Hi,
Another attempt.
With this patch I get noticeable improvements in streaming write +16%!
(streaming write throughput is close to streaming read :-)
dbench results are mixed - slightly worse than plain test5-1...
It now survives mmap002, as opposed to vmfix-2.x :-) there were
bugs of cause.
* Basic idea in this patch is to keep free pages of zones in the
range [pages_high ... pages_low]. Kswapd will only run until
one zone gets pages_high. In this situation pages from all zones
are free able.
* In addition kswapd will run if any zone has less than pages_low.
* Actually implemented by using three values in zone_wake_kswapd
0 = zone initially above pages_high, allocs allowed until zone
gets < pages_low
1 = zone < pages_low
-1 = additional alloc done after zone become < pages_low
Most of the time there will only be one zone to with
zone_wake_kswapd zero. This zone will get the allocs until it
also gets < pages_low, then kswapd starts and runs until any
zone gets > pages_high - it will probably be another zone. Now
that one gets the allocs, ...
* There are some additional stuff that needs cleaning / further
investigations.
/RogerL
-- Home page: http://www.norran.net/nra02596/
--- linux/mm/vmscan.c.orig Sat Jul 15 23:44:34 2000 +++ linux/mm/vmscan.c Wed Jul 19 20:27:20 2000 @@ -419,7 +419,7 @@ out: } /* - * Check if there is any memory pressure (free_pages < pages_low) + * Check if there is any memory pressure (zone_wake_kswapd < 0) */ static inline int memory_pressure(void) { @@ -430,7 +430,7 @@ static inline int memory_pressure(void) for(i = 0; i < MAX_NR_ZONES; i++) { zone_t *zone = pgdat->node_zones+ i; if (zone->size && - zone->free_pages < zone->pages_low) + zone->zone_wake_kswapd < 0) return 1; } pgdat = pgdat->node_next; @@ -440,24 +440,31 @@ static inline int memory_pressure(void) } /* - * Check if there recently has been memory pressure (zone_wake_kswapd) + * Check if any zone have recently been critical (low_on_memory) + * any zone with current memory pressure (zone_wake_kswapd < 0) + * all zones have recently had memory_pressure (zone_wake_kswapd) */ static inline int keep_kswapd_awake(void) { + int all_recent = 1; pg_data_t *pgdat = pgdat_list; do { int i; for(i = 0; i < MAX_NR_ZONES; i++) { zone_t *zone = pgdat->node_zones+ i; - if (zone->size && - zone->zone_wake_kswapd) - return 1; + if (zone->size) { + if (zone->zone_wake_kswapd < 0 || + zone->low_on_memory) + return 1; + if (!zone->zone_wake_kswapd) + all_recent = 0; + } } pgdat = pgdat->node_next; } while (pgdat); - return 0; + return all_recent; } /* @@ -484,7 +491,7 @@ static int do_try_to_free_pages(unsigned priority = 64; do { - if (current->need_resched) { + if ((gfp_mask & __GFP_IO) && current->need_resched) { schedule(); /* time has passed - pressure too? */ if (!memory_pressure()) @@ -496,9 +503,6 @@ static int do_try_to_free_pages(unsigned goto done; } - /* not (been) low on memory - it is - * pointless to try to swap out. - */ if (!keep_kswapd_awake()) goto done; @@ -516,12 +520,10 @@ static int do_try_to_free_pages(unsigned * In the inner funtions there is a comment: * "To help debugging, a zero exit status indicates * all slabs were released." (-arca?) - * lets handle it in a primitive but working way... + * remove it... to visualize the problem. * if (count <= 0) * goto done; */ - if (!keep_kswapd_awake()) - goto done; while (shm_swap(priority, gfp_mask)) { if (!--count) @@ -596,10 +598,7 @@ int kswapd(void *unused) for (;;) { if (!keep_kswapd_awake()) { - /* wake up regulary to do an early attempt too free - * pages - pages will not actually be freed. - */ - interruptible_sleep_on_timeout(&kswapd_wait, HZ); + interruptible_sleep_on(&kswapd_wait); } do_try_to_free_pages(GFP_KSWAPD); @@ -628,24 +627,30 @@ int try_to_free_pages(unsigned int gfp_m if (gfp_mask & __GFP_WAIT) { current->state = TASK_RUNNING; current->flags |= PF_MEMALLOC; - retval = do_try_to_free_pages(gfp_mask); + do { + retval = do_try_to_free_pages(gfp_mask); + } while (!retval); current->flags &= ~PF_MEMALLOC; } - else { - /* make sure kswapd runs */ - if (waitqueue_active(&kswapd_wait)) - wake_up_interruptible(&kswapd_wait); - } + + /* someone needed memory that kswapd had not provided + * make sure kswapd runs, should not happen often */ + if (waitqueue_active(&kswapd_wait)) + wake_up_interruptible(&kswapd_wait); return retval; } static int __init