[PATCH 18/21] NFSv4/pnfs: Grow the deviceid cache hash table

The global deviceid cache has 32 buckets shared by every server and
layout type.  pNFS striping deployments put hundreds to ~1000
distinct deviceIDs in play per mount (mirrors x stripes), a load
factor of ~31: resolution and re-resolution bursts walk long chains.

Grow to 256 buckets (2KB fixed): load factor ~4 at 1000 devices.
The existing byte-folding hash distributes fine at any power-of-two
size (the multiplier is odd), so bucket count is the only lever.
Alternatives considered and set aside: per-client tables do not help
the headline single-mount striping case, and a resizing rhashtable
is unwarranted for a burst-only (never per-I/O) lookup path.

Assisted-by: Claude:claude-fable-5
Signed-off-by: Benjamin Coddington <bcodding-F/[email protected]>
---
 fs/nfs/pnfs_dev.c | 10 +++++++++-
 1 file changed, 9 insertions(+), 1 deletion(-)

diff --git a/fs/nfs/pnfs_dev.c b/fs/nfs/pnfs_dev.c
index 0c39ae8325b3..e09f25a26004 100644
--- a/fs/nfs/pnfs_dev.c
+++ b/fs/nfs/pnfs_dev.c
@@ -40,8 +40,16 @@
 
 /*
  * Device ID RCU cache. A device ID is unique per server and layout type.
+ *
+ * 256 buckets keep chains short for pNFS striping deployments with
+ * hundreds to ~1000 distinct deviceIDs (load factor ~4 at 1000) at a
+ * fixed cost of 2KB.  The hash folds every deviceID byte through
+ * x = x * 37 + byte; 37 is odd, so the low-bit mask distributes
+ * deviceIDs differing in any single byte perfectly at any
+ * power-of-two table size -- bucket count, not the hash, is the
+ * scaling lever.
  */
-#define NFS4_DEVICE_ID_HASH_BITS	5
+#define NFS4_DEVICE_ID_HASH_BITS	8
 #define NFS4_DEVICE_ID_HASH_SIZE	(1 << NFS4_DEVICE_ID_HASH_BITS)
 #define NFS4_DEVICE_ID_HASH_MASK	(NFS4_DEVICE_ID_HASH_SIZE - 1)
 
-- 
2.53.0