clusters */ 124 125 /* requested values */ 126 u_int r_objtotal; 127 u_int r_objsize; 128 }; 129 130 #define NMA_LOCK_T NM_MTX_T 131 132 133 struct netmap_mem_ops { 134 int (*nmd_get_lut)(struct netmap_mem_d *, struct netmap_lut*); 135 int (*nmd_get_info)(struct netmap_mem_d *, u_int *size, 136 u_int *memflags, uint16_t *id); 137 138 vm_paddr_t (*nmd_ofstophys)(struct netmap_mem_d *, vm_ooffset_t); 139 int (*nmd_config)(struct netmap_mem_d *); 140 int (*nmd_finalize)(struct netmap_mem_d *); 141 void (*nmd_deref)(struct netmap_mem_d *); 142 ssize_t (*nmd_if_offset)(struct netmap_mem_d *, const void *vaddr); 143 void (*nmd_delete)(struct netmap_mem_d *); 144 145 struct netmap_if * (*nmd_if_new)(struct netmap_adapter *, 146 struct netmap_priv_d *); 147 void (*nmd_if_delete)(struct netmap_adapter *, struct netmap_if *); 148 int (*nmd_rings_create)(struct netmap_adapter *); 149 void (*nmd_rings_delete)(struct netmap_adapter *); 150 }; 151 152 struct netmap_mem_d { 153 NMA_LOCK_T nm_mtx; /* protect the allocator */ 154 u_int nm_totalsize; /* shorthand */ 155 156 u_int flags; 157 #define NETMAP_MEM_FINALIZED 0x1 /* preallocation done */ 158 #define NETMAP_MEM_HIDDEN 0x8 /* beeing prepared */ 159 int lasterr; /* last error for curr config */ 160 int active; /* active users */ 161 int refcount; 162 /* the three allocators */ 163 struct netmap_obj_pool pools[NETMAP_POOLS_NR]; 164 165 nm_memid_t nm_id; /* allocator identifier */ 166 int nm_grp; /* iommu groupd id */ 167 168 /* list of all existing allocators, sorted by nm_id */ 169 struct netmap_mem_d *prev, *next; 170 171 struct netmap_mem_ops *ops; 172 173 struct netmap_obj_params params[NETMAP_POOLS_NR]; 174 175 #define NM_MEM_NAMESZ 16 176 char name[NM_MEM_NAMESZ]; 177 }; 178 179 /* 180 * XXX need to fix the case of t0 == void 181 */ 182 #define NMD_DEFCB(t0, name) \ 183 t0 \ 184 netmap_mem_##name(struct netmap_mem_d *nmd) \ 185 { \ 186 return nmd->ops->nmd_##name(nmd); \ 187 } 188 189 #define NMD_DEFCB1(t0, name, t1) \ 190 t0 \ 191 netmap_mem_##name(struct netmap_mem_d *nmd, t1 a1) \ 192 { \ 193 return nmd->ops->nmd_##name(nmd, a1); \ 194 } 195 196 #define NMD_DEFCB3(t0, name, t1, t2, t3) \ 197 t0 \ 198 netmap_mem_##name(struct netmap_mem_d *nmd, t1 a1, t2 a2, t3 a3) \ 199 { \ 200 return nmd->ops->nmd_##name(nmd, a1, a2, a3); \ 201 } 202 203 #define NMD_DEFNACB(t0, name) \ 204 t0 \ 205 netmap_mem_##name(struct netmap_adapter *na) \ 206 { \ 207 return na->nm_mem->ops->nmd_##name(na); \ 208 } 209 210 #define NMD_DEFNACB1(t0, name, t1) \ 211 t0 \ 212 netmap_mem_##name(struct netmap_adapter *na, t1 a1) \ 213 { \ 214 return na->nm_mem->ops->nmd_##name(na, a1); \ 215 } 216 217 NMD_DEFCB1(int, get_lut, struct netmap_lut *); 218 NMD_DEFCB3(int, get_info, u_int *, u_int *, uint16_t *); 219 NMD_DEFCB1(vm_paddr_t, ofstophys, vm_ooffset_t); 220 static int netmap_mem_config(struct netmap_mem_d *); 221 NMD_DEFCB(int, config); 222 NMD_DEFCB1(ssize_t, if_offset, const void *); 223 NMD_DEFCB(void, delete); 224 225 NMD_DEFNACB1(struct netmap_if *, if_new, struct netmap_priv_d *); 226 NMD_DEFNACB1(void, if_delete, struct netmap_if *); 227 NMD_DEFNACB(int, rings_create); 228 NMD_DEFNACB(void, rings_delete); 229 230 static int netmap_mem_map(struct netmap_obj_pool *, struct netmap_adapter *); 231 static int netmap_mem_unmap(struct netmap_obj_pool *, struct netmap_adapter *); 232 static int nm_mem_assign_group(struct netmap_mem_d *, struct device *); 233 static void nm_mem_release_id(struct netmap_mem_d *); 234 235 nm_memid_t 236 netmap_mem_get_id(struct netmap_mem_d *nmd) 237 { 238 return nmd->nm_id; 239 } 240 241 #define NMA_LOCK_INIT(n) NM_MTX_INIT((n)->nm_mtx) 242 #define NMA_LOCK_DESTROY(n) NM_MTX_DESTROY((n)->nm_mtx) 243 #define NMA_LOCK(n) NM_MTX_LOCK((n)->nm_mtx) 244 #define NMA_UNLOCK(n) NM_MTX_UNLOCK((n)->nm_mtx) 245 246 #ifdef NM_DEBUG_MEM_PUTGET 247 #define NM_DBG_REFC(nmd, func, line) \ 248 nm_prinf("%s:%d mem[%d] -> %d\n", func, line, (nmd)->nm_id, (nmd)->refcount); 249 #else 250 #define NM_DBG_REFC(nmd, func, line) 251 #endif 252 253 /* circular list of all existing allocators */ 254 static struct netmap_mem_d *netmap_last_mem_d = &nm_mem; 255 NM_MTX_T nm_mem_list_lock; 256 257 struct netmap_mem_d * 258 __netmap_mem_get(struct netmap_mem_d *nmd, const char *func, int line) 259 { 260 NM_MTX_LOCK(nm_mem_list_lock); 261 nmd->refcount++; 262 NM_DBG_REFC(nmd, func, line); 263 NM_MTX_UNLOCK(nm_mem_list_lock); 264 return nmd; 265 } 266 267 void 268 __netmap_mem_put(struct netmap_mem_d *nmd, const char *func, int line) 269 { 270 int last; 271 NM_MTX_LOCK(nm_mem_list_lock); 272 last = (--nmd->refcount == 0); 273 if (last) 274 nm_mem_release_id(nmd); 275 NM_DBG_REFC(nmd, func, line); 276 NM_MTX_UNLOCK(nm_mem_list_lock); 277 if (last) 278 netmap_mem_delete(nmd); 279 } 280 281 int 282 netmap_mem_finalize(struct netmap_mem_d *nmd, struct netmap_adapter *na) 283 { 284 if (nm_mem_assign_group(nmd, na->pdev) < 0) { 285 return ENOMEM; 286 } else { 287 NMA_LOCK(nmd); 288 nmd->lasterr = nmd->ops->nmd_finalize(nmd); 289 NMA_UNLOCK(nmd); 290 } 291 292 if (!nmd->lasterr && na->pdev) 293 netmap_mem_map(&nmd->pools[NETMAP_BUF_POOL], na); 294 295 return nmd->lasterr; 296 } 297 298 void 299 netmap_mem_deref(struct netmap_mem_d *nmd, struct netmap_adapter *na) 300 { 301 NMA_LOCK(nmd); 302 netmap_mem_unmap(&nmd->pools[NETMAP_BUF_POOL], na); 303 if (nmd->active == 1) { 304 u_int i; 305 306 /* 307 * Reset the allocator when it falls out of use so that any 308 * pool resources leaked by unclean application exits are 309 * reclaimed. 310 */ 311 for (i = 0; i < NETMAP_POOLS_NR; i++) { 312 struct netmap_obj_pool *p; 313 u_int j; 314 315 p = &nmd->pools[i]; 316 p->objfree = p->objtotal; 317 /* 318 * Reproduce the net effect of the M_ZERO malloc() 319 * and marking of free entries in the bitmap that 320 * occur in finalize_obj_allocator() 321 */ 322 memset(p->bitmap, 323 '\0', 324 sizeof(uint32_t) * ((p->objtotal + 31) / 32)); 325 326 /* 327 * Set all the bits in the bitmap that have 328 * corresponding buffers to 1 to indicate they are 329 * free. 330 */ 331 for (j = 0; j < p->objtotal; j++) { 332 if (p->lut[j].vaddr != NULL) { 333 p->bitmap[ (j>>5) ] |= ( 1 << (j & 31) ); 334 } 335 } 336 } 337 338 /* 339 * Per netmap_mem_finalize_all(), 340 * buffers 0 and 1 are reserved 341 */ 342 nmd->pools[NETMAP_BUF_POOL].objfree -= 2; 343 if (nmd->pools[NETMAP_BUF_POOL].bitmap) { 344 /* XXX This check is a workaround that prevents a 345 * NULL pointer crash which currently happens only 346 * with ptnetmap guests. 347 * Removed shared-info --> is the bug still there? */ 348 nmd->pools[NETMAP_BUF_POOL].bitmap[0] = ~3; 349 } 350 } 351 nmd->ops->nmd_deref(nmd); 352 353 NMA_UNLOCK(nmd); 354 } 355 356 357 /* accessor functions */ 358 static int 359 netmap_mem2_get_lut(struct netmap_mem_d *nmd, struct netmap_lut *lut) 360 { 361 lut->lut = nmd->pools[NETMAP_BUF_POOL].lut; 362 lut->objtotal = nmd->pools[NETMAP_BUF_POOL].objtotal; 363 lut->objsize = nmd->pools[NETMAP_BUF_POOL]._objsize; 364 365 return 0; 366 } 367 368 static struct netmap_obj_params netmap_min_priv_params[NETMAP_POOLS_NR] = { 369 [NETMAP_IF_POOL] = { 370 .size = 1024, 371 .num = 2, 372 }, 373 [NETMAP_RING_POOL] = { 374 .size = 5*PAGE_SIZE, 375 .num = 4, 376 }, 377 [NETMAP_BUF_POOL] = { 378 .size = 2048, 379 .num = 4098, 380 }, 381 }; 382 383 384 /* 385 * nm_mem is the memory allocator used for all physical interfaces 386 * running in netmap mode. 387 * Virtual (VALE) ports will have each its own allocator. 