Commit 7078202e authored by Artem Bityutskiy's avatar Artem Bityutskiy

UBIFS: document dark_wm and dead_wm better

Just add more commentaries. Also some commentary fixes for
lprops flags.
Signed-off-by: default avatarArtem Bityutskiy <Artem.Bityutskiy@nokia.com>
parent a50412e3
...@@ -31,6 +31,26 @@ ...@@ -31,6 +31,26 @@
* to be reused. Garbage collection will cause the number of dirty index nodes * to be reused. Garbage collection will cause the number of dirty index nodes
* to grow, however sufficient space is reserved for the index to ensure the * to grow, however sufficient space is reserved for the index to ensure the
* commit will never run out of space. * commit will never run out of space.
*
* Notes about dead watermark. At current UBIFS implementation we assume that
* LEBs which have less than @c->dead_wm bytes of free + dirty space are full
* and not worth garbage-collecting. The dead watermark is one min. I/O unit
* size, or min. UBIFS node size, depending on what is greater. Indeed, UBIFS
* Garbage Collector has to synchronize the GC head's write buffer before
* returning, so this is about wasting one min. I/O unit. However, UBIFS GC can
* actually reclaim even very small pieces of dirty space by garbage collecting
* enough dirty LEBs, but we do not bother doing this at this implementation.
*
* Notes about dark watermark. The results of GC work depends on how big are
* the UBIFS nodes GC deals with. Large nodes make GC waste more space. Indeed,
* if GC move data from LEB A to LEB B and nodes in LEB A are large, GC would
* have to waste large pieces of free space at the end of LEB B, because nodes
* from LEB A would not fit. And the worst situation is when all nodes are of
* maximum size. So dark watermark is the amount of free + dirty space in LEB
* which are guaranteed to be reclaimable. If LEB has less space, the GC migh
* be unable to reclaim it. So, LEBs with free + dirty greater than dark
* watermark are "good" LEBs from GC's point of few. The other LEBs are not so
* good, and GC takes extra care when moving them.
*/ */
#include <linux/pagemap.h> #include <linux/pagemap.h>
......
...@@ -573,15 +573,8 @@ static int init_constants_early(struct ubifs_info *c) ...@@ -573,15 +573,8 @@ static int init_constants_early(struct ubifs_info *c)
c->ranges[UBIFS_IDX_NODE].max_len = INT_MAX; c->ranges[UBIFS_IDX_NODE].max_len = INT_MAX;
/* /*
* Initialize dead and dark LEB space watermarks. * Initialize dead and dark LEB space watermarks. See gc.c for comments
* * about these values.
* Dead space is the space which cannot be used. Its watermark is
* equivalent to min. I/O unit or minimum node size if it is greater
* then min. I/O unit.
*
* Dark space is the space which might be used, or might not, depending
* on which node should be written to the LEB. Its watermark is
* equivalent to maximum UBIFS node size.
*/ */
c->dead_wm = ALIGN(MIN_WRITE_SZ, c->min_io_size); c->dead_wm = ALIGN(MIN_WRITE_SZ, c->min_io_size);
c->dark_wm = ALIGN(UBIFS_MAX_NODE_SZ, c->min_io_size); c->dark_wm = ALIGN(UBIFS_MAX_NODE_SZ, c->min_io_size);
......
...@@ -426,9 +426,9 @@ struct ubifs_unclean_leb { ...@@ -426,9 +426,9 @@ struct ubifs_unclean_leb {
* LEB properties flags. * LEB properties flags.
* *
* LPROPS_UNCAT: not categorized * LPROPS_UNCAT: not categorized
* LPROPS_DIRTY: dirty > 0, not index * LPROPS_DIRTY: dirty > free, dirty >= @c->dead_wm, not index
* LPROPS_DIRTY_IDX: dirty + free > @c->min_idx_node_sze and index * LPROPS_DIRTY_IDX: dirty + free > @c->min_idx_node_sze and index
* LPROPS_FREE: free > 0, not empty, not index * LPROPS_FREE: free > 0, dirty < @c->dead_wm, not empty, not index
* LPROPS_HEAP_CNT: number of heaps used for storing categorized LEBs * LPROPS_HEAP_CNT: number of heaps used for storing categorized LEBs
* LPROPS_EMPTY: LEB is empty, not taken * LPROPS_EMPTY: LEB is empty, not taken
* LPROPS_FREEABLE: free + dirty == leb_size, not index, not taken * LPROPS_FREEABLE: free + dirty == leb_size, not index, not taken
......
Markdown is supported
0%
or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment