Re: [PATCH v3 5/6] zswap: replace tree in zswap with radix tree in zbud

[Krzysztof Kozlowski]

On Wed, 2013-10-09 at 12:16 -0500, Seth Jennings wrote:
> On Wed, Oct 09, 2013 at 10:30:22AM -0500, Seth Jennings wrote:
> > In my approach, I was also looking at allowing the zbud pools to use
> > HIGHMEM pages, since the handle is no longer an address.  This requires
> > the pages that are being mapped to be kmapped (atomic) which will
> > disable preemption.  This isn't an additional overhead since the
> > map/unmap corresponds with a compress/decompress operation at the zswap
> > level which uses per-cpu variables that disable preemption already.
> 
> On second though, lets not mess with the HIGHMEM page support for now.
> Turns out it is tricker than I thought since the unbuddied lists are
> linked through the zbud header stored in the page.  But we can still
> disable preemption to allow per-cpu tracking of the current mapping and
> avoid a lookup (and races) in zbud_unmap().

This tracking of current mapping could solve another problem I
encountered with new one-radix-tree approach with storage of duplicated
entries.

The problem is in zbud_unmap() API using offset to unmap (if duplicated
entries are overwritten):
 - thread 1: zswap_fronstwap_load() of some offset
   - zbud_map() maps this offset -> zhdr1
 - thread 2: zswap_frontswap_store() stores new data for this offset 
   - zbud_alloc() allocated new zhdr2 and replaces zhdr1 in radix tree 
     under this offset
   - new compressed data is stored by zswap
 - thread 1: tries to zbud_unmap() of this offset, but now the old
   zhdr1 is not present in radix tree so unmap will either fail or use
   zhdr2 which is wrong

To solve this issue I experimented with unmapping by zbud_mapped_entry
instead of offset (so zbud_unmap() won't search zbud_header in radix
tree at all):
##########################
int zbud_unmap(struct zbud_pool *pool, pgoff_t offset,
		struct zbud_mapped_entry *entry)
{
	struct zbud_header *zhdr = handle_to_zbud_header((unsigned
long)entry->addr);

	VM_BUG_ON((offset != zhdr->first_offset) && (offset !=
zhdr->last_offset));
	spin_lock(&pool->lock);
	if (put_map_count(zhdr, offset)) {
		/* Racing zbud_free() could not free the offset because
		 * it was still mapped so it is our job to free. */
		zbud_header_free(pool, zhdr, offset);
		spin_unlock(&pool->lock);
		return -EFAULT;
	}
	put_zbud_page(zhdr);
	spin_unlock(&pool->lock);
	return 0;
}
##########################

However getting rid of first/last_map_count seems much more simpler! 

Best regards,
Krzysztof


--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/