388 */ 389 extern struct netmap_mem_ops netmap_mem_global_ops; /* forward */ 390 struct netmap_mem_d nm_mem = { /* Our memory allocator. */ 391 .pools = { 392 [NETMAP_IF_POOL] = { 393 .name = "netmap_if", 394 .objminsize = sizeof(struct netmap_if), 395 .objmaxsize = 4096, 396 .nummin = 10, /* don't be stingy */ 397 .nummax = 10000, /* XXX very large */ 398 }, 399 [NETMAP_RING_POOL] = { 400 .name = "netmap_ring", 401 .objminsize = sizeof(struct netmap_ring), 402 .objmaxsize = 32*PAGE_SIZE, 403 .nummin = 2, 404 .nummax = 1024, 405 }, 406 [NETMAP_BUF_POOL] = { 407 .name = "netmap_buf", 408 .objminsize = 64, 409 .objmaxsize = 65536, 410 .nummin = 4, 411 .nummax = 1000000, /* one million! */ 412 }, 413 }, 414 415 .params = { 416 [NETMAP_IF_POOL] = { 417 .size = 1024, 418 .num = 100, 419 }, 420 [NETMAP_RING_POOL] = { 421 .size = 9*PAGE_SIZE, 422 .num = 200, 423 }, 424 [NETMAP_BUF_POOL] = { 425 .size = 2048, 426 .num = NETMAP_BUF_MAX_NUM, 427 }, 428 }, 429 430 .nm_id = 1, 431 .nm_grp = -1, 432 433 .prev = &nm_mem, 434 .next = &nm_mem, 435 436 .ops = &netmap_mem_global_ops, 437 438 .name = "1" 439 }; 440 441 442 /* blueprint for the private memory allocators */ 443 extern struct netmap_mem_ops netmap_mem_private_ops; /* forward */ 444 /* XXX clang is not happy about using name as a print format */ 445 static const struct netmap_mem_d nm_blueprint = { 446 .pools = { 447 [NETMAP_IF_POOL] = { 448 .name = "%s_if", 449 .objminsize = sizeof(struct netmap_if), 450 .objmaxsize = 4096, 451 .nummin = 1, 452 .nummax = 100, 453 }, 454 [NETMAP_RING_POOL] = { 455 .name = "%s_ring", 456 .objminsize = sizeof(struct netmap_ring), 457 .objmaxsize = 32*PAGE_SIZE, 458 .nummin = 2, 459 .nummax = 1024, 460 }, 461 [NETMAP_BUF_POOL] = { 462 .name = "%s_buf", 463 .objminsize = 64, 464 .objmaxsize = 65536, 465 .nummin = 4, 466 .nummax = 1000000, /* one million! */ 467 }, 468 }, 469 470 .nm_grp = -1, 471 472 .flags = NETMAP_MEM_PRIVATE, 473 474 .ops = &netmap_mem_global_ops, 475 }; 476 477 /* memory allocator related sysctls */ 478 479 #define STRINGIFY(x) #x 480 481 482 #define DECLARE_SYSCTLS(id, name) \ 483 SYSBEGIN(mem2_ ## name); \ 484 SYSCTL_INT(_dev_netmap, OID_AUTO, name##_size, \ 485 CTLFLAG_RW, &nm_mem.params[id].size, 0, "Requested size of netmap " STRINGIFY(name) "s"); \ 486 SYSCTL_INT(_dev_netmap, OID_AUTO, name##_curr_size, \ 487 CTLFLAG_RD, &nm_mem.pools[id]._objsize, 0, "Current size of netmap " STRINGIFY(name) "s"); \ 488 SYSCTL_INT(_dev_netmap, OID_AUTO, name##_num, \ 489 CTLFLAG_RW, &nm_mem.params[id].num, 0, "Requested number of netmap " STRINGIFY(name) "s"); \ 490 SYSCTL_INT(_dev_netmap, OID_AUTO, name##_curr_num, \ 491 CTLFLAG_RD, &nm_mem.pools[id].objtotal, 0, "Current number of netmap " STRINGIFY(name) "s"); \ 492 SYSCTL_INT(_dev_netmap, OID_AUTO, priv_##name##_size, \ 493 CTLFLAG_RW, &netmap_min_priv_params[id].size, 0, \ 494 "Default size of private netmap " STRINGIFY(name) "s"); \ 495 SYSCTL_INT(_dev_netmap, OID_AUTO, priv_##name##_num, \ 496 CTLFLAG_RW, &netmap_min_priv_params[id].num, 0, \ 497 "Default number of private netmap " STRINGIFY(name) "s"); \ 498 SYSEND 499 500 SYSCTL_DECL(_dev_netmap); 501 DECLARE_SYSCTLS(NETMAP_IF_POOL, if); 502 DECLARE_SYSCTLS(NETMAP_RING_POOL, ring); 503 DECLARE_SYSCTLS(NETMAP_BUF_POOL, buf); 504 505 /* call with nm_mem_list_lock held */ 506 static int 507 nm_mem_assign_id_locked(struct netmap_mem_d *nmd) 508 { 509 nm_memid_t id; 510 struct netmap_mem_d *scan = netmap_last_mem_d; 511 int error = ENOMEM; 512 513 do { 514 /* we rely on unsigned wrap around */ 515 id = scan->nm_id + 1; 516 if (id == 0) /* reserve 0 as error value */ 517 id = 1; 518 scan = scan->next; 519 if (id != scan->nm_id) { 520 nmd->nm_id = id; 521 nmd->prev = scan->prev; 522 nmd->next = scan; 523 scan->prev->next = nmd; 524 scan->prev = nmd; 525 netmap_last_mem_d = nmd; 526 nmd->refcount = 1; 527 NM_DBG_REFC(nmd, __FUNCTION__, __LINE__); 528 error = 0; 529 break; 530 } 531 } while (scan != netmap_last_mem_d); 532 533 return error; 534 } 535 536 /* call with nm_mem_list_lock *not* held */ 537 static int 538 nm_mem_assign_id(struct netmap_mem_d *nmd) 539 { 540 int ret; 541 542 NM_MTX_LOCK(nm_mem_list_lock); 543 ret = nm_mem_assign_id_locked(nmd); 544 NM_MTX_UNLOCK(nm_mem_list_lock); 545 546 return ret; 547 } 548 549 /* call with nm_mem_list_lock held */ 550 static void 551 nm_mem_release_id(struct netmap_mem_d *nmd) 552 { 553 nmd->prev->next = nmd->next; 554 nmd->next->prev = nmd->prev; 555 556 if (netmap_last_mem_d == nmd) 557 netmap_last_mem_d = nmd->prev; 558 559 nmd->prev = nmd->next = NULL; 560 } 561 562 struct netmap_mem_d * 563 netmap_mem_find(nm_memid_t id) 564 { 565 struct netmap_mem_d *nmd; 566 567 NM_MTX_LOCK(nm_mem_list_lock); 568 nmd = netmap_last_mem_d; 569 do { 570 if (!(nmd->flags & NETMAP_MEM_HIDDEN) && nmd->nm_id == id) { 571 nmd->refcount++; 572 NM_DBG_REFC(nmd, __FUNCTION__, __LINE__); 573 NM_MTX_UNLOCK(nm_mem_list_lock); 574 return nmd; 575 } 576 nmd = nmd->next; 577 } while (nmd != netmap_last_mem_d); 578 NM_MTX_UNLOCK(nm_mem_list_lock); 579 return NULL; 580 } 581 582 static int 583 nm_mem_assign_group(struct netmap_mem_d *nmd, struct device *dev) 584 { 585 int err = 0, id; 586 id = nm_iommu_group_id(dev); 587 if (netmap_verbose) 588 D("iommu_group %d", id); 589 590 NMA_LOCK(nmd); 591 592 if (nmd->nm_grp < 0) 593 nmd->nm_grp = id; 594 595 if (nmd->nm_grp != id) 596 nmd->lasterr = err = ENOMEM; 597 598 NMA_UNLOCK(nmd); 599 return err; 600 } 601 602 /* 603 * First, find the allocator that contains the requested offset, 604 * then locate the cluster through a lookup table. 605 */ 606 static vm_paddr_t 607 netmap_mem2_ofstophys(struct netmap_mem_d* nmd, vm_ooffset_t offset) 608 { 609 int i; 610 vm_ooffset_t o = offset; 611 vm_paddr_t pa; 612 struct netmap_obj_pool *p; 613 614 NMA_LOCK(nmd); 615 p = nmd->pools; 616 617 for (i = 0; i < NETMAP_POOLS_NR; offset -= p[i].memtotal, i++) { 618 if (offset >= p[i].memtotal) 619 continue; 620 // now lookup the cluster's address 621 #ifndef _WIN32 622 pa = vtophys(p[i].lut[offset / p[i]._objsize].vaddr) + 623 offset % p[i]._objsize; 624 #else 625 pa = vtophys(p[i].lut[offset / p[i]._objsize].vaddr); 626 pa.QuadPart += offset % p[i]._objsize; 627 #endif 628 NMA_UNLOCK(nmd); 629 return pa; 630 } 631 /* this is only in case of errors */ 632 D("invalid ofs 0x%x out of 0x%x 0x%x 0x%x", (u_int)o, 633 p[NETMAP_IF_POOL].memtotal, 634 p[NETMAP_IF_POOL].memtotal 635 + p[NETMAP_RING_POOL].memtotal, 636 p[NETMAP_IF_POOL].memtotal 637 + p[NETMAP_RING_POOL].memtotal 638 + p[NETMAP_BUF_POOL].memtotal); 639 NMA_UNLOCK(nmd); 640 #ifndef _WIN32 641 return 0; // XXX bad address 642 #else 643 vm_paddr_t res; 644 res.QuadPart = 0; 645 return res; 646 #endif 647 } 648 649 #ifdef _WIN32 650 651 /* 652 * win32_build_virtual_memory_for_userspace 653 * 654 * This function get all the object making part of the pools and maps 655 * a contiguous virtual memory space for the userspace 656 * It works this way 657 * 1 - allocate a Memory Descriptor List wide as the sum 658 * of the memory needed for the pools 659 * 2 - cycle all the objects in every pool and for every object do 660 * 661 * 2a - cycle all the objects in every pool, get the list 662 * of the physical address descriptors 663 * 2b - calculate the offset in the array of pages desciptor in the 664 * main MDL 665 * 2c - copy the descriptors of the object in the main MDL 666 * 667 * 3 - return the resulting MDL that needs to be mapped in userland 668 * 669 * In this way we will have an MDL that describes all the memory for the 670 * objects in a single object 671 */ 672 673 PMDL 674 win32_build_user_vm_map(struct netmap_mem_d* nmd) 675 { 676 int i, j; 677 u_int memsize, memflags, ofs = 0; 678 PMDL mainMdl, tempMdl; 679 680 if (netmap_mem_get_info(nmd, &memsize, &memflags, NULL)) { 681 D("memory not finalised yet"); 682 return NULL; 683 } 684 685 mainMdl = IoAllocateMdl(NULL, memsize, FALSE, FALSE, NULL); 686 if (mainMdl == NULL) { 687 D("failed to allocate mdl"); 688 return NULL; 689 } 690 691 NMA_LOCK(nmd); 692 for (i = 0; i < NETMAP_POOLS_NR; i++) { 693 struct netmap_obj_pool *p = &nmd->pools[i]; 694 int clsz = p->_clustsize; 695 int clobjs = p->_clustentries; /* objects per cluster */ 696 int mdl_len = sizeof(PFN_NUMBER) * BYTES_TO_PAGES(clsz); 697 PPFN_NUMBER pSrc, pDst; 698 699 /* each pool has a different cluster size so we need to reallocate */ 700 tempMdl = IoAllocateMdl(p->lut[0].vaddr, clsz, FALSE, FALSE, NULL); 701 if (tempMdl == NULL) { 702 NMA_UNLOCK(nmd); 703 D("fail to allocate tempMdl"); 704 IoFreeMdl(mainMdl); 705 return NULL; 706 } 707 pSrc = MmGetMdlPfnArray(tempMdl); 708 /* create one entry per cluster, the lut[] has one entry per object */ 709 for (j = 0; j < p->numclusters; j++, ofs += clsz) { 710 pDst = &MmGetMdlPfnArray(mainMdl)[BYTES_TO_PAGES(ofs)]; 711 MmInitializeMdl(tempMdl, p->lut[j*clobjs].vaddr, clsz); 712 MmBuildMdlForNonPagedPool(tempMdl); /* compute physical page addresses */ 713 RtlCopyMemory(pDst, pSrc, mdl_len); /* copy the page descriptors */ 714 mainMdl->MdlFlags = tempMdl->MdlFlags; /* XXX what is in here ? */ 715 } 716 IoFreeMdl(tempMdl); 717 } 718 NMA_UNLOCK(nmd); 719 return mainMdl; 720 } 721 722 #endif /* _WIN32 */ 723 724 /* 725 * helper function for OS-specific mmap routines (currently only windows). 726 * Given an nmd and a pool index, returns the cluster size and number of clusters. 727 * Returns 0 if memory is finalised and the pool is valid, otherwise 1. 728 * It should be called under NMA_LOCK(nmd) otherwise the underlying info can change. 729 */ 730 731 int 732 netmap_mem2_get_pool_info(struct netmap_mem_d* nmd, u_int pool, u_int *clustsize, u_int *numclusters) 733 { 734 if (!nmd || !clustsize || !numclusters || pool >= NETMAP_POOLS_NR) 735 return 1; /* invalid arguments */ 736 // NMA_LOCK_ASSERT(nmd); 737 if (!(nmd->flags & NETMAP_MEM_FINALIZED)) { 738 *clustsize = *numclusters = 0; 739 return 1; /* not ready yet */ 740 } 741 *clustsize = nmd->pools[pool]._clustsize; 742 *numclusters = nmd->pools[pool].numclusters; 743 return 0; /* success */ 744 } 745 746 static int 747 netmap_mem2_get_info(struct netmap_mem_d* nmd, u_int* size, u_int *memflags, 748 nm_memid_t *id) 749 { 750 int error = 0; 751 NMA_LOCK(nmd); 752 error = netmap_mem_config(nmd); 753 if (error) 754 goto out; 755 if (size) { 756 if (nmd->flags & NETMAP_MEM_FINALIZED) { 757 *size = nmd->nm_totalsize; 758 } else { 759 int i; 760 *size = 0; 761 for (i = 0; i < NETMAP_POOLS_NR; i++) { 762 struct netmap_obj_pool *p = nmd->pools + i; 763 *size += (p->_numclusters * p->_clustsize); 764 } 765 } 766 } 767 if (memflags) 768 *memflags = nmd->flags; 769 if (id) 770 *id = nmd->nm_id; 771 out: 772 NMA_UNLOCK(nmd); 773 return error; 774 } 775 776 /* 777 * we store objects by kernel address, need to find the offset 778 * within the pool to export the value to userspace. 779 * Algorithm: scan until we find the cluster, then add the 780 * actual offset in the cluster 781 */ 782 static ssize_t 783 netmap_obj_offset(struct netmap_obj_pool *p, const void *vaddr) 784 { 785 int i, k = p->_clustentries, n = p->objtotal; 786 ssize_t ofs = 0; 787 788 for (i = 0; i < n; i += k, ofs += p->_clustsize) { 789 const char *base = p->lut[i].vaddr; 790 ssize_t relofs = (const char *) vaddr - base; 791 792 if (relofs < 0 || relofs >= p->_clustsize) 793 continue; 794 795 ofs = ofs + relofs; 796 ND("%s: return offset %d (cluster %d) for pointer %p", 797 p->name, ofs, i, vaddr); 798 return ofs; 799 } 800 D("address %p is not contained inside any cluster (%s)", 801 vaddr, p->name); 802 return 0; /* An error occurred */ 803 } 804 805 /* Helper functions which convert virtual addresses to offsets */ 806 #define netmap_if_offset(n, v) \ 807 netmap_obj_offset(&(n)->pools[NETMAP_IF_POOL], (v)) 808 809 #define netmap_ring_offset(n, v) \ 810 ((n)->pools[NETMAP_IF_POOL].memtotal + \ 811 netmap_obj_offset(&(n)->pools[NETMAP_RING_POOL], (v))) 812 813 static ssize_t 814 netmap_mem2_if_offset(struct netmap_mem_d *nmd, const void *addr) 815 { 816 ssize_t v; 817 NMA_LOCK(nmd); 818 v = netmap_if_offset(nmd, addr); 819 NMA_UNLOCK(nmd); 820 return v; 821 } 822 823 /* 824 * report the index, and use start position as a hint, 825 * otherwise buffer allocation becomes terribly expensive. 826 */ 827 static void * 828 netmap_obj_malloc(struct netmap_obj_pool *p, u_int len, uint32_t *start, uint32_t *index) 829 { 830 uint32_t i = 0; /* index in the bitmap */ 831 uint32_t mask, j = 0; /* slot counter */ 832 void *vaddr = NULL; 833 834 if (len > p->_objsize) { 835 D("%s request size %d too large", p->name, len); 836 // XXX cannot reduce the size 837 return NULL; 838 } 839 840 if (p->objfree == 0) { 841 D("no more %s objects", p->name); 842 return NULL; 843 } 844 if (start) 845 i = *start; 846 847 /* termination is guaranteed by p->free, but better check bounds on i */ 848 while (vaddr == NULL && i < p->bitmap_slots) { 849 uint32_t cur = p->bitmap[i]; 850 if (cur == 0) { /* bitmask is fully used */ 851 i++; 852 continue; 853 } 854 /* locate a slot */ 855 for (j = 0, mask = 1; (cur & mask) == 0; j++, mask <<= 1) 856 ; 857 858 p->bitmap[i] &= ~mask; /* mark object as in use */ 859 p->objfree--; 860 861 vaddr = p->lut[i * 32 + j].vaddr; 862 if (index) 863 *index = i * 32 + j; 864 } 865 ND("%s allocator: allocated object @ [%d][%d]: vaddr %p",p->name, i, j, vaddr); 866 867 if (start) 868 *start = i; 869 return vaddr; 870 } 871 872 873 /* 874 * free by index, not by address. 875 * XXX should we also cleanup the content ? 876 */ 877 static int 878 netmap_obj_free(struct netmap_obj_pool *p, uint32_t j) 879 { 880 uint32_t *ptr, mask; 881 882 if (j >= p->objtotal) { 883 D("invalid index %u, max %u", j, p->objtotal); 884 return 1; 885 } 886 ptr = &p->bitmap[j / 32]; 887 mask = (1 << (j % 32)); 888 if (*ptr & mask) { 889 D("ouch, double free on buffer %d", j); 890 return 1; 891 } else { 892 *ptr |= mask; 893 p->objfree++; 894 return 0; 895 } 896 } 897 898 /* 899 * free by address. This is slow but is only used for a few 900 * objects (rings, nifp) 901 */ 902 static void 903 netmap_obj_free_va(struct netmap_obj_pool *p, void *vaddr) 904 { 905 u_int i, j, n = p->numclusters; 906 907 for (i = 0, j = 0; i < n; i++, j += p->_clustentries) { 908 void *base = p->lut[i * p->_clustentries].vaddr; 909 ssize_t relofs = (ssize_t) vaddr - (ssize_t) base; 910 911 /* Given address, is out of the scope of the current cluster.*/ 912 if (vaddr < base || relofs >= p->_clustsize) 913 continue; 914 915 j = j + relofs / p->_objsize; 916 /* KASSERT(j != 0, ("Cannot free object 0")); */ 917 netmap_obj_free(p, j); 918 return; 919 } 920 D("address %p is not contained inside any cluster (%s)", 921 vaddr, p->name); 922 } 923 924 #define netmap_mem_bufsize(n) \ 925 ((n)->pools[NETMAP_BUF_POOL]._objsize) 926 927 #define netmap_if_malloc(n, len) netmap_obj_malloc(&(n)->pools[NETMAP_IF_POOL], len, NULL, NULL) 928 #define netmap_if_free(n, v) netmap_obj_free_va(&(n)->pools[NETMAP_IF_POOL], (v)) 929 #define netmap_ring_malloc(n, len) netmap_obj_malloc(&(n)->pools[NETMAP_RING_POOL], len, NULL, NULL) 930 #define netmap_ring_free(n, v) netmap_obj_free_va(&(n)->pools[NETMAP_RING_POOL], (v)) 931 #define netmap_buf_malloc(n, _pos, _index) \ 932 netmap_obj_malloc(&(n)->pools[NETMAP_BUF_POOL], netmap_mem_bufsize(n), _pos, _index) 933 934 935 #if 0 // XXX unused 936 /* Return the index associated to the given packet buffer */ 937 #define netmap_buf_index(n, v) \ 938 (netmap_obj_offset(&(n)->pools[NETMAP_BUF_POOL], (v)) / NETMAP_BDG_BUF_SIZE(n)) 939 #endif 940 941 /* 942 * allocate extra buffers in a linked list. 943 * returns the actual number. 