Containers: Pagecache accounting and control subsystem (v1)

mm/pagecache_acct.c +3 -1

@@ -29,6 +29,7 @@
 #include <linux/uaccess.h>
 #include <asm/div64.h>
 #include <linux/pagecache_acct.h>
+#include <linux/memcontrol.h>
 
 /*
  * Convert unit from pages to kilobytes

@@ -337,12 +338,20 @@ int pagecache_acct_cont_overlimit(struct
 		return 0;
 }
 
-extern unsigned long shrink_all_pagecache_memory(unsigned long nr_pages);
+extern unsigned long shrink_container_memory(unsigned int memory_type,
+				unsigned long nr_pages, void *container);
 
 int pagecache_acct_shrink_used(unsigned long nr_pages)
 {
 	unsigned long ret = 0;
 	atomic_inc(&reclaim_count);

Vaidyanathan Srinivasan <[email protected]> writes in gmane.linux.kernel,gmane.linux.kernel.mm:

Quoted 18 lines

mm/pagecache_acct.c +3 -1

@@ -29,6 +29,7 @@
 #include <linux/uaccess.h>
 #include <asm/div64.h>
 #include <linux/pagecache_acct.h>
+#include <linux/memcontrol.h>
 
 /*
  * Convert unit from pages to kilobytes

@@ -337,12 +338,20 @@ int pagecache_acct_cont_overlimit(struct
 		return 0;
 }
 
-extern unsigned long shrink_all_pagecache_memory(unsigned long nr_pages);
+extern unsigned long shrink_container_memory(unsigned int memory_type,
+				unsigned long nr_pages, void *container);
 

Quoted 4 lines

mm/pagecache_acct.c +0 -0

 int pagecache_acct_shrink_used(unsigned long nr_pages)
 {
 	unsigned long ret = 0;
 	atomic_inc(&reclaim_count);

mm/pagecache_acct.c int pagecache_acct_cont_overlimit(struct · line 357

Quoted 10 lines

mm/pagecache_acct.c +7 -0

+
+	/* Don't call reclaim for each page above limit */
+	if (nr_pages > NR_PAGES_RECLAIM_THRESHOLD) {
+		ret += shrink_container_memory(
+				RECLAIM_PAGECACHE_MEMORY, nr_pages, NULL);
+	}
+
 	return 0;
 }
 

'ret' is not used ?

/ Kari Hurtta

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

The reclaim code is similar to RSS memory controller. Scan control is slightly different since we are targeting different type of pages.

Additionally no mapped pages are touched when scanning for pagecache pages.

RSS memory controller and pagecache controller share common code in reclaim and hence pagecache controller patches are dependent on RSS memory controller patch even though the features are independently configurable at compile time.

mm/vmscan.c +14 -1

@@ -43,6 +43,7 @@
 
 #include <linux/swapops.h>
 #include <linux/memcontrol.h>
+#include <linux/pagecache_acct.h>
 
 #include "internal.h"
 

@@ -70,6 +71,8 @@ struct scan_control {
 
        struct container *container;    /* Used by containers for reclaiming */
                                        /* pages when the limit is exceeded  */
+       int reclaim_pagecache_only;     /* Set when called from
+                                          pagecache controller */
 };
 
 /*

@@ -474,6 +477,15 @@ static unsigned long shrink_page_list(st
                        goto keep;
 
                VM_BUG_ON(PageActive(page));
+               /* Take it easy if we are doing only pagecache pages */
+               if (sc->reclaim_pagecache_only) {
+                       /* Check if this is a pagecache page they are not mapped */
+                       if (page_mapped(page))
+                               goto keep_locked;
+                       /* Check if this container has exceeded pagecache limit */
+                       if (!pagecache_acct_page_overlimit(page))
+                               goto keep_locked;
+               }
 
                sc->nr_scanned++;
 

@@ -522,7 +534,8 @@ static unsigned long shrink_page_list(st
                }
 
                if (PageDirty(page)) {
-                       if (referenced)
+                       /* Reclaim even referenced pagecache pages if over limit */
+                       if (!pagecache_acct_page_overlimit(page) && referenced)
                                goto keep_locked;
                        if (!may_enter_fs)
                                goto keep_locked;

Quoted 61 lines · Vaidyanathan Srinivasan

The reclaim code is similar to RSS memory controller. Scan control is slightly different since we are targeting different type of pages.

