regcache: Push async I/O request down into the rbtree cache

Currently the regcache core unconditionally enables async I/O for all cache
types, causing problems for the maple tree cache which dynamically allocates
the buffers used to write registers to the device since async requires the
buffers to be kept around until the I/O has been completed.

This use of async I/O is mainly for the rbtree cache which stores data in
a format directly usable for regmap_raw_write(), though there is a special
case for single register writes which would also have allowed it to be used
with the flat cache. It is a bit of a landmine for other caches since it
implicitly converts sync operations to async, and with modern hardware it
is not clear that async I/O is actually a performance win as shown by the
performance work David Jander did with SPI. In multi core systems the cost
of managing concurrency ends up swamping the performance benefit and almost
all modern systems are multi core.

Address this by pushing the enablement of async I/O down into the rbtree
cache where it is actively used, avoiding surprises for other cache
implementations.

Reported-by: Charles Keepax <ckeepax@opensource.cirrus.com>
Fixes: bfa0b38c1483 ("regmap: maple: Implement block sync for the maple tree cache")
Reviewed-by: Charles Keepax <ckeepax@opensource.cirrus.com>
Tested-by: Charles Keepax <ckeepax@opensource.cirrus.com>
Signed-off-by: Mark Brown <broonie@kernel.org>
Link: https://lore.kernel.org/r/20230719-regcache-async-rbtree-v1-1-b03d30cf1daf@kernel.org
Signed-off-by: Mark Brown <broonie@kernel.org>

diff --git a/drivers/base/regmap/regcache-rbtree.c b/drivers/base/regmap/regcache-rbtree.c
index fabf87058d80bc0610f2e7cff25b29c91821cb56..584bcc55f56e36f3e81b63558e3d0172d2fb4b5c 100644 (file)
--- a/drivers/base/regmap/regcache-rbtree.c
+++ b/drivers/base/regmap/regcache-rbtree.c
@@ -471,6 +471,8 @@ static int regcache_rbtree_sync(struct regmap *map, unsigned int min,
        unsigned int start, end;
        int ret;
 
+       map->async = true;
+
        rbtree_ctx = map->cache;
        for (node = rb_first(&rbtree_ctx->root); node; node = rb_next(node)) {
                rbnode = rb_entry(node, struct regcache_rbtree_node, node);
@@ -499,6 +501,8 @@ static int regcache_rbtree_sync(struct regmap *map, unsigned int min,
                        return ret;
        }
 
+       map->async = false;
+
        return regmap_async_complete(map);
 }
 

diff --git a/drivers/base/regmap/regcache.c b/drivers/base/regmap/regcache.c
index 28bc3ae9458acb94ff07e8e20adf12b79c29a01b..7d3e47436056e57780beaa3ddf792a014833f606 100644 (file)
--- a/drivers/base/regmap/regcache.c
+++ b/drivers/base/regmap/regcache.c
@@ -368,8 +368,6 @@ int regcache_sync(struct regmap *map)
        if (!map->cache_dirty)
                goto out;
 
-       map->async = true;
-
        /* Apply any patch first */
        map->cache_bypass = true;
        for (i = 0; i < map->patch_regs; i++) {
@@ -392,7 +390,6 @@ int regcache_sync(struct regmap *map)
 
 out:
        /* Restore the bypass state */
-       map->async = false;
        map->cache_bypass = bypass;
        map->no_sync_defaults = false;
        map->unlock(map->lock_arg);