Commit 630922e8 authored by Kozlowski Mateusz's avatar Kozlowski Mateusz Committed by Jim Harris
Browse files

ftl: Add lazy unmap process 



Since only L2P pages as a whole are marked as invalid during trim, the
specific L2P entries won't be updated until someone touches that page.
The unmap process will slowly invalidate pages during runtime, by paging
them in. This will allow compaction and relocation to benefit from the
trim as the user data gets invalidated.

Signed-off-by: default avatarKozlowski Mateusz <mateusz.kozlowski@intel.com>
Signed-off-by: default avatarArtur Paszkiewicz <artur.paszkiewicz@intel.com>
Change-Id: I239b9adf0aaaeac58f440145f4ab78b0d78d98b0
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/13381


Tested-by: default avatarSPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: default avatarJim Harris <james.r.harris@intel.com>
Reviewed-by: default avatarBen Walker <benjamin.walker@intel.com>
parent b3e5d8a7

lib/ftl/ftl_l2p_cache.c

+74 −0
Original line number Diff line number Diff line
@@ -121,6 +121,17 @@ struct ftl_l2p_cache { 
	struct ftl_mempool *page_sets_pool;

	TAILQ_HEAD(, ftl_l2p_page_set) deferred_page_set_list; /* for deferred page sets */



	/* Process unmap in backgorund */

	struct {

#define FTL_L2P_MAX_LAZY_UNMAP_QD 1

		/* Unmap queue depth */

		uint32_t qd;

		/* Currently processed page */

		uint64_t page_no;

		/* Context for page pinning */

		struct ftl_l2p_pin_ctx pin_ctx;

	} lazy_unmap;



	/* This is a context for a management process */

	struct ftl_l2p_cache_process_ctx mctx;
@@ -1476,6 +1487,68 @@ ftl_l2p_cache_process_eviction(struct spdk_ftl_dev *dev, struct ftl_l2p_cache *c 
	}

}



static void

ftl_l2p_lazy_unmap_process_cb(struct spdk_ftl_dev *dev, int status, struct ftl_l2p_pin_ctx *pin_ctx)

{

	struct ftl_l2p_cache *cache = dev->l2p;



	cache->lazy_unmap.qd--;



	/* We will retry on next ftl_l2p_lazy_unmap_process */

	if (spdk_unlikely(status != 0)) {

		return;

	}



	if (ftl_l2p_cache_running(cache)) {

		ftl_l2p_cache_get(dev, pin_ctx->lba);

	}



	ftl_l2p_cache_unpin(dev, pin_ctx->lba, pin_ctx->count);

}



static void

ftl_l2p_lazy_unmap_process(struct spdk_ftl_dev *dev)

{

	struct ftl_l2p_cache *cache = dev->l2p;

	struct ftl_l2p_pin_ctx *pin_ctx;

	uint64_t page_no;



	if (spdk_likely(!dev->unmap_in_progress)) {

		return;

	}



	if (cache->lazy_unmap.qd == FTL_L2P_MAX_LAZY_UNMAP_QD) {

		return;

	}



	page_no = ftl_bitmap_find_first_set(dev->unmap_map, cache->lazy_unmap.page_no, UINT64_MAX);

	if (page_no == UINT64_MAX) {

		cache->lazy_unmap.page_no = 0;



		/* Check unmap map from beginning to detect unprocessed unmaps */

		page_no = ftl_bitmap_find_first_set(dev->unmap_map, cache->lazy_unmap.page_no, UINT64_MAX);

		if (page_no == UINT64_MAX) {

			dev->unmap_in_progress = false;

			return;

		}

	}



	cache->lazy_unmap.page_no = page_no;



	pin_ctx = &cache->lazy_unmap.pin_ctx;



	cache->lazy_unmap.qd++;

	assert(cache->lazy_unmap.qd <= FTL_L2P_MAX_LAZY_UNMAP_QD);

	assert(page_no < cache->num_pages);



	pin_ctx->lba = page_no * cache->lbas_in_page;

	pin_ctx->count = 1;

	pin_ctx->cb = ftl_l2p_lazy_unmap_process_cb;

	pin_ctx->cb_ctx = pin_ctx;



	ftl_l2p_cache_pin(dev, pin_ctx);

}



void

ftl_l2p_cache_process(struct spdk_ftl_dev *dev)

{
@@ -1493,4 +1566,5 @@ ftl_l2p_cache_process(struct spdk_ftl_dev *dev) 
	}



	ftl_l2p_cache_process_eviction(dev, cache);

	ftl_l2p_lazy_unmap_process(dev);

}