1 /* 2 * CDDL HEADER START 3 * 4 * The contents of this file are subject to the terms of the 5 * Common Development and Distribution License (the "License"). 6 * You may not use this file except in compliance with the License. 7 * 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9 * or http://www.opensolaris.org/os/licensing. 10 * See the License for the specific language governing permissions 11 * and limitations under the License. 12 * 13 * When distributing Covered Code, include this CDDL HEADER in each 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15 * If applicable, add the following below this CDDL HEADER, with the 16 * fields enclosed by brackets "[]" replaced with your own identifying 17 * information: Portions Copyright [yyyy] [name of copyright owner] 18 * 19 * CDDL HEADER END 20 */ 21 /* 22 * Copyright (c) 1998, 2010, Oracle and/or its affiliates. All rights reserved. 23 */ 24 25 #include <sys/types.h> 26 #include <sys/t_lock.h> 27 #include <sys/param.h> 28 #include <sys/sysmacros.h> 29 #include <sys/tuneable.h> 30 #include <sys/systm.h> 31 #include <sys/vm.h> 32 #include <sys/kmem.h> 33 #include <sys/vmem.h> 34 #include <sys/mman.h> 35 #include <sys/cmn_err.h> 36 #include <sys/debug.h> 37 #include <sys/dumphdr.h> 38 #include <sys/bootconf.h> 39 #include <sys/lgrp.h> 40 #include <vm/seg_kmem.h> 41 #include <vm/hat.h> 42 #include <vm/page.h> 43 #include <vm/vm_dep.h> 44 #include <vm/faultcode.h> 45 #include <sys/promif.h> 46 #include <vm/seg_kp.h> 47 #include <sys/bitmap.h> 48 #include <sys/mem_cage.h> 49 50 #ifdef __sparc 51 #include <sys/ivintr.h> 52 #include <sys/panic.h> 53 #endif 54 55 /* 56 * seg_kmem is the primary kernel memory segment driver. It 57 * maps the kernel heap [kernelheap, ekernelheap), module text, 58 * and all memory which was allocated before the VM was initialized 59 * into kas. 60 * 61 * Pages which belong to seg_kmem are hashed into &kvp vnode at 62 * an offset equal to (u_offset_t)virt_addr, and have p_lckcnt >= 1. 63 * They must never be paged out since segkmem_fault() is a no-op to 64 * prevent recursive faults. 65 * 66 * Currently, seg_kmem pages are sharelocked (p_sharelock == 1) on 67 * __x86 and are unlocked (p_sharelock == 0) on __sparc. Once __x86 68 * supports relocation the #ifdef kludges can be removed. 69 * 70 * seg_kmem pages may be subject to relocation by page_relocate(), 71 * provided that the HAT supports it; if this is so, segkmem_reloc 72 * will be set to a nonzero value. All boot time allocated memory as 73 * well as static memory is considered off limits to relocation. 74 * Pages are "relocatable" if p_state does not have P_NORELOC set, so 75 * we request P_NORELOC pages for memory that isn't safe to relocate. 76 * 77 * The kernel heap is logically divided up into four pieces: 78 * 79 * heap32_arena is for allocations that require 32-bit absolute 80 * virtual addresses (e.g. code that uses 32-bit pointers/offsets). 81 * 82 * heap_core is for allocations that require 2GB *relative* 83 * offsets; in other words all memory from heap_core is within 84 * 2GB of all other memory from the same arena. This is a requirement 85 * of the addressing modes of some processors in supervisor code. 86 * 87 * heap_arena is the general heap arena. 88 * 89 * static_arena is the static memory arena. Allocations from it 90 * are not subject to relocation so it is safe to use the memory 91 * physical address as well as the virtual address (e.g. the VA to 92 * PA translations are static). Caches may import from static_arena; 93 * all other static memory allocations should use static_alloc_arena. 94 * 95 * On some platforms which have limited virtual address space, seg_kmem 96 * may share [kernelheap, ekernelheap) with seg_kp; if this is so, 97 * segkp_bitmap is non-NULL, and each bit represents a page of virtual 98 * address space which is actually seg_kp mapped. 99 */ 100 101 extern ulong_t *segkp_bitmap; /* Is set if segkp is from the kernel heap */ 102 103 char *kernelheap; /* start of primary kernel heap */ 104 char *ekernelheap; /* end of primary kernel heap */ 105 struct seg kvseg; /* primary kernel heap segment */ 106 struct seg kvseg_core; /* "core" kernel heap segment */ 107 struct seg kzioseg; /* Segment for zio mappings */ 108 vmem_t *heap_arena; /* primary kernel heap arena */ 109 vmem_t *heap_core_arena; /* core kernel heap arena */ 110 char *heap_core_base; /* start of core kernel heap arena */ 111 char *heap_lp_base; /* start of kernel large page heap arena */ 112 char *heap_lp_end; /* end of kernel large page heap arena */ 113 vmem_t *hat_memload_arena; /* HAT translation data */ 114 struct seg kvseg32; /* 32-bit kernel heap segment */ 115 vmem_t *heap32_arena; /* 32-bit kernel heap arena */ 116 vmem_t *heaptext_arena; /* heaptext arena */ 117 struct as kas; /* kernel address space */ 118 int segkmem_reloc; /* enable/disable relocatable segkmem pages */ 119 vmem_t *static_arena; /* arena for caches to import static memory */ 120 vmem_t *static_alloc_arena; /* arena for allocating static memory */ 121 vmem_t *zio_arena = NULL; /* arena for allocating zio memory */ 122 vmem_t *zio_alloc_arena = NULL; /* arena for allocating zio memory */ 123 124 /* 125 * seg_kmem driver can map part of the kernel heap with large pages. 126 * Currently this functionality is implemented for sparc platforms only. 127 * 128 * The large page size "segkmem_lpsize" for kernel heap is selected in the 129 * platform specific code. It can also be modified via /etc/system file. 130 * Setting segkmem_lpsize to PAGESIZE in /etc/system disables usage of large 131 * pages for kernel heap. "segkmem_lpshift" is adjusted appropriately to 132 * match segkmem_lpsize. 133 * 134 * At boot time we carve from kernel heap arena a range of virtual addresses 135 * that will be used for large page mappings. This range [heap_lp_base, 136 * heap_lp_end) is set up as a separate vmem arena - "heap_lp_arena". We also 137 * create "kmem_lp_arena" that caches memory already backed up by large 138 * pages. kmem_lp_arena imports virtual segments from heap_lp_arena. 