vm_pageout.c (c98bb15d01876b54224f52232cee5ba1b5b0ce6f) vm_pageout.c (44f1c916109d4d88941d257b7c4c96c26ab55477)
1/*-
2 * Copyright (c) 1991 Regents of the University of California.
3 * All rights reserved.
4 * Copyright (c) 1994 John S. Dyson
5 * All rights reserved.
6 * Copyright (c) 1994 David Greenman
7 * All rights reserved.
8 * Copyright (c) 2005 Yahoo! Technologies Norway AS

--- 664 unchanged lines hidden (view full) ---

673 }
674
675 /*
676 * Make the next scan start on the next domain.
677 */
678 initial_dom = atomic_fetchadd_int(&start_dom, 1) % vm_ndomains;
679
680 inactl = 0;
1/*-
2 * Copyright (c) 1991 Regents of the University of California.
3 * All rights reserved.
4 * Copyright (c) 1994 John S. Dyson
5 * All rights reserved.
6 * Copyright (c) 1994 David Greenman
7 * All rights reserved.
8 * Copyright (c) 2005 Yahoo! Technologies Norway AS

--- 664 unchanged lines hidden (view full) ---

673 }
674
675 /*
676 * Make the next scan start on the next domain.
677 */
678 initial_dom = atomic_fetchadd_int(&start_dom, 1) % vm_ndomains;
679
680 inactl = 0;
681 inactmax = cnt.v_inactive_count;
681 inactmax = vm_cnt.v_inactive_count;
682 actl = 0;
682 actl = 0;
683 actmax = tries < 2 ? 0 : cnt.v_active_count;
683 actmax = tries < 2 ? 0 : vm_cnt.v_active_count;
684 dom = initial_dom;
685
686 /*
687 * Scan domains in round-robin order, first inactive queues,
688 * then active. Since domain usually owns large physically
689 * contiguous chunk of memory, it makes sense to completely
690 * exhaust one domain before switching to next, while growing
691 * the pool of contiguous physical pages.

--- 613 unchanged lines hidden (view full) ---

1305 TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_marker, plinks.q);
1306 }
1307 vm_pagequeue_unlock(pq);
1308
1309 /*
1310 * Compute the number of pages we want to try to move from the
1311 * active queue to the inactive queue.
1312 */
684 dom = initial_dom;
685
686 /*
687 * Scan domains in round-robin order, first inactive queues,
688 * then active. Since domain usually owns large physically
689 * contiguous chunk of memory, it makes sense to completely
690 * exhaust one domain before switching to next, while growing
691 * the pool of contiguous physical pages.

--- 613 unchanged lines hidden (view full) ---

1305 TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_marker, plinks.q);
1306 }
1307 vm_pagequeue_unlock(pq);
1308
1309 /*
1310 * Compute the number of pages we want to try to move from the
1311 * active queue to the inactive queue.
1312 */
1313 page_shortage = cnt.v_inactive_target - cnt.v_inactive_count +
1313 page_shortage = vm_cnt.v_inactive_target - vm_cnt.v_inactive_count +
1314 vm_paging_target() + deficit + addl_page_shortage;
1315
1316 pq = &vmd->vmd_pagequeues[PQ_ACTIVE];
1317 vm_pagequeue_lock(pq);
1318 maxscan = pq->pq_cnt;
1319
1320 /*
1321 * If we're just idle polling attempt to visit every

--- 249 unchanged lines hidden (view full) ---

1571 } else
1572 PROC_UNLOCK(p);
1573 }
1574 sx_sunlock(&allproc_lock);
1575 if (bigproc != NULL) {
1576 killproc(bigproc, "out of swap space");
1577 sched_nice(bigproc, PRIO_MIN);
1578 PROC_UNLOCK(bigproc);
1314 vm_paging_target() + deficit + addl_page_shortage;
1315
1316 pq = &vmd->vmd_pagequeues[PQ_ACTIVE];
1317 vm_pagequeue_lock(pq);
1318 maxscan = pq->pq_cnt;
1319
1320 /*
1321 * If we're just idle polling attempt to visit every

--- 249 unchanged lines hidden (view full) ---

1571 } else
1572 PROC_UNLOCK(p);
1573 }
1574 sx_sunlock(&allproc_lock);
1575 if (bigproc != NULL) {
1576 killproc(bigproc, "out of swap space");
1577 sched_nice(bigproc, PRIO_MIN);
1578 PROC_UNLOCK(bigproc);
1579 wakeup(&cnt.v_free_count);
1579 wakeup(&vm_cnt.v_free_count);
1580 }
1581}
1582
1583static void
1584vm_pageout_worker(void *arg)
1585{
1586 struct vm_domain *domain;
1587 int domidx;