944 */ 945 uint32_t 946 netmap_extra_alloc(struct netmap_adapter *na, uint32_t *head, uint32_t n) 947 { 948 struct netmap_mem_d *nmd = na->nm_mem; 949 uint32_t i, pos = 0; /* opaque, scan position in the bitmap */ 950 951 NMA_LOCK(nmd); 952 953 *head = 0; /* default, 'null' index ie empty list */ 954 for (i = 0 ; i < n; i++) { 955 uint32_t cur = *head; /* save current head */ 956 uint32_t *p = netmap_buf_malloc(nmd, &pos, head); 957 if (p == NULL) { 958 D("no more buffers after %d of %d", i, n); 959 *head = cur; /* restore */ 960 break; 961 } 962 ND(5, "allocate buffer %d -> %d", *head, cur); 963 *p = cur; /* link to previous head */ 964 } 965 966 NMA_UNLOCK(nmd); 967 968 return i; 969 } 970 971 static void 972 netmap_extra_free(struct netmap_adapter *na, uint32_t head) 973 { 974 struct lut_entry *lut = na->na_lut.lut; 975 struct netmap_mem_d *nmd = na->nm_mem; 976 struct netmap_obj_pool *p = &nmd->pools[NETMAP_BUF_POOL]; 977 uint32_t i, cur, *buf; 978 979 ND("freeing the extra list"); 980 for (i = 0; head >=2 && head < p->objtotal; i++) { 981 cur = head; 982 buf = lut[head].vaddr; 983 head = *buf; 984 *buf = 0; 985 if (netmap_obj_free(p, cur)) 986 break; 987 } 988 if (head != 0) 989 D("breaking with head %d", head); 990 if (netmap_verbose) 991 D("freed %d buffers", i); 992 } 993 994 995 /* Return nonzero on error */ 996 static int 997 netmap_new_bufs(struct netmap_mem_d *nmd, struct netmap_slot *slot, u_int n) 998 { 999 struct netmap_obj_pool *p = &nmd->pools[NETMAP_BUF_POOL]; 1000 u_int i = 0; /* slot counter */ 1001 uint32_t pos = 0; /* slot in p->bitmap */ 1002 uint32_t index = 0; /* buffer index */ 1003 1004 for (i = 0; i < n; i++) { 1005 void *vaddr = netmap_buf_malloc(nmd, &pos, &index); 1006 if (vaddr == NULL) { 1007 D("no more buffers after %d of %d", i, n); 1008 goto cleanup; 1009 } 1010 slot[i].buf_idx = index; 1011 slot[i].len = p->_objsize; 1012 slot[i].flags = 0; 1013 } 1014 1015 ND("allocated %d buffers, %d available, first at %d", n, p->objfree, pos); 1016 return (0); 1017 1018 cleanup: 1019 while (i > 0) { 1020 i--; 1021 netmap_obj_free(p, slot[i].buf_idx); 1022 } 1023 bzero(slot, n * sizeof(slot[0])); 1024 return (ENOMEM); 1025 } 1026 1027 static void 1028 netmap_mem_set_ring(struct netmap_mem_d *nmd, struct netmap_slot *slot, u_int n, uint32_t index) 1029 { 1030 struct netmap_obj_pool *p = &nmd->pools[NETMAP_BUF_POOL]; 1031 u_int i; 1032 1033 for (i = 0; i < n; i++) { 1034 slot[i].buf_idx = index; 1035 slot[i].len = p->_objsize; 1036 slot[i].flags = 0; 1037 } 1038 } 1039 1040 1041 static void 1042 netmap_free_buf(struct netmap_mem_d *nmd, uint32_t i) 1043 { 1044 struct netmap_obj_pool *p = &nmd->pools[NETMAP_BUF_POOL]; 1045 1046 if (i < 2 || i >= p->objtotal) { 1047 D("Cannot free buf#%d: should be in [2, %d[", i, p->objtotal); 1048 return; 1049 } 1050 netmap_obj_free(p, i); 1051 } 1052 1053 1054 static void 1055 netmap_free_bufs(struct netmap_mem_d *nmd, struct netmap_slot *slot, u_int n) 1056 { 1057 u_int i; 1058 1059 for (i = 0; i < n; i++) { 1060 if (slot[i].buf_idx > 2) 1061 netmap_free_buf(nmd, slot[i].buf_idx); 1062 } 1063 } 1064 1065 static void 1066 netmap_reset_obj_allocator(struct netmap_obj_pool *p) 1067 { 1068 1069 if (p == NULL) 1070 return; 1071 if (p->bitmap) 1072 nm_os_free(p->bitmap); 1073 p->bitmap = NULL; 1074 if (p->lut) { 1075 u_int i; 1076 1077 /* 1078 * Free each cluster allocated in 1079 * netmap_finalize_obj_allocator(). The cluster start 1080 * addresses are stored at multiples of p->_clusterentries 1081 * in the lut. 1082 */ 1083 for (i = 0; i < p->objtotal; i += p->_clustentries) { 1084 if (p->lut[i].vaddr) 1085 contigfree(p->lut[i].vaddr, p->_clustsize, M_NETMAP); 1086 } 1087 bzero(p->lut, sizeof(struct lut_entry) * p->objtotal); 1088 #ifdef linux 1089 vfree(p->lut); 1090 #else 1091 nm_os_free(p->lut); 1092 #endif 1093 } 1094 p->lut = NULL; 1095 p->objtotal = 0; 1096 p->memtotal = 0; 1097 p->numclusters = 0; 1098 p->objfree = 0; 1099 } 1100 1101 /* 1102 * Free all resources related to an allocator. 1103 */ 1104 static void 1105 netmap_destroy_obj_allocator(struct netmap_obj_pool *p) 1106 { 1107 if (p == NULL) 1108 return; 1109 netmap_reset_obj_allocator(p); 1110 } 1111 1112 /* 1113 * We receive a request for objtotal objects, of size objsize each. 1114 * Internally we may round up both numbers, as we allocate objects 1115 * in small clusters multiple of the page size. 1116 * We need to keep track of objtotal and clustentries, 1117 * as they are needed when freeing memory. 1118 * 1119 * XXX note -- userspace needs the buffers to be contiguous, 1120 * so we cannot afford gaps at the end of a cluster. 1121 */ 1122 1123 1124 /* call with NMA_LOCK held */ 1125 static int 1126 netmap_config_obj_allocator(struct netmap_obj_pool *p, u_int objtotal, u_int objsize) 1127 { 1128 int i; 1129 u_int clustsize; /* the cluster size, multiple of page size */ 1130 u_int clustentries; /* how many objects per entry */ 1131 1132 /* we store the current request, so we can 1133 * detect configuration changes later */ 1134 p->r_objtotal = objtotal; 1135 p->r_objsize = objsize; 1136 1137 #define MAX_CLUSTSIZE (1<<22) // 4 MB 1138 #define LINE_ROUND NM_CACHE_ALIGN // 64 1139 if (objsize >= MAX_CLUSTSIZE) { 1140 /* we could do it but there is no point */ 1141 D("unsupported allocation for %d bytes", objsize); 1142 return EINVAL; 1143 } 1144 /* make sure objsize is a multiple of LINE_ROUND */ 1145 i = (objsize & (LINE_ROUND - 1)); 1146 if (i) { 1147 D("XXX aligning object by %d bytes", LINE_ROUND - i); 1148 objsize += LINE_ROUND - i; 1149 } 1150 if (objsize < p->objminsize || objsize > p->objmaxsize) { 1151 D("requested objsize %d out of range [%d, %d]", 1152 objsize, p->objminsize, p->objmaxsize); 1153 return EINVAL; 1154 } 1155 if (objtotal < p->nummin || objtotal > p->nummax) { 1156 D("requested objtotal %d out of range [%d, %d]", 1157 objtotal, p->nummin, p->nummax); 1158 return EINVAL; 1159 } 1160 /* 1161 * Compute number of objects using a brute-force approach: 1162 * given a max cluster size, 1163 * we try to fill it with objects keeping track of the 1164 * wasted space to the next page boundary. 1165 */ 1166 for (clustentries = 0, i = 1;; i++) { 1167 u_int delta, used = i * objsize; 1168 if (used > MAX_CLUSTSIZE) 1169 break; 1170 delta = used % PAGE_SIZE; 1171 if (delta == 0) { // exact solution 1172 clustentries = i; 1173 break; 1174 } 1175 } 1176 /* exact solution not found */ 1177 if (clustentries == 0) { 1178 D("unsupported allocation for %d bytes", objsize); 1179 return EINVAL; 1180 } 1181 /* compute clustsize */ 1182 clustsize = clustentries * objsize; 1183 if (netmap_verbose) 1184 D("objsize %d clustsize %d objects %d", 1185 objsize, clustsize, clustentries); 1186 1187 /* 1188 * The number of clusters is n = ceil(objtotal/clustentries) 1189 * objtotal' = n * clustentries 1190 */ 1191 p->_clustentries = clustentries; 1192 p->_clustsize = clustsize; 1193 p->_numclusters = (objtotal + clustentries - 1) / clustentries; 1194 1195 /* actual values (may be larger than requested) */ 1196 p->_objsize = objsize; 1197 p->_objtotal = p->_numclusters * clustentries; 1198 1199 return 0; 1200 } 1201 1202 static struct lut_entry * 1203 nm_alloc_lut(u_int nobj) 1204 { 1205 size_t n = sizeof(struct lut_entry) * nobj; 1206 struct lut_entry *lut; 1207 #ifdef linux 1208 lut = vmalloc(n); 1209 #else 1210 lut = nm_os_malloc(n); 1211 #endif 1212 return lut; 1213 } 1214 1215 /* call with NMA_LOCK held */ 1216 static int 1217 netmap_finalize_obj_allocator(struct netmap_obj_pool *p) 1218 { 1219 int i; /* must be signed */ 1220 size_t n; 1221 1222 /* optimistically assume we have enough memory */ 1223 p->numclusters = p->_numclusters; 1224 p->objtotal = p->_objtotal; 1225 1226 p->lut = nm_alloc_lut(p->objtotal); 1227 if (p->lut == NULL) { 1228 D("Unable to create lookup table for '%s'", p->name); 1229 goto clean; 1230 } 1231 1232 /* Allocate the bitmap */ 1233 n = (p->objtotal + 31) / 32; 1234 p->bitmap = nm_os_malloc(sizeof(uint32_t) * n); 1235 if (p->bitmap == NULL) { 1236 D("Unable to create bitmap (%d entries) for allocator '%s'", (int)n, 1237 p->name); 1238 goto clean; 1239 } 1240 p->bitmap_slots = n; 1241 1242 /* 1243 * Allocate clusters, init pointers and bitmap 1244 */ 1245 1246 n = p->_clustsize; 1247 for (i = 0; i < (int)p->objtotal;) { 1248 int lim = i + p->_clustentries; 1249 char *clust; 1250 1251 /* 1252 * XXX Note, we only need contigmalloc() for buffers attached 1253 * to native interfaces. In all other cases (nifp, netmap rings 1254 * and even buffers for VALE ports or emulated interfaces) we 1255 * can live with standard malloc, because the hardware will not 1256 * access the pages directly. 