139 */ 140 141 size_t segkmem_lpsize; 142 static uint_t segkmem_lpshift = PAGESHIFT; 143 int segkmem_lpszc = 0; 144 145 size_t segkmem_kmemlp_quantum = 0x400000; /* 4MB */ 146 size_t segkmem_heaplp_quantum; 147 vmem_t *heap_lp_arena; 148 static vmem_t *kmem_lp_arena; 149 static vmem_t *segkmem_ppa_arena; 150 static segkmem_lpcb_t segkmem_lpcb; 151 152 /* 153 * We use "segkmem_kmemlp_max" to limit the total amount of physical memory 154 * consumed by the large page heap. By default this parameter is set to 1/8 of 155 * physmem but can be adjusted through /etc/system either directly or 156 * indirectly by setting "segkmem_kmemlp_pcnt" to the percent of physmem 157 * we allow for large page heap. 158 */ 159 size_t segkmem_kmemlp_max; 160 static uint_t segkmem_kmemlp_pcnt; 161 162 /* 163 * Getting large pages for kernel heap could be problematic due to 164 * physical memory fragmentation. That's why we allow to preallocate 165 * "segkmem_kmemlp_min" bytes at boot time. 166 */ 167 static size_t segkmem_kmemlp_min; 168 169 /* 170 * Throttling is used to avoid expensive tries to allocate large pages 171 * for kernel heap when a lot of succesive attempts to do so fail. 172 */ 173 static ulong_t segkmem_lpthrottle_max = 0x400000; 174 static ulong_t segkmem_lpthrottle_start = 0x40; 175 static ulong_t segkmem_use_lpthrottle = 1; 176 177 /* 178 * Freed pages accumulate on a garbage list until segkmem is ready, 179 * at which point we call segkmem_gc() to free it all. 180 */ 181 typedef struct segkmem_gc_list { 182 struct segkmem_gc_list *gc_next; 183 vmem_t *gc_arena; 184 size_t gc_size; 185 } segkmem_gc_list_t; 186 187 static segkmem_gc_list_t *segkmem_gc_list; 188 189 /* 190 * Allocations from the hat_memload arena add VM_MEMLOAD to their 191 * vmflags so that segkmem_xalloc() can inform the hat layer that it needs 192 * to take steps to prevent infinite recursion. HAT allocations also 193 * must be non-relocatable to prevent recursive page faults. 194 */ 195 static void * 196 hat_memload_alloc(vmem_t *vmp, size_t size, int flags) 197 { 198 flags |= (VM_MEMLOAD | VM_NORELOC); 199 return (segkmem_alloc(vmp, size, flags)); 200 } 201 202 /* 203 * Allocations from static_arena arena (or any other arena that uses 204 * segkmem_alloc_permanent()) require non-relocatable (permanently 205 * wired) memory pages, since these pages are referenced by physical 206 * as well as virtual address. 207 */ 208 void * 209 segkmem_alloc_permanent(vmem_t *vmp, size_t size, int flags) 210 { 211 return (segkmem_alloc(vmp, size, flags | VM_NORELOC)); 212 } 213 214 /* 215 * Initialize kernel heap boundaries. 216 */ 217 void 218 kernelheap_init( 219 void *heap_start, 220 void *heap_end, 221 char *first_avail, 222 void *core_start, 223 void *core_end) 224 { 225 uintptr_t textbase; 226 size_t core_size; 227 size_t heap_size; 228 vmem_t *heaptext_parent; 229 size_t heap_lp_size = 0; 230 #ifdef __sparc 231 size_t kmem64_sz = kmem64_aligned_end - kmem64_base; 232 #endif /* __sparc */ 233 234 kernelheap = heap_start; 235 ekernelheap = heap_end; 236 237 #ifdef __sparc 238 heap_lp_size = (((uintptr_t)heap_end - (uintptr_t)heap_start) / 4); 239 /* 240 * Bias heap_lp start address by kmem64_sz to reduce collisions 241 * in 4M kernel TSB between kmem64 area and heap_lp 242 */ 243 kmem64_sz = P2ROUNDUP(kmem64_sz, MMU_PAGESIZE256M); 244 if (kmem64_sz <= heap_lp_size / 2) 245 heap_lp_size -= kmem64_sz; 246 heap_lp_base = ekernelheap - heap_lp_size; 247 heap_lp_end = heap_lp_base + heap_lp_size; 248 #endif /* __sparc */ 249 250 /* 251 * If this platform has a 'core' heap area, then the space for 252 * overflow module text should be carved out of the end of that 253 * heap. Otherwise, it gets carved out of the general purpose 254 * heap. 255 */ 256 core_size = (uintptr_t)core_end - (uintptr_t)core_start; 257 if (core_size > 0) { 258 ASSERT(core_size >= HEAPTEXT_SIZE); 259 textbase = (uintptr_t)core_end - HEAPTEXT_SIZE; 260 core_size -= HEAPTEXT_SIZE; 261 } 262 #ifndef __sparc 263 else { 264 ekernelheap -= HEAPTEXT_SIZE; 265 textbase = (uintptr_t)ekernelheap; 266 } 267 #endif 268 269 heap_size = (uintptr_t)ekernelheap - (uintptr_t)kernelheap; 270 heap_arena = vmem_init("heap", kernelheap, heap_size, PAGESIZE, 271 segkmem_alloc, segkmem_free); 272 273 if (core_size > 0) { 274 heap_core_arena = vmem_create("heap_core", core_start, 275 core_size, PAGESIZE, NULL, NULL, NULL, 0, VM_SLEEP); 276 heap_core_base = core_start; 277 } else { 278 heap_core_arena = heap_arena; 279 heap_core_base = kernelheap; 280 } 281 282 /* 283 * reserve space for the large page heap. If large pages for kernel 284 * heap is enabled large page heap arean will be created later in the 285 * boot sequence in segkmem_heap_lp_init(). Otherwise the allocated 286 * range will be returned back to the heap_arena. 287 */ 288 if (heap_lp_size) { 289 (void) vmem_xalloc(heap_arena, heap_lp_size, PAGESIZE, 0, 0, 290 heap_lp_base, heap_lp_end, 291 VM_NOSLEEP | VM_BESTFIT | VM_PANIC); 292 } 293 294 /* 295 * Remove the already-spoken-for memory range [kernelheap, first_avail). 296 */ 297 (void) vmem_xalloc(heap_arena, first_avail - kernelheap, PAGESIZE, 298 0, 0, kernelheap, first_avail, VM_NOSLEEP | VM_BESTFIT | VM_PANIC); 299 300 #ifdef __sparc 301 heap32_arena = vmem_create("heap32", (void *)SYSBASE32, 302 SYSLIMIT32 - SYSBASE32 - HEAPTEXT_SIZE, PAGESIZE, NULL, 303 NULL, NULL, 0, VM_SLEEP); 304 /* 305 * Prom claims the physical and virtual resources used by panicbuf 306 * and inter_vec_table. So reserve space for panicbuf, intr_vec_table, 307 * reserved interrupt vector data structures from 32-bit heap. 308 */ 309 (void) vmem_xalloc(heap32_arena, PANICBUFSIZE, PAGESIZE, 0, 0, 310 panicbuf, panicbuf + PANICBUFSIZE, 311 VM_NOSLEEP | VM_BESTFIT | VM_PANIC); 312 313 (void) vmem_xalloc(heap32_arena, IVSIZE, PAGESIZE, 0, 0, 314 intr_vec_table, (caddr_t)intr_vec_table + IVSIZE, 315 VM_NOSLEEP | VM_BESTFIT | VM_PANIC); 316 317 textbase = SYSLIMIT32 - HEAPTEXT_SIZE; 318 heaptext_parent = NULL; 319 #else /* __sparc */ 320 heap32_arena = heap_core_arena; 321 heaptext_parent = heap_core_arena; 322 #endif /* __sparc */ 323 324 heaptext_arena = vmem_create("heaptext", (void *)textbase, 325 HEAPTEXT_SIZE, PAGESIZE, NULL, NULL, heaptext_parent, 0, VM_SLEEP); 326 327 /* 328 * Create a set of arenas for memory with static translations 329 * (e.g. VA -> PA translations cannot change). Since using 330 * kernel pages by physical address implies it isn't safe to 331 * walk across page boundaries, the static_arena quantum must 332 * be PAGESIZE. Any kmem caches that require static memory 333 * should source from static_arena, while direct allocations 334 * should only use static_alloc_arena. 