While testing NFSD for-next, I noticed svc_generic_init_request()
was an unexpected hot spot on NFSv4 workloads. Drilling into the
perf report, it shows that the hot path in there is:
1208 memset(rqstp->rq_argp, 0, procp->pc_argsize);
1209 memset(rqstp->rq_resp, 0, procp->pc_ressize);
For an NFSv4 COMPOUND,
procp->pc_argsize = sizeof(nfsd4_compoundargs),
struct nfsd4_compoundargs on my system is more than 17KB! This is
due to the size of the iops field:
struct nfsd4_op iops[8];
Each struct nfsd4_op contains a union of the arguments for each
NFSv4 operation. Each argument is typically less than 128 bytes
except that struct nfsd4_copy and struct nfsd4_copy_notify are both
larger than 2KB each.
I'm not yet totally convinced this series never orphans memory, but
it does reduce the size of nfsd4_compoundargs to just over 4KB. This
is still due to struct nfsd4_copy being almost 500 bytes. I don't
see more low-hanging fruit there, though.
---
Chuck Lever (11):
NFSD: Shrink size of struct nfsd4_copy_notify
NFSD: Shrink size of struct nfsd4_copy
NFSD: Reorder the fields in struct nfsd4_op
NFSD: Make nfs4_put_copy() static
NFSD: Make boolean fields in struct nfsd4_copy into atomic bit flags
NFSD: Refactor nfsd4_cleanup_inter_ssc() (1/2)
NFSD: Refactor nfsd4_cleanup_inter_ssc() (2/2)
NFSD: Refactor nfsd4_do_copy()
NFSD: Remove kmalloc from nfsd4_do_async_copy()
NFSD: Add nfsd4_send_cb_offload()
NFSD: Move copy offload callback arguments into a separate structure
fs/nfsd/nfs4callback.c | 37 +++++----
fs/nfsd/nfs4proc.c | 165 +++++++++++++++++++++--------------------
fs/nfsd/nfs4xdr.c | 30 +++++---
fs/nfsd/state.h | 1 -
fs/nfsd/xdr4.h | 54 ++++++++++----
5 files changed, 163 insertions(+), 124 deletions(-)
--
Chuck Lever
