1 /* 2 * High memory handling common code and variables. 3 * 4 * (C) 1999 Andrea Arcangeli, SuSE GmbH, andrea@suse.de 5 * Gerhard Wichert, Siemens AG, Gerhard.Wichert@pdb.siemens.de 6 * 7 * 8 * Redesigned the x86 32-bit VM architecture to deal with 9 * 64-bit physical space. With current x86 CPUs this 10 * means up to 64 Gigabytes physical RAM. 11 * 12 * Rewrote high memory support to move the page cache into 13 * high memory. Implemented permanent (schedulable) kmaps 14 * based on Linus' idea. 15 * 16 * Copyright (C) 1999 Ingo Molnar <mingo@redhat.com> 17 */ 18 19 #include <linux/mm.h> 20 #include <linux/module.h> 21 #include <linux/swap.h> 22 #include <linux/bio.h> 23 #include <linux/pagemap.h> 24 #include <linux/mempool.h> 25 #include <linux/blkdev.h> 26 #include <linux/init.h> 27 #include <linux/hash.h> 28 #include <linux/highmem.h> 29 #include <asm/tlbflush.h> 30 31 static mempool_t *page_pool, *isa_page_pool; 32 33 static void *page_pool_alloc_isa(gfp_t gfp_mask, void *data) 34 { 35 return alloc_page(gfp_mask | GFP_DMA); 36 } 37 38 static void page_pool_free(void *page, void *data) 39 { 40 __free_page(page); 41 } 42 43 /* 44 * Virtual_count is not a pure "count". 45 * 0 means that it is not mapped, and has not been mapped 46 * since a TLB flush - it is usable. 47 * 1 means that there are no users, but it has been mapped 48 * since the last TLB flush - so we can't use it. 49 * n means that there are (n-1) current users of it. 50 */ 51 #ifdef CONFIG_HIGHMEM 52 53 static void *page_pool_alloc(gfp_t gfp_mask, void *data) 54 { 55 return alloc_page(gfp_mask); 56 } 57 58 static int pkmap_count[LAST_PKMAP]; 59 static unsigned int last_pkmap_nr; 60 static __cacheline_aligned_in_smp DEFINE_SPINLOCK(kmap_lock); 61 62 pte_t * pkmap_page_table; 63 64 static DECLARE_WAIT_QUEUE_HEAD(pkmap_map_wait); 65 66 static void flush_all_zero_pkmaps(void) 67 { 68 int i; 69 70 flush_cache_kmaps(); 71 72 for (i = 0; i < LAST_PKMAP; i++) { 73 struct page *page; 74 75 /* 76 * zero means we don't have anything to do, 77 * >1 means that it is still in use. Only 78 * a count of 1 means that it is free but 79 * needs to be unmapped 80 */ 81 if (pkmap_count[i] != 1) 82 continue; 83 pkmap_count[i] = 0; 84 85 /* sanity check */ 86 if (pte_none(pkmap_page_table[i])) 87 BUG(); 88 89 /* 90 * Don't need an atomic fetch-and-clear op here; 91 * no-one has the page mapped, and cannot get at 92 * its virtual address (and hence PTE) without first 93 * getting the kmap_lock (which is held here). 94 * So no dangers, even with speculative execution. 95 */ 96 page = pte_page(pkmap_page_table[i]); 97 pte_clear(&init_mm, (unsigned long)page_address(page), 98 &pkmap_page_table[i]); 99 100 set_page_address(page, NULL); 101 } 102 flush_tlb_kernel_range(PKMAP_ADDR(0), PKMAP_ADDR(LAST_PKMAP)); 103 } 104 105 static inline unsigned long map_new_virtual(struct page *page) 106 { 107 unsigned long vaddr; 108 int count; 109 110 start: 111 count = LAST_PKMAP; 112 /* Find an empty entry */ 113 for (;;) { 114 last_pkmap_nr = (last_pkmap_nr + 1) & LAST_PKMAP_MASK; 115 if (!last_pkmap_nr) { 116 flush_all_zero_pkmaps(); 