Don't set/test/wait-for radix tree tags if no capability

While debugging a problem [in the out-of-tree migration cache], I noticed a lot of radix-tree tag activity for address spaces that have the BDI_CAP_NO_{ACCT_DIRTY|WRITEBACK} capability flags set--effectively disabling these capabilities--in their backing device. Altho' functionally benign, I believe that this unnecessary overhead. Seeking contrary opinions.

Would this patch, if correct, be worthwhile? Seems not to cause any problem, and does eliminate the tag handling for swap space and similar address spaces.

---

page-writeback functions are setting/testing radix-tree tags for mappings whose backing device has the corresponding "capabilities" turned off via the BDI_CAP_NO_{ACCT_DIRTY|WRITEBACK} flags.

This patch makes all setting/getting/waiting-on the radix-tree tags conditional on the corresponding mapping capability, eliminating a lot of unnecessary radix tree traversal--e.g., for swap cache.

Note: perhaps all of the tags should be conditional on the '_WRITEBACK capability alone, because if the mapping doesn't support writeback, there is no need to track dirty pages in the cache, whether or not dirty page accounting is enabled. However, currently, all mappings that have one of the capability flags set also have the other flag set. Might not always be the case, tho.

 mm/filemap.c        |    2 +-
 mm/page-writeback.c |   21 ++++++++++++---------
 2 files changed, 13 insertions(+), 10 deletions(-)

Index: linux-2.6.18-rc6-mm2/mm/page-writeback.c
===================================================================
@@ -766,11 +766,13 @@ int __set_page_dirty_nobuffers(struct pa
 			mapping2 = page_mapping(page);
 			if (mapping2) { /* Race with truncate? */
 				BUG_ON(mapping2 != mapping);
-				if (mapping_cap_account_dirty(mapping))
+				if (mapping_cap_account_dirty(mapping)) {
 					__inc_zone_page_state(page,
 								NR_FILE_DIRTY);
-				radix_tree_tag_set(&mapping->page_tree,
-					page_index(page), PAGECACHE_TAG_DIRTY);
+					radix_tree_tag_set(&mapping->page_tree,
+							page_index(page),
+							PAGECACHE_TAG_DIRTY);
+				}
 			}
 			write_unlock_irq(&mapping->tree_lock);
 			if (mapping->host) {

@@ -855,9 +857,10 @@ int test_clear_page_dirty(struct page *p
 	if (mapping) {
 		write_lock_irqsave(&mapping->tree_lock, flags);
 		if (TestClearPageDirty(page)) {
-			radix_tree_tag_clear(&mapping->page_tree,
-						page_index(page),
-						PAGECACHE_TAG_DIRTY);
+			if (mapping_cap_account_dirty(mapping))
+				radix_tree_tag_clear(&mapping->page_tree,
+							page_index(page),
+							PAGECACHE_TAG_DIRTY);
 			write_unlock_irqrestore(&mapping->tree_lock, flags);
 			/*
 			 * We can continue to use `mapping' here because the

@@ -919,7 +922,7 @@ int test_clear_page_writeback(struct pag
 
 		write_lock_irqsave(&mapping->tree_lock, flags);
 		ret = TestClearPageWriteback(page);
-		if (ret)
+		if (ret && mapping_cap_writeback_dirty(mapping))
 			radix_tree_tag_clear(&mapping->page_tree,
 						page_index(page),
 						PAGECACHE_TAG_WRITEBACK);

@@ -940,11 +943,11 @@ int test_set_page_writeback(struct page 
 
 		write_lock_irqsave(&mapping->tree_lock, flags);
 		ret = TestSetPageWriteback(page);
-		if (!ret)
+		if (!ret && mapping_cap_writeback_dirty(mapping))
 			radix_tree_tag_set(&mapping->page_tree,
 						page_index(page),
 						PAGECACHE_TAG_WRITEBACK);
-		if (!PageDirty(page))
+		if (!PageDirty(page) && mapping_cap_account_dirty(mapping))
 			radix_tree_tag_clear(&mapping->page_tree,
 						page_index(page),
 						PAGECACHE_TAG_DIRTY);

