4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
6 * Swap reorganised 29.12.95, Stephen Tweedie.
7 * kswapd added: 7.1.96 sct
8 * Removed kswapd_ctl limits, and swap out as many pages as needed
9 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10 * Version: $Id: vmscan.c,v 1.5 1998/02/23 22:14:28 sct Exp $
11 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
14 #include <linux/slab.h>
15 #include <linux/kernel_stat.h>
16 #include <linux/swap.h>
17 #include <linux/swapctl.h>
18 #include <linux/smp_lock.h>
19 #include <linux/pagemap.h>
20 #include <linux/init.h>
21 #include <linux/highmem.h>
22 #include <linux/file.h>
24 #include <asm/pgalloc.h>
27 * The swap-out functions return 1 if they successfully
28 * threw something out, and we got a free page. It returns
29 * zero if it couldn't do anything, and any other value
30 * indicates it decreased rss, but the page was shared.
32 * NOTE! If it sleeps, it *must* return 1 to make sure we
33 * don't continue with the swap-out. Otherwise we may be
34 * using a process that no longer actually exists (it might
35 * have died while we slept).
37 static int try_to_swap_out(struct mm_struct
* mm
, struct vm_area_struct
* vma
, unsigned long address
, pte_t
* page_table
, int gfp_mask
)
42 int (*swapout
)(struct page
*, struct file
*);
45 if (!pte_present(pte
))
48 if ((page
-mem_map
>= max_mapnr
) || PageReserved(page
))
52 /* Don't look at this pte if it's been accessed recently. */
55 * Transfer the "accessed" bit from the page
56 * tables to the global page map.
58 set_pte(page_table
, pte_mkold(pte
));
59 set_bit(PG_referenced
, &page
->flags
);
67 * Is the page already in the swap cache? If so, then
68 * we can just drop our reference to it without doing
69 * any IO - it's already up-to-date on disk.
71 * Return 0, as we didn't actually free any real
72 * memory, and we should just continue our scan.
74 if (PageSwapCache(page
)) {
75 entry
.val
= page
->index
;
76 swap_duplicate(entry
);
77 set_pte(page_table
, swp_entry_to_pte(entry
));
80 flush_tlb_page(vma
, address
);
86 * Is it a clean page? Then it must be recoverable
87 * by just paging it in again, and we can just drop
90 * However, this won't actually free any real
91 * memory, as the page will just be in the page cache
92 * somewhere, and as such we should just continue
95 * Basically, this just makes it possible for us to do
96 * some real work in the future in "shrink_mmap()".
98 if (!pte_dirty(pte
)) {
99 flush_cache_page(vma
, address
);
100 pte_clear(page_table
);
105 * Don't go down into the swap-out stuff if
106 * we cannot do I/O! Avoid recursing on FS
109 if (!(gfp_mask
& __GFP_IO
))
113 * Ok, it's really dirty. That means that
114 * we should either create a new swap cache
115 * entry for it, or we should write it back
116 * to its own backing store.
118 * Note that in neither case do we actually
119 * know that we make a page available, but
120 * as we potentially sleep we can no longer
121 * continue scanning, so we migth as well
122 * assume we free'd something.
124 * NOTE NOTE NOTE! This should just set a
125 * dirty bit in 'page', and just drop the
126 * pte. All the hard work would be done by
129 * That would get rid of a lot of problems.
131 flush_cache_page(vma
, address
);
132 if (vma
->vm_ops
&& (swapout
= vma
->vm_ops
->swapout
)) {
134 struct file
*file
= vma
->vm_file
;
135 if (file
) get_file(file
);
136 pte_clear(page_table
);
138 flush_tlb_page(vma
, address
);
139 vmlist_access_unlock(vma
->vm_mm
);
140 error
= swapout(page
, file
);
141 if (file
) fput(file
);
143 goto out_free_success
;
149 * This is a dirty, swappable page. First of all,
150 * get a suitable swap entry for it, and make sure
151 * we have the swap cache set up to associate the
152 * page with that swap entry.
154 entry
= acquire_swap_entry(page
);
156 goto out_failed
; /* No swap space left */
158 if (!(page
= prepare_highmem_swapout(page
)))
161 swap_duplicate(entry
); /* One for the process, one for the swap cache */
163 /* This will also lock the page */
164 add_to_swap_cache(page
, entry
);
165 /* Put the swap entry into the pte after the page is in swapcache */
167 set_pte(page_table
, swp_entry_to_pte(entry
));
168 flush_tlb_page(vma
, address
);
169 vmlist_access_unlock(vma
->vm_mm
);
171 /* OK, do a physical asynchronous write to swap. */
172 rw_swap_page(WRITE
, page
, 0);
185 * A new implementation of swap_out(). We do not swap complete processes,
186 * but only a small number of blocks, before we continue with the next
187 * process. The number of blocks actually swapped is determined on the
188 * number of page faults, that this process actually had in the last time,
189 * so we won't swap heavily used processes all the time ...
191 * Note: the priority argument is a hint on much CPU to waste with the
192 * swap block search, not a hint, of how much blocks to swap with
195 * (C) 1993 Kai Petzke, wpp@marie.physik.tu-berlin.de
198 static inline int swap_out_pmd(struct mm_struct
* mm
, struct vm_area_struct
* vma
, pmd_t
*dir
, unsigned long address
, unsigned long end
, int gfp_mask
)
201 unsigned long pmd_end
;
211 pte
= pte_offset(dir
, address
);
213 pmd_end
= (address
+ PMD_SIZE
) & PMD_MASK
;
219 vma
->vm_mm
->swap_address
= address
+ PAGE_SIZE
;
220 result
= try_to_swap_out(mm
, vma
, address
, pte
, gfp_mask
);
225 address
+= PAGE_SIZE
;
227 } while (address
&& (address
< end
));
231 static inline int swap_out_pgd(struct mm_struct
* mm
, struct vm_area_struct
* vma
, pgd_t
*dir
, unsigned long address
, unsigned long end
, int gfp_mask
)
234 unsigned long pgd_end
;
244 pmd
= pmd_offset(dir
, address
);
246 pgd_end
= (address
+ PGDIR_SIZE
) & PGDIR_MASK
;
247 if (pgd_end
&& (end
> pgd_end
))
251 int result
= swap_out_pmd(mm
, vma
, pmd
, address
, end
, gfp_mask
);
256 address
= (address
+ PMD_SIZE
) & PMD_MASK
;
258 } while (address
&& (address
< end
));
262 static int swap_out_vma(struct mm_struct
* mm
, struct vm_area_struct
* vma
, unsigned long address
, int gfp_mask
)
267 /* Don't swap out areas which are locked down */
268 if (vma
->vm_flags
& VM_LOCKED
)
271 pgdir
= pgd_offset(vma
->vm_mm
, address
);
277 int result
= swap_out_pgd(mm
, vma
, pgdir
, address
, end
, gfp_mask
);
282 address
= (address
+ PGDIR_SIZE
) & PGDIR_MASK
;
284 } while (address
&& (address
< end
));
288 static int swap_out_mm(struct mm_struct
* mm
, int gfp_mask
)
290 unsigned long address
;
291 struct vm_area_struct
* vma
;
294 * Go through process' page directory.
296 address
= mm
->swap_address
;
299 * Find the proper vm-area after freezing the vma chain
302 vmlist_access_lock(mm
);
303 vma
= find_vma(mm
, address
);
305 if (address
< vma
->vm_start
)
306 address
= vma
->vm_start
;
309 int result
= swap_out_vma(mm
, vma
, address
, gfp_mask
);
315 address
= vma
->vm_start
;
318 vmlist_access_unlock(mm
);
320 /* We didn't find anything for the process */
322 mm
->swap_address
= 0;
327 * Select the task with maximal swap_cnt and try to swap out a page.
328 * N.B. This function returns only 0 or 1. Return values != 1 from
329 * the lower level routines result in continued processing.
331 int swap_out(unsigned int priority
, int gfp_mask
)
333 struct task_struct
* p
;
340 * We make one or two passes through the task list, indexed by
342 * Pass 1: select the swappable task with maximal RSS that has
343 * not yet been swapped out.
344 * Pass 2: re-assign rss swap_cnt values, then select as above.
346 * With this approach, there's no need to remember the last task
347 * swapped out. If the swap-out fails, we clear swap_cnt so the
348 * task won't be selected again until all others have been tried.
350 * Think of swap_cnt as a "shadow rss" - it tells us which process
351 * we want to page out (always try largest first).
353 counter
= nr_threads
/ (priority
+1);
357 for (; counter
>= 0; counter
--) {
358 unsigned long max_cnt
= 0;
359 struct mm_struct
*best
= NULL
;
362 read_lock(&tasklist_lock
);
363 p
= init_task
.next_task
;
364 for (; p
!= &init_task
; p
= p
->next_task
) {
365 struct mm_struct
*mm
= p
->mm
;
367 if (!p
->swappable
|| !mm
)
371 /* Refresh swap_cnt? */
373 mm
->swap_cnt
= mm
->rss
;
374 if (mm
->swap_cnt
> max_cnt
) {
375 max_cnt
= mm
->swap_cnt
;
381 /* we just assigned swap_cnt, normalise values */
383 p
= init_task
.next_task
;
384 for (; p
!= &init_task
; p
= p
->next_task
) {
386 struct mm_struct
*mm
= p
->mm
;
387 if (!p
->swappable
|| !mm
|| mm
->rss
<= 0)
389 /* small processes are swapped out less */
390 while ((mm
->swap_cnt
<< 2 * (i
+ 1) < max_cnt
)
393 mm
->swap_cnt
+= i
; /* if swap_cnt reaches 0 */
394 /* we're big -> hog treatment */
399 read_unlock(&tasklist_lock
);
409 atomic_inc(&best
->mm_count
);
410 ret
= swap_out_mm(best
, gfp_mask
);
417 kill_proc(pid
, SIGBUS
, 1);
428 * We need to make the locks finer granularity, but right
429 * now we need this so that we can do page allocations
430 * without holding the kernel lock etc.