117 count = LAST_PKMAP; 118 } 119 if (!pkmap_count[last_pkmap_nr]) 120 break; /* Found a usable entry */ 121 if (--count) 122 continue; 123 124 /* 125 * Sleep for somebody else to unmap their entries 126 */ 127 { 128 DECLARE_WAITQUEUE(wait, current); 129 130 __set_current_state(TASK_UNINTERRUPTIBLE); 131 add_wait_queue(&pkmap_map_wait, &wait); 132 spin_unlock(&kmap_lock); 133 schedule(); 134 remove_wait_queue(&pkmap_map_wait, &wait); 135 spin_lock(&kmap_lock); 136 137 /* Somebody else might have mapped it while we slept */ 138 if (page_address(page)) 139 return (unsigned long)page_address(page); 140 141 /* Re-start */ 142 goto start; 143 } 144 } 145 vaddr = PKMAP_ADDR(last_pkmap_nr); 146 set_pte_at(&init_mm, vaddr, 147 &(pkmap_page_table[last_pkmap_nr]), mk_pte(page, kmap_prot)); 148 149 pkmap_count[last_pkmap_nr] = 1; 150 set_page_address(page, (void *)vaddr); 151 152 return vaddr; 153 } 154 155 void fastcall *kmap_high(struct page *page) 156 { 157 unsigned long vaddr; 158 159 /* 160 * For highmem pages, we can't trust "virtual" until 161 * after we have the lock. 162 * 163 * We cannot call this from interrupts, as it may block 164 */ 165 spin_lock(&kmap_lock); 166 vaddr = (unsigned long)page_address(page); 167 if (!vaddr) 168 vaddr = map_new_virtual(page); 169 pkmap_count[PKMAP_NR(vaddr)]++; 170 if (pkmap_count[PKMAP_NR(vaddr)] < 2) 171 BUG(); 172 spin_unlock(&kmap_lock); 173 return (void*) vaddr; 174 } 175 176 EXPORT_SYMBOL(kmap_high); 177 178 void fastcall kunmap_high(struct page *page) 179 { 180 unsigned long vaddr; 181 unsigned long nr; 182 int need_wakeup; 183 184 spin_lock(&kmap_lock); 185 vaddr = (unsigned long)page_address(page); 186 if (!vaddr) 187 BUG(); 188 nr = PKMAP_NR(vaddr); 189 190 /* 191 * A count must never go down to zero 192 * without a TLB flush! 193 */ 194 need_wakeup = 0; 195 switch (--pkmap_count[nr]) { 196 case 0: 197 BUG(); 198 case 1: 199 /* 200 * Avoid an unnecessary wake_up() function call. 201 * The common case is pkmap_count[] == 1, but 202 * no waiters. 203 * The tasks queued in the wait-queue are guarded 204 * by both the lock in the wait-queue-head and by 205 * the kmap_lock. As the kmap_lock is held here, 206 * no need for the wait-queue-head's lock. Simply 207 * test if the queue is empty. 208 */ 209 need_wakeup = waitqueue_active(&pkmap_map_wait); 210 } 211 spin_unlock(&kmap_lock); 212 213 /* do wake-up, if needed, race-free outside of the spin lock */ 214 if (need_wakeup) 215 wake_up(&pkmap_map_wait); 216 } 217 218 EXPORT_SYMBOL(kunmap_high); 219 220 #define POOL_SIZE 64 221 222 static __init int init_emergency_pool(void) 223 { 224 struct sysinfo i; 225 si_meminfo(&i); 226 si_swapinfo(&i); 227 228 if (!i.totalhigh) 229 return 0; 230 231 page_pool = mempool_create(POOL_SIZE, page_pool_alloc, page_pool_free, NULL); 232 if (!page_pool) 233 BUG(); 234 printk("highmem bounce pool size: %d pages\n", POOL_SIZE); 235 236 return 0; 237 } 238 239 __initcall(init_emergency_pool); 240 241 /* 242 * highmem version, map in to vec 243 */ 244 static void bounce_copy_vec(struct bio_vec *to, unsigned char *vfrom) 245 { 246 unsigned