target/rd: improve sg_table lookup scalability

Sequential scan of rd_dev->sg_table_array in rd_get_sg_table is
a serious I/O performance bottleneck for large rd LUNs. Fix this
by computing the sg_table index directly from page offset because
all sg_tables (except the last one) have the same number of pages.

Tested with 90 GiB rd_mcp LUN, where the patch improved maximal
random R/W IOPS by more than 100-150%, depending on actual
hardware and SAN setup.

Signed-off-by: Martin Svec<martin.svec@zoner.cz>
Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>

diff --git a/drivers/target/target_core_rd.c b/drivers/target/target_core_rd.c
index 0457de362e6801e934656c21d399fc3f1127781f..b0fff52c990e517722f88085ef04ef61432777e6 100644 (file)
--- a/drivers/target/target_core_rd.c
+++ b/drivers/target/target_core_rd.c
@@ -256,10 +256,12 @@ static void rd_free_device(struct se_device *dev)
 
 static struct rd_dev_sg_table *rd_get_sg_table(struct rd_dev *rd_dev, u32 page)
 {
-       u32 i;
        struct rd_dev_sg_table *sg_table;
+       u32 i, sg_per_table = (RD_MAX_ALLOCATION_SIZE /
+                               sizeof(struct scatterlist));
 
-       for (i = 0; i < rd_dev->sg_table_count; i++) {
+       i = page / sg_per_table;
+       if (i < rd_dev->sg_table_count) {
                sg_table = &rd_dev->sg_table_array[i];
                if ((sg_table->page_start_offset <= page) &&
                    (sg_table->page_end_offset >= page))