335 */ 336 static_arena = vmem_create("static", NULL, 0, PAGESIZE, 337 segkmem_alloc_permanent, segkmem_free, heap_arena, 0, VM_SLEEP); 338 static_alloc_arena = vmem_create("static_alloc", NULL, 0, 339 sizeof (uint64_t), vmem_alloc, vmem_free, static_arena, 340 0, VM_SLEEP); 341 342 /* 343 * Create an arena for translation data (ptes, hmes, or hblks). 344 * We need an arena for this because hat_memload() is essential 345 * to vmem_populate() (see comments in common/os/vmem.c). 346 * 347 * Note: any kmem cache that allocates from hat_memload_arena 348 * must be created as a KMC_NOHASH cache (i.e. no external slab 349 * and bufctl structures to allocate) so that slab creation doesn't 350 * require anything more than a single vmem_alloc(). 351 */ 352 hat_memload_arena = vmem_create("hat_memload", NULL, 0, PAGESIZE, 353 hat_memload_alloc, segkmem_free, heap_arena, 0, 354 VM_SLEEP | VMC_POPULATOR | VMC_DUMPSAFE); 355 } 356 357 void 358 boot_mapin(caddr_t addr, size_t size) 359 { 360 caddr_t eaddr; 361 page_t *pp; 362 pfn_t pfnum; 363 364 if (page_resv(btop(size), KM_NOSLEEP) == 0) 365 panic("boot_mapin: page_resv failed"); 366 367 for (eaddr = addr + size; addr < eaddr; addr += PAGESIZE) { 368 pfnum = va_to_pfn(addr); 369 if (pfnum == PFN_INVALID) 370 continue; 371 if ((pp = page_numtopp_nolock(pfnum)) == NULL) 372 panic("boot_mapin(): No pp for pfnum = %lx", pfnum); 373 374 /* 375 * must break up any large pages that may have constituent 376 * pages being utilized for BOP_ALLOC()'s before calling 377 * page_numtopp().The locking code (ie. page_reclaim()) 378 * can't handle them 379 */ 380 if (pp->p_szc != 0) 381 page_boot_demote(pp); 382 383 pp = page_numtopp(pfnum, SE_EXCL); 384 if (pp == NULL || PP_ISFREE(pp)) 385 panic("boot_alloc: pp is NULL or free"); 386 387 /* 388 * If the cage is on but doesn't yet contain this page, 389 * mark it as non-relocatable. 390 */ 391 if (kcage_on && !PP_ISNORELOC(pp)) { 392 PP_SETNORELOC(pp); 393 PLCNT_XFER_NORELOC(pp); 394 } 395 396 (void) page_hashin(pp, &kvp, (u_offset_t)(uintptr_t)addr, NULL); 397 pp->p_lckcnt = 1; 398 #if defined(__x86) 399 page_downgrade(pp); 400 #else 401 page_unlock(pp); 402 #endif 403 } 404 } 405 406 /* 407 * Get pages from boot and hash them into the kernel's vp. 408 * Used after page structs have been allocated, but before segkmem is ready. 409 */ 410 void * 411 boot_alloc(void *inaddr, size_t size, uint_t align) 412 { 413 caddr_t addr = inaddr; 414 415 if (bootops == NULL) 416 prom_panic("boot_alloc: attempt to allocate memory after " 417 "BOP_GONE"); 418 419 size = ptob(btopr(size)); 420 #ifdef __sparc 421 if (bop_alloc_chunk(addr, size, align) != (caddr_t)addr) 422 panic("boot_alloc: bop_alloc_chunk failed"); 423 #else 424 if (BOP_ALLOC(bootops, addr, size, align) != addr) 425 panic("boot_alloc: BOP_ALLOC failed"); 426 #endif 427 boot_mapin((caddr_t)addr, size); 428 return (addr); 429 } 430 431 /*ARGSUSED*/ 432 static faultcode_t 433 segkmem_fault(struct hat *hat, struct seg *seg, caddr_t addr, size_t size, 434 enum fault_type type, enum seg_rw rw) 435 { 436 pgcnt_t npages; 437 spgcnt_t pg; 438 page_t *pp; 439 struct vnode *vp = seg->s_data; 440 441 ASSERT(RW_READ_HELD(&seg->s_as->a_lock)); 442 443 if (seg->s_as != &kas || size > seg->s_size || 444 addr < seg->s_base || addr + size > seg->s_base + seg->s_size) 445 panic("segkmem_fault: bad args"); 446 447 /* 448 * If it is one of segkp pages, call segkp_fault. 449 */ 450 if (segkp_bitmap && seg == &kvseg && 451 BT_TEST(segkp_bitmap, btop((uintptr_t)(addr - seg->s_base)))) 452 return (segop_fault(hat, segkp, addr, size, type, rw)); 453 454 if (rw != S_READ && rw != S_WRITE && rw != S_OTHER) 455 return (FC_NOSUPPORT); 456 457 npages = btopr(size); 458 459 switch (type) { 460 case F_SOFTLOCK: /* lock down already-loaded translations */ 461 for (pg = 0; pg < npages; pg++) { 462 pp = page_lookup(vp, (u_offset_t)(uintptr_t)addr, 463 SE_SHARED); 464 if (pp == NULL) { 465 /* 466 * Hmm, no page. Does a kernel mapping 467 * exist for it? 468 */ 469 if (!hat_probe(kas.a_hat, addr)) { 470 addr -= PAGESIZE; 471 while (--pg >= 0) { 472 pp = page_find(vp, (u_offset_t) 473 (uintptr_t)addr); 474 if (pp) 475 page_unlock(pp); 476 addr -= PAGESIZE; 477 } 478 return (FC_NOMAP); 479 } 480 } 481 addr += PAGESIZE; 482 } 483 if (rw == S_OTHER) 484 hat_reserve(seg->s_as, addr, size); 485 return (0); 486 case F_SOFTUNLOCK: 487 while (npages--) { 488 pp = page_find(vp, (u_offset_t)(uintptr_t)addr); 489 if (pp) 490 page_unlock(pp); 491 addr += PAGESIZE; 492 } 493 return (0); 494 default: 495 return (FC_NOSUPPORT); 496 } 497 /*NOTREACHED*/ 498 } 499 500 static int 501 segkmem_setprot(struct seg *seg, caddr_t addr, size_t size, uint_t prot) 502 { 503 ASSERT(RW_LOCK_HELD(&seg->s_as->a_lock)); 504 505 if (seg->s_as != &kas || size > seg->s_size || 506 addr < seg->s_base || addr + size > seg->s_base + seg->s_size) 507 panic("segkmem_setprot: bad args"); 508 509 /* 510 * If it is one of segkp pages, call segkp. 511 */ 512 if (segkp_bitmap && seg == &kvseg && 513 BT_TEST(segkp_bitmap, btop((uintptr_t)(addr - seg->s_base)))) 514 return (segop_setprot(segkp, addr, size, prot)); 515 516 if (prot == 0) 517 hat_unload(kas.a_hat, addr, size, HAT_UNLOAD); 518 else 519 hat_chgprot(kas.a_hat, addr, size, prot); 520 return (0); 521 } 522 523 /* 524 * This is a dummy segkmem function overloaded to call segkp 525 * when segkp is under the heap. 526 */ 527 /* ARGSUSED */ 528 static int 529 segkmem_checkprot(struct seg *seg, caddr_t addr, size_t size, uint_t prot) 530 { 531 ASSERT(RW_LOCK_HELD(&seg->s_as->a_lock)); 532 533 if (seg->s_as != &kas) 534 panic("segkmem badop"); 535 536 /* 537 * If it is one of segkp pages, call into segkp. 