long flags; 247 unsigned char *vto; 248 249 local_irq_save(flags); 250 vto = kmap_atomic(to->bv_page, KM_BOUNCE_READ); 251 memcpy(vto + to->bv_offset, vfrom, to->bv_len); 252 kunmap_atomic(vto, KM_BOUNCE_READ); 253 local_irq_restore(flags); 254 } 255 256 #else /* CONFIG_HIGHMEM */ 257 258 #define bounce_copy_vec(to, vfrom) \ 259 memcpy(page_address((to)->bv_page) + (to)->bv_offset, vfrom, (to)->bv_len) 260 261 #endif 262 263 #define ISA_POOL_SIZE 16 264 265 /* 266 * gets called "every" time someone init's a queue with BLK_BOUNCE_ISA 267 * as the max address, so check if the pool has already been created. 268 */ 269 int init_emergency_isa_pool(void) 270 { 271 if (isa_page_pool) 272 return 0; 273 274 isa_page_pool = mempool_create(ISA_POOL_SIZE, page_pool_alloc_isa, page_pool_free, NULL); 275 if (!isa_page_pool) 276 BUG(); 277 278 printk("isa bounce pool size: %d pages\n", ISA_POOL_SIZE); 279 return 0; 280 } 281 282 /* 283 * Simple bounce buffer support for highmem pages. Depending on the 284 * queue gfp mask set, *to may or may not be a highmem page. kmap it 285 * always, it will do the Right Thing 286 */ 287 static void copy_to_high_bio_irq(struct bio *to, struct bio *from) 288 { 289 unsigned char *vfrom; 290 struct bio_vec *tovec, *fromvec; 291 int i; 292 293 __bio_for_each_segment(tovec, to, i, 0) { 294 fromvec = from->bi_io_vec + i; 295 296 /* 297 * not bounced 298 */ 299 if (tovec->bv_page == fromvec->bv_page) 300 continue; 301 302 /* 303 * fromvec->bv_offset and fromvec->bv_len might have been 304 * modified by the block layer, so use the original copy, 305 * bounce_copy_vec already uses tovec->bv_len 306 */ 307 vfrom = page_address(fromvec->bv_page) + tovec->bv_offset; 308 309 flush_dcache_page(tovec->bv_page); 310 bounce_copy_vec(tovec, vfrom); 311 } 312 } 313 314 static void bounce_end_io(struct bio *bio, mempool_t *pool, int err) 315 { 316 struct bio *bio_orig = bio->bi_private; 317 struct bio_vec *bvec, *org_vec; 318 int i; 319 320 if (test_bit(BIO_EOPNOTSUPP, &bio->bi_flags)) 321 set_bit(BIO_EOPNOTSUPP, &bio_orig->bi_flags); 322 323 /* 324 * free up bounce indirect pages used 325 */ 326 __bio_for_each_segment(bvec, bio, i, 0) { 327 org_vec = bio_orig->bi_io_vec + i; 328 if (bvec->bv_page == org_vec->bv_page) 329 continue; 330 331 mempool_free(bvec->bv_page, pool); 332 dec_page_state(nr_bounce); 333 } 334 335 bio_endio(bio_orig, bio_orig->bi_size, err); 336 bio_put(bio); 337 } 338 339 static int bounce_end_io_write(struct bio *bio, unsigned int bytes_done,int err) 340 { 341 if (bio->bi_size) 342 return 1; 343 344 bounce_end_io(bio, page_pool, err); 345 return 0; 346 } 347 348 static int bounce_end_io_write_isa(struct bio *bio, unsigned int bytes_done, int err) 349 { 350 if (bio->bi_size) 351 return 1; 352 353 bounce_end_io(bio, isa_page_pool, err); 354 return 0; 355 } 356 357 static void __bounce_end_io_read(struct bio *bio, mempool_t *pool, int err) 358 { 359 struct bio *bio_orig = bio->bi_private; 360 361 if (test_bit(BIO_UPTODATE, &bio->bi_flags)) 362 copy_to_high_bio_irq(bio_orig, bio); 363 364 bounce_end_io(bio, pool, err); 365 } 366 367 