Re: [PATCH] xprtrdma: silence frame size warning

[Trond Myklebust <trond.myklebust@primarydata.com>]

On Jan 13, 2014, at 10:45, Paul Bolle <pebolle@tiscali.nl> wrote:

> Building verbs.o on 32 bits x86, with CONFIG_FRAME_WARN set to 1024, its
> default value, triggers this GCC warning:
>    net/sunrpc/xprtrdma/verbs.c: In function ‘rpcrdma_register_default_external’:
>    net/sunrpc/xprtrdma/verbs.c:1774:1: warning: the frame size of 1040 bytes is larger than 1024 bytes [-Wframe-larger-than=]
> 
> Silence this warning by allocating "ipb" dynamically.
> 
> Signed-off-by: Paul Bolle <pebolle@tiscali.nl>
> ---
> 0) Compile tested only (on 32 bits x86). I don't have access to
> Infiniband hardware.
> 
> 1) Please note that this is not a new warning. The oldest build log I
> have still available on this machine is for a v3.8 rc, and it already
> showed this warning. 
> 
> 2) I do hope my choice for the GFP_KERNEL flag is correct here.
> 
> net/sunrpc/xprtrdma/verbs.c | 6 +++++-
> 1 file changed, 5 insertions(+), 1 deletion(-)
> 
> diff --git a/net/sunrpc/xprtrdma/verbs.c b/net/sunrpc/xprtrdma/verbs.c
> index 93726560..939ccc8 100644
> --- a/net/sunrpc/xprtrdma/verbs.c
> +++ b/net/sunrpc/xprtrdma/verbs.c
> @@ -1736,11 +1736,14 @@ rpcrdma_register_default_external(struct rpcrdma_mr_seg *seg,
> 	int mem_priv = (writing ? IB_ACCESS_REMOTE_WRITE :
> 				  IB_ACCESS_REMOTE_READ);
> 	struct rpcrdma_mr_seg *seg1 = seg;
> -	struct ib_phys_buf ipb[RPCRDMA_MAX_DATA_SEGS];
> +	struct ib_phys_buf *ipb;
> 	int len, i, rc = 0;
> 
> 	if (*nsegs > RPCRDMA_MAX_DATA_SEGS)
> 		*nsegs = RPCRDMA_MAX_DATA_SEGS;
> +	ipb = kmalloc(sizeof(*ipb) * *nsegs, GFP_KERNEL);
> +	if (ipb == NULL)
> +		return -ENOMEM;
> 	for (len = 0, i = 0; i < *nsegs;) {
> 		rpcrdma_map_one(ia, seg, writing);
> 		ipb[i].addr = seg->mr_dma;
> @@ -1770,6 +1773,7 @@ rpcrdma_register_default_external(struct rpcrdma_mr_seg *seg,
> 		seg1->mr_len = len;
> 	}
> 	*nsegs = i;
> +	kfree(ipb);
> 	return rc;
> }

Hi Paul,

Unfortunately, the above could be called from a file write back context, so we cannot use GFP_KERNEL allocations. In fact, I suspect we should only use GFP_ATOMIC, and then have the rpc_task delay and retry if the allocation fails.

The problem is that it looks to me as if xprt_rdma_send_request will currently fail, if we return an error from rpcrdma_register_default_external.

Chuck, will you be able to look into the above issue as part of your RDMA work?

Cheers
  Trond

--
Trond Myklebust
Linux NFS client maintainer