538 */ 539 if (segkp_bitmap && seg == &kvseg && 540 BT_TEST(segkp_bitmap, btop((uintptr_t)(addr - seg->s_base)))) 541 return (segop_checkprot(segkp, addr, size, prot)); 542 543 panic("segkmem badop"); 544 return (0); 545 } 546 547 /* 548 * This is a dummy segkmem function overloaded to call segkp 549 * when segkp is under the heap. 550 */ 551 /* ARGSUSED */ 552 static int 553 segkmem_kluster(struct seg *seg, caddr_t addr, ssize_t delta) 554 { 555 ASSERT(RW_LOCK_HELD(&seg->s_as->a_lock)); 556 557 if (seg->s_as != &kas) 558 panic("segkmem badop"); 559 560 /* 561 * If it is one of segkp pages, call into segkp. 562 */ 563 if (segkp_bitmap && seg == &kvseg && 564 BT_TEST(segkp_bitmap, btop((uintptr_t)(addr - seg->s_base)))) 565 return (segop_kluster(segkp, addr, delta)); 566 567 panic("segkmem badop"); 568 return (0); 569 } 570 571 static void 572 segkmem_xdump_range(void *arg, void *start, size_t size) 573 { 574 struct as *as = arg; 575 caddr_t addr = start; 576 caddr_t addr_end = addr + size; 577 578 while (addr < addr_end) { 579 pfn_t pfn = hat_getpfnum(kas.a_hat, addr); 580 if (pfn != PFN_INVALID && pfn <= physmax && pf_is_memory(pfn)) 581 dump_addpage(as, addr, pfn); 582 addr += PAGESIZE; 583 dump_timeleft = dump_timeout; 584 } 585 } 586 587 static void 588 segkmem_dump_range(void *arg, void *start, size_t size) 589 { 590 caddr_t addr = start; 591 caddr_t addr_end = addr + size; 592 593 /* 594 * If we are about to start dumping the range of addresses we 595 * carved out of the kernel heap for the large page heap walk 596 * heap_lp_arena to find what segments are actually populated 597 */ 598 if (SEGKMEM_USE_LARGEPAGES && 599 addr == heap_lp_base && addr_end == heap_lp_end && 600 vmem_size(heap_lp_arena, VMEM_ALLOC) < size) { 601 vmem_walk(heap_lp_arena, VMEM_ALLOC | VMEM_REENTRANT, 602 segkmem_xdump_range, arg); 603 } else { 604 segkmem_xdump_range(arg, start, size); 605 } 606 } 607 608 static void 609 segkmem_dump(struct seg *seg) 610 { 611 /* 612 * The kernel's heap_arena (represented by kvseg) is a very large 613 * VA space, most of which is typically unused. To speed up dumping 614 * we use vmem_walk() to quickly find the pieces of heap_arena that 615 * are actually in use. We do the same for heap32_arena and 616 * heap_core. 617 * 618 * We specify VMEM_REENTRANT to vmem_walk() because dump_addpage() 619 * may ultimately need to allocate memory. Reentrant walks are 620 * necessarily imperfect snapshots. The kernel heap continues 621 * to change during a live crash dump, for example. For a normal 622 * crash dump, however, we know that there won't be any other threads 623 * messing with the heap. Therefore, at worst, we may fail to dump 624 * the pages that get allocated by the act of dumping; but we will 625 * always dump every page that was allocated when the walk began. 626 * 627 * The other segkmem segments are dense (fully populated), so there's 628 * no need to use this technique when dumping them. 629 * 630 * Note: when adding special dump handling for any new sparsely- 631 * populated segments, be sure to add similar handling to the ::kgrep 632 * code in mdb. 633 */ 634 if (seg == &kvseg) { 635 vmem_walk(heap_arena, VMEM_ALLOC | VMEM_REENTRANT, 636 segkmem_dump_range, seg->s_as); 637 #ifndef __sparc 638 vmem_walk(heaptext_arena, VMEM_ALLOC | VMEM_REENTRANT, 639 segkmem_dump_range, seg->s_as); 640 #endif 641 } else if (seg == &kvseg_core) { 642 vmem_walk(heap_core_arena, VMEM_ALLOC | VMEM_REENTRANT, 643 segkmem_dump_range, seg->s_as); 644 } else if (seg == &kvseg32) { 645 vmem_walk(heap32_arena, VMEM_ALLOC | VMEM_REENTRANT, 646 segkmem_dump_range, seg->s_as); 647 vmem_walk(heaptext_arena, VMEM_ALLOC | VMEM_REENTRANT, 648 segkmem_dump_range, seg->s_as); 649 } else if (seg == &kzioseg) { 650 /* 651 * We don't want to dump pages attached to kzioseg since they 652 * contain file data from ZFS. If this page's segment is 653 * kzioseg return instead of writing it to the dump device. 654 */ 655 return; 656 } else { 657 segkmem_dump_range(seg->s_as, seg->s_base, seg->s_size); 658 } 659 } 660 661 /* 662 * lock/unlock kmem pages over a given range [addr, addr+len). 663 * Returns a shadow list of pages in ppp. If there are holes 664 * in the range (e.g. some of the kernel mappings do not have 665 * underlying page_ts) returns ENOTSUP so that as_pagelock() 666 * will handle the range via as_fault(F_SOFTLOCK). 667 */ 668 /*ARGSUSED*/ 669 static int 670 segkmem_pagelock(struct seg *seg, caddr_t addr, size_t len, 671 page_t ***ppp, enum lock_type type, enum seg_rw rw) 672 { 673 page_t **pplist, *pp; 674 pgcnt_t npages; 675 spgcnt_t pg; 676 size_t nb; 677 struct vnode *vp = seg->s_data; 678 679 ASSERT(ppp != NULL); 680 681 /* 682 * If it is one of segkp pages, call into segkp. 683 */ 684 if (segkp_bitmap && seg == &kvseg && 685 BT_TEST(segkp_bitmap, btop((uintptr_t)(addr - seg->s_base)))) 686 return (segop_pagelock(segkp, addr, len, ppp, type, rw)); 687 688 npages = btopr(len); 689 nb = sizeof (page_t *) * npages; 690 691 if (type == L_PAGEUNLOCK) { 692 pplist = *ppp; 693 ASSERT(pplist != NULL); 694 695 for (pg = 0; pg < npages; pg++) { 696 pp = pplist[pg]; 697 page_unlock(pp); 698 } 699 kmem_free(pplist, nb); 700 return (0); 701 } 702 703 ASSERT(type == L_PAGELOCK); 704 705 pplist = kmem_alloc(nb, KM_NOSLEEP); 706 if (pplist == NULL) { 707 *ppp = NULL; 708 return (ENOTSUP); /* take the slow path */ 709 } 710 711 for (pg = 0; pg < npages; pg++) { 712 pp = page_lookup(vp, (u_offset_t)(uintptr_t)addr, SE_SHARED); 713 if (pp == NULL) { 714 while (--pg >= 0) 715 page_unlock(pplist[pg]); 716 kmem_free(pplist, nb); 717 *ppp = NULL; 718 return (ENOTSUP); 719 } 720 pplist[pg] = pp; 721 addr += PAGESIZE; 722 } 723 724 *ppp = pplist; 725 return (0); 726 } 727 728 /* 729 * This is a dummy segkmem function overloaded to call segkp 730 * when segkp is under the heap. 731 */ 732 /* ARGSUSED */ 733 static int 734 segkmem_getmemid(struct seg *seg, caddr_t addr, memid_t *memidp) 735 { 736 ASSERT(RW_LOCK_HELD(&seg->s_as->a_lock)); 737 738 if (seg->s_as != &kas) 739 panic("segkmem badop"); 740 741 /* 742 * If it is one of segkp pages, call into segkp. 