Hi,
The following patch addresses two issues:
- Buffer cache pages in the inactive lists are not getting their age
increased if they get touched by getblk (which will set the referenced bit
on the page). page_launder() simply cleans the referenced bit on such
pages and moves them to the active list. To resume: buffercache pages
suffer more pressure from VM than pagecache pages. That is horrible for
performance.
- When there is no memory available on the system for normal allocations
(GFP_KERNEL), the tasks may loop in try_to_free_pages() (which is here
called by __alloc_pages()) without blocking:
- GFP_BUFFER allocations will _never_ block on IO inside
try_to_free_pages(). They will keep looping inside __alloc_pages()
until they get a free page.
- __GFP_IO|__GFP_WAIT allocations may not find any way to block on
IO inside try_to_free_pages() in case we already have other tasks
inside there (kswapd will be there in such condition, for sure).
To address this issues, I've changed the __alloc_pages() loop to act as
following:
for normal allocations we sleep on the kswapd wait queue _only_ if not
able to make any progress with try_to_free_pages.
for GFP_BUFFER allocations we fail in case we're not able to make progress
ourselves (try_to_free_pages() returns zero in that case).
create_buffers() is able to deal with that nicely. (it sleeps until there
are "enough" free buffer_head's available)
Comments (especially about the second issue) are welcome.
And here is the patch:
diff -Nur --exclude-from=exclude linux.orig/include/linux/swap.h
linux/include/linux/swap.h
--- linux.orig/include/linux/swap.h Fri May 11 11:17:25 2001
+++ linux/include/linux/swap.h Fri May 11 11:21:26 2001
@@ -125,7 +125,7 @@
extern int page_launder(int, int);
extern int free_shortage(void);
extern int inactive_shortage(void);
-extern void wakeup_kswapd(void);
+extern void wakeup_kswapd(int);
extern int try_to_free_pages(unsigned int gfp_mask);
/* linux/mm/page_io.c */
diff -Nur --exclude-from=exclude linux.orig/mm/page_alloc.c linux/mm/page_alloc.c
--- linux.orig/mm/page_alloc.c Fri May 11 11:17:19 2001
+++ linux/mm/page_alloc.c Fri May 11 11:20:31 2001
@@ -364,7 +364,7 @@
* - if we don't have __GFP_IO set, kswapd may be
* able to free some memory we can't free ourselves
*/
- wakeup_kswapd();
+ wakeup_kswapd(0);
if (gfp_mask & __GFP_WAIT) {
__set_current_state(TASK_RUNNING);
current->policy |= SCHED_YIELD;
@@ -444,9 +444,21 @@
* the inactive clean list. (done by page_launder)
*/
if (gfp_mask & __GFP_WAIT) {
- memory_pressure++;
- try_to_free_pages(gfp_mask);
- goto try_again;
+ /*
+ * In case we fail doing any progress freeing pages,
+ * wait for kswapd if possible.
+ */
+ if (try_to_free_pages(gfp_mask))
+ goto try_again;
+ else if (gfp_mask & __GFP_IO) {
+ wakeup_kswapd(1);
+ memory_pressure++;
+ goto try_again;
+ } else
+ /*
+ * This is a GFP_BUFFER allocation.
+ */
+ return NULL;
}
}
diff -Nur --exclude-from=exclude linux.orig/mm/vmscan.c linux/mm/vmscan.c
--- linux.orig/mm/vmscan.c Fri May 11 11:17:19 2001
+++ linux/mm/vmscan.c Fri May 11 13:21:06 2001
@@ -350,7 +350,7 @@
}
/* Page is or was in use? Move it to the active list. */
- if (PageTestandClearReferenced(page) || page->age > 0 ||
+ if (PageReferenced(page) || page->age > 0 ||
(!page->buffers && page_count(page) > 1)) {
del_page_from_inactive_clean_list(page);
add_page_to_active_list(page);
@@ -461,7 +461,7 @@
}
/* Page is or was in use? Move it to the active list. */
- if (PageTestandClearReferenced(page) || page->age > 0 ||
+ if (PageReferenced(page) || page->age > 0 ||
(!page->buffers && page_count(page) > 1) ||
page_ramdisk(page)) {
del_page_from_inactive_dirty_list(page);
@@ -1003,6 +1003,7 @@
recalculate_vm_stats();
}
+ wake_up_all(&kswapd_done);
run_task_queue(&tq_disk);
/*
@@ -1033,10 +1034,33 @@
}
}
-void wakeup_kswapd(void)
+void wakeup_kswapd(int block)
{
- if (current != kswapd_task)
- wake_up_process(kswapd_task);
+ DECLARE_WAITQUEUE(wait, current);
+
+ if (current == kswapd_task)
+ return;
+
+ if (!block) {
+ if (waitqueue_active(&kswapd_wait))
+ wake_up(&kswapd_wait);
+ return;
+ }
+
+ /*
+ * Kswapd could wake us up before we get a chance
+ * to sleep, so we have to be very careful here to
+ * prevent SMP races...
+ */
+ __set_current_state(TASK_UNINTERRUPTIBLE);
+ add_wait_queue(&kswapd_done, &wait);
+
+ if (waitqueue_active(&kswapd_wait))
+ wake_up(&kswapd_wait);
+ schedule();
+
+ remove_wait_queue(&kswapd_done, &wait);
+ __set_current_state(TASK_RUNNING);
}
/*
-
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/
