| Message ID | 1371821563-5784-5-git-send-email-simon.derr@bull.net (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
Hi, Simon Derr wrote on Fri, Jun 21, 2013 : > -#define P9_RDMA_MAXSIZE (4*4096) /* Min SGE is 4, so we can > - * safely advertise a maxsize > - * of 64k */ > +#define P9_RDMA_MAXSIZE (1024*1024) /* 1MB */ May I push my luck and suggest we use 1024*1024 + 24? The idea would be to be able to do read and write requests of exactly 1MB, which need the extra bit because write has a 23 bytes overhead (read only has 11) - but it will also need the vfs calls to ask for the exact size that will no longer be obvious from the msize I guess that as long as we don't have this it would be negative to bring this further up - comments welcome before I look into it. That being said, we've been running many tests here with all these patches for a while and there doesn't seem to be any problem. Regards,
diff --git a/net/9p/trans_rdma.c b/net/9p/trans_rdma.c index b1dfdf2..b8b66d3 100644 --- a/net/9p/trans_rdma.c +++ b/net/9p/trans_rdma.c @@ -57,9 +57,7 @@ #define P9_RDMA_IRD 0 #define P9_RDMA_ORD 0 #define P9_RDMA_TIMEOUT 30000 /* 30 seconds */ -#define P9_RDMA_MAXSIZE (4*4096) /* Min SGE is 4, so we can - * safely advertise a maxsize - * of 64k */ +#define P9_RDMA_MAXSIZE (1024*1024) /* 1MB */ /** * struct p9_trans_rdma - RDMA transport instance
The current value is too low to get good performance. Signed-off-by: Simon Derr <simon.derr@bull.net> --- net/9p/trans_rdma.c | 4 +--- 1 files changed, 1 insertions(+), 3 deletions(-)