743 */ 744 if (segkp_bitmap && seg == &kvseg && 745 BT_TEST(segkp_bitmap, btop((uintptr_t)(addr - seg->s_base)))) 746 return (segop_getmemid(segkp, addr, memidp)); 747 748 panic("segkmem badop"); 749 return (0); 750 } 751 752 /*ARGSUSED*/ 753 static int 754 segkmem_capable(struct seg *seg, segcapability_t capability) 755 { 756 if (capability == S_CAPABILITY_NOMINFLT) 757 return (1); 758 return (0); 759 } 760 761 static const struct seg_ops segkmem_ops = { 762 .fault = segkmem_fault, 763 .setprot = segkmem_setprot, 764 .checkprot = segkmem_checkprot, 765 .kluster = segkmem_kluster, 766 .dump = segkmem_dump, 767 .pagelock = segkmem_pagelock, 768 .getmemid = segkmem_getmemid, 769 .capable = segkmem_capable, 770 }; 771 772 int 773 segkmem_zio_create(struct seg *seg) 774 { 775 ASSERT(seg->s_as == &kas && RW_WRITE_HELD(&kas.a_lock)); 776 seg->s_ops = &segkmem_ops; 777 seg->s_data = &zvp; 778 kas.a_size += seg->s_size; 779 return (0); 780 } 781 782 int 783 segkmem_create(struct seg *seg) 784 { 785 ASSERT(seg->s_as == &kas && RW_WRITE_HELD(&kas.a_lock)); 786 seg->s_ops = &segkmem_ops; 787 seg->s_data = &kvp; 788 kas.a_size += seg->s_size; 789 return (0); 790 } 791 792 /*ARGSUSED*/ 793 page_t * 794 segkmem_page_create(void *addr, size_t size, int vmflag, void *arg) 795 { 796 struct seg kseg; 797 int pgflags; 798 struct vnode *vp = arg; 799 800 if (vp == NULL) 801 vp = &kvp; 802 803 kseg.s_as = &kas; 804 pgflags = PG_EXCL; 805 806 if (segkmem_reloc == 0 || (vmflag & VM_NORELOC)) 807 pgflags |= PG_NORELOC; 808 if ((vmflag & VM_NOSLEEP) == 0) 809 pgflags |= PG_WAIT; 810 if (vmflag & VM_PANIC) 811 pgflags |= PG_PANIC; 812 if (vmflag & VM_PUSHPAGE) 813 pgflags |= PG_PUSHPAGE; 814 if (vmflag & VM_NORMALPRI) { 815 ASSERT(vmflag & VM_NOSLEEP); 816 pgflags |= PG_NORMALPRI; 817 } 818 819 return (page_create_va(vp, (u_offset_t)(uintptr_t)addr, size, 820 pgflags, &kseg, addr)); 821 } 822 823 /* 824 * Allocate pages to back the virtual address range [addr, addr + size). 825 * If addr is NULL, allocate the virtual address space as well. 826 */ 827 void * 828 segkmem_xalloc(vmem_t *vmp, void *inaddr, size_t size, int vmflag, uint_t attr, 829 page_t *(*page_create_func)(void *, size_t, int, void *), void *pcarg) 830 { 831 page_t *ppl; 832 caddr_t addr = inaddr; 833 pgcnt_t npages = btopr(size); 834 int allocflag; 835 836 if (inaddr == NULL && (addr = vmem_alloc(vmp, size, vmflag)) == NULL) 837 return (NULL); 838 839 ASSERT(((uintptr_t)addr & PAGEOFFSET) == 0); 840 841 if (page_resv(npages, vmflag & VM_KMFLAGS) == 0) { 842 if (inaddr == NULL) 843 vmem_free(vmp, addr, size); 844 return (NULL); 845 } 846 847 ppl = page_create_func(addr, size, vmflag, pcarg); 848 if (ppl == NULL) { 849 if (inaddr == NULL) 850 vmem_free(vmp, addr, size); 851 page_unresv(npages); 852 return (NULL); 853 } 854 855 /* 856 * Under certain conditions, we need to let the HAT layer know 857 * that it cannot safely allocate memory. Allocations from 858 * the hat_memload vmem arena always need this, to prevent 859 * infinite recursion. 860 * 861 * In addition, the x86 hat cannot safely do memory 862 * allocations while in vmem_populate(), because there 863 * is no simple bound on its usage. 864 */ 865 if (vmflag & VM_MEMLOAD) 866 allocflag = HAT_NO_KALLOC; 867 #if defined(__x86) 868 else if (vmem_is_populator()) 869 allocflag = HAT_NO_KALLOC; 870 #endif 871 else 872 allocflag = 0; 873 874 while (ppl != NULL) { 875 page_t *pp = ppl; 876 page_sub(&ppl, pp); 877 ASSERT(page_iolock_assert(pp)); 878 ASSERT(PAGE_EXCL(pp)); 879 page_io_unlock(pp); 880 hat_memload(kas.a_hat, (caddr_t)(uintptr_t)pp->p_offset, pp, 881 (PROT_ALL & ~PROT_USER) | HAT_NOSYNC | attr, 882 HAT_LOAD_LOCK | allocflag); 883 pp->p_lckcnt = 1; 884 #if defined(__x86) 885 page_downgrade(pp); 886 #else 887 if (vmflag & SEGKMEM_SHARELOCKED) 888 page_downgrade(pp); 889 else 890 page_unlock(pp); 891 #endif 892 } 893 894 return (addr); 895 } 896 897 static void * 898 segkmem_alloc_vn(vmem_t *vmp, size_t size, int vmflag, struct vnode *vp) 899 { 900 void *addr; 901 segkmem_gc_list_t *gcp, **prev_gcpp; 902 903 ASSERT(vp != NULL); 904 905 if (kvseg.s_base == NULL) { 906 #ifndef __sparc 907 if (bootops->bsys_alloc == NULL) 908 halt("Memory allocation between bop_alloc() and " 909 "kmem_alloc().\n"); 910 #endif 911 912 /* 913 * There's not a lot of memory to go around during boot, 914 * so recycle it if we can. 915 */ 916 for (prev_gcpp = &segkmem_gc_list; (gcp = *prev_gcpp) != NULL; 917 prev_gcpp = &gcp->gc_next) { 918 if (gcp->gc_arena == vmp && gcp->gc_size == size) { 919 *prev_gcpp = gcp->gc_next; 920 return (gcp); 921 } 922 } 923 924 addr = vmem_alloc(vmp, size, vmflag | VM_PANIC); 925 if (boot_alloc(addr, size, BO_NO_ALIGN) != addr) 926 panic("segkmem_alloc: boot_alloc failed"); 927 return (addr); 928 } 929 return (segkmem_xalloc(vmp, NULL, size, vmflag, 0, 930 segkmem_page_create, vp)); 931 } 932 933 void * 934 segkmem_alloc(vmem_t *vmp, size_t size, int vmflag) 935 { 936 return (segkmem_alloc_vn(vmp, size, vmflag, &kvp)); 937 } 938 939 void * 940 segkmem_zio_alloc(vmem_t *vmp, size_t size, int vmflag) 941 { 942 return (segkmem_alloc_vn(vmp, size, vmflag, &zvp)); 943 } 944 945 /* 946 * Any changes to this routine must also be carried over to 947 * devmap_free_pages() in the seg_dev driver. This is because 948 * we currently don't have a special kernel segment for non-paged 949 * kernel memory that is exported by drivers to user space. 