2.4 fix for write throttling on x86 >1G
Subsystems (2)
Hello Marcelo,
I've got a fix for you on 2.4. I got reports of stalls with heavy writes on 2.4. There was a mistake in nr_free_buffer_pages. That function is definitely meant _not_ to take highmem into account (dirty cache cannot spread over highmem in 2.4 [even when on top of fs]). For unknown reasons it was actually taking highmem into account. The code was obviously meant to not take inot account see the GFP_USER and zonelist, except it wasn't using the zonelist. That is a severe problem because there will be no write throttling at all, and no bdflush wakeup either.
This should fix it, though my compiler fails to compile 2.4, so it's not immediate to verify it. If any problem showup I'll post a followup.
This is a noop for all systems <800M (1G shouldn't be noticeable either). This is why most people can't notice.
Thanks.
mm/page_alloc.c renamed from mm/page_alloc.c.~1~ +1 -1
@@ -656,7 +656,7 @@ unsigned int nr_free_buffer_pages (void) class_idx =zone_idx (zone); sum += zone->nr_cache_pages; -for (zone = pgdat->node_zones; zone < pgdat->node_zones +MAX_NR_ZONES ; zone++) { +for (; zone; zone = *zonep++) {int free = zone->free_pages - zone->watermarks[class_idx].high;if (free <=0 )continue ;
Hi Andrea!
Quoted 4 lines · Andrea Arcangeli
Hello Marcelo,
I've got a fix for you on 2.4. I got reports of stalls with heavy writes on 2.4.
Out of curiosity, that was SuSE not mainline ?
Quoted 6 lines
There was a mistake in
nr_free_buffer_pages. That function is definitely meant _not_ to take highmem into account (dirty cache cannot spread over highmem in 2.4 [even when on top of fs]). For unknown reasons it was actually taking highmem into account. The code was obviously meant to not take inot account see theGFP_USERand zonelist, except it wasn't using the zonelist.
True, initialization of "zone" variable in nr_free_buffer_pages() is un-nice.
Quoted 8 lines
That is a severe problem because
there will be no write throttling at all, and no bdflush wakeup either.This should fix it, though my compiler fails to compile 2.4, so it's not immediate to verify it. If any problem showup I'll post a followup.
This is a noop for all systems <800M (1G shouldn't be noticeable either). This is why most people can't notice.
Do we really want to limit dirty cache to low mem on HIGHIO capable machines? I'm afraid doing so might hurt performance on such systems.
I think it might be wise to have nr_free_buffer_pages() take highmem into account if CONFIG_HIGHIO is set ?
Quoted 11 lines
mm/page_alloc.c renamed from mm/page_alloc.c.~1~ +1
-1@@ -656,7 +656,7 @@ unsigned int nr_free_buffer_pages (void) class_idx =zone_idx (zone); sum += zone->nr_cache_pages; -for (zone = pgdat->node_zones; zone < pgdat->node_zones +MAX_NR_ZONES ; zone++) { +for (; zone; zone = *zonep++) {int free = zone->free_pages - zone->watermarks[class_idx].high;if (free <=0 )continue ;
Hello Marcelo,
Quoted 1 line · Marcelo Tosatti
Out of curiosity, that was SuSE not mainline ?
yep.
Quoted 5 lines
Do we really want to limit dirty cache to low mem on HIGHIO capable machines? I'm afraid doing so might hurt performance on such systems.
I think it might be wise to have
nr_free_buffer_pages()take highmem into account ifCONFIG_HIGHIOis set ?
The problem is the buffercache/blkdev-pagecache: it simply can't go in highmem. A similar fix happened recently in 2.6 for the same reasons, but in 2.6 we allowed it with some logic specific for the blkdev-pagecache.
nr_free_buffer_pages() was never intended to take highmem into account, that's why there's the GFP_USER thing already, except we didn't loop into the zonelist, so I didn't try to make a fix similar to 2.6.
Marcelo Tosatti
AppliedQuoted 10 lines · Andrea Arcangeli
Hello Marcelo,
Quoted 1 line · Marcelo Tosatti
Out of curiosity, that was SuSE not mainline ?
yep.
Quoted 3 lines
Do we really want to limit dirty cache to low mem on HIGHIO capable machines? I'm afraid doing so might hurt performance on such systems.
Quoted 7 lines
Quoted 2 lines
I think it might be wise to have
nr_free_buffer_pages()take highmem into account ifCONFIG_HIGHIOis set ?The problem is the buffercache/blkdev-pagecache: it simply can't go in highmem. A similar fix happened recently in 2.6 for the same reasons, but in 2.6 we allowed it with some logic specific for the blkdev-pagecache.
Right, I dont think it is easy nor wanted to make that distiction in v2.4.
Quoted 3 lines
nr_free_buffer_pages()was never intended to take highmem into account, that's why there's theGFP_USERthing already, except we didn't loop into the zonelist, so I didn't try to make a fix similar to 2.6.
Hopefully it is not a big deal to not-allow >1GB dirty pagecache on v2.4.
Applied, thanks.