1257 */ 1258 clust = contigmalloc(n, M_NETMAP, M_NOWAIT | M_ZERO, 1259 (size_t)0, -1UL, PAGE_SIZE, 0); 1260 if (clust == NULL) { 1261 /* 1262 * If we get here, there is a severe memory shortage, 1263 * so halve the allocated memory to reclaim some. 1264 */ 1265 D("Unable to create cluster at %d for '%s' allocator", 1266 i, p->name); 1267 if (i < 2) /* nothing to halve */ 1268 goto out; 1269 lim = i / 2; 1270 for (i--; i >= lim; i--) { 1271 p->bitmap[ (i>>5) ] &= ~( 1 << (i & 31) ); 1272 if (i % p->_clustentries == 0 && p->lut[i].vaddr) 1273 contigfree(p->lut[i].vaddr, 1274 n, M_NETMAP); 1275 p->lut[i].vaddr = NULL; 1276 } 1277 out: 1278 p->objtotal = i; 1279 /* we may have stopped in the middle of a cluster */ 1280 p->numclusters = (i + p->_clustentries - 1) / p->_clustentries; 1281 break; 1282 } 1283 /* 1284 * Set bitmap and lut state for all buffers in the current 1285 * cluster. 1286 * 1287 * [i, lim) is the set of buffer indexes that cover the 1288 * current cluster. 1289 * 1290 * 'clust' is really the address of the current buffer in 1291 * the current cluster as we index through it with a stride 1292 * of p->_objsize. 1293 */ 1294 for (; i < lim; i++, clust += p->_objsize) { 1295 p->bitmap[ (i>>5) ] |= ( 1 << (i & 31) ); 1296 p->lut[i].vaddr = clust; 1297 p->lut[i].paddr = vtophys(clust); 1298 } 1299 } 1300 p->objfree = p->objtotal; 1301 p->memtotal = p->numclusters * p->_clustsize; 1302 if (p->objfree == 0) 1303 goto clean; 1304 if (netmap_verbose) 1305 D("Pre-allocated %d clusters (%d/%dKB) for '%s'", 1306 p->numclusters, p->_clustsize >> 10, 1307 p->memtotal >> 10, p->name); 1308 1309 return 0; 1310 1311 clean: 1312 netmap_reset_obj_allocator(p); 1313 return ENOMEM; 1314 } 1315 1316 /* call with lock held */ 1317 static int 1318 netmap_mem_params_changed(struct netmap_obj_params* p) 1319 { 1320 int i, rv = 0; 1321 1322 for (i = 0; i < NETMAP_POOLS_NR; i++) { 1323 if (p[i].last_size != p[i].size || p[i].last_num != p[i].num) { 1324 p[i].last_size = p[i].size; 1325 p[i].last_num = p[i].num; 1326 rv = 1; 1327 } 1328 } 1329 return rv; 1330 } 1331 1332 static void 1333 netmap_mem_reset_all(struct netmap_mem_d *nmd) 1334 { 1335 int i; 1336 1337 if (netmap_verbose) 1338 D("resetting %p", nmd); 1339 for (i = 0; i < NETMAP_POOLS_NR; i++) { 1340 netmap_reset_obj_allocator(&nmd->pools[i]); 1341 } 1342 nmd->flags &= ~NETMAP_MEM_FINALIZED; 1343 } 1344 1345 static int 1346 netmap_mem_unmap(struct netmap_obj_pool *p, struct netmap_adapter *na) 1347 { 1348 int i, lim = p->_objtotal; 1349 1350 if (na == NULL || na->pdev == NULL) 1351 return 0; 1352 1353 #if defined(__FreeBSD__) 1354 (void)i; 1355 (void)lim; 1356 D("unsupported on FreeBSD"); 1357 1358 #elif defined(_WIN32) 1359 (void)i; 1360 (void)lim; 1361 D("unsupported on Windows"); //XXX_ale, really? 1362 #else /* linux */ 1363 for (i = 2; i < lim; i++) { 1364 netmap_unload_map(na, (bus_dma_tag_t) na->pdev, &p->lut[i].paddr); 1365 } 1366 #endif /* linux */ 1367 1368 return 0; 1369 } 1370 1371 static int 1372 netmap_mem_map(struct netmap_obj_pool *p, struct netmap_adapter *na) 1373 { 1374 #if defined(__FreeBSD__) 1375 D("unsupported on FreeBSD"); 1376 #elif defined(_WIN32) 1377 D("unsupported on Windows"); //XXX_ale, really? 1378 #else /* linux */ 1379 int i, lim = p->_objtotal; 1380 1381 if (na->pdev == NULL) 1382 return 0; 1383 1384 for (i = 2; i < lim; i++) { 1385 netmap_load_map(na, (bus_dma_tag_t) na->pdev, &p->lut[i].paddr, 1386 p->lut[i].vaddr); 1387 } 1388 #endif /* linux */ 1389 1390 return 0; 1391 } 1392 1393 static int 1394 netmap_mem_finalize_all(struct netmap_mem_d *nmd) 1395 { 1396 int i; 1397 if (nmd->flags & NETMAP_MEM_FINALIZED) 1398 return 0; 1399 nmd->lasterr = 0; 1400 nmd->nm_totalsize = 0; 1401 for (i = 0; i < NETMAP_POOLS_NR; i++) { 1402 nmd->lasterr = netmap_finalize_obj_allocator(&nmd->pools[i]); 1403 if (nmd->lasterr) 1404 goto error; 1405 nmd->nm_totalsize += nmd->pools[i].memtotal; 1406 } 1407 /* buffers 0 and 1 are reserved */ 1408 nmd->pools[NETMAP_BUF_POOL].objfree -= 2; 1409 nmd->pools[NETMAP_BUF_POOL].bitmap[0] = ~3; 1410 nmd->flags |= NETMAP_MEM_FINALIZED; 1411 1412 if (netmap_verbose) 1413 D("interfaces %d KB, rings %d KB, buffers %d MB", 1414 nmd->pools[NETMAP_IF_POOL].memtotal >> 10, 1415 nmd->pools[NETMAP_RING_POOL].memtotal >> 10, 1416 nmd->pools[NETMAP_BUF_POOL].memtotal >> 20); 1417 1418 if (netmap_verbose) 1419 D("Free buffers: %d", nmd->pools[NETMAP_BUF_POOL].objfree); 1420 1421 1422 return 0; 1423 error: 1424 netmap_mem_reset_all(nmd); 1425 return nmd->lasterr; 1426 } 1427 1428 /* 1429 * allocator for private memory 1430 */ 1431 static struct netmap_mem_d * 1432 _netmap_mem_private_new(struct netmap_obj_params *p, int *perr) 1433 { 1434 struct netmap_mem_d *d = NULL; 1435 int i, err = 0; 1436 1437 d = nm_os_malloc(sizeof(struct netmap_mem_d)); 1438 if (d == NULL) { 1439 err = ENOMEM; 1440 goto error; 1441 } 1442 1443 *d = nm_blueprint; 1444 1445 err = nm_mem_assign_id(d); 1446 if (err) 1447 goto error; 1448 snprintf(d->name, NM_MEM_NAMESZ, "%d", d->nm_id); 1449 1450 for (i = 0; i < NETMAP_POOLS_NR; i++) { 1451 snprintf(d->pools[i].name, NETMAP_POOL_MAX_NAMSZ, 1452 nm_blueprint.pools[i].name, 1453 d->name); 1454 d->params[i].num = p[i].num; 1455 d->params[i].size = p[i].size; 1456 } 1457 1458 NMA_LOCK_INIT(d); 1459 1460 err = netmap_mem_config(d); 1461 if (err) 1462 goto error; 1463 1464 d->flags &= ~NETMAP_MEM_FINALIZED; 1465 1466 return d; 1467 1468 error: 1469 netmap_mem_delete(d); 1470 if (perr) 1471 *perr = err; 1472 return NULL; 1473 } 1474 1475 struct netmap_mem_d * 1476 netmap_mem_private_new(u_int txr, u_int txd, u_int rxr, u_int rxd, 1477 u_int extra_bufs, u_int npipes, int *perr) 1478 { 1479 struct netmap_mem_d *d = NULL; 1480 struct netmap_obj_params p[NETMAP_POOLS_NR]; 1481 int i, err = 0; 1482 u_int v, maxd; 1483 /* account for the fake host rings */ 1484 txr++; 1485 rxr++; 1486 1487 /* copy the min values */ 1488 for (i = 0; i < NETMAP_POOLS_NR; i++) { 1489 p[i] = netmap_min_priv_params[i]; 1490 } 1491 1492 /* possibly increase them to fit user request */ 1493 v = sizeof(struct netmap_if) + sizeof(ssize_t) * (txr + rxr); 1494 if (p[NETMAP_IF_POOL].size < v) 1495 p[NETMAP_IF_POOL].size = v; 1496 v = 2 + 4 * npipes; 1497 if (p[NETMAP_IF_POOL].num < v) 1498 p[NETMAP_IF_POOL].num = v; 1499 maxd = (txd > rxd) ? txd : rxd; 1500 v = sizeof(struct netmap_ring) + sizeof(struct netmap_slot) * maxd; 1501 if (p[NETMAP_RING_POOL].size < v) 1502 p[NETMAP_RING_POOL].size = v; 1503 /* each pipe endpoint needs two tx rings (1 normal + 1 host, fake) 1504 * and two rx rings (again, 1 normal and 1 fake host) 1505 */ 1506 v = txr + rxr + 8 * npipes; 1507 if (p[NETMAP_RING_POOL].num < v) 1508 p[NETMAP_RING_POOL].num = v; 1509 /* for each pipe we only need the buffers for the 4 "real" rings. 