Additionally no mapped pages are touched when scanning for pagecache pages.

RSS memory controller and pagecache controller share common code in reclaim and hence pagecache controller patches are dependent on RSS memory controller patch even though the features are independently configurable at compile time.

mm/vmscan.c +17 -1

@@ -43,6 +43,7 @@
 
 #include <linux/swapops.h>
 #include <linux/memcontrol.h>
+#include <linux/pagecache_acct.h>
 
 #include "internal.h"
 

@@ -70,6 +71,8 @@ struct scan_control {
 
        struct container *container;    /* Used by containers for reclaiming */
                                        /* pages when the limit is exceeded  */
+       int reclaim_pagecache_only;     /* Set when called from
+                                          pagecache controller */
 };
 
 /*

@@ -474,6 +477,15 @@ static unsigned long shrink_page_list(st
                        goto keep;
 
                VM_BUG_ON(PageActive(page));
+               /* Take it easy if we are doing only pagecache pages */
+               if (sc->reclaim_pagecache_only) {
+                       /* Check if this is a pagecache page they are not mapped */
+                       if (page_mapped(page))
+                               goto keep_locked;
+                       /* Check if this container has exceeded pagecache limit */
+                       if (!pagecache_acct_page_overlimit(page))
+                               goto keep_locked;
+               }
 
                sc->nr_scanned++;
 

@@ -522,7 +534,8 @@ static unsigned long shrink_page_list(st
                }
 
                if (PageDirty(page)) {
-                       if (referenced)
+                       /* Reclaim even referenced pagecache pages if over limit */
+                       if (!pagecache_acct_page_overlimit(page) && referenced)
                                goto keep_locked;
                        if (!may_enter_fs)
                                goto keep_locked;

@@ -869,6 +882,13 @@ force_reclaim_mapped:
                cond_resched();
                page = lru_to_page(&l_hold);
                list_del(&page->lru);
+               /* While reclaiming pagecache make it easy */
+               if (sc->reclaim_pagecache_only) {
+                       if (page_mapped(page) || !pagecache_acct_page_overlimit(page)) {

mm/vmscan.c force_reclaim_mapped: · line 891

Quoted 4 lines

mm/vmscan.c +4 -0

+                               list_add(&page->lru, &l_active);
+                               continue;
+                       }
+               }

Please correct me if I'm wrong.
Here, if page type is mapped or not overlimit, why add it back to active list? Did shrink_page_list() is called by shrink_inactive_list()?

Correct, shrink_page_list() is called from shrink_inactive_list() but the above code is patched in shrink_active_list(). The 'force_reclaim_mapped' label is from function shrink_active_list() and not in shrink_page_list() as it may seem in the patch file.

While removing pages from active_list, we want to select only pagecache pages and leave the remaining in the active_list. page_mapped() pages are _not_ of interest to pagecache controller (they will be taken care by rss controller) and hence we put it back.
Also if the pagecache controller is below limit, no need to reclaim
so we put back all pages and come out.

Quoted 11 lines · Vaidyanathan Srinivasan

Correct, shrink_page_list() is called from shrink_inactive_list() but the above code is patched in shrink_active_list(). The 'force_reclaim_mapped' label is from function shrink_active_list() and not in shrink_page_list() as it may seem in the patch file.

While removing pages from active_list, we want to select only pagecache pages and leave the remaining in the active_list. page_mapped() pages are _not_ of interest to pagecache controller (they will be taken care by rss controller) and hence we put it back.
Also if the pagecache controller is below limit, no need to reclaim
so we put back all pages and come out.

Oh, I just read the patch, not apply it to my local tree, I'm working on 2.6.19 now.
So the question is, when vfs pagecache limit is hit, the current implementation just reclaim few pages, so it's quite possible the limit is hit again, and hence the reclaim code will be called again and again, that will impact application performance.

Quoted 13 lines · Aubrey Li

Quoted 11 lines · Vaidyanathan Srinivasan

Correct, shrink_page_list() is called from shrink_inactive_list() but the above code is patched in shrink_active_list(). The 'force_reclaim_mapped' label is from function shrink_active_list() and not in shrink_page_list() as it may seem in the patch file.