--- 19 unchanged lines hidden (view full) ---

1607 * not clear vm_pages_needed until we reach our target,
1608 * otherwise we may be woken up over and over again and
1609 * waste a lot of cpu.
1610 */
1611 mtx_lock(&vm_page_queue_free_mtx);
1612 if (vm_pages_needed && !vm_page_count_min()) {
1613 if (!vm_paging_needed())
1614 vm_pages_needed = 0;
1580 }
1581}
1582
1583static void
1584vm_pageout_worker(void *arg)
1585{
1586 struct vm_domain *domain;
1587 int domidx;

--- 19 unchanged lines hidden (view full) ---

1607 * not clear vm_pages_needed until we reach our target,
1608 * otherwise we may be woken up over and over again and
1609 * waste a lot of cpu.
1610 */
1611 mtx_lock(&vm_page_queue_free_mtx);
1612 if (vm_pages_needed && !vm_page_count_min()) {
1613 if (!vm_paging_needed())
1614 vm_pages_needed = 0;
1615 wakeup(&cnt.v_free_count);
1615 wakeup(&vm_cnt.v_free_count);
1616 }
1617 if (vm_pages_needed) {
1618 /*
1619 * Still not done, take a second pass without waiting
1620 * (unlimited dirty cleaning), otherwise sleep a bit
1621 * and try again.
1622 */
1623 if (domain->vmd_pass > 1)

--- 6 unchanged lines hidden (view full) ---

1630 * stats.
1631 */
1632 domain->vmd_pass = 0;
1633 msleep(&vm_pages_needed, &vm_page_queue_free_mtx,
1634 PVM, "psleep", hz);
1635
1636 }
1637 if (vm_pages_needed) {
1616 }
1617 if (vm_pages_needed) {
1618 /*
1619 * Still not done, take a second pass without waiting
1620 * (unlimited dirty cleaning), otherwise sleep a bit
1621 * and try again.
1622 */
1623 if (domain->vmd_pass > 1)

--- 6 unchanged lines hidden (view full) ---

1630 * stats.
1631 */
1632 domain->vmd_pass = 0;
1633 msleep(&vm_pages_needed, &vm_page_queue_free_mtx,
1634 PVM, "psleep", hz);
1635
1636 }
1637 if (vm_pages_needed) {
1638 cnt.v_pdwakeups++;
1638 vm_cnt.v_pdwakeups++;
1639 domain->vmd_pass++;
1640 }
1641 mtx_unlock(&vm_page_queue_free_mtx);
1642 vm_pageout_scan(domain, domain->vmd_pass);
1643 }
1644}
1645
1646/*

--- 4 unchanged lines hidden (view full) ---

1651{
1652#if MAXMEMDOM > 1
1653 int error, i;
1654#endif
1655
1656 /*
1657 * Initialize some paging parameters.
1658 */
1639 domain->vmd_pass++;
1640 }
1641 mtx_unlock(&vm_page_queue_free_mtx);
1642 vm_pageout_scan(domain, domain->vmd_pass);
1643 }
1644}
1645
1646/*