1510 * On the other end, the pipe ring dimension may be different from 1511 * the parent port ring dimension. As a compromise, we allocate twice the 1512 * space actually needed if the pipe rings were the same size as the parent rings 1513 */ 1514 v = (4 * npipes + rxr) * rxd + (4 * npipes + txr) * txd + 2 + extra_bufs; 1515 /* the +2 is for the tx and rx fake buffers (indices 0 and 1) */ 1516 if (p[NETMAP_BUF_POOL].num < v) 1517 p[NETMAP_BUF_POOL].num = v; 1518 1519 if (netmap_verbose) 1520 D("req if %d*%d ring %d*%d buf %d*%d", 1521 p[NETMAP_IF_POOL].num, 1522 p[NETMAP_IF_POOL].size, 1523 p[NETMAP_RING_POOL].num, 1524 p[NETMAP_RING_POOL].size, 1525 p[NETMAP_BUF_POOL].num, 1526 p[NETMAP_BUF_POOL].size); 1527 1528 d = _netmap_mem_private_new(p, perr); 1529 if (d == NULL) 1530 goto error; 1531 1532 return d; 1533 error: 1534 netmap_mem_delete(d); 1535 if (perr) 1536 *perr = err; 1537 return NULL; 1538 } 1539 1540 1541 /* call with lock held */ 1542 static int 1543 netmap_mem2_config(struct netmap_mem_d *nmd) 1544 { 1545 int i; 1546 1547 if (nmd->active) 1548 /* already in use, we cannot change the configuration */ 1549 goto out; 1550 1551 if (!netmap_mem_params_changed(nmd->params)) 1552 goto out; 1553 1554 ND("reconfiguring"); 1555 1556 if (nmd->flags & NETMAP_MEM_FINALIZED) { 1557 /* reset previous allocation */ 1558 for (i = 0; i < NETMAP_POOLS_NR; i++) { 1559 netmap_reset_obj_allocator(&nmd->pools[i]); 1560 } 1561 nmd->flags &= ~NETMAP_MEM_FINALIZED; 1562 } 1563 1564 for (i = 0; i < NETMAP_POOLS_NR; i++) { 1565 nmd->lasterr = netmap_config_obj_allocator(&nmd->pools[i], 1566 nmd->params[i].num, nmd->params[i].size); 1567 if (nmd->lasterr) 1568 goto out; 1569 } 1570 1571 out: 1572 1573 return nmd->lasterr; 1574 } 1575 1576 static int 1577 netmap_mem2_finalize(struct netmap_mem_d *nmd) 1578 { 1579 int err; 1580 1581 /* update configuration if changed */ 1582 if (netmap_mem2_config(nmd)) 1583 goto out1; 1584 1585 nmd->active++; 1586 1587 if (nmd->flags & NETMAP_MEM_FINALIZED) { 1588 /* may happen if config is not changed */ 1589 ND("nothing to do"); 1590 goto out; 1591 } 1592 1593 if (netmap_mem_finalize_all(nmd)) 1594 goto out; 1595 1596 nmd->lasterr = 0; 1597 1598 out: 1599 if (nmd->lasterr) 1600 nmd->active--; 1601 out1: 1602 err = nmd->lasterr; 1603 1604 return err; 1605 1606 } 1607 1608 static void 1609 netmap_mem2_delete(struct netmap_mem_d *nmd) 1610 { 1611 int i; 1612 1613 for (i = 0; i < NETMAP_POOLS_NR; i++) { 1614 netmap_destroy_obj_allocator(&nmd->pools[i]); 1615 } 1616 1617 NMA_LOCK_DESTROY(nmd); 1618 if (nmd != &nm_mem) 1619 nm_os_free(nmd); 1620 } 1621 1622 int 1623 netmap_mem_init(void) 1624 { 1625 NM_MTX_INIT(nm_mem_list_lock); 1626 NMA_LOCK_INIT(&nm_mem); 1627 netmap_mem_get(&nm_mem); 1628 return (0); 1629 } 1630 1631 void 1632 netmap_mem_fini(void) 1633 { 1634 netmap_mem_put(&nm_mem); 1635 } 1636 1637 static void 1638 netmap_free_rings(struct netmap_adapter *na) 1639 { 1640 enum txrx t; 1641 1642 for_rx_tx(t) { 1643 u_int i; 1644 for (i = 0; i < nma_get_nrings(na, t) + 1; i++) { 1645 struct netmap_kring *kring = &NMR(na, t)[i]; 1646 struct netmap_ring *ring = kring->ring; 1647 1648 if (ring == NULL || kring->users > 0 || (kring->nr_kflags & NKR_NEEDRING)) { 1649 ND("skipping ring %s (ring %p, users %d)", 1650 kring->name, ring, kring->users); 1651 continue; 1652 } 1653 if (i != nma_get_nrings(na, t) || na->na_flags & NAF_HOST_RINGS) 1654 netmap_free_bufs(na->nm_mem, ring->slot, kring->nkr_num_slots); 1655 netmap_ring_free(na->nm_mem, ring); 1656 kring->ring = NULL; 1657 } 1658 } 1659 } 1660 1661 /* call with NMA_LOCK held * 1662 * 1663 * Allocate netmap rings and buffers for this card 1664 * The rings are contiguous, but have variable size. 1665 * The kring array must follow the layout described 1666 * in netmap_krings_create(). 1667 */ 1668 static int 1669 netmap_mem2_rings_create(struct netmap_adapter *na) 1670 { 1671 enum txrx t; 1672 1673 NMA_LOCK(na->nm_mem); 1674 1675 for_rx_tx(t) { 1676 u_int i; 1677 1678 for (i = 0; i <= nma_get_nrings(na, t); i++) { 1679 struct netmap_kring *kring = &NMR(na, t)[i]; 1680 struct netmap_ring *ring = kring->ring; 1681 u_int len, ndesc; 1682 1683 if (ring || (!kring->users && !(kring->nr_kflags & NKR_NEEDRING))) { 1684 /* uneeded, or already created by somebody else */ 1685 ND("skipping ring %s", kring->name); 1686 continue; 1687 } 1688 ndesc = kring->nkr_num_slots; 1689 len = sizeof(struct netmap_ring) + 1690 ndesc * sizeof(struct netmap_slot); 1691 ring = netmap_ring_malloc(na->nm_mem, len); 1692 if (ring == NULL) { 1693 D("Cannot allocate %s_ring", nm_txrx2str(t)); 1694 goto cleanup; 1695 } 1696 ND("txring at %p", ring); 1697 kring->ring = ring; 1698 *(uint32_t *)(uintptr_t)&ring->num_slots = ndesc; 1699 *(int64_t *)(uintptr_t)&ring->buf_ofs = 1700 (na->nm_mem->pools[NETMAP_IF_POOL].memtotal + 1701 na->nm_mem->pools[NETMAP_RING_POOL].memtotal) - 1702 netmap_ring_offset(na->nm_mem, ring); 1703 1704 /* copy values from kring */ 1705 ring->head = kring->rhead; 1706 ring->cur = kring->rcur; 1707 ring->tail = kring->rtail; 1708 *(uint16_t *)(uintptr_t)&ring->nr_buf_size = 1709 netmap_mem_bufsize(na->nm_mem); 1710 ND("%s h %d c %d t %d", kring->name, 1711 ring->head, ring->cur, ring->tail); 1712 ND("initializing slots for %s_ring", nm_txrx2str(txrx)); 1713 if (i != nma_get_nrings(na, t) || (na->na_flags & NAF_HOST_RINGS)) { 1714 /* this is a real ring */ 1715 if (netmap_new_bufs(na->nm_mem, ring->slot, ndesc)) { 1716 D("Cannot allocate buffers for %s_ring", nm_txrx2str(t)); 1717 goto cleanup; 1718 } 1719 } else { 1720 /* this is a fake ring, set all indices to 0 */ 1721 netmap_mem_set_ring(na->nm_mem, ring->slot, ndesc, 0); 1722 } 1723 /* ring info */ 1724 *(uint16_t *)(uintptr_t)&ring->ringid = kring->ring_id; 1725 *(uint16_t *)(uintptr_t)&ring->dir = kring->tx; 1726 } 1727 } 1728 1729 NMA_UNLOCK(na->nm_mem); 1730 1731 return 0; 1732 1733 cleanup: 1734 netmap_free_rings(na); 1735 1736 NMA_UNLOCK(na->nm_mem); 1737 1738 return ENOMEM; 1739 } 1740 1741 static void 1742 netmap_mem2_rings_delete(struct netmap_adapter *na) 1743 { 1744 /* last instance, release bufs and rings */ 1745 NMA_LOCK(na->nm_mem); 1746 1747 netmap_free_rings(na); 1748 1749 NMA_UNLOCK(na->nm_mem); 1750 } 1751 1752 1753 /* call with NMA_LOCK held */ 1754 /* 1755 * Allocate the per-fd structure netmap_if. 1756 * 1757 * We assume that the configuration stored in na 1758 * (number of tx/rx rings and descs) does not change while 1759 * the interface is in netmap mode. 1760 */ 1761 static struct netmap_if * 1762 netmap_mem2_if_new(struct netmap_adapter *na, struct netmap_priv_d *priv) 1763 { 1764 struct netmap_if *nifp; 1765 ssize_t base; /* handy for relative offsets between rings and nifp */ 1766 u_int i, len, n[NR_TXRX], ntot; 1767 enum txrx t; 1768 1769 ntot = 0; 1770 for_rx_tx(t) { 1771 /* account for the (eventually fake) host rings */ 1772 n[t] = nma_get_nrings(na, t) + 1; 1773 ntot += n[t]; 1774 } 1775 /* 1776 * the descriptor is followed inline by an array of offsets 1777 * to the tx and rx rings in the shared memory region. 1778 */ 1779 1780 NMA_LOCK(na->nm_mem); 1781 1782 len = sizeof(struct netmap_if) + (ntot * sizeof(ssize_t)); 1783 nifp = netmap_if_malloc(na->nm_mem, len); 1784 if (nifp == NULL) { 1785 NMA_UNLOCK(na->nm_mem); 1786 return NULL; 1787 } 1788 1789 /* initialize base fields -- override const */ 1790 *(u_int *)(uintptr_t)&nifp->ni_tx_rings = na->num_tx_rings; 1791 *(u_int *)(uintptr_t)&nifp->ni_rx_rings = na->num_rx_rings; 1792 strncpy(nifp->ni_name, na->name, (size_t)IFNAMSIZ); 1793 1794 /* 1795 * fill the slots for the rx and tx rings. They contain the offset 1796 * between the ring and nifp, so the information is usable in 1797 * userspace to reach the ring from the nifp. 1798 */ 1799 base = netmap_if_offset(na->nm_mem, nifp); 1800 for (i = 0; i < n[NR_TX]; i++) { 1801 /* XXX instead of ofs == 0 maybe use the offset of an error 1802 * ring, like we do for buffers? */ 1803 ssize_t ofs = 0; 1804 1805 if (na->tx_rings[i].ring != NULL && i >= priv->np_qfirst[NR_TX] 1806 && i < priv->np_qlast[NR_TX]) { 1807 ofs = netmap_ring_offset(na->nm_mem, 1808 na->tx_rings[i].ring) - base; 1809 } 1810 *(ssize_t *)(uintptr_t)&nifp->ring_ofs[i] = ofs; 1811 } 1812 for (i = 0; i < n[NR_RX]; i++) { 1813 /* XXX instead of ofs == 0 maybe use the offset of an error 1814 * ring, like we do for buffers? */ 1815 ssize_t ofs = 0; 1816 1817 if (na->rx_rings[i].ring != NULL && i >= priv->np_qfirst[NR_RX] 1818 && i < priv->np_qlast[NR_RX]) { 1819 ofs = netmap_ring_offset(na->nm_mem, 1820 na->rx_rings[i].ring) - base; 1821 } 1822 *(ssize_t *)(uintptr_t)&nifp->ring_ofs[i+n[NR_TX]] = ofs; 1823 } 1824 1825 NMA_UNLOCK(na->nm_mem); 1826 1827 return (nifp); 1828 } 1829 1830 static void 1831 netmap_mem2_if_delete(struct netmap_adapter *na, struct netmap_if *nifp) 1832 { 1833 if (nifp == NULL) 1834 /* nothing to do */ 1835 return; 1836 NMA_LOCK(na->nm_mem); 1837 if (nifp->ni_bufs_head) 1838 netmap_extra_free(na, nifp->ni_bufs_head); 1839 netmap_if_free(na->nm_mem, nifp); 1840 1841 NMA_UNLOCK(na->nm_mem); 1842 } 1843 1844 static void 1845 netmap_mem2_deref(struct netmap_mem_d *nmd) 1846 { 1847 1848 nmd->active--; 1849 if (!nmd->active) 1850 nmd->nm_grp = -1; 1851 if (netmap_verbose) 1852 D("active = %d", nmd->active); 1853 1854 } 1855 1856 struct netmap_mem_ops netmap_mem_global_ops = { 1857 .nmd_get_lut = netmap_mem2_get_lut, 1858 .nmd_get_info = netmap_mem2_get_info, 1859 .nmd_ofstophys = netmap_mem2_ofstophys, 1860 .nmd_config = netmap_mem2_config, 1861 .nmd_finalize = netmap_mem2_finalize, 1862 .nmd_deref = netmap_mem2_deref, 1863 .nmd_delete = netmap_mem2_delete, 1864 .nmd_if_offset = netmap_mem2_if_offset, 1865 .nmd_if_new = netmap_mem2_if_new, 1866 .nmd_if_delete = netmap_mem2_if_delete, 1867 .nmd_rings_create = netmap_mem2_rings_create, 1868 .nmd_rings_delete = netmap_mem2_rings_delete 1869 }; 1870 1871 int 1872 netmap_mem_pools_info_get(struct nmreq *nmr, struct netmap_mem_d *nmd) 1873 { 1874 uintptr_t *pp = (uintptr_t *)&nmr->nr_arg1; 1875 struct netmap_pools_info *upi = (struct netmap_pools_info *)(*pp); 1876 struct netmap_pools_info pi; 1877 unsigned int memsize; 1878 uint16_t memid; 1879 int ret; 1880 1881 ret = netmap_mem_get_info(nmd, &memsize, NULL, &memid); 1882 if (ret) { 1883 return ret; 1884 } 1885 1886 pi.memsize = memsize; 1887 pi.memid = memid; 1888 NMA_LOCK(nmd); 1889 pi.if_pool_offset = 0; 1890 pi.if_pool_objtotal = nmd->pools[NETMAP_IF_POOL].objtotal; 1891 pi.if_pool_objsize = nmd->pools[NETMAP_IF_POOL]._objsize; 1892 1893 pi.ring_pool_offset = nmd->pools[NETMAP_IF_POOL].memtotal; 1894 pi.ring_pool_objtotal = nmd->pools[NETMAP_RING_POOL].objtotal; 1895 pi.ring_pool_objsize = nmd->pools[NETMAP_RING_POOL]._objsize; 1896 1897 pi.buf_pool_offset = nmd->pools[NETMAP_IF_POOL].memtotal + 1898 nmd->pools[NETMAP_RING_POOL].memtotal; 1899 pi.buf_pool_objtotal = nmd->pools[NETMAP_BUF_POOL].objtotal; 1900 pi.buf_pool_objsize = nmd->pools[NETMAP_BUF_POOL]._objsize; 1901 NMA_UNLOCK(nmd); 1902 1903 ret = copyout(&pi, upi, sizeof(pi)); 1904 if (ret) { 1905 return ret; 1906 } 1907 1908 return 0; 1909 } 1910 1911 #ifdef WITH_PTNETMAP_GUEST 1912 struct mem_pt_if { 1913 struct mem_pt_if *next; 1914 struct ifnet *ifp; 1915 unsigned int nifp_offset; 1916 }; 1917 1918 /* Netmap allocator for ptnetmap guests. */ 1919 struct netmap_mem_ptg { 1920 struct netmap_mem_d up; 1921 1922 vm_paddr_t nm_paddr; /* physical address in the guest */ 1923 void *nm_addr; /* virtual address in the guest */ 1924 struct netmap_lut buf_lut; /* lookup table for BUF pool in the guest */ 1925 nm_memid_t host_mem_id; /* allocator identifier in the host */ 1926 struct ptnetmap_memdev *ptn_dev;/* ptnetmap memdev */ 1927 struct mem_pt_if *pt_ifs; /* list of interfaces in passthrough */ 1928 }; 1929 1930 /* Link a passthrough interface to a passthrough netmap allocator. */ 1931 static int 1932 netmap_mem_pt_guest_ifp_add(struct netmap_mem_d *nmd, struct ifnet *ifp, 1933 unsigned int nifp_offset) 1934 { 1935 struct netmap_mem_ptg *ptnmd = (struct netmap_mem_ptg *)nmd; 1936 struct mem_pt_if *ptif = nm_os_malloc(sizeof(*ptif)); 1937 1938 if (!ptif) { 1939 return ENOMEM; 1940 } 1941 1942 NMA_LOCK(nmd); 1943 1944 ptif->ifp = ifp; 1945 ptif->nifp_offset = nifp_offset; 1946 1947 if (ptnmd->pt_ifs) { 1948 ptif->next = ptnmd->pt_ifs; 1949 } 1950 ptnmd->pt_ifs = ptif; 1951 1952 NMA_UNLOCK(nmd); 1953 1954 D("added (ifp=%p,nifp_offset=%u)", ptif->ifp, ptif->nifp_offset); 1955 1956 return 0; 1957 } 1958 1959 /* Called with NMA_LOCK(nmd) held. */ 1960 static struct mem_pt_if * 1961 netmap_mem_pt_guest_ifp_lookup(struct netmap_mem_d *nmd, struct ifnet *ifp) 1962 { 1963 struct netmap_mem_ptg *ptnmd = (struct netmap_mem_ptg *)nmd; 1964 struct mem_pt_if *curr; 1965 1966 for (curr = ptnmd->pt_ifs; curr; curr = curr->next) { 1967 if (curr->ifp == ifp) { 1968 return curr; 1969 } 1970 } 1971 1972 return NULL; 1973 } 1974 1975 /* Unlink a passthrough interface from a passthrough netmap allocator. */ 1976 int 1977 netmap_mem_pt_guest_ifp_del(struct netmap_mem_d *nmd, struct ifnet *ifp) 1978 { 1979 struct netmap_mem_ptg *ptnmd = (struct netmap_mem_ptg *)nmd; 1980 struct mem_pt_if *prev = NULL; 1981 struct mem_pt_if *curr; 1982 int ret = -1; 1983 1984 NMA_LOCK(nmd); 1985 1986 for (curr = ptnmd->pt_ifs; curr; curr = curr->next) { 1987 if (curr->ifp == ifp) { 1988 if (prev) { 1989 prev->next = curr->next; 1990 } else { 1991 ptnmd->pt_ifs = curr->next; 1992 } 1993 D("removed (ifp=%p,nifp_offset=%u)", 1994 curr->ifp, curr->nifp_offset); 1995 nm_os_free(curr); 1996 ret = 0; 1997 break; 1998 } 1999 prev = curr; 2000 } 2001 2002 NMA_UNLOCK(nmd); 2003 2004 return ret; 2005 } 2006 2007 static int 2008 netmap_mem_pt_guest_get_lut(struct netmap_mem_d *nmd, struct netmap_lut *lut) 2009 { 2010 struct netmap_mem_ptg *ptnmd = (struct netmap_mem_ptg *)nmd; 2011 2012 if (!(nmd->flags & NETMAP_MEM_FINALIZED)) { 2013 return EINVAL; 2014 } 2015 2016 *lut = ptnmd->buf_lut; 2017 return 0; 2018 } 2019 2020 static int 2021 netmap_mem_pt_guest_get_info(struct netmap_mem_d *nmd, u_int *size, 2022 u_int *memflags, uint16_t *id) 2023 { 2024 int error = 0; 2025 2026 NMA_LOCK(nmd); 2027 2028 error = nmd->ops->nmd_config(nmd); 2029 if (error) 2030 goto out; 2031 2032 if (size) 2033 *size = nmd->nm_totalsize; 2034 if (memflags) 2035 *memflags = nmd->flags; 2036 if (id) 2037 *id = nmd->nm_id; 2038 2039 out: 2040 NMA_UNLOCK(nmd); 2041 2042 return error; 2043 } 2044 2045 static vm_paddr_t 2046 netmap_mem_pt_guest_ofstophys(struct netmap_mem_d *nmd, vm_ooffset_t off) 2047 { 2048 struct netmap_mem_ptg *ptnmd = (struct netmap_mem_ptg *)nmd; 2049 vm_paddr_t paddr; 2050 /* if the offset is valid, just return csb->base_addr + off */ 2051 paddr = (vm_paddr_t)(ptnmd->nm_paddr + off); 2052 ND("off %lx padr %lx", off, (unsigned long)paddr); 2053 return paddr; 2054 } 2055 2056 static int 2057 netmap_mem_pt_guest_config(struct netmap_mem_d *nmd) 2058 { 2059 /* nothing to do, we are configured on creation 2060 * and configuration never changes thereafter 2061 */ 2062 return 0; 2063 } 2064 2065 static int 2066 netmap_mem_pt_guest_finalize(struct netmap_mem_d *nmd) 2067 { 2068 struct netmap_mem_ptg *ptnmd = (struct netmap_mem_ptg *)nmd; 2069 uint64_t mem_size; 2070 uint32_t bufsize; 2071 uint32_t nbuffers; 2072 uint32_t poolofs; 2073 vm_paddr_t paddr; 2074 char *vaddr; 2075 int i; 2076 int error = 0; 2077 2078 nmd->active++; 2079 2080 if (nmd->flags & NETMAP_MEM_FINALIZED) 2081 goto out; 2082 2083 if (ptnmd->ptn_dev == NULL) { 2084 D("ptnetmap memdev not attached"); 2085 error = ENOMEM; 2086 goto err; 2087 } 2088 /* Map memory through ptnetmap-memdev BAR. */ 2089 error = nm_os_pt_memdev_iomap(ptnmd->ptn_dev, &ptnmd->nm_paddr, 2090 &ptnmd->nm_addr, &mem_size); 2091 if (error) 2092 goto err; 2093 2094 /* Initialize the lut using the information contained in the 2095 * ptnetmap memory device. */ 2096 bufsize = nm_os_pt_memdev_ioread(ptnmd->ptn_dev, 2097 PTNET_MDEV_IO_BUF_POOL_OBJSZ); 2098 nbuffers = nm_os_pt_memdev_ioread(ptnmd->ptn_dev, 2099 PTNET_MDEV_IO_BUF_POOL_OBJNUM); 2100 2101 /* allocate the lut */ 2102 if (ptnmd->buf_lut.lut == NULL) { 2103 D("allocating lut"); 2104 ptnmd->buf_lut.lut = nm_alloc_lut(nbuffers); 2105 if (ptnmd->buf_lut.lut == NULL) { 2106 D("lut allocation failed"); 2107 return ENOMEM; 2108 } 2109 } 2110 2111 /* we have physically contiguous memory mapped through PCI BAR */ 2112 poolofs = nm_os_pt_memdev_ioread(ptnmd->ptn_dev, 2113 PTNET_MDEV_IO_BUF_POOL_OFS); 2114 vaddr = (char *)(ptnmd->nm_addr) + poolofs; 2115 paddr = ptnmd->nm_paddr + poolofs; 2116 2117 for (i = 0; i < nbuffers; i++) { 2118 ptnmd->buf_lut.lut[i].vaddr = vaddr; 2119 ptnmd->buf_lut.lut[i].paddr = paddr; 2120 vaddr += bufsize; 2121 paddr += bufsize; 2122 } 2123 2124 