[Devel] [patch05/05]: Cotnainers(V2)- Over the limit memory handler
Rohit Seth
rohitseth at google.com
Tue Sep 19 19:23:31 PDT 2006
This patch implements the handler for cases when container(s) go over
their limit.
Signed-off-by: Rohit Seth <rohitseth at google.com>
mm/container_mm.c | 512 +++++++++++++++++++++++++++++++++++++++++++++++++++++++++
1 files changed, 512 insertions(+)
--- linux-2.6.18-rc6-mm2.org/mm/container_mm.c 1969-12-31 16:00:00.000000000 -0800
+++ linux-2.6.18-rc6-mm2.ctn/mm/container_mm.c 2006-09-19 18:23:11.000000000 -0700
@@ -0,0 +1,512 @@
+/*
+ * Copyright (c) 2006 Rohit Seth <rohitseth at google.com>
+ *
+ * Basic mm reclaimer for containers.
+ */
+
+#include <linux/mm.h>
+#include <linux/sched.h>
+#include <linux/container.h>
+#include <linux/rmap.h>
+#include <linux/mm_inline.h>
+#include <linux/writeback.h>
+#include <linux/swap.h>
+#include <linux/pagemap.h>
+#include <linux/pagevec.h>
+#include <linux/writeback.h>
+
+#if 0
+#define Cprintk(x...) do { printk(x); } while (0)
+#else
+#define Cprintk(x...)
+#endif
+
+/*
+ * Count of task's vmas scanned for active page in one shot.
+ */
+#define MAX_VMA_COUNT 5
+
+/*
+ * Max number of active pages that get deactivated in each iteration
+ * of page_limit hit.
+ */
+#define DEACTIVATE_COUNT 1000
+
+/*
+ * File size smaller than this number of pages is skipped in the
+ * first scan for inactivating pages. This is also used for scanning
+ * tasks with anon page count greater than this.
+ */
+#define PAGE_THRESHOLD_LIMIT 50
+
+/*
+ * Check if container is under its threshold for memory. This threshold is
+ * currently set at about 12% below the limit. Used to find out if we need to
+ * deactivate more pages.
+ */
+
+static int container_under_threshold(struct container_struct *ctn, long nr)
+{
+ int ret = 0;
+ long count = ctn->page_limit - (ctn->page_limit >> 3);
+
+ if ((atomic_long_read(&ctn->num_file_pages) +
+ atomic_long_read(&ctn->num_anon_pages) - nr) < count)
+ ret = 1;
+
+ return ret;
+}
+
+/*
+ * Internal function that builds a list of vmas belonging to mm whose pages
+ * are going to be scanned to be deactivated. We only look for anonymous vmas
+ * It returns the number of vmas populated in vmas array.
+ */
+static int get_task_vmas(struct mm_struct *mm, struct vm_area_struct **vmas,
+ struct vm_area_struct *start_vma)
+{
+ struct vm_area_struct *tmp;
+ int i = 0;
+
+
+ if (start_vma == NULL)
+ tmp = mm->mmap;
+ else
+ tmp = start_vma->vm_next;
+
+ while ((i < MAX_VMA_COUNT) && (tmp != NULL)) {
+ vmas[i] = tmp;
+ if (!(tmp->vm_flags & (VM_IO | VM_RESERVED))) {
+ if (tmp->anon_vma != NULL)
+ i++;
+ }
+ tmp = tmp->vm_next;
+ }
+ return i;
+}
+
+/*
+ * This function iterates over vma's pages. It reinitializes the container
+ * pointer for a page if it belongs to ctn. It decrements the anon_page and
+ * active_page count for the container ctn.
+ * We only check for anonymous pages as the file pages get their container
+ * inheritance from mapping.
+ */
+static long anonpages_init(struct container_struct *ctn,
+ struct vm_area_struct *vma)
+{
+ struct page *page;
+ struct zone *z;
+ unsigned long addr, end;
+ long ret = 0, active = 0;
+
+ addr = vma->vm_start;
+ end = vma->vm_end;
+
+ while (addr < end) {
+ page = follow_page(vma, addr, 0);
+ if ((page) && (page->ctn == ctn)) {
+ z = page_zone(page);
+ spin_lock_irq(&z->lru_lock);
+ if (page->ctn == ctn) {
+ if (PageAnon(page)) {
+ page->ctn = NULL;
+ atomic_long_dec(&ctn->num_anon_pages);
+ ret++;
+ if (PageActive(page) && PageLRU(page))
+ active++;
+ }
+ }
+ spin_unlock_irq(&z->lru_lock);
+ }
+ addr += PAGE_SIZE;
+ }
+ atomic_long_sub(active, &ctn->num_active_pages);
+ return ret;
+}
+
+/*
+ * This function is called when a task is explicitly requested to be moved out
+ * of its container through /configfs/containers/<container>/rmtask interface.