static int bounce_end_io_read(struct bio *bio, unsigned int bytes_done, int err) 368 { 369 if (bio->bi_size) 370 return 1; 371 372 __bounce_end_io_read(bio, page_pool, err); 373 return 0; 374 } 375 376 static int bounce_end_io_read_isa(struct bio *bio, unsigned int bytes_done, int err) 377 { 378 if (bio->bi_size) 379 return 1; 380 381 __bounce_end_io_read(bio, isa_page_pool, err); 382 return 0; 383 } 384 385 static void __blk_queue_bounce(request_queue_t *q, struct bio **bio_orig, 386 mempool_t *pool) 387 { 388 struct page *page; 389 struct bio *bio = NULL; 390 int i, rw = bio_data_dir(*bio_orig); 391 struct bio_vec *to, *from; 392 393 bio_for_each_segment(from, *bio_orig, i) { 394 page = from->bv_page; 395 396 /* 397 * is destination page below bounce pfn? 398 */ 399 if (page_to_pfn(page) < q->bounce_pfn) 400 continue; 401 402 /* 403 * irk, bounce it 404 */ 405 if (!bio) 406 bio = bio_alloc(GFP_NOIO, (*bio_orig)->bi_vcnt); 407 408 to = bio->bi_io_vec + i; 409 410 to->bv_page = mempool_alloc(pool, q->bounce_gfp); 411 to->bv_len = from->bv_len; 412 to->bv_offset = from->bv_offset; 413 inc_page_state(nr_bounce); 414 415 if (rw == WRITE) { 416 char *vto, *vfrom; 417 418 flush_dcache_page(from->bv_page); 419 vto = page_address(to->bv_page) + to->bv_offset; 420 vfrom = kmap(from->bv_page) + from->bv_offset; 421 memcpy(vto, vfrom, to->bv_len); 422 kunmap(from->bv_page); 423 } 424 } 425 426 /* 427 * no pages bounced 428 */ 429 if (!bio) 430 return; 431 432 /* 433 * at least one page was bounced, fill in possible non-highmem 434 * pages 435 */ 436 __bio_for_each_segment(from, *bio_orig, i, 0) { 437 to = bio_iovec_idx(bio, i); 438 if (!to->bv_page) { 439 to->bv_page = from->bv_page; 440 to->bv_len = from->bv_len; 441 to->bv_offset = from->bv_offset; 442 } 443 } 444 445 bio->bi_bdev = (*bio_orig)->bi_bdev; 446 bio->bi_flags |= (1 << BIO_BOUNCED); 447 bio->bi_sector = (*bio_orig)->bi_sector; 448 bio->bi_rw = (*bio_orig)->bi_rw; 449 450 bio->bi_vcnt = (*bio_orig)->bi_vcnt; 451 bio->bi_idx = (*bio_orig)->bi_idx; 452 bio->bi_size = (*bio_orig)->bi_size; 453 454 if (pool == page_pool) { 455 bio->bi_end_io = bounce_end_io_write; 456 if (rw == READ) 457 bio->bi_end_io = bounce_end_io_read; 458 } else { 459 bio->bi_end_io = bounce_end_io_write_isa; 460 if (rw == READ) 461 bio->bi_end_io = bounce_end_io_read_isa; 462 } 463 464 bio->bi_private = *bio_orig; 465 *bio_orig = bio; 466 } 467 468 void blk_queue_bounce(request_queue_t *q, struct bio **bio_orig) 469 { 470 mempool_t *pool; 471 472 /* 473 * for non-isa bounce case, just check if the bounce pfn is equal 474 * to or bigger than the highest pfn in the system -- in that case, 475 * don't waste time iterating over bio segments 476 */ 477 if (!