950 */ 951 static void 952 segkmem_free_vn(vmem_t *vmp, void *inaddr, size_t size, struct vnode *vp, 953 void (*func)(page_t *)) 954 { 955 page_t *pp; 956 caddr_t addr = inaddr; 957 caddr_t eaddr; 958 pgcnt_t npages = btopr(size); 959 960 ASSERT(((uintptr_t)addr & PAGEOFFSET) == 0); 961 ASSERT(vp != NULL); 962 963 if (kvseg.s_base == NULL) { 964 segkmem_gc_list_t *gc = inaddr; 965 gc->gc_arena = vmp; 966 gc->gc_size = size; 967 gc->gc_next = segkmem_gc_list; 968 segkmem_gc_list = gc; 969 return; 970 } 971 972 hat_unload(kas.a_hat, addr, size, HAT_UNLOAD_UNLOCK); 973 974 for (eaddr = addr + size; addr < eaddr; addr += PAGESIZE) { 975 #if defined(__x86) 976 pp = page_find(vp, (u_offset_t)(uintptr_t)addr); 977 if (pp == NULL) 978 panic("segkmem_free: page not found"); 979 if (!page_tryupgrade(pp)) { 980 /* 981 * Some other thread has a sharelock. Wait for 982 * it to drop the lock so we can free this page. 983 */ 984 page_unlock(pp); 985 pp = page_lookup(vp, (u_offset_t)(uintptr_t)addr, 986 SE_EXCL); 987 } 988 #else 989 pp = page_lookup(vp, (u_offset_t)(uintptr_t)addr, SE_EXCL); 990 #endif 991 if (pp == NULL) 992 panic("segkmem_free: page not found"); 993 /* Clear p_lckcnt so page_destroy() doesn't update availrmem */ 994 pp->p_lckcnt = 0; 995 if (func) 996 func(pp); 997 else 998 page_destroy(pp, 0); 999 } 1000 if (func == NULL) 1001 page_unresv(npages); 1002 1003 if (vmp != NULL) 1004 vmem_free(vmp, inaddr, size); 1005 1006 } 1007 1008 void 1009 segkmem_xfree(vmem_t *vmp, void *inaddr, size_t size, void (*func)(page_t *)) 1010 { 1011 segkmem_free_vn(vmp, inaddr, size, &kvp, func); 1012 } 1013 1014 void 1015 segkmem_free(vmem_t *vmp, void *inaddr, size_t size) 1016 { 1017 segkmem_free_vn(vmp, inaddr, size, &kvp, NULL); 1018 } 1019 1020 void 1021 segkmem_zio_free(vmem_t *vmp, void *inaddr, size_t size) 1022 { 1023 segkmem_free_vn(vmp, inaddr, size, &zvp, NULL); 1024 } 1025 1026 void 1027 segkmem_gc(void) 1028 { 1029 ASSERT(kvseg.s_base != NULL); 1030 while (segkmem_gc_list != NULL) { 1031 segkmem_gc_list_t *gc = segkmem_gc_list; 1032 segkmem_gc_list = gc->gc_next; 1033 segkmem_free(gc->gc_arena, gc, gc->gc_size); 1034 } 1035 } 1036 1037 /* 1038 * Legacy entry points from here to end of file. 1039 */ 1040 void 1041 segkmem_mapin(struct seg *seg, void *addr, size_t size, uint_t vprot, 1042 pfn_t pfn, uint_t flags) 1043 { 1044 hat_unload(seg->s_as->a_hat, addr, size, HAT_UNLOAD_UNLOCK); 1045 hat_devload(seg->s_as->a_hat, addr, size, pfn, vprot, 1046 flags | HAT_LOAD_LOCK); 1047 } 1048 1049 void 1050 segkmem_mapout(struct seg *seg, void *addr, size_t size) 1051 { 1052 hat_unload(seg->s_as->a_hat, addr, size, HAT_UNLOAD_UNLOCK); 1053 } 1054 1055 void * 1056 kmem_getpages(pgcnt_t npages, int kmflag) 1057 { 1058 return (kmem_alloc(ptob(npages), kmflag)); 1059 } 1060 1061 void 1062 kmem_freepages(void *addr, pgcnt_t npages) 1063 { 1064 kmem_free(addr, ptob(npages)); 1065 } 1066 1067 /* 1068 * segkmem_page_create_large() allocates a large page to be used for the kmem 1069 * caches. If kpr is enabled we ask for a relocatable page unless requested 1070 * otherwise. If kpr is disabled we have to ask for a non-reloc page 1071 */ 1072 static page_t * 1073 segkmem_page_create_large(void *addr, size_t size, int vmflag, void *arg) 1074 { 1075 int pgflags; 1076 1077 pgflags = PG_EXCL; 1078 1079 if (segkmem_reloc == 0 || (vmflag & VM_NORELOC)) 1080 pgflags |= PG_NORELOC; 1081 if (!(vmflag & VM_NOSLEEP)) 1082 pgflags |= PG_WAIT; 1083 if (vmflag & VM_PUSHPAGE) 1084 pgflags |= PG_PUSHPAGE; 1085 if (vmflag & VM_NORMALPRI) 1086 pgflags |= PG_NORMALPRI; 1087 1088 return (page_create_va_large(&kvp, (u_offset_t)(uintptr_t)addr, size, 1089 pgflags, &kvseg, addr, arg)); 1090 } 1091 1092 /* 1093 * Allocate a large page to back the virtual address range 1094 * [addr, addr + size). If addr is NULL, allocate the virtual address 1095 * space as well. 1096 */ 1097 static void * 1098 segkmem_xalloc_lp(vmem_t *vmp, void *inaddr, size_t size, int vmflag, 1099 uint_t attr, page_t *(*page_create_func)(void *, size_t, int, void *), 1100 void *pcarg) 1101 { 1102 caddr_t addr = inaddr, pa; 1103 size_t lpsize = segkmem_lpsize; 1104 pgcnt_t npages = btopr(size); 1105 pgcnt_t nbpages = btop(lpsize); 1106 pgcnt_t nlpages = size >> segkmem_lpshift; 1107 size_t ppasize = nbpages * sizeof (page_t *); 1108 page_t *pp, *rootpp, **ppa, *pplist = NULL; 1109 int i; 1110 1111 vmflag |= VM_NOSLEEP; 1112 1113 if (page_resv(npages, vmflag & VM_KMFLAGS) == 0) { 1114 return (NULL); 1115 } 1116 1117 /* 1118 * allocate an array we need for hat_memload_array. 1119 * we use a separate arena to avoid recursion. 1120 * we will not need this array when hat_memload_array learns pp++ 1121 */ 1122 if ((ppa = vmem_alloc(segkmem_ppa_arena, ppasize, vmflag)) == NULL) { 1123 goto fail_array_alloc; 1124 } 1125 1126 if (inaddr == NULL && (addr = vmem_alloc(vmp, size, vmflag)) == NULL) 1127 goto fail_vmem_alloc; 1128 1129 ASSERT(((uintptr_t)addr & (lpsize - 1)) == 0); 1130 1131 /* create all the pages */ 1132 for (pa = addr, i = 0; i < nlpages; i++, pa += lpsize) { 1133 if ((pp = page_create_func(pa, lpsize, vmflag, pcarg)) == NULL) 1134 goto fail_page_create; 1135 page_list_concat(&pplist, &pp); 1136 } 1137 1138 /* at this point we have all the resource to complete the request */ 1139 while ((rootpp = pplist) != NULL) { 1140 for (i = 0; i < nbpages; i++) { 1141 ASSERT(pplist != NULL); 1142 pp = pplist; 1143 page_sub(&pplist, pp); 1144 ASSERT(page_iolock_assert(pp)); 1145 page_io_unlock(pp); 1146 ppa[i] = pp; 1147 } 1148 /* 1149 * Load the locked entry. It's OK to preload the entry into the 1150 * TSB since we now support large mappings in the kernel TSB. 1151 */ 1152 hat_memload_array(kas.a_hat, 1153 (caddr_t)(uintptr_t)rootpp->p_offset, lpsize, 1154 ppa, (PROT_ALL & ~PROT_USER) | HAT_NOSYNC | attr, 1155 HAT_LOAD_LOCK); 1156 1157 for (--i; i >= 0; --i) { 1158 ppa[i]->p_lckcnt = 1; 1159 page_unlock(ppa[i]); 1160 } 1161 } 1162 1163 vmem_free(segkmem_ppa_arena, ppa, ppasize); 1164 return (addr); 1165 1166 fail_page_create: 1167 while ((rootpp = pplist) != NULL) { 1168 for (i = 0, pp = pplist; i < nbpages; i++, pp = pplist) { 1169 ASSERT(pp != NULL); 1170 page_sub(&pplist, pp); 1171 ASSERT(page_iolock_assert(pp)); 1172 page_io_unlock(pp); 1173 } 1174 page_destroy_pages(rootpp); 1175 } 1176 1177 if (inaddr == NULL) 1178 vmem_free(vmp, addr, size); 1179 1180 fail_vmem_alloc: 1181 vmem_free(segkmem_ppa_arena, ppa, ppasize); 1182 1183 fail_array_alloc: 1184 page_unresv(npages); 1185 1186 return (NULL); 1187 } 1188 1189 static void 1190 segkmem_free_one_lp(caddr_t addr, size_t size) 1191 { 1192 page_t *pp, *rootpp = NULL; 1193 pgcnt_t pgs_left = btopr(size); 1194 1195 ASSERT(size == segkmem_lpsize); 1196 1197 hat_unload(kas.a_hat, addr, size, HAT_UNLOAD_UNLOCK); 1198 1199 for (; pgs_left > 0; addr += PAGESIZE, pgs_left--) { 1200 pp = page_lookup(&kvp, (u_offset_t)(uintptr_t)addr, SE_EXCL); 1201 if (pp == NULL) 1202 panic("segkmem_free_one_lp: page not found"); 1203 ASSERT(PAGE_EXCL(pp)); 1204 pp->p_lckcnt = 0; 1205 if (rootpp == NULL) 1206 rootpp = pp; 1207 } 1208 ASSERT(rootpp != NULL); 1209 page_destroy_pages(rootpp); 1210 1211 /* page_unresv() is done by the caller */ 1212 } 1213 1214 /* 1215 * This function is called to import new spans into the vmem arenas like 1216 * kmem_default_arena and kmem_oversize_arena. It first tries to import 1217 * spans from large page arena - kmem_lp_arena. In order to do this it might 1218 * have to "upgrade the requested size" to kmem_lp_arena quantum. If 1219 * it was not able to satisfy the upgraded request it then calls regular 1220 * segkmem_alloc() that satisfies the request by importing from "*vmp" arena 1221 */ 1222 /*ARGSUSED*/ 1223 void * 1224 segkmem_alloc_lp(vmem_t *vmp, size_t *sizep, size_t align, int vmflag) 1225 { 1226 size_t size; 1227 kthread_t *t = curthread; 1228 segkmem_lpcb_t *lpcb = &segkmem_lpcb; 1229 1230 ASSERT(sizep != NULL); 1231 1232 size = *sizep; 1233 1234 if (lpcb->lp_uselp && !(t->t_flag & T_PANIC) && 1235 !(vmflag & SEGKMEM_SHARELOCKED)) { 1236 1237 size_t kmemlp_qnt = segkmem_kmemlp_quantum; 1238 size_t asize = P2ROUNDUP(size, kmemlp_qnt); 1239 void *addr = NULL; 1240 ulong_t *lpthrtp = &lpcb->lp_throttle; 1241 ulong_t lpthrt = *lpthrtp; 1242 int dowakeup = 0; 1243 int doalloc = 1; 1244 1245 ASSERT(kmem_lp_arena != NULL); 1246 ASSERT(asize >= size); 1247 1248 if (lpthrt != 0) { 1249 /* try to update the throttle value */ 1250 lpthrt = atomic_inc_ulong_nv(lpthrtp); 1251 if (lpthrt >= segkmem_lpthrottle_max) { 1252 lpthrt = atomic_cas_ulong(lpthrtp, lpthrt, 1253 segkmem_lpthrottle_max / 4); 1254 } 1255 1256 /* 1257 * when we get above throttle start do an exponential 1258 * backoff at trying large pages and reaping 1259 */ 1260 if (lpthrt > segkmem_lpthrottle_start && 1261 !ISP2(lpthrt)) { 1262 lpcb->allocs_throttled++; 1263 lpthrt--; 1264 if (ISP2(lpthrt)) 1265 kmem_reap(); 1266 return (segkmem_alloc(vmp, size, vmflag)); 1267 } 1268 } 1269 1270 if (!(vmflag & VM_NOSLEEP) && 1271 segkmem_heaplp_quantum >= (8 * kmemlp_qnt) && 1272 vmem_size(kmem_lp_arena, VMEM_FREE) <= kmemlp_qnt && 1273 asize < (segkmem_heaplp_quantum - kmemlp_qnt)) { 1274 1275 /* 1276 * we are low on free memory in kmem_lp_arena 1277 * we let only one guy to allocate heap_lp 1278 * quantum size chunk that everybody is going to 1279 * share 1280 */ 1281 mutex_enter(&lpcb->lp_lock); 1282 1283 if (lpcb->lp_wait) { 1284 1285 /* we are not the first one - wait */ 1286 cv_wait(&lpcb->lp_cv, &lpcb->lp_lock); 1287 if (vmem_size(kmem_lp_arena, VMEM_FREE) < 1288 kmemlp_qnt) { 1289 doalloc = 0; 1290 } 1291 } else if (vmem_size(kmem_lp_arena, VMEM_FREE) <= 1292 kmemlp_qnt) { 1293 1294 /* 1295 * we are the first one, make sure we import 1296 * a large page 1297 */ 1298 if (asize == kmemlp_qnt) 1299 asize += kmemlp_qnt; 1300 dowakeup = 1; 1301 lpcb->lp_wait = 1; 1302 } 1303 1304 mutex_exit(&lpcb->lp_lock); 1305 } 1306 1307 /* 1308 * VM_ABORT flag prevents sleeps in vmem_xalloc when 1309 * large pages are not available. In that case this allocation 1310 * attempt will fail and we will retry allocation with small 1311 * pages. We also do not want to panic if this allocation fails 1312 * because we are going to retry. 1313 */ 1314 if (doalloc) { 1315 addr = vmem_alloc(kmem_lp_arena, asize, 1316 (vmflag | VM_ABORT) & ~VM_PANIC); 1317 1318 if (dowakeup) { 1319 mutex_enter(&lpcb->lp_lock); 1320 ASSERT(lpcb->lp_wait != 0); 1321 lpcb->lp_wait = 0; 1322 cv_broadcast(&lpcb->lp_cv); 1323 mutex_exit(&lpcb->lp_lock); 1324 } 1325 } 1326 1327 if (addr != NULL) { 1328 *sizep = asize; 1329 *lpthrtp = 0; 1330 return (addr); 1331 } 1332 1333 if (vmflag & VM_NOSLEEP) 1334 lpcb->nosleep_allocs_failed++; 1335 else 1336 lpcb->sleep_allocs_failed++; 1337 lpcb->alloc_bytes_failed += size; 1338 1339 /* if large page throttling is not started yet do it */ 1340 if (segkmem_use_lpthrottle && lpthrt == 0) { 1341 lpthrt = atomic_cas_ulong(lpthrtp, lpthrt, 1); 1342 } 1343 } 1344 return (segkmem_alloc(vmp, size, vmflag)); 1345 } 1346 1347 void 1348 segkmem_free_lp(vmem_t *vmp, void *inaddr, size_t size) 1349 { 1350 if (kmem_lp_arena == NULL || !IS_KMEM_VA_LARGEPAGE((caddr_t)inaddr)) { 1351 segkmem_free(vmp, inaddr, size); 1352 } else { 1353 vmem_free(kmem_lp_arena, inaddr, size); 1354 } 1355 } 1356 1357 /* 1358 * segkmem_alloc_lpi() imports virtual memory from large page heap arena 1359 * into kmem_lp arena. In the process it maps the imported segment with 1360 * large pages 1361 */ 1362 static void * 1363 segkmem_alloc_lpi(vmem_t *vmp, size_t size, int vmflag) 1364 { 1365 segkmem_lpcb_t *lpcb = &segkmem_lpcb; 1366 void *addr; 1367 1368 ASSERT(size != 0); 1369 ASSERT(vmp == heap_lp_arena); 1370 1371 /* do not allow large page heap grow beyound limits */ 1372 if (vmem_size(vmp, VMEM_ALLOC) >= segkmem_kmemlp_max) { 1373 lpcb->allocs_limited++; 1374 return (NULL); 1375 } 1376 1377 addr = segkmem_xalloc_lp(vmp, NULL, size, vmflag, 0, 1378 segkmem_page_create_large, NULL); 1379 return (addr); 1380 } 1381 1382 /* 1383 * segkmem_free_lpi() returns virtual memory back into large page heap arena 1384 * from kmem_lp arena. Beore doing this it unmaps the segment and frees 1385 * large pages used to map it. 1386 */ 1387 static void 1388 segkmem_free_lpi(vmem_t *vmp, void *inaddr, size_t size) 1389 { 1390 pgcnt_t nlpages = size >> segkmem_lpshift; 1391 size_t lpsize = segkmem_lpsize; 1392 caddr_t addr = inaddr; 1393 pgcnt_t npages = btopr(size); 1394 int i; 1395 1396 ASSERT(vmp == heap_lp_arena); 1397 ASSERT(IS_KMEM_VA_LARGEPAGE(addr)); 1398 ASSERT(((uintptr_t)inaddr & (lpsize - 1)) == 0); 1399 1400 for (i = 0; i < nlpages; i++) { 1401 segkmem_free_one_lp(addr, lpsize); 1402 addr += lpsize; 1403 } 1404 1405 page_unresv(npages); 1406 1407 vmem_free(vmp, inaddr, size); 1408 } 1409 1410 /* 1411 * This function is called at system boot time by kmem_init right after 1412 * /etc/system file has been read. It checks based on hardware configuration 1413 * and /etc/system settings if system is going to use large pages. The 1414 * initialiazation necessary to actually start using large pages 1415 * happens later in the process after segkmem_heap_lp_init() is called. 