kswapd_init(void) { - printk("Starting kswapd v1.6\n"); + printk("Starting kswapd v1.7\n"); swap_setup(); kernel_thread(kswapd, NULL, CLONE_FS | CLONE_FILES | CLONE_SIGHAND); return 0; } module_init(kswapd_init) + + + + --- linux/mm/page_alloc.c.orig Sat Jul 15 23:44:46 2000 +++ linux/mm/page_alloc.c Wed Jul 19 19:48:57 2000 @@ -30,7 +30,7 @@ pg_data_t *pgdat_list; static char *zone_names[MAX_NR_ZONES] = { "DMA", "Normal", "HighMem" }; static int zone_balance_ratio[MAX_NR_ZONES] = { 128, 128, 128, }; -static int zone_balance_min[MAX_NR_ZONES] = { 10 , 10, 10, }; +static int zone_balance_min[MAX_NR_ZONES] = { 10 , 10, 10, }; static int zone_balance_max[MAX_NR_ZONES] = { 255 , 255, 255, }; /* @@ -141,8 +141,13 @@ void __free_pages_ok (struct page *page, spin_unlock_irqrestore(&zone->lock, flags); + if (zone->free_pages > zone->pages_high) { - zone->zone_wake_kswapd = 0; + zone->zone_wake_kswapd = 0; + } + else if (zone->zone_wake_kswapd < 0 && + zone->free_pages > zone->pages_low) { + zone->zone_wake_kswapd = 1; zone->low_on_memory = 0; } } @@ -217,7 +222,7 @@ static struct page * rmqueue(zone_t *zon */ struct page * __alloc_pages(zonelist_t *zonelist, unsigned long order) { - zone_t **zone = zonelist->zones; + zone_t **zone; extern wait_queue_head_t kswapd_wait; /* @@ -228,21 +233,6 @@ struct page * __alloc_pages(zonelist_t * * in a higher zone fails. */ - for (;;) { - zone_t *z = *(zone++); - if (!z) - break; - if (!z->size) - BUG(); - - /* If there are zones with a lot of free memory allocate from them */ - if (z->free_pages > z->pages_high) { - struct page *page = rmqueue(z, order); - if (page) - return page; - } - } - zone = zonelist->zones; for (;;) { zone_t *z = *(zone++); @@ -256,6 +246,16 @@ struct page * __alloc_pages(zonelist_t * struct page *page = rmqueue(z, order); if (z->free_pages < z->pages_low) { z->zone_wake_kswapd = 1; + + /* Usually zone_wake_kswapd is set to -1 + * on second alloc below pages_low + * but if this was a big one + * - do not let it pass unnoticed + */ + if (z->free_pages < z->pages_low - MAX_ORDER) { + z->zone_wake_kswapd = -1; + } + if (waitqueue_active(&kswapd_wait)) wake_up_interruptible(&kswapd_wait); } @@ -264,6 +264,21 @@ struct page * __alloc_pages(zonelist_t * } } + /* Three possibilities to get here + * - Previous alloc_pages resulted in last zone set to have + * zone_wake_kswapd and start it. kswapd has not been able + * to release enough pages so that one zone does not have + * zone_wake_kswapd set. + * - Different sets of zones (zonelist) + * previous did not have all zones with zone_wake_kswapd but + * this one has... should kswapd be woken up? it will run once. + * - SMP race, kswapd went to sleep slightly after it as running + * in 'if (waitqueue_active(...))' above. + * + anyway the test is very cheap to do... + */ + if (waitqueue_active(&kswapd_wait)) + wake_up_interruptible(&kswapd_wait); + /* * Ok, we don't have any zones that don't need some * balancing.. See if we have any that aren't critical.. @@ -275,8 +290,17 @@ struct page * __alloc_pages(zonelist_t * break; if (!z->low_on_memory) { struct page *page = rmqueue(z, order); - if (z->free_pages < z->pages_min) - z->low_on_memory = 1; + if (z->free_pages < z->pages_low) { + z->zone_wake_kswapd = -1; + + if (z->free_pages < z->pages_min) + z->low_on_memory = 1; + + /* make kswapd notice new condition */ + if (waitqueue_active(&kswapd_wait)) + wake_up_interruptible(&kswapd_wait); + } + if (page) return page; } @@ -385,7 +409,10 @@ void show_free_areas_core(int nid) zone_t *zone = NODE_DATA(nid)->node_zones + type; unsigned long nr, total, flags; - printk(" %s: ", zone->name); + printk(" %s %ld (%ld %ld %ld) %d %d: ", + zone->name, zone->free_pages, + zone->pages_min, zone->pages_low, zone->pages_high, + zone->zone_wake_kswapd, zone->low_on_memory); total = 0; if (zone->size) { @@ -573,8 +600,8 @@ void __init free_area_init_core(int nid, else if (mask > zone_balance_max[j]) mask = zone_balance_max[j]; zone->pages_min = mask; - zone->pages_low = mask*2; - zone->pages_high = mask*3; + zone->pages_low = mask*3; + zone->pages_high = mask*4; zone->low_on_memory = 0; zone->zone_wake_kswapd = 0; zone->zone_mem_map = mem_map + offset;
- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.rutgers.edu Please read the FAQ at http://www.tux.org/lkml/
This archive was generated by hypermail 2b29 : Sun Jul 23 2000 - 21:00:12 EST