ptnmd->buf_lut.objtotal = nbuffers; 2125 ptnmd->buf_lut.objsize = bufsize; 2126 nmd->nm_totalsize = (unsigned int)mem_size; 2127 2128 nmd->flags |= NETMAP_MEM_FINALIZED; 2129 out: 2130 return 0; 2131 err: 2132 nmd->active--; 2133 return error; 2134 } 2135 2136 static void 2137 netmap_mem_pt_guest_deref(struct netmap_mem_d *nmd) 2138 { 2139 struct netmap_mem_ptg *ptnmd = (struct netmap_mem_ptg *)nmd; 2140 2141 nmd->active--; 2142 if (nmd->active <= 0 && 2143 (nmd->flags & NETMAP_MEM_FINALIZED)) { 2144 nmd->flags &= ~NETMAP_MEM_FINALIZED; 2145 /* unmap ptnetmap-memdev memory */ 2146 if (ptnmd->ptn_dev) { 2147 nm_os_pt_memdev_iounmap(ptnmd->ptn_dev); 2148 } 2149 ptnmd->nm_addr = NULL; 2150 ptnmd->nm_paddr = 0; 2151 } 2152 } 2153 2154 static ssize_t 2155 netmap_mem_pt_guest_if_offset(struct netmap_mem_d *nmd, const void *vaddr) 2156 { 2157 struct netmap_mem_ptg *ptnmd = (struct netmap_mem_ptg *)nmd; 2158 2159 return (const char *)(vaddr) - (char *)(ptnmd->nm_addr); 2160 } 2161 2162 static void 2163 netmap_mem_pt_guest_delete(struct netmap_mem_d *nmd) 2164 { 2165 if (nmd == NULL) 2166 return; 2167 if (netmap_verbose) 2168 D("deleting %p", nmd); 2169 if (nmd->active > 0) 2170 D("bug: deleting mem allocator with active=%d!", nmd->active); 2171 if (netmap_verbose) 2172 D("done deleting %p", nmd); 2173 NMA_LOCK_DESTROY(nmd); 2174 nm_os_free(nmd); 2175 } 2176 2177 static struct netmap_if * 2178 netmap_mem_pt_guest_if_new(struct netmap_adapter *na, struct netmap_priv_d *priv) 2179 { 2180 struct netmap_mem_ptg *ptnmd = (struct netmap_mem_ptg *)na->nm_mem; 2181 struct mem_pt_if *ptif; 2182 struct netmap_if *nifp = NULL; 2183 2184 NMA_LOCK(na->nm_mem); 2185 2186 ptif = netmap_mem_pt_guest_ifp_lookup(na->nm_mem, na->ifp); 2187 if (ptif == NULL) { 2188 D("Error: interface %p is not in passthrough", na->ifp); 2189 goto out; 2190 } 2191 2192 nifp = (struct netmap_if *)((char *)(ptnmd->nm_addr) + 2193 ptif->nifp_offset); 2194 NMA_UNLOCK(na->nm_mem); 2195 out: 2196 return nifp; 2197 } 2198 2199 static void 2200 netmap_mem_pt_guest_if_delete(struct netmap_adapter *na, struct netmap_if *nifp) 2201 { 2202 struct mem_pt_if *ptif; 2203 2204 NMA_LOCK(na->nm_mem); 2205 ptif = netmap_mem_pt_guest_ifp_lookup(na->nm_mem, na->ifp); 2206 if (ptif == NULL) { 2207 D("Error: interface %p is not in passthrough", na->ifp); 2208 } 2209 NMA_UNLOCK(na->nm_mem); 2210 } 2211 2212 static int 2213 netmap_mem_pt_guest_rings_create(struct netmap_adapter *na) 2214 { 2215 struct netmap_mem_ptg *ptnmd = (struct netmap_mem_ptg *)na->nm_mem; 2216 struct mem_pt_if *ptif; 2217 struct netmap_if *nifp; 2218 int i, error = -1; 2219 2220 NMA_LOCK(na->nm_mem); 2221 2222 ptif = netmap_mem_pt_guest_ifp_lookup(na->nm_mem, na->ifp); 2223 if (ptif == NULL) { 2224 D("Error: interface %p is not in passthrough", na->ifp); 2225 goto out; 2226 } 2227 2228 2229 /* point each kring to the corresponding backend ring */ 2230 nifp = (struct netmap_if *)((char *)ptnmd->nm_addr + ptif->nifp_offset); 2231 for (i = 0; i <= na->num_tx_rings; i++) { 2232 struct netmap_kring *kring = na->tx_rings + i; 2233 if (kring->ring) 2234 continue; 2235 kring->ring = (struct netmap_ring *) 2236 ((char *)nifp + nifp->ring_ofs[i]); 2237 } 2238 for (i = 0; i <= na->num_rx_rings; i++) { 2239 struct netmap_kring *kring = na->rx_rings + i; 2240 if (kring->ring) 2241 continue; 2242 kring->ring = (struct netmap_ring *) 2243 ((char *)nifp + 2244 nifp->ring_ofs[i + na->num_tx_rings + 1]); 2245 } 2246 2247 error = 0; 2248 out: 2249 NMA_UNLOCK(na->nm_mem); 2250 2251 return error; 2252 } 2253 2254 static void 2255 netmap_mem_pt_guest_rings_delete(struct netmap_adapter *na) 2256 { 2257 /* TODO: remove?? */ 2258 #if 0 2259 struct netmap_mem_ptg *ptnmd = (struct netmap_mem_ptg *)na->nm_mem; 2260 struct mem_pt_if *ptif = netmap_mem_pt_guest_ifp_lookup(na->nm_mem, 2261 na->ifp); 2262 #endif 2263 } 2264 2265 static struct netmap_mem_ops netmap_mem_pt_guest_ops = { 2266 .nmd_get_lut = netmap_mem_pt_guest_get_lut, 2267 .nmd_get_info = netmap_mem_pt_guest_get_info, 2268 .nmd_ofstophys = netmap_mem_pt_guest_ofstophys, 2269 .nmd_config = netmap_mem_pt_guest_config, 2270 .nmd_finalize = netmap_mem_pt_guest_finalize, 2271 .nmd_deref = netmap_mem_pt_guest_deref, 2272 .nmd_if_offset = netmap_mem_pt_guest_if_offset, 2273 .nmd_delete = netmap_mem_pt_guest_delete, 2274 .nmd_if_new = netmap_mem_pt_guest_if_new, 2275 .nmd_if_delete = netmap_mem_pt_guest_if_delete, 2276 .nmd_rings_create = netmap_mem_pt_guest_rings_create, 2277 .nmd_rings_delete = netmap_mem_pt_guest_rings_delete 2278 }; 2279 2280 /* Called with nm_mem_list_lock held. */ 2281 static struct netmap_mem_d * 2282 netmap_mem_pt_guest_find_memid(nm_memid_t mem_id) 2283 { 2284 struct netmap_mem_d *mem = NULL; 2285 struct netmap_mem_d *scan = netmap_last_mem_d; 2286 2287 do { 2288 /* find ptnetmap allocator through host ID */ 2289 if (scan->ops->nmd_deref == netmap_mem_pt_guest_deref && 2290 ((struct netmap_mem_ptg *)(scan))->host_mem_id == mem_id) { 2291 mem = scan; 2292 mem->refcount++; 2293 NM_DBG_REFC(mem, __FUNCTION__, __LINE__); 2294 break; 2295 } 2296 scan = scan->next; 2297 } while (scan != netmap_last_mem_d); 2298 2299 return mem; 2300 } 2301 2302 /* Called with nm_mem_list_lock held. */ 2303 static struct netmap_mem_d * 2304 netmap_mem_pt_guest_create(nm_memid_t mem_id) 2305 { 2306 struct netmap_mem_ptg *ptnmd; 2307 int err = 0; 2308 2309 ptnmd = nm_os_malloc(sizeof(struct netmap_mem_ptg)); 2310 if (ptnmd == NULL) { 2311 err = ENOMEM; 2312 goto error; 2313 } 2314 2315 ptnmd->up.ops = &netmap_mem_pt_guest_ops; 2316 ptnmd->host_mem_id = mem_id; 2317 ptnmd->pt_ifs = NULL; 2318 2319 /* Assign new id in the guest (We have the lock) */ 2320 err = nm_mem_assign_id_locked(&ptnmd->up); 2321 if (err) 2322 goto error; 2323 2324 ptnmd->up.flags &= ~NETMAP_MEM_FINALIZED; 2325 ptnmd->up.flags |= NETMAP_MEM_IO; 2326 2327 NMA_LOCK_INIT(&ptnmd->up); 2328 2329 snprintf(ptnmd->up.name, NM_MEM_NAMESZ, "%d", ptnmd->up.nm_id); 2330 2331 2332 return &ptnmd->up; 2333 error: 2334 netmap_mem_pt_guest_delete(&ptnmd->up); 2335 return NULL; 2336 } 2337 2338 /* 2339 * find host id in guest allocators and create guest allocator 2340 * if it is not there 2341 */ 2342 static struct netmap_mem_d * 2343 netmap_mem_pt_guest_get(nm_memid_t mem_id) 2344 { 2345 struct netmap_mem_d *nmd; 2346 2347 NM_MTX_LOCK(nm_mem_list_lock); 2348 nmd = netmap_mem_pt_guest_find_memid(mem_id); 2349 if (nmd == NULL) { 2350 nmd = netmap_mem_pt_guest_create(mem_id); 2351 } 2352 NM_MTX_UNLOCK(nm_mem_list_lock); 2353 2354 return nmd; 2355 } 2356 2357 /* 2358 * The guest allocator can be created by ptnetmap_memdev (during the device 2359 * attach) or by ptnetmap device (ptnet), during the netmap_attach. 2360 * 2361 * The order is not important (we have different order in LINUX and FreeBSD). 2362 * The first one, creates the device, and the second one simply attaches it. 2363 */ 2364 2365 /* Called when ptnetmap_memdev is attaching, to attach a new allocator in 2366 * the guest */ 2367 struct netmap_mem_d * 2368 netmap_mem_pt_guest_attach(struct ptnetmap_memdev *ptn_dev, nm_memid_t mem_id) 2369 { 2370 struct netmap_mem_d *nmd; 2371 struct netmap_mem_ptg *ptnmd; 2372 2373 nmd = netmap_mem_pt_guest_get(mem_id); 2374 2375 /* assign this device to the guest allocator */ 2376 if (nmd) { 2377 ptnmd = (struct netmap_mem_ptg *)nmd; 2378 ptnmd->ptn_dev = ptn_dev; 2379 } 2380 2381 return nmd; 2382 } 2383 2384 /* Called when ptnet device is attaching */ 2385 struct netmap_mem_d * 2386 netmap_mem_pt_guest_new(struct ifnet *ifp, 2387 unsigned int nifp_offset, 2388 unsigned int memid) 2389 { 2390 struct netmap_mem_d *nmd; 2391 2392 if (ifp == NULL) { 2393 return NULL; 2394 } 2395 2396 nmd = netmap_mem_pt_guest_get((nm_memid_t)memid); 2397 2398 if (nmd) { 2399 netmap_mem_pt_guest_ifp_add(nmd, ifp, nifp_offset); 2400 } 2401 2402 return nmd; 2403 } 2404 2405 #endif /* WITH_PTNETMAP_GUEST */ 2406