diff mbox series

[-,2/2] nfsd: degraded slot-count more gracefully as allocation nears exhaustion.

Message ID 87zhiz8oea.fsf@notabene.neil.brown.name (mailing list archive)
State New, archived
Headers show
Series [1/2,-,vers2] nfsd: handle drc over-allocation gracefully. | expand

Commit Message

NeilBrown Sept. 20, 2019, 6:36 a.m. UTC
This original code in nfsd4_get_drc_mem() would hand out 30
slots (approximately NFSD_MAX_MEM_PER_SESSION bytes at slightly
over 2K per slot) to each requesting client until it ran out
of space, then it would possibly give one last client a reduced
allocation, then fail the allocation.

Since commit de766e570413 ("nfsd: give out fewer session slots as
limit approaches") the last 90 slots to be given to about 12
clients with quickly reducing slot counts (better than just 3
clients).  This still seems unnecessarily hasty.
A subsequent patch allows over-allocation so every client gets
at least one slot, but that might be a bit restrictive.

The requested number of nfsd threads is the best guide we have to the
expected number of clients, so use that - if it is at least 8.

256 threads on a 256Meg machine - which is a lot for a tiny machine -
would result in nfsd_drc_max_mem being 2Meg, so 8K (3 slots) would be
available for the first client, and over 200 clients would get more
than 1 slot.  So I don't think this change will be too debilitating on
poorly configured machines, though it does mean that a sensible
configuration is a little more important.

Signed-off-by: NeilBrown <neilb@suse.de>
 fs/nfsd/nfs4state.c | 15 +++++++++++----
 1 file changed, 11 insertions(+), 4 deletions(-)
diff mbox series


diff --git a/fs/nfsd/nfs4state.c b/fs/nfsd/nfs4state.c
index 084a30d77359..c49af2ba3924 100644
--- a/fs/nfsd/nfs4state.c
+++ b/fs/nfsd/nfs4state.c
@@ -1563,11 +1563,12 @@  static inline u32 slot_bytes(struct nfsd4_channel_attrs *ca)
  * re-negotiate active sessions and reduce their slot usage to make
  * room for new connections. For now we just fail the create session.
-static u32 nfsd4_get_drc_mem(struct nfsd4_channel_attrs *ca)
+static u32 nfsd4_get_drc_mem(struct nfsd4_channel_attrs *ca, struct nfsd_net *nn)
 	u32 slotsize = slot_bytes(ca);
 	u32 num = ca->maxreqs;
 	unsigned long avail, total_avail;
+	unsigned int scale_factor;
 	if (nfsd_drc_max_mem > nfsd_drc_mem_used)
@@ -1581,12 +1582,18 @@  static u32 nfsd4_get_drc_mem(struct nfsd4_channel_attrs *ca)
 		total_avail = 0;
 	avail = min((unsigned long)NFSD_MAX_MEM_PER_SESSION, total_avail);
-	 * Never use more than a third of the remaining memory,
+	 * Never use more than a fraction of the remaining memory,
 	 * unless it's the only way to give this client a slot.
+	 * The chosen fraction is either 1/8 or 1/number of threads,
+	 * whichever is smaller.  This ensures there are adequate
+	 * slots to support multiple clients per thread.
 	 * Give the client one slot even that would require
 	 * over-allocation, it is better than failure.
-	avail = clamp_t(unsigned long, avail, slotsize, total_avail/3);
+	scale_factor = max_t(unsigned int, 8, nn->nfsd_serv->sv_nrthreads);
+	avail = clamp_t(unsigned long, avail, slotsize,
+			total_avail/scale_factor);
 	num = min_t(int, num, avail / slotsize);
 	num = max_t(int, num, 1);
 	nfsd_drc_mem_used += num * slotsize;
@@ -3183,7 +3190,7 @@  static __be32 check_forechannel_attrs(struct nfsd4_channel_attrs *ca, struct nfs
 	 * Note that we always allow at least one slot, because our
 	 * accounting is soft and provides no guarantees either way.
-	ca->maxreqs = nfsd4_get_drc_mem(ca);
+	ca->maxreqs = nfsd4_get_drc_mem(ca, nn);
 	return nfs_ok;