432 * We want to try to free "count" pages, and we need to
433 * cluster them so that we get good swap-out behaviour. See
434 * the "free_memory()" macro for details.
436 static int do_try_to_free_pages(unsigned int gfp_mask
, zone_t
*zone
)
439 int count
= SWAP_CLUSTER_MAX
;
441 /* Always trim SLAB caches when memory gets low. */
442 kmem_cache_reap(gfp_mask
);
446 while (shrink_mmap(priority
, gfp_mask
, zone
)) {
452 /* Try to get rid of some shared memory pages.. */
453 if (gfp_mask
& __GFP_IO
) {
455 * don't be too light against the d/i cache since
456 * shrink_mmap() almost never fail when there's
457 * really plenty of memory free.
459 count
-= shrink_dcache_memory(priority
, gfp_mask
, zone
);
460 count
-= shrink_icache_memory(priority
, gfp_mask
, zone
);
463 while (shm_swap(priority
, gfp_mask
, zone
)) {
469 /* Then, try to page stuff out.. */
470 while (swap_out(priority
, gfp_mask
)) {
474 } while (--priority
>= 0);
477 return priority
>= 0;
480 DECLARE_WAIT_QUEUE_HEAD(kswapd_wait
);
483 * The background pageout daemon, started as a kernel thread
484 * from the init process.
486 * This basically trickles out pages so that we have _some_
487 * free memory available even if there is no other activity
488 * that frees anything up. This is needed for things like routing
489 * etc, where we otherwise might have all activity going on in
490 * asynchronous contexts that cannot page things out.
492 * If there are applications that are active memory-allocators
493 * (most normal use), this basically shouldn't matter.
495 int kswapd(void *unused
)
498 struct task_struct
*tsk
= current
;
504 strcpy(tsk
->comm
, "kswapd");
505 sigfillset(&tsk
->blocked
);
508 * Tell the memory management that we're a "memory allocator",
509 * and that if we need more memory we should get access to it
510 * regardless (see "__alloc_pages()"). "kswapd" should
511 * never get caught in the normal page freeing logic.
513 * (Kswapd normally doesn't need memory anyway, but sometimes
514 * you need a small amount of memory in order to be able to
515 * page out something else, and this flag essentially protects
516 * us from recursively trying to free more memory as we're
517 * trying to free the first piece of memory in the first place).
519 tsk
->flags
|= PF_MEMALLOC
;
523 * If we actually get into a low-memory situation,
524 * the processes needing more memory will wake us
525 * up on a more timely basis.
529 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
530 int count
= SWAP_CLUSTER_MAX
;
531 zone
= pgdat
->node_zones
+ i
;
533 if (tsk
->need_resched
)
535 if ((!zone
->size
) || (!zone
->zone_wake_kswapd
))
537 do_try_to_free_pages(GFP_KSWAPD
, zone
);
538 } while (zone
->free_pages
< zone
->pages_low
&&
541 pgdat
= pgdat
->node_next
;
543 run_task_queue(&tq_disk
);
544 tsk
->state
= TASK_INTERRUPTIBLE
;
545 interruptible_sleep_on(&kswapd_wait
);
550 * Called by non-kswapd processes when they want more
553 * In a perfect world, this should just wake up kswapd
554 * and return. We don't actually want to swap stuff out
555 * from user processes, because the locking issues are
556 * nasty to the extreme (file write locks, and MM locking)
558 * One option might be to let kswapd do all the page-out
559 * and VM page table scanning that needs locking, and this
560 * process thread could do just the mmap shrink stage that
561 * can be done by just dropping cached pages without having
562 * any deadlock issues.
564 int try_to_free_pages(unsigned int gfp_mask
, zone_t
*zone
)
568 if (gfp_mask
& __GFP_WAIT
) {
569 current
->flags
|= PF_MEMALLOC
;
570 retval
= do_try_to_free_pages(gfp_mask
, zone
);
571 current
->flags
&= ~PF_MEMALLOC
;
576 static int __init
kswapd_init(void)
578 printk("Starting kswapd v1.6\n");
580 kernel_thread(kswapd
, NULL
, CLONE_FS
| CLONE_FILES
| CLONE_SIGHAND
);
584 module_init(kswapd_init
)