(q->bounce_gfp & GFP_DMA)) { 478 if (q->bounce_pfn >= blk_max_pfn) 479 return; 480 pool = page_pool; 481 } else { 482 BUG_ON(!isa_page_pool); 483 pool = isa_page_pool; 484 } 485 486 /* 487 * slow path 488 */ 489 __blk_queue_bounce(q, bio_orig, pool); 490 } 491 492 EXPORT_SYMBOL(blk_queue_bounce); 493 494 #if defined(HASHED_PAGE_VIRTUAL) 495 496 #define PA_HASH_ORDER 7 497 498 /* 499 * Describes one page->virtual association 500 */ 501 struct page_address_map { 502 struct page *page; 503 void *virtual; 504 struct list_head list; 505 }; 506 507 /* 508 * page_address_map freelist, allocated from page_address_maps. 509 */ 510 static struct list_head page_address_pool; /* freelist */ 511 static spinlock_t pool_lock; /* protects page_address_pool */ 512 513 /* 514 * Hash table bucket 515 */ 516 static struct page_address_slot { 517 struct list_head lh; /* List of page_address_maps */ 518 spinlock_t lock; /* Protect this bucket's list */ 519 } ____cacheline_aligned_in_smp page_address_htable[1<<PA_HASH_ORDER]; 520 521 static struct page_address_slot *page_slot(struct page *page) 522 { 523 return &page_address_htable[hash_ptr(page, PA_HASH_ORDER)]; 524 } 525 526 void *page_address(struct page *page) 527 { 528 unsigned long flags; 529 void *ret; 530 struct page_address_slot *pas; 531 532 if (!PageHighMem(page)) 533 return lowmem_page_address(page); 534 535 pas = page_slot(page); 536 ret = NULL; 537 spin_lock_irqsave(&pas->lock, flags); 538 if (!list_empty(&pas->lh)) { 539 struct page_address_map *pam; 540 541 list_for_each_entry(pam, &pas->lh, list) { 542 if (pam->page == page) { 543 ret = pam->virtual; 544 goto done; 545 } 546 } 547 } 548 done: 549 spin_unlock_irqrestore(&pas->lock, flags); 550 return ret; 551 } 552 553 EXPORT_SYMBOL(page_address); 554 555 void set_page_address(struct page *page, void *virtual) 556 { 557 unsigned long flags; 558 struct page_address_slot *pas; 559 struct page_address_map *pam; 560 561 BUG_ON(!PageHighMem(page)); 562 563 pas = page_slot(page); 564 if (virtual) { /* Add */ 565 BUG_ON(list_empty(&page_address_pool)); 566 567 spin_lock_irqsave(&pool_lock, flags); 568 pam = list_entry(page_address_pool.next, 569 struct page_address_map, list); 570 list_del(&pam->list); 571 spin_unlock_irqrestore(&pool_lock, flags); 572 573 pam->page = page; 574 pam->virtual = virtual; 575 576 spin_lock_irqsave(&pas->lock, flags); 577 list_add_tail(&pam->list, &pas->lh); 578 spin_unlock_irqrestore(&pas->lock, flags); 579 } else { /* Remove */ 580 spin_lock_irqsave(&pas->lock, flags); 581 list_for_each_entry(pam, &pas->lh, list) { 582 if (pam->page == page) { 583 list_del(&pam->list); 584 spin_unlock_irqrestore(&pas->lock, flags); 585 spin_lock_irqsave(&pool_lock, flags); 586 list_add_tail(&pam->list, &page_address_pool); 587 spin_unlock_irqrestore(&pool_lock, flags); 588 goto done; 589 } 590 } 591 spin_unlock_irqrestore(&pas->lock, flags); 592 } 593 done: 594 return; 595 } 596 597 static struct page_address_map page_address_maps[LAST_PKMAP]; 598 599 void __init page_address_init(void) 600 { 601 int i; 602 603 INIT_LIST_HEAD(&page_address_pool); 604 for (i = 0; i < ARRAY_SIZE(page_address_maps); i++) 605 list_add(&page_address_maps[i].list, &page_address_pool); 606 for (i = 0; i < ARRAY_SIZE(page_address_htable); i++) { 607 INIT_LIST_HEAD(&page_address_htable[i].lh); 608 spin_lock_init(&page_address_htable[i].lock); 609 } 610 spin_lock_init(&pool_lock); 611 } 612 613 #endif /* defined(CONFIG_HIGHMEM) && !defined(WANT_PAGE_VIRTUAL) */ 614