Skip to contentSkip to conversation

first boot allocator to not scan hole on every allocation]

PATCH Applied Jes Sorensen 3 participants 4 messages Dec 11, 2006 – Dec 14, 2006

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 first boot allocator to not scan hole on every allocation]

Hi,

Alex suggested I send this one straight to xen-devel given it's generic code.

I would claim it's harmless to architectures which have memory starting 0x00 and it's absolutely vital for those architectures that don't.

Cheers,
Jes

Hi,

Ok, I finally figured out why the boot allocator took .0172 seconds for each single page allocation on a small Altix (a partitioned system would be even worse). I posted a patch for this earlier, but it was wrong (by a factor of 0x40 :-( ).

This patch goes into common/page_alloc.c but it should be obviously correct.

It's really vital for the SN2 port to get this one in.

Cheers,
Jes

Keep track of which page is the first available to the boot allocator to avoid scanning the hole "0 ... first_pg" on every allocation, on platforms that do not have physical memory starting at address 0x00.

This reduces the Xen boot time by roughly 150 seconds on a small Altix with 6GB of RAM.

xen/common/page_alloc.c +12 -3
diff -r 66609699c9d0 xen/common/page_alloc.c
@@ -137,6 +137,8 @@ static void map_alloc(unsigned long firs
 }
 
 
+static unsigned long first_pg = -1UL;
+
 static void map_free(unsigned long first_page, unsigned long nr_pages)
 {
     unsigned long start_off, end_off, curr_idx, end_idx;
@@ -152,6 +154,13 @@ static void map_free(unsigned long first
     start_off = first_page & (PAGES_PER_MAPWORD-1);
     end_idx   = (first_page + nr_pages) / PAGES_PER_MAPWORD;
     end_off   = (first_page + nr_pages) & (PAGES_PER_MAPWORD-1);
+
+    /*
+     * Set first_pg to the first entry initialized so the boot
+     * allocator can avoid scanning the hole from 0 for no reason
+     */
+    if (first_page < first_pg)
+        first_pg = first_page;
 
     if ( curr_idx == end_idx )
     {
@@ -258,7 +267,7 @@ unsigned long alloc_boot_pages(unsigned 
 {
     unsigned long pg, i = 0;
 
-    for ( pg = 0; (pg + nr_pfns) < max_page; pg += pfn_align )
+    for ( pg = first_pg; (pg + nr_pfns) < max_page; pg += pfn_align )
     {
         i = alloc_boot_pages_at(nr_pfns, pg);
         if (i != 0)
@@ -301,7 +310,7 @@ void end_boot_allocator(void)
                 INIT_LIST_HEAD(&heap[i][j][k]);
 
     /* Pages that are free now go to the domain sub-allocator. */
-    for ( i = 0; i < max_page; i++ )
+    for ( i = first_pg; i < max_page; i++ )
     {
         curr_free = next_free;
         next_free = !allocated_in_map(i+1);
@@ -482,7 +491,7 @@ void scrub_heap_pages(void)
 
     printk("Scrubbing Free RAM: ");
 
-    for ( pfn = 0; pfn < max_page; pfn++ )
+    for ( pfn = first_pg; pfn < max_page; pfn++ )
     {
         /* Every 100MB, print a progress dot. */
         if ( (pfn % ((100*1024*1024)/PAGE_SIZE)) == 0 )

_______________________________________________
Xen-devel mailing list
[email protected]
http://lists.xensource.com/xen-devel

#1 lore raw

Hi Keir,

Do you have any objections to this patch? If not, I'd like to add it
to the xen-ia64-unstable.hg tree and let it get pulled into xen-unstable after 3.0.4. Let us know, thanks,

Alex

Quoted 103 lines · Jes Sorensen

Hi,

Alex suggested I send this one straight to xen-devel given it's generic code.

I would claim it's harmless to architectures which have memory starting 0x00 and it's absolutely vital for those architectures that don't.

