Hi,
Basically this new swap_writepage function looks for dirty swapcache pages
which may be contiguous (reverse and forward searching wrt to the physical
address of the page being passed to swap_writepage) and builds a page list
which is written "at once".
The patch is against test13pre3.
Comments are welcome. (especially about the __find_page_nolock
modification)
diff -Nur --exclude-from=exclude linux.orig/include/linux/mm.h linux/include/linux/mm.h
--- linux.orig/include/linux/mm.h Thu Dec 21 04:34:19 2000
+++ linux/include/linux/mm.h Thu Dec 21 03:56:30 2000
@@ -451,6 +451,8 @@
extern int filemap_swapout(struct page *, struct file *);
extern int filemap_sync(struct vm_area_struct *, unsigned long, size_t, unsigned int);
extern struct page *filemap_nopage(struct vm_area_struct *, unsigned long, int);
+extern inline struct page * ___find_page_nolock(struct address_space *, unsigned long, struct page *);
+
/*
* GFP bitmasks..
diff -Nur --exclude-from=exclude linux.orig/include/linux/swap.h linux/include/linux/swap.h
--- linux.orig/include/linux/swap.h Thu Dec 21 04:34:19 2000
+++ linux/include/linux/swap.h Thu Dec 21 04:16:42 2000
@@ -166,6 +166,8 @@
extern unsigned long swap_cache_find_total;
extern unsigned long swap_cache_find_success;
#endif
+
+extern struct swap_info_struct swap_info[MAX_SWAPFILES];
/*
* Work out if there are any other processes sharing this page, ignoring
diff -Nur --exclude-from=exclude linux.orig/mm/filemap.c linux/mm/filemap.c
--- linux.orig/mm/filemap.c Thu Dec 21 04:34:20 2000
+++ linux/mm/filemap.c Thu Dec 21 03:53:41 2000
@@ -242,7 +242,7 @@
spin_unlock(&pagecache_lock);
}
-static inline struct page * __find_page_nolock(struct address_space *mapping, unsigned long offset, struct page *page)
+inline struct page * ___find_page_nolock(struct address_space *mapping, unsigned long offset, struct page *page)
{
goto inside;
@@ -250,12 +250,22 @@
page = page->next_hash;
inside:
if (!page)
- goto not_found;
+ return NULL;
if (page->mapping != mapping)
continue;
if (page->index == offset)
break;
}
+ return page;
+}
+
+static inline struct page * __find_page_nolock(struct address_space *mapping, unsigned long offset, struct page *page)
+{
+ page = ___find_page_nolock(mapping, offset, page);
+
+ if(!page)
+ return NULL;
+
/*
* Touching the page may move it to the active list.
* If we end up with too few inactive pages, we wake
@@ -264,7 +274,7 @@
age_page_up(page);
if (inactive_shortage() > inactive_target / 2 && free_shortage())
wakeup_kswapd(0);
-not_found:
+
return page;
}
diff -Nur --exclude-from=exclude linux.orig/mm/swap_state.c linux/mm/swap_state.c
--- linux.orig/mm/swap_state.c Mon Dec 4 19:22:02 2000
+++ linux/mm/swap_state.c Thu Dec 21 04:23:47 2000
@@ -5,6 +5,8 @@
* Swap reorganised 29.12.95, Stephen Tweedie
*
* Rewritten to use page cache, (C) 1998 Stephen Tweedie
+ *
+ * 21/12/2000 Added swap write clustering. Marcelo Tosatti
*/
#include <linux/mm.h>
@@ -17,9 +19,95 @@
#include <asm/pgtable.h>
+static inline struct page * swap_page_dirty(unsigned long, unsigned long, struct swap_info_struct *);
+
+#define SWAP_WRITE_CLUSTER (1 << page_cluster)
+
static int swap_writepage(struct page *page)
{
- rw_swap_page(WRITE, page, 0);
+ unsigned long page_offset, curr, offset, i, type;
+ struct swap_info_struct *swap;
+ swp_entry_t entry;
+ struct page *cpages[SWAP_WRITE_CLUSTER*2];
+ int count, first;
+
+ entry.val = page->index;
+
+ type = SWP_TYPE(entry);
+
+ swap = &swap_info[type];
+
+ /* If swap area is not a real device, do not try to write cluster. */
+ if(!swap->swap_device) {
+ rw_swap_page(WRITE, page, 0);
+ return 0;
+ }
+
+ page_offset = offset = SWP_OFFSET(entry);
+ cpages[SWAP_WRITE_CLUSTER] = page;
+ count = 1;
+ first = SWAP_WRITE_CLUSTER;
+ curr = 1;
+
+ spin_lock(&pagecache_lock);
+ swap_device_lock(swap);
+
+ /*
+ * Search for clusterable dirty swap pages.