Index: linux-2.6.18-rc6-mm2/mm/filemap.c
===================================================================
@@ -258,7 +258,7 @@ int wait_on_page_writeback_range(struct 
 	int ret = 0;
 	pgoff_t index;
 
-	if (end < start)
+	if (end < start || !mapping_cap_writeback_dirty(mapping))
 		return 0;
 
 	pagevec_init(&pvec, 0);

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

I do not think that not wanting accounting for dirty pages means that we should not mark those dirty. If we do this then filesystems will not be able to find the dirty pags for writeout.

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

That's why I asked, and why I noted that maybe setting the dirty tags should be gated by the 'No writeback' capability, rather than the "No dirty accounting" capability. But then, maybe "no writeback" doesn't really mean that the address space/backing device doesn't do writeback.

The 'no writeback' capability is set for things like: configfs, hugetlbfs, dlmfs, ramfs, cpuset, sysfs, shmem segs, swap, ... And, as I mentioned, the 'no dirty accounting' capability happens to be set for all file systems that set 'no writeback'. However, I agree that we shouldn't count on this.

So, do the file systems need to writeout dirty pages for these file systems using the radix tree tags? Just looking where the tags are queried [radix_tree_gang_lookup_tag()], it appears that tags are only used by "real" file systems, despite a call from pagevec_lookup_tag() that resides in mm/swap.c. And, it appears that the 'no writeback' capability flag will prevent writeback in some cases. Not sure if it catches all.

If we can't gate setting the flags based on the existing capabilities, maybe we want to define a new cap flag--e.g., BDI_CAP_NO_TAGS--for use by file systems that don't need the tags set? Not sure it's worth it, but could eliminate some cache pollution.

Lee

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

I agree you do need to check it out carefully, but it sounds very reasonable to me to avoid that radix tree tag overhead on the whole class of storageless filesystems (and swap plays by those same rules, despite that it does have backing storage).

If it checks out right at present, I think you can "count on this", just so long as you insert a suitable BUG_ON somewhere to alert us if some later mod unconsciously changes the situation (e.g. we find we do need more dirty page tracking on those currently exempt).

A related saving you can make there, I believe, is to add another .set_page_dirty variant which does nothing(?) more than SetPageDirty - swap and tmpfs and probably all those you mentioned above don't really want to do any more than that there - or didn't two or three years ago, when I had a patch for that but got diverted - the situation may have changed significantly since, and no longer be an option.

You are shamelessly parading your naivete, expecting mm/swap.c to have something to do with swap. Perhaps it did once upon a time (ooh, swap_setup does have something to do with swap), and it still has a lot to do with the page LRU lists (which are pointless unless you're swapping in the wider sense); but really it's just mm/misc.c.

Hugh

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

Yes. I considered that at first, when I saw where the tags were getting
set. But in looking at the existing set_page_dirty variants, it
occurred to me that the BDI capability flags were telling us what file systems needed dirty page tracking and came up with this patch.

However, a separate function for those file systems might be a bit more efficient in the paths that call set_page_dirty(). There might still some benefit to be had in short circuiting some of the wait-on and gang lookups on the "other side". But, perhaps those functions never get called for the file systems we're discussing.

I think I'll rework the patch to gate all of the tags on the NO_WRITEBACK_DIRTY capability instead of the ACCOUNT_DIRTY cap. Along the way, I'll look into which of these changes could be dropped by adding the new set_page_dirty op function. As time permits, of course...

So, I guess I shouldn't worry too much about why swapin_readahead() is in mm/memory.c instead of one of the mm/swap*.c files, huh?

Lee

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

Don't worry too much about it, indeed; though I agree it'd be much nicer tucked away somewhere else, near read_swap_cache_async or near valid_swaphandles or the three brought together.

(But if you do want to worry about swapin_readahead, just savour the absurdity of its NUMA next_vma code: just what is the likelihood that swap allocation will match a task's address layout?)

Hugh

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>