On 2009-05-05 23:28, Benny Halevy wrote:
> On May. 05, 2009, 22:41 +0300, Trond Myklebust <trond.myklebust@xxxxxxxxxx> wrote:
>> On Tue, 2009-05-05 at 15:39 -0400, Trond Myklebust wrote:
>>> On Tue, 2009-05-05 at 15:34 -0400, Trond Myklebust wrote:
>>>> On Fri, 2009-05-01 at 23:14 +0300, Benny Halevy wrote:
>>>>> struct nfs41_exchange_id_res is currently allocated on the stack
>>>>> insanely taking over 2K of stack space due to the NFS4_OPAQUE_LIMIT (1K(
>>>>> byte arrays embedded in server_owner and server_scope.
>>>>> Since these are not in use yet, this patch gets rid of them for the
>>>>> time being.
>>>>>
>>>>> Signed-off-by: Benny Halevy <bhalevy@xxxxxxxxxxx>
>>>>> ---
>>>>> fs/nfs/nfs4xdr.c | 27 ++++++++++++++-------------
>>>>> include/linux/nfs_xdr.h | 3 ---
>>>>> 2 files changed, 14 insertions(+), 16 deletions(-)
>>>>>
>>>>> diff --git a/fs/nfs/nfs4xdr.c b/fs/nfs/nfs4xdr.c
>>>>> index 80af0ae..3350d19 100644
>>>>> --- a/fs/nfs/nfs4xdr.c
>>>>> +++ b/fs/nfs/nfs4xdr.c
>>>>> @@ -4185,8 +4185,8 @@ static int decode_delegreturn(struct xdr_stream *xdr)
>>>>> static int decode_exchange_id(struct xdr_stream *xdr,
>>>>> struct nfs41_exchange_id_res *res)
>>>>> {
>>>>> - uint32_t *p;
>>>>> - int status, dummy;
>>>>> + uint32_t *p, dummy;
>> Oh, and 'p' _always_ has to be of type '__be32', otherwise the 'sparse'
>> checker will yell at you.
>
> Thanks! I'll send a fixed version
> of this patch and also look into the rest of the xdr code.
>
>>>>> + int status;
>>>>> struct nfs_client *clp = res->client;
>>>>>
>>>>> status = decode_op_hdr(xdr, OP_EXCHANGE_ID);
>>>>> @@ -4204,25 +4204,26 @@ static int decode_exchange_id(struct xdr_stream *xdr,
>>>>> if (dummy != SP4_NONE)
>>>>> return -EIO;
>>>>>
>>>>> - /* minor_id */
>>>>> + /* Throw away minor_id */
>>>>> READ_BUF(8);
>>>>> - READ64(res->server_owner.minor_id);
>>>>> + p += 8;
>>>> ^^^^^^^^ Err... This isn't the same thing at all!
>
> Ouch, of course. What did I smoke that day?
>
>>>> You're suddenly skipping 10=words instead of the original 2. READ_BUF()
>>>> will already take care of updating the 'p' pointer.
>
> Which p?
> It takes care of argp->p, not the local 'p' variable, doesn't it.
Grr, I meant "xdr->p" of course, via xdr_inline_decode
(argp is the server's READ_BUF, sigh)
> p += 2 has an equivalent side effect on 'p' as doing READ64.
> I can do "p = argp;" instead though to reset 'p' onto
p = xdr->p; ...
Benny
> the current xdr stream "head".
>
>>> BTW, this is exactly why the whole READ_BUF(), READ*(), WRITE*() macro
>>> crap is so utterly broken. The magic that happens to the 'p' pointer is
>>> completely opaque to someone unfamiliar with the code.
>
> I completely agree. We're dealing with bits and bytes (or 32 bit words
> actually) at the wrong abstraction layer.
>
>>>>>
>>>>> - /* Major id */
>>>>> + /* Throw away Major id */
>>>>> READ_BUF(4);
>>>>> - READ32(res->server_owner.major_id_sz);
>>>>> - READ_BUF(res->server_owner.major_id_sz);
>>>>> - COPYMEM(res->server_owner.major_id, res->server_owner.major_id_sz);
>>>>> + READ32(dummy);
>>>>> + READ_BUF(dummy);
>>>>> + p += XDR_QUADLEN(dummy);
>>>> ^^^^^^^^^^^^^^^^^^^^^^^^^ Ditto. You're skipping 2*dummy words.
>
> Why?
> READ_BUF increments argp->p in XDR_QUADLEN(dummy) words
> and the local p should be adjusted correspondingly.
> This used to happen as the side effect of COPY_MEM.
>
>>>>>
>>>>> - /* server_scope */
>>>>> + /* Throw away server_scope */
>>>>> READ_BUF(4);
>>>>> - READ32(res->server_scope.server_scope_sz);
>>>>> - READ_BUF(res->server_scope.server_scope_sz);
>>>>> - COPYMEM(res->server_scope.server_scope,
>>>>> - res->server_scope.server_scope_sz);
>>>>> + READ32(dummy);
>>>>> + READ_BUF(dummy);
>>>>> + p += XDR_QUADLEN(dummy);
>>>> ^^^^^^^^^^^^^^^^^^^^^^^^ Ditto
>
> Ditto
>
>>>>> +
>>>>> /* Throw away Implementation id array */
>>>>> READ_BUF(4);
>>>>> READ32(dummy);
>>>>> + READ_BUF(dummy);
>>>>> p += XDR_QUADLEN(dummy);
>>>> ^^^^^^^^^^^^^^^^^^^^^^^^ Ditto
>>>>
>
> and here too...
>
> BTW, Calling READ_BUF here was missing before, so this could be put
> in a patch of its own, and adjusting p is not really necessary since
> we're about to exit the function...
>
> In any case, are we going to squash these fixes into the respective
> queued patch, or would like to start accumulating the patches for
> 2.6.31 without rebasing?
>
> Benny
>
>>>>>
>>>>> return 0;
>>>>> diff --git a/include/linux/nfs_xdr.h b/include/linux/nfs_xdr.h
>>>>> index 071a6d1..62f63fb 100644
>>>>> --- a/include/linux/nfs_xdr.h
>>>>> +++ b/include/linux/nfs_xdr.h
>>>>> @@ -925,9 +925,6 @@ struct server_scope {
>>>>> struct nfs41_exchange_id_res {
>>>>> struct nfs_client *client;
>>>>> u32 flags;
>>>>> - struct server_owner server_owner;
>>>>> - struct server_scope server_scope;
>>>>> - struct nfs_impl_id4 impl_id;
>>>>> };
>>>>>
>>>>> struct nfs41_create_session_args {
>>
--
To unsubscribe from this list: send the line "unsubscribe linux-next" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