+ * We iterate over task's vmas and reinitialize the container pointer in
+ * every anon page mapping of this task.
+ */
+long anonpages_sub(struct task_struct *task, struct container_struct *ctn)
+{
+ struct vm_area_struct *vmas[MAX_VMA_COUNT];
+ struct vm_area_struct *start_vma = NULL;
+ struct mm_struct *mm;
+ long ret = 0;
+ int count, i;
+
+ mm = get_task_mm(task);
+ if (!mm)
+ return -1;
+ Cprintk("anonpages_sub scanning for task %d... ", task->pid);
+ /*
+ * No more new page faults beyond this point till the time
+ * we finish scanning.
+ */
+ down_write(&mm->mmap_sem);
+
+ /*
+ * Iterate over the vmas to deactivate pages.
+ */
+ while ((count = get_task_vmas(mm, vmas, start_vma)) > 0) {
+ i = 0;
+ while (i < count)
+ ret += anonpages_init(ctn, vmas[i++]);
+
+ if (count == MAX_VMA_COUNT)
+ /*
+ * Make get_task_vmas start from after the last
+ * scanned vma.
+ */
+ start_vma = vmas[MAX_VMA_COUNT-1];
+ else
+ break;
+ }
+ up_write(&mm->mmap_sem);
+ mmput(mm);
+ Cprintk("returning %ld \n", ret);
+ return ret;
+}
+
+/*
+ * Write back some of the pages belonging to this container. container
+ * mutex is already held.
+ */
+void writeback_container_file_pages(struct container_struct *ctn)
+{
+ struct address_space *mapping;
+ long flushed = 0;
+ long limit = (ctn->page_limit >> 1);
+ long invalidated = 0;
+ long temp;
+ int ret;
+ struct writeback_control wbc = {
+ .sync_mode = WB_SYNC_ALL,
+ .range_start = 0,
+ .range_end = LLONG_MAX,
+ };
+
+ list_for_each_entry(mapping, &ctn->mappings, ctn_mapping_list) {
+ if (mapping->nrpages > limit)
+ temp = limit;
+ else
+ temp = mapping->nrpages;
+ wbc.nr_to_write = temp;
+ ret = do_writepages(mapping, &wbc);
+ flushed += (temp - wbc.nr_to_write);
+ invalidated += invalidate_inode_pages(mapping);
+ if (flushed > DEACTIVATE_COUNT)
+ break;
+ }
+ Cprintk("Writeback %ld %ld\n", flushed, invalidated);
+}
+
+/*
+ * If the page is active then this function marks it inactive, clears the
+ * reference bit (so as to make it a better target for reclaim) and also
+ * puts it on inactive list. This function also decrements the zone's
+ * active number of pages.
+ */
+static int deactivate_container_page(struct page *page)
+{
+ struct zone *z;
+ int ret = 0, tmp;
+
+ if (!PageActive(page))
+ goto out;
+ if (!PageLRU(page))
+ goto out;
+
+ z = page_zone(page);
+ spin_lock_irq(&z->lru_lock);
+ if (PageActive(page) && PageLRU(page)) {
+ /*
+ * Not calling del_page_from_active_list so as to reduce
+ * the number of atomic dec. We do one atomic_long_sub
+ * after required number of pages have been deactivated.
+ */
+ list_del(&page->lru);
+ ClearPageActive(page);
+ z->nr_active--;
+ ret++;
+ add_page_to_inactive_list(z, page);
+ }
+ spin_unlock_irq(&z->lru_lock);
+out:
+ tmp = page_referenced(page, 0);
+ return ret;
+}
+
+#ifdef CONFIG_SWAP
+/*
+ * This function tries to inactive any (anonymous) active pages belonging to the vma.
