Return-Path: Received: from mx1.redhat.com ([209.132.183.28]:45181 "EHLO mx1.redhat.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1751249AbcCVUvv (ORCPT ); Tue, 22 Mar 2016 16:51:51 -0400 Date: Tue, 22 Mar 2016 16:51:47 -0400 (EDT) From: Benjamin Coddington To: "J. Bruce Fields" cc: Jeff Layton , linux-nfs@vger.kernel.org Subject: Re: [PATCH v2] nfsd: use short read rather than i_size to set eof In-Reply-To: <20160322185336.GD4083@fieldses.org> Message-ID: References: <20160321213655.GB807@fieldses.org> <82e4c7a9756b21a4645421d04735ccf491b4296a.1458571329.git.bcodding@redhat.com> <20160322164624.GB4083@fieldses.org> <20160322185336.GD4083@fieldses.org> MIME-Version: 1.0 Content-Type: text/plain; charset=US-ASCII Sender: linux-nfs-owner@vger.kernel.org List-ID: On Tue, 22 Mar 2016, J. Bruce Fields wrote: > Possibly overkill, but I think I'll fold in something like this if you > don't see an objection. No objection. To answer your earlier question (does the client even send a read if it would extend past the size returned from a recent getattr?): Looks to me (by testing with a server that never sets eof) the answer is no. The client won't issue reads beyond the end of the file. Also the previous patch's subject is now a little misleading. Maybe s/rather than/as well as/ would do. Ben > commit 58e18a2a14a0 > Author: J. Bruce Fields > Date: Tue Mar 22 14:08:11 2016 -0400 > > nfsd: document read eof logic > > The choice of checks here is a little subtle, let's document this for > posterity. > > Signed-off-by: J. Bruce Fields > > diff --git a/fs/nfsd/nfs3proc.c b/fs/nfsd/nfs3proc.c > index 83c9abb33e8b..df0f0a86f21d 100644 > --- a/fs/nfsd/nfs3proc.c > +++ b/fs/nfsd/nfs3proc.c > @@ -168,8 +168,7 @@ nfsd3_proc_read(struct svc_rqst *rqstp, struct nfsd3_readargs *argp, > &resp->count); > if (nfserr == 0) { > struct inode *inode = d_inode(resp->fh.fh_dentry); > - resp->eof = (cnt > resp->count) || > - ((argp->offset + resp->count) >= inode->i_size); > + resp->eof = nfsd_eof_on_read(cnt, resp->count, argp->offset, inode->i_size); > } > > RETURN_STATUS(nfserr); > diff --git a/fs/nfsd/nfs4xdr.c b/fs/nfsd/nfs4xdr.c > index 90232bd7e498..9df898ba648f 100644 > --- a/fs/nfsd/nfs4xdr.c > +++ b/fs/nfsd/nfs4xdr.c > @@ -3387,8 +3387,8 @@ static __be32 nfsd4_encode_splice_read( > return nfserr; > } > > - eof = (len > maxcount) || > - ((read->rd_offset + maxcount >= d_inode(read->rd_fhp->fh_dentry)->i_size)); > + eof = nfsd_eof_on_read(len, maxcount, read->rd_offset, > + d_inode(read->rd_fhp->fh_dentry)->i_size); > > *(p++) = htonl(eof); > *(p++) = htonl(maxcount); > @@ -3465,8 +3465,8 @@ static __be32 nfsd4_encode_readv(struct nfsd4_compoundres *resp, > return nfserr; > xdr_truncate_encode(xdr, starting_len + 8 + ((maxcount+3)&~3)); > > - eof = (len > maxcount) || > - ((read->rd_offset + maxcount >= d_inode(read->rd_fhp->fh_dentry)->i_size)); > + eof = nfsd_eof_on_read(len, maxcount, read->rd_offset, > + d_inode(read->rd_fhp->fh_dentry)->i_size); > > tmp = htonl(eof); > write_bytes_to_xdr_buf(xdr->buf, starting_len , &tmp, 4); > diff --git a/fs/nfsd/vfs.h b/fs/nfsd/vfs.h > index c11ba316f23f..6244e073c137 100644 > --- a/fs/nfsd/vfs.h > +++ b/fs/nfsd/vfs.h > @@ -139,4 +139,24 @@ static inline int nfsd_create_is_exclusive(int createmode) > || createmode == NFS4_CREATE_EXCLUSIVE4_1; > } > > +static inline bool nfsd_eof_on_read(long requested, long read, > + loff_t offset, loff_t size) > +{ > + /* We assume a short read means eof: */ > + if (requested > read) > + return true; > + /* > + * A non-short read might also reach end of file. The spec > + * still requires us to set eof in that case. > + * > + * Further operations may have modified the file size since > + * the read, so the following check is not atomic with the read. > + * The only case we've seen that cause a problem for a client > + * is the case where the read returned a count of 0 without > + * setting eof. That case was fixed by the addition of the > + * above check. > + */ > + return (offset + read >= size); > +} > + > #endif /* LINUX_NFSD_VFS_H */ >