+ */
+
+ while (count < SWAP_WRITE_CLUSTER) {
+ struct page *p = NULL;
+
+ if(offset <= 0)
+ break;
+
+ offset = page_offset - curr;
+ p = swap_page_dirty(offset, type, swap);
+
+ if(!p)
+ break;
+
+ cpages[--first] = p;
+
+ ClearPageDirty(p);
+ curr++;
+ count++;
+ }
+
+ curr = 1;
+
+ while (count < SWAP_WRITE_CLUSTER) {
+ struct page *p = NULL;
+ offset = page_offset + curr;
+
+ if(offset >= swap->max)
+ break;
+
+ p = swap_page_dirty(offset, type, swap);
+ if(!p)
+ break;
+
+ cpages[first + count] = p;
+ ClearPageDirty(p);
+ curr++;
+ count++;
+ }
+
+ swap_device_unlock(swap);
+ spin_unlock(&pagecache_lock);
+
+ /* Now write out everything we were able to cluster */
+
+ for(i=0; i<count; i++) {
+ int pos = first + i;
+ rw_swap_page(WRITE, cpages[pos], 0);
+
+ /* If the page is a clustered one, decrement its counter */
+ if(pos != SWAP_WRITE_CLUSTER)
+ page_cache_release(cpages[pos]);
+ }
+
return 0;
}
@@ -36,6 +124,43 @@
0, /* nrpages */
&swap_aops,
};
+
+static inline struct page * swap_page_dirty(unsigned long offset, unsigned long type,
+ struct swap_info_struct *swapdev)
+{
+ struct page *page, **hash;
+ swp_entry_t entry;
+ unsigned long index;
+
+ entry = SWP_ENTRY(type, offset);
+
+ index = entry.val;
+
+ if(!swapdev->swap_map[offset])
+ return NULL;
+
+ if(swapdev->swap_map[offset] == SWAP_MAP_BAD)
+ return NULL;
+
+ hash = page_hash(&swapper_space, index);
+
+ page = ___find_page_nolock(&swapper_space, index, *hash);
+
+ if(!page)
+ return NULL;
+
+ if(TryLockPage(page))
+ return NULL;
+
+ page_cache_get(page);
+
+ if(PageDirty(page) && PageSwapCache(page))
+ return page;
+
+ UnlockPage(page);
+ page_cache_release(page);
+ return NULL;
+}
#ifdef SWAP_CACHE_INFO
unsigned long swap_cache_add_total;
Marcelo Tosatti wrote:
>
> Hi,
>
> Basically this new swap_writepage function looks for dirty swapcache pages
> which may be contiguous (reverse and forward searching wrt to the physical
> address of the page being passed to swap_writepage) and builds a page list
> which is written "at once".
>
> The patch is against test13pre3.
>
> Comments are welcome. (especially about the __find_page_nolock
> modification)
>
Have you any benchmarks for this?
On Thu, 21 Dec 2000, Andrew Morton wrote:
> Marcelo Tosatti wrote:
> >
> > Hi,
> >
> > Basically this new swap_writepage function looks for dirty swapcache pages
> > which may be contiguous (reverse and forward searching wrt to the physical
> > address of the page being passed to swap_writepage) and builds a page list
> > which is written "at once".
> >
> > The patch is against test13pre3.
> >
> > Comments are welcome. (especially about the __find_page_nolock
> > modification)
> >
>
> Have you any benchmarks for this?
Not yet.
Under some stress tests on a 16mb machine, around 30% of the clustered
swapouts were reaching the limit of pages, which was 16.
Anyway, I hope to do some useful benchmarking today.
On Thu, 21 Dec 2000, Marcelo Tosatti wrote:
>
> On Thu, 21 Dec 2000, Andrew Morton wrote:
>
> > Marcelo Tosatti wrote:
> > >
> > > Hi,
> > >
> > > Basically this new swap_writepage function looks for dirty swapcache pages
> > > which may be contiguous (reverse and forward searching wrt to the physical
> > > address of the page being passed to swap_writepage) and builds a page list
> > > which is written "at once".
> > >
> > > The patch is against test13pre3.
> > >
> > > Comments are welcome. (especially about the __find_page_nolock
> > > modification)
> > >
> >
> > Have you any benchmarks for this?
>
> Not yet.
>
> Under some stress tests on a 16mb machine, around 30% of the clustered
> swapouts were reaching the limit of pages, which was 16.
While running X, Netscape and dbench (2 threads) on the 16mb machine the
number of successful write request merges increased by around 1000
(without the patch it was an average of 4000 write merges).
The dbench throughtput increased from 2.0Mb/s to 2.3Mb/s on average.
(Note: the swap partition was on the same disk as the rest of the system)