--- 4 unchanged lines hidden (view full) ---

1651{
1652#if MAXMEMDOM > 1
1653 int error, i;
1654#endif
1655
1656 /*
1657 * Initialize some paging parameters.
1658 */
1659 cnt.v_interrupt_free_min = 2;
1660 if (cnt.v_page_count < 2000)
1659 vm_cnt.v_interrupt_free_min = 2;
1660 if (vm_cnt.v_page_count < 2000)
1661 vm_pageout_page_count = 8;
1662
1663 /*
1664 * v_free_reserved needs to include enough for the largest
1665 * swap pager structures plus enough for any pv_entry structs
1666 * when paging.
1667 */
1661 vm_pageout_page_count = 8;
1662
1663 /*
1664 * v_free_reserved needs to include enough for the largest
1665 * swap pager structures plus enough for any pv_entry structs
1666 * when paging.
1667 */
1668 if (cnt.v_page_count > 1024)
1669 cnt.v_free_min = 4 + (cnt.v_page_count - 1024) / 200;
1668 if (vm_cnt.v_page_count > 1024)
1669 vm_cnt.v_free_min = 4 + (vm_cnt.v_page_count - 1024) / 200;
1670 else
1670 else
1671 cnt.v_free_min = 4;
1672 cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE +
1673 cnt.v_interrupt_free_min;
1674 cnt.v_free_reserved = vm_pageout_page_count +
1675 cnt.v_pageout_free_min + (cnt.v_page_count / 768);
1676 cnt.v_free_severe = cnt.v_free_min / 2;
1677 cnt.v_free_target = 4 * cnt.v_free_min + cnt.v_free_reserved;
1678 cnt.v_free_min += cnt.v_free_reserved;
1679 cnt.v_free_severe += cnt.v_free_reserved;
1680 cnt.v_inactive_target = (3 * cnt.v_free_target) / 2;
1681 if (cnt.v_inactive_target > cnt.v_free_count / 3)
1682 cnt.v_inactive_target = cnt.v_free_count / 3;
1671 vm_cnt.v_free_min = 4;
1672 vm_cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE +
1673 vm_cnt.v_interrupt_free_min;
1674 vm_cnt.v_free_reserved = vm_pageout_page_count +
1675 vm_cnt.v_pageout_free_min + (vm_cnt.v_page_count / 768);
1676 vm_cnt.v_free_severe = vm_cnt.v_free_min / 2;
1677 vm_cnt.v_free_target = 4 * vm_cnt.v_free_min + vm_cnt.v_free_reserved;
1678 vm_cnt.v_free_min += vm_cnt.v_free_reserved;
1679 vm_cnt.v_free_severe += vm_cnt.v_free_reserved;
1680 vm_cnt.v_inactive_target = (3 * vm_cnt.v_free_target) / 2;
1681 if (vm_cnt.v_inactive_target > vm_cnt.v_free_count / 3)
1682 vm_cnt.v_inactive_target = vm_cnt.v_free_count / 3;
1683
1684 /*
1685 * Set the default wakeup threshold to be 10% above the minimum
1686 * page limit. This keeps the steady state out of shortfall.
1687 */
1683
1684 /*
1685 * Set the default wakeup threshold to be 10% above the minimum
1686 * page limit. This keeps the steady state out of shortfall.
1687 */
1688 vm_pageout_wakeup_thresh = (cnt.v_free_min / 10) * 11;
1688 vm_pageout_wakeup_thresh = (vm_cnt.v_free_min / 10) * 11;
1689
1690 /*
1691 * Set interval in seconds for active scan. We want to visit each
1692 * page at least once every ten minutes. This is to prevent worst
1693 * case paging behaviors with stale active LRU.
1694 */
1695 if (vm_pageout_update_period == 0)
1696 vm_pageout_update_period = 600;
1697
1698 /* XXX does not really belong here */
1699 if (vm_page_max_wired == 0)
1689
1690 /*
1691 * Set interval in seconds for active scan. We want to visit each
1692 * page at least once every ten minutes. This is to prevent worst
1693 * case paging behaviors with stale active LRU.
1694 */
1695 if (vm_pageout_update_period == 0)
1696 vm_pageout_update_period = 600;
1697
1698 /* XXX does not really belong here */
1699 if (vm_page_max_wired == 0)
1700 vm_page_max_wired = cnt.v_free_count / 3;
1700 vm_page_max_wired = vm_cnt.v_free_count / 3;
1701
1702 swap_pager_swap_init();
1703#if MAXMEMDOM > 1
1704 for (i = 1; i < vm_ndomains; i++) {
1705 error = kthread_add(vm_pageout_worker, (void *)(uintptr_t)i,
1706 curproc, NULL, 0, 0, "dom%d", i);
1707 if (error != 0) {
1708 panic("starting pageout for domain %d, error %d\n",
1709 i, error);
1710 }
1711 }
1712#endif
1713 vm_pageout_worker((void *)(uintptr_t)0);
1714}
1715
1716/*
1717 * Unless the free page queue lock is held by the caller, this function
1718 * should be regarded as advisory. Specifically, the caller should
1701
1702 swap_pager_swap_init();
1703#if MAXMEMDOM > 1
1704 for (i = 1; i < vm_ndomains; i++) {
1705 error = kthread_add(vm_pageout_worker, (void *)(uintptr_t)i,
1706 curproc, NULL, 0, 0, "dom%d", i);
1707 if (error != 0) {
1708 panic("starting pageout for domain %d, error %d\n",
1709 i, error);
1710 }
1711 }
1712#endif
1713 vm_pageout_worker((void *)(uintptr_t)0);
1714}
1715
1716/*
1717 * Unless the free page queue lock is held by the caller, this function
1718 * should be regarded as advisory. Specifically, the caller should
1719 * not msleep() on &cnt.v_free_count following this function unless
1719 * not msleep() on &vm_cnt.v_free_count following this function unless
1720 * the free page queue lock is held until the msleep() is performed.
1721 */
1722void
1723pagedaemon_wakeup(void)
1724{
1725
1726 if (!vm_pages_needed && curthread->td_proc != pageproc) {
1727 vm_pages_needed = 1;

--- 149 unchanged lines hidden --- 		1720 * the free page queue lock is held until the msleep() is performed.
1721 */
1722void
1723pagedaemon_wakeup(void)
1724{
1725
1726 if (!vm_pages_needed && curthread->td_proc != pageproc) {
1727 vm_pages_needed = 1;

--- 149 unchanged lines hidden ---