1416 */ 1417 int 1418 segkmem_lpsetup() 1419 { 1420 int use_large_pages = 0; 1421 1422 #ifdef __sparc 1423 1424 size_t memtotal = physmem * PAGESIZE; 1425 1426 if (heap_lp_base == NULL) { 1427 segkmem_lpsize = PAGESIZE; 1428 return (0); 1429 } 1430 1431 /* get a platform dependent value of large page size for kernel heap */ 1432 segkmem_lpsize = get_segkmem_lpsize(segkmem_lpsize); 1433 1434 if (segkmem_lpsize <= PAGESIZE) { 1435 /* 1436 * put virtual space reserved for the large page kernel 1437 * back to the regular heap 1438 */ 1439 vmem_xfree(heap_arena, heap_lp_base, 1440 heap_lp_end - heap_lp_base); 1441 heap_lp_base = NULL; 1442 heap_lp_end = NULL; 1443 segkmem_lpsize = PAGESIZE; 1444 return (0); 1445 } 1446 1447 /* set heap_lp quantum if necessary */ 1448 if (segkmem_heaplp_quantum == 0 || !ISP2(segkmem_heaplp_quantum) || 1449 P2PHASE(segkmem_heaplp_quantum, segkmem_lpsize)) { 1450 segkmem_heaplp_quantum = segkmem_lpsize; 1451 } 1452 1453 /* set kmem_lp quantum if necessary */ 1454 if (segkmem_kmemlp_quantum == 0 || !ISP2(segkmem_kmemlp_quantum) || 1455 segkmem_kmemlp_quantum > segkmem_heaplp_quantum) { 1456 segkmem_kmemlp_quantum = segkmem_heaplp_quantum; 1457 } 1458 1459 /* set total amount of memory allowed for large page kernel heap */ 1460 if (segkmem_kmemlp_max == 0) { 1461 if (segkmem_kmemlp_pcnt == 0 || segkmem_kmemlp_pcnt > 100) 1462 segkmem_kmemlp_pcnt = 12; 1463 segkmem_kmemlp_max = (memtotal * segkmem_kmemlp_pcnt) / 100; 1464 } 1465 segkmem_kmemlp_max = P2ROUNDUP(segkmem_kmemlp_max, 1466 segkmem_heaplp_quantum); 1467 1468 /* fix lp kmem preallocation request if necesssary */ 1469 if (segkmem_kmemlp_min) { 1470 segkmem_kmemlp_min = P2ROUNDUP(segkmem_kmemlp_min, 1471 segkmem_heaplp_quantum); 1472 if (segkmem_kmemlp_min > segkmem_kmemlp_max) 1473 segkmem_kmemlp_min = segkmem_kmemlp_max; 1474 } 1475 1476 use_large_pages = 1; 1477 segkmem_lpszc = page_szc(segkmem_lpsize); 1478 segkmem_lpshift = page_get_shift(segkmem_lpszc); 1479 1480 #endif 1481 return (use_large_pages); 1482 } 1483 1484 void 1485 segkmem_zio_init(void *zio_mem_base, size_t zio_mem_size) 1486 { 1487 ASSERT(zio_mem_base != NULL); 1488 ASSERT(zio_mem_size != 0); 1489 1490 /* 1491 * To reduce VA space fragmentation, we set up quantum caches for the 1492 * smaller sizes; we chose 32k because that translates to 128k VA 1493 * slabs, which matches nicely with the common 128k zio_data bufs. 1494 */ 1495 zio_arena = vmem_create("zfs_file_data", zio_mem_base, zio_mem_size, 1496 PAGESIZE, NULL, NULL, NULL, 32 * 1024, VM_SLEEP); 1497 1498 zio_alloc_arena = vmem_create("zfs_file_data_buf", NULL, 0, PAGESIZE, 1499 segkmem_zio_alloc, segkmem_zio_free, zio_arena, 0, VM_SLEEP); 1500 1501 ASSERT(zio_arena != NULL); 1502 ASSERT(zio_alloc_arena != NULL); 1503 } 1504 1505 #ifdef __sparc 1506 1507 1508 static void * 1509 segkmem_alloc_ppa(vmem_t *vmp, size_t size, int vmflag) 1510 { 1511 size_t ppaquantum = btopr(segkmem_lpsize) * sizeof (page_t *); 1512 void *addr; 1513 1514 if (ppaquantum <= PAGESIZE) 1515 return (segkmem_alloc(vmp, size, vmflag)); 1516 1517 ASSERT((size & (ppaquantum - 1)) == 0); 1518 1519 addr = vmem_xalloc(vmp, size, ppaquantum, 0, 0, NULL, NULL, vmflag); 1520 if (addr != NULL && segkmem_xalloc(vmp, addr, size, vmflag, 0, 1521 segkmem_page_create, NULL) == NULL) { 1522 vmem_xfree(vmp, addr, size); 1523 addr = NULL; 1524 } 1525 1526 return (addr); 1527 } 1528 1529 static void 1530 segkmem_free_ppa(vmem_t *vmp, void *addr, size_t size) 1531 { 1532 size_t ppaquantum = btopr(segkmem_lpsize) * sizeof (page_t *); 1533 1534 ASSERT(addr != NULL); 1535 1536 if (ppaquantum <= PAGESIZE) { 1537 segkmem_free(vmp, addr, size); 1538 } else { 1539 segkmem_free(NULL, addr, size); 1540 vmem_xfree(vmp, addr, size); 1541 } 1542 } 1543 1544 void 1545 segkmem_heap_lp_init() 1546 { 1547 segkmem_lpcb_t *lpcb = &segkmem_lpcb; 1548 size_t heap_lp_size = heap_lp_end - heap_lp_base; 1549 size_t lpsize = segkmem_lpsize; 1550 size_t ppaquantum; 1551 void *addr; 1552 1553 if (segkmem_lpsize <= PAGESIZE) { 1554 ASSERT(heap_lp_base == NULL); 1555 ASSERT(heap_lp_end == NULL); 1556 return; 1557 } 1558 1559 ASSERT(segkmem_heaplp_quantum >= lpsize); 1560 ASSERT((segkmem_heaplp_quantum & (lpsize - 1)) == 0); 1561 ASSERT(lpcb->lp_uselp == 0); 1562 ASSERT(heap_lp_base != NULL); 1563 ASSERT(heap_lp_end != NULL); 1564 ASSERT(heap_lp_base < heap_lp_end); 1565 ASSERT(heap_lp_arena == NULL); 1566 ASSERT(((uintptr_t)heap_lp_base & (lpsize - 1)) == 0); 1567 ASSERT(((uintptr_t)heap_lp_end & (lpsize - 1)) == 0); 1568 1569 /* create large page heap arena */ 1570 heap_lp_arena = vmem_create("heap_lp", heap_lp_base, heap_lp_size, 1571 segkmem_heaplp_quantum, NULL, NULL, NULL, 0, VM_SLEEP); 1572 1573 ASSERT(heap_lp_arena != NULL); 1574 1575 /* This arena caches memory already mapped by large pages */ 1576 kmem_lp_arena = vmem_create("kmem_lp", NULL, 0, segkmem_kmemlp_quantum, 1577 segkmem_alloc_lpi, segkmem_free_lpi, heap_lp_arena, 0, VM_SLEEP); 1578 1579 ASSERT(kmem_lp_arena != NULL); 1580 1581 mutex_init(&lpcb->lp_lock, NULL, MUTEX_DEFAULT, NULL); 1582 cv_init(&lpcb->lp_cv, NULL, CV_DEFAULT, NULL); 1583 1584 /* 1585 * this arena is used for the array of page_t pointers necessary 1586 * to call hat_mem_load_array 1587 */ 1588 ppaquantum = btopr(lpsize) * sizeof (page_t *); 1589 segkmem_ppa_arena = vmem_create("segkmem_ppa", NULL, 0, ppaquantum, 1590 segkmem_alloc_ppa, segkmem_free_ppa, heap_arena, ppaquantum, 1591 VM_SLEEP); 1592 1593 ASSERT(segkmem_ppa_arena != NULL); 1594 1595 /* prealloacate some memory for the lp kernel heap */ 1596 if (segkmem_kmemlp_min) { 1597 1598 ASSERT(P2PHASE(segkmem_kmemlp_min, 1599 segkmem_heaplp_quantum) == 0); 1600 1601 if ((addr = segkmem_alloc_lpi(heap_lp_arena, 1602 segkmem_kmemlp_min, VM_SLEEP)) != NULL) { 1603 1604 addr = vmem_add(kmem_lp_arena, addr, 1605 segkmem_kmemlp_min, VM_SLEEP); 1606 ASSERT(addr != NULL); 1607 } 1608 } 1609 1610 lpcb->lp_uselp = 1; 1611 } 1612 1613 #endif