While removing pages from active_list, we want to select only pagecache pages and leave the remaining in the active_list. page_mapped() pages are _not_ of interest to pagecache controller (they will be taken care by rss controller) and hence we put it back.
Also if the pagecache controller is below limit, no need to reclaim
so we put back all pages and come out.

Quoted 6 lines

Oh, I just read the patch, not apply it to my local tree, I'm working on 2.6.19 now.
So the question is, when vfs pagecache limit is hit, the current implementation just reclaim few pages, so it's quite possible the limit is hit again, and hence the reclaim code will be called again and again, that will impact application performance.

Yes, you are correct. So if we start reclaiming one page at a time, then the cost of reclaim is very high and we would be calling the reclaim code too often. Hence we have a 'buffer zone' or 'reclaim threshold' or 'push back' around the limit. In the patch we have a 64 page (256KB) NR_PAGES_RECLAIM_THRESHOLD:

int pagecache_acct_shrink_used(unsigned long nr_pages)

 {
        unsigned long ret = 0;
        atomic_inc(&reclaim_count);

mm/pagecache_acct.c +7 -0

+
+       /* Don't call reclaim for each page above limit */
+       if (nr_pages > NR_PAGES_RECLAIM_THRESHOLD) {
+               ret += shrink_container_memory(
+                               RECLAIM_PAGECACHE_MEMORY, nr_pages, NULL);
+       }
+
        return 0;
 }

Hence we do not call the reclaimer if the threshold is exceeded by just 1 page... we wait for 64 pages or 256KB of pagecache memory to go
overlimit and then call the reclaimer which will reclaim all 64 pages
in one shot.

This prevents the reclaim code from being called too often and it also keeps the cost of reclaim low.

In future patches we are planing to have a percentage based reclaim threshold so that it would scale well with the container size.

Quoted 45 lines · Vaidyanathan Srinivasan

Quoted 12 lines · Aubrey Li

Quoted 11 lines · Vaidyanathan Srinivasan

Correct, shrink_page_list() is called from shrink_inactive_list() but the above code is patched in shrink_active_list(). The 'force_reclaim_mapped' label is from function shrink_active_list() and not in shrink_page_list() as it may seem in the patch file.

While removing pages from active_list, we want to select only pagecache pages and leave the remaining in the active_list. page_mapped() pages are _not_ of interest to pagecache controller (they will be taken care by rss controller) and hence we put it back.
Also if the pagecache controller is below limit, no need to reclaim
so we put back all pages and come out.

Quoted 6 lines

Oh, I just read the patch, not apply it to my local tree, I'm working on 2.6.19 now.
So the question is, when vfs pagecache limit is hit, the current implementation just reclaim few pages, so it's quite possible the limit is hit again, and hence the reclaim code will be called again and again, that will impact application performance.

Yes, you are correct. So if we start reclaiming one page at a time, then the cost of reclaim is very high and we would be calling the reclaim code too often. Hence we have a 'buffer zone' or 'reclaim threshold' or 'push back' around the limit. In the patch we have a 64 page (256KB) NR_PAGES_RECLAIM_THRESHOLD:

int pagecache_acct_shrink_used(unsigned long nr_pages)

 {
        unsigned long ret = 0;
        atomic_inc(&reclaim_count);

mm/pagecache_acct.c +7 -0

+
+       /* Don't call reclaim for each page above limit */
+       if (nr_pages > NR_PAGES_RECLAIM_THRESHOLD) {
+               ret += shrink_container_memory(
+                               RECLAIM_PAGECACHE_MEMORY, nr_pages, NULL);
+       }
+
        return 0;
 }

Hence we do not call the reclaimer if the threshold is exceeded by just 1 page... we wait for 64 pages or 256KB of pagecache memory to go
overlimit and then call the reclaimer which will reclaim all 64 pages
in one shot.

Quoted 6 lines

This prevents the reclaim code from being called too often and it also keeps the cost of reclaim low.

In future patches we are planing to have a percentage based reclaim threshold so that it would scale well with the container size.

Actually it's not a good idea IMHO. No matter how big the threshold is, it's not suitable. If it's too small, application performance will be impacted seriously after pagecache limit is hit. If it's too large, Limiting pagecache is useless.

Why not reclaim pages as much as possible when the pagecache limit is hit?