+ */
+static int deactivate_pages(struct container_struct *ctn, struct vm_area_struct *vma)
+{
+ struct page *page;
+ unsigned long addr, end;
+ int deactivated = 0;
+
+ addr = vma->vm_start;
+ end = vma->vm_end;
+
+ while (addr < end) {
+ page = follow_page(vma, addr, 0);
+ if (page && (page->ctn == ctn) && PageAnon(page)) {
+ deactivated += deactivate_container_page(page);
+ if (deactivated >= DEACTIVATE_COUNT)
+ break; /* We have deactivated enough pages. */
+ }
+ addr += PAGE_SIZE;
+ }
+ atomic_long_sub(deactivated, &ctn->num_active_pages);
+ return deactivated;
+}
+
+/*
+ * This is the core of container anonymous page reclaimer. It gets called
+ * whenever page_limit of a container is hit and we are not able to
+ * deactivate enough file backed pages belonging to this container.
+ * This function traverses over the task list belonging to this container
+ * and collects vma for each task till the time it can deactivate at least
+ * DEACTIVATE_COUNT pages.
+ * We use the get_task_mm and mmput to avoid handling of freed up mm.
+ */
+static int sync_container_anon_pages(struct container_struct *ctn)
+{
+ struct vm_area_struct *vmas[MAX_VMA_COUNT];
+ struct vm_area_struct *start_vma;
+ struct task_struct *tsk;
+ struct mm_struct *mm;
+ int i, nr = 0, count;
+ int vma_scanned = 0;
+ int small_task = 0;
+
+again:
+ list_for_each_entry(tsk, &ctn->tasks, ctn_task_list) {
+ start_vma = NULL;
+
+ if (tsk->flags & PF_EXITING)
+ continue;
+
+ mm = get_task_mm(tsk);
+ if (!mm)
+ continue;
+ if (!small_task)
+ if (get_mm_counter(mm, anon_rss) < PAGE_THRESHOLD_LIMIT)
+ goto next_task;
+ down_read(&mm->mmap_sem);
+ /*
+ * Iterate over the anonymous vmas to deactivate
+ * pages.
+ */
+ while ((count = get_task_vmas(mm, vmas, start_vma)) > 0) {
+ i = 0;
+
+ vma_scanned += count;
+ while ((i < count) && (nr < DEACTIVATE_COUNT))
+ nr += deactivate_pages(ctn, vmas[i++]);
+
+ if (nr >= DEACTIVATE_COUNT)
+ break;
+
+ if (count == MAX_VMA_COUNT)
+ /*
+ * Make get_task_vmas start from after the last
+ * scanned vma.
+ */
+ start_vma = vmas[MAX_VMA_COUNT-1];
+ else
+ break;
+ }
+ up_read(&mm->mmap_sem);
+next_task:
+ mmput(mm);
+
+ /*
+ * Break out of the loop if we have deactivated enough pages.
+ * This should get the currently used pages to some what below
+ * the limit (currently set at about 12%).
+ */
+ if (container_under_threshold(ctn, nr))
+ goto out;
+ }
+ if (!small_task) {
+ small_task = 1;
+ goto again;
+ }
+out:
+ Cprintk("vma scanned %d Anon returning: %d\n", vma_scanned, nr);
+ return nr;
+
+}
+
+#endif /*CONFIG_SWAP */
+
+/*
+ * This function deactivates any of the active pages present in page
+ * array pl. count is the number of pages contained in the array.
+ */
+int deactivate_page_array(struct page **page_array, int count)
+{
+ int i = 0;
+ int ret = 0;
+ struct page *page;
+
+ while (i < count) {
+ page = page_array[i++];
+ ret += deactivate_container_page(page);
+ put_page(page);
+ }
+ return ret;
+}
+
+/*
+ * Find all the pages belonging to mapping and reinitialize
+ * its container pointer to point to ctn.
+ */
+long filepages_to_new_container(struct address_space *mapping,
+ struct container_struct *ctn)
+{
+ struct pagevec pvec;
+ long ret = 0;
+ int i;
+ pgoff_t next = 0;
+
+again:
+ pagevec_init(&pvec, 0);
+ while (pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
+ ret += pagevec_count(&pvec);
+ for (i=0; i< pagevec_count(&pvec); i++) {
+ struct page *page = pvec.pages[i];
+ pgoff_t page_index = page->index;
+
+ BUG_ON(page->ctn != ctn);
+ page->ctn = ctn;
+ if (page_index > next)
+ next = page_index;
+ next++;
+ }
+ }
+ if (ret < mapping->nrpages)
+ goto again;
+ return ret;
+}
+
+/*
+ * This is the core of container file page reclaimer. It gets called
+ * whenever page_limit of a container is hit.