Cheers,
Jes
email message attachment ([patch] first boot allocator to not scan hole on every allocation)

Quoted 78 lines

-------- Forwarded Message --------
From: Jes Sorensen <[email protected]>
To: [email protected]
Cc: Alex Williamson <[email protected]>
Subject: [patch] first boot allocator to not scan hole on every allocation
Date: Mon, 11 Dec 2006 16:01:35 +0100

Hi,

Ok, I finally figured out why the boot allocator took .0172 seconds for each single page allocation on a small Altix (a partitioned system would be even worse). I posted a patch for this earlier, but it was wrong (by a factor of 0x40 :-( ).

This patch goes into common/page_alloc.c but it should be obviously correct.

It's really vital for the SN2 port to get this one in.

Cheers,
Jes

plain text document attachment (xen-boot-alloc-skip-hole.diff) Keep track of which page is the first available to the boot allocator to avoid scanning the hole "0 ... first_pg" on every allocation, on platforms that do not have physical memory starting at address 0x00.

This reduces the Xen boot time by roughly 150 seconds on a small Altix with 6GB of RAM.

diff -r 66609699c9d0 xen/common/page_alloc.c

xen/common/page_alloc.c +11 -2
@@ -137,6 +137,8 @@ static void map_alloc(unsigned long firs
 }
 
 
+static unsigned long first_pg = -1UL;
+
 static void map_free(unsigned long first_page, unsigned long nr_pages)
 {
     unsigned long start_off, end_off, curr_idx, end_idx;
@@ -152,6 +154,13 @@ static void map_free(unsigned long first
     start_off = first_page & (PAGES_PER_MAPWORD-1);
     end_idx   = (first_page + nr_pages) / PAGES_PER_MAPWORD;
     end_off   = (first_page + nr_pages) & (PAGES_PER_MAPWORD-1);
+
+    /*
+     * Set first_pg to the first entry initialized so the boot
+     * allocator can avoid scanning the hole from 0 for no reason
+     */
+    if (first_page < first_pg)
+        first_pg = first_page;
 
     if ( curr_idx == end_idx )
     {
@@ -258,7 +267,7 @@ unsigned long alloc_boot_pages(unsigned 
 {
     unsigned long pg, i = 0;
 
-    for ( pg = 0; (pg + nr_pfns) < max_page; pg += pfn_align )
+    for ( pg = first_pg; (pg + nr_pfns) < max_page; pg += pfn_align )
     {
         i = alloc_boot_pages_at(nr_pfns, pg);
         if (i != 0)
@@ -301,7 +310,7 @@ void end_boot_allocator(void)
                 INIT_LIST_HEAD(&heap[i][j][k]);
 
     /* Pages that are free now go to the domain sub-allocator. */
-    for ( i = 0; i < max_page; i++ )
+    for ( i = first_pg; i < max_page; i++ )
     {
         curr_free = next_free;
         next_free = !allocated_in_map(i+1);

xen/common/page_alloc.c void scrub_heap_pages(void) · line 497

Quoted 8 lines
xen/common/page_alloc.c +1 -1
@@ -482,7 +491,7 @@ void scrub_heap_pages(void)
 
     printk("Scrubbing Free RAM: ");
 
-    for ( pfn = 0; pfn < max_page; pfn++ )
+    for ( pfn = first_pg; pfn < max_page; pfn++ )
     {
         /* Every 100MB, print a progress dot. */
         if ( (pfn % ((100*1024*1024)/PAGE_SIZE)) == 0 )

_______________________________________________
Xen-devel mailing list
[email protected]
http://lists.xensource.com/xen-devel

Signature · 1 line
Alex Williamson                             HP Open Source & Linux Org.
Applied
#2 lore raw
Quoted 3 lines · Alex Williamson

Do you have any objections to this patch? If not, I'd like to add it
to the xen-ia64-unstable.hg tree and let it get pulled into xen-unstable after 3.0.4. Let us know, thanks,

I applied it to xen-unstable today.

-- Keir

#3 lore raw
Quoted 9 lines · Keir Fraser
Quoted 3 lines · Alex Williamson

Do you have any objections to this patch? If not, I'd like to add it
to the xen-ia64-unstable.hg tree and let it get pulled into xen-unstable after 3.0.4. Let us know, thanks,

I applied it to xen-unstable today.

Even better, thanks,

Alex

Signature · 1 line
Alex Williamson                             HP Open Source & Linux Org.