+ * This function traverses over the list of files belonging to this container
+ * and collects pages for each file till the time it can deactivate at least
+ * DEACTIVATE_COUNT pages.
+ * It is a two pass algorithm. First it looks for files that have larger than
+ * PAGE_THRESHOLD_LIMIT pages mapped. If we are not able to deactivate
+ * DEACTIVATE_COUNT number of pages then we do a second pass. This second
+ * pass scans smaller files only.
+ */
+static int sync_container_file_pages(struct container_struct *ctn)
+{
+ struct address_space *map;
+ struct pagevec pvec;
+ pgoff_t next = 0;
+ int small_files = 0;
+ int i;
+ int ret = 0;
+ int files_scanned = 0;
+ unsigned long invalidated = 0;
+
+again:
+ list_for_each_entry(map, &ctn->mappings, ctn_mapping_list) {
+ if (map->nrpages == 0)
+ continue;
+ if (!small_files) { /* Look for big files */
+ if (map->nrpages < PAGE_THRESHOLD_LIMIT)
+ continue;
+ }
+ /*
+ * If the number of page cache pages itself is close to the
+ * limit then invalidate some of these pages here.
+ */
+ if (atomic_long_read(&ctn->num_file_pages) > (ctn->page_limit >> 1))
+ invalidated += invalidate_inode_pages(map);
+
+ pagevec_init(&pvec, 0);
+ /*
+ * In the second iteration, deactivate more pages from each
+ * file.
+ */
+ while ((ret < (DEACTIVATE_COUNT * (small_files +1))) &&
+ pagevec_lookup(&pvec, map, next, PAGEVEC_SIZE)) {
+ for (i=0; i< pagevec_count(&pvec); i++) {
+ struct page *page = pvec.pages[i];
+ pgoff_t page_index = page->index;
+
+ if (page_index > next)
+ next = page_index;
+ next++;
+ }
+ ret += deactivate_page_array(pvec.pages, pagevec_count(&pvec));
+ }
+ files_scanned++;
+ /*
+ * Break out of the loop if we have deactivated enough pages.
+ */
+ if (container_under_threshold(ctn, ret))
+ goto out;
+ }
+
+ if (!small_files) {
+ /* Iterate over smaller files also as we are not
+ * able to deactivate enough pages.
+ */
+ small_files = 1;
+ /*
+ * It will block all resource management part for this container. That is
+ * okay as we are not finding anything that we can easily deactivate.
+ */
+ goto again;
+ }
+out:
+ atomic_long_sub(ret, &ctn->num_active_pages);
+ Cprintk("File returning: %d, invalidated %ld\n", ret, invalidated);
+ return ret;
+}
+
+/*
+ * This function is called when container's page limit is hit. In this case
+ * we move the pages (charged against this container) from active list
+ * to head of inactive list. So that if there is any memory pressure
+ * then these are the prime candidates for freeing. It first tries to inactivate
+ * file pages and then goes into inactivating anonymous pages.
+ * XXX: More smarts should be put here to get to the right pages
+ * faster and not start iterating over the same range again.
+ * Container's mutex is already held. This is in context of kcontainerd.
+ */
+void container_over_pagelimit(struct container_struct *ctn)
+{
+ int ret =0;
+ long anon_pages, file_pages;
+ long tmp;
+ /*
+ * We have the mutex so container can not be freed and no new
+ * resource can be added while over the limit handler is working
+ * its way through.
+ */
+ file_pages = atomic_long_read(&ctn->num_file_pages);
+ anon_pages = atomic_long_read(&ctn->num_anon_pages);
+ tmp = file_pages + anon_pages;
+ if (tmp > ctn->page_limit) {
+ if (file_pages > (ctn->page_limit >> 1))
+ ret += sync_container_file_pages(ctn);
+#ifdef CONFIG_SWAP
+ anon_pages = atomic_long_read(&ctn->num_anon_pages);
+ if ((total_swap_pages > 0) &&
+ (anon_pages > (ctn->page_limit >> 6)))
+ ret += sync_container_anon_pages(ctn);
+#endif
+ }
+}
More information about the Devel
mailing list