2008-06-13 21:17:45

by [email protected]

[permalink] [raw]
Subject: Re: [PATCH 08/37] SUNRPC: Use GFP_NOFS when allocating credentials

On Thu, Jun 12, 2008 at 03:22:00PM -0400, Trond Myklebust wrote:
> Since the credentials may be allocated during the call to rpc_new_task(),
> which again may be called by a memory allocator...

Most of these are only called in gssd process's context when it
performs the downcall. Which doesn't change the fact that they're
needed for the rpc_new_task() to succeed, OK, so this isn't an objection
to this patch--but is there any plan for how we'll deal with e.g. memory
allocations that gssd does on its own?

--b.

>
> Signed-off-by: Trond Myklebust <[email protected]>
> ---
>
> net/sunrpc/auth_gss/auth_gss.c | 10 +++++-----
> net/sunrpc/auth_gss/gss_krb5_mech.c | 4 ++--
> net/sunrpc/auth_gss/gss_spkm3_mech.c | 4 ++--
> net/sunrpc/auth_gss/gss_spkm3_token.c | 2 +-
> net/sunrpc/auth_unix.c | 2 +-
> 5 files changed, 11 insertions(+), 11 deletions(-)
>
> diff --git a/net/sunrpc/auth_gss/auth_gss.c b/net/sunrpc/auth_gss/auth_gss.c
> index cc12d5f..bf7585b 100644
> --- a/net/sunrpc/auth_gss/auth_gss.c
> +++ b/net/sunrpc/auth_gss/auth_gss.c
> @@ -146,7 +146,7 @@ simple_get_netobj(const void *p, const void *end, struct xdr_netobj *dest)
> q = (const void *)((const char *)p + len);
> if (unlikely(q > end || q < p))
> return ERR_PTR(-EFAULT);
> - dest->data = kmemdup(p, len, GFP_KERNEL);
> + dest->data = kmemdup(p, len, GFP_NOFS);
> if (unlikely(dest->data == NULL))
> return ERR_PTR(-ENOMEM);
> dest->len = len;
> @@ -171,7 +171,7 @@ gss_alloc_context(void)
> {
> struct gss_cl_ctx *ctx;
>
> - ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
> + ctx = kzalloc(sizeof(*ctx), GFP_NOFS);
> if (ctx != NULL) {
> ctx->gc_proc = RPC_GSS_PROC_DATA;
> ctx->gc_seq = 1; /* NetApp 6.4R1 doesn't accept seq. no. 0 */
> @@ -341,7 +341,7 @@ gss_alloc_msg(struct gss_auth *gss_auth, uid_t uid)
> {
> struct gss_upcall_msg *gss_msg;
>
> - gss_msg = kzalloc(sizeof(*gss_msg), GFP_KERNEL);
> + gss_msg = kzalloc(sizeof(*gss_msg), GFP_NOFS);
> if (gss_msg != NULL) {
> INIT_LIST_HEAD(&gss_msg->list);
> rpc_init_wait_queue(&gss_msg->rpc_waitqueue, "RPCSEC_GSS upcall waitq");
> @@ -503,7 +503,7 @@ gss_pipe_downcall(struct file *filp, const char __user *src, size_t mlen)
> if (mlen > MSG_BUF_MAXSIZE)
> goto out;
> err = -ENOMEM;
> - buf = kmalloc(mlen, GFP_KERNEL);
> + buf = kmalloc(mlen, GFP_NOFS);
> if (!buf)
> goto out;
>
> @@ -806,7 +806,7 @@ gss_create_cred(struct rpc_auth *auth, struct auth_cred *acred, int flags)
> dprintk("RPC: gss_create_cred for uid %d, flavor %d\n",
> acred->uid, auth->au_flavor);
>
> - if (!(cred = kzalloc(sizeof(*cred), GFP_KERNEL)))
> + if (!(cred = kzalloc(sizeof(*cred), GFP_NOFS)))
> goto out_err;
>
> rpcauth_init_cred(&cred->gc_base, acred, auth, &gss_credops);
> diff --git a/net/sunrpc/auth_gss/gss_krb5_mech.c b/net/sunrpc/auth_gss/gss_krb5_mech.c
> index 60c3dba..ef45eba 100644
> --- a/net/sunrpc/auth_gss/gss_krb5_mech.c
> +++ b/net/sunrpc/auth_gss/gss_krb5_mech.c
> @@ -70,7 +70,7 @@ simple_get_netobj(const void *p, const void *end, struct xdr_netobj *res)
> q = (const void *)((const char *)p + len);
> if (unlikely(q > end || q < p))
> return ERR_PTR(-EFAULT);
> - res->data = kmemdup(p, len, GFP_KERNEL);
> + res->data = kmemdup(p, len, GFP_NOFS);
> if (unlikely(res->data == NULL))
> return ERR_PTR(-ENOMEM);
> res->len = len;
> @@ -131,7 +131,7 @@ gss_import_sec_context_kerberos(const void *p,
> struct krb5_ctx *ctx;
> int tmp;
>
> - if (!(ctx = kzalloc(sizeof(*ctx), GFP_KERNEL)))
> + if (!(ctx = kzalloc(sizeof(*ctx), GFP_NOFS)))
> goto out_err;
>
> p = simple_get_bytes(p, end, &ctx->initiate, sizeof(ctx->initiate));
> diff --git a/net/sunrpc/auth_gss/gss_spkm3_mech.c b/net/sunrpc/auth_gss/gss_spkm3_mech.c
> index 5deb4b6..035e1dd 100644
> --- a/net/sunrpc/auth_gss/gss_spkm3_mech.c
> +++ b/net/sunrpc/auth_gss/gss_spkm3_mech.c
> @@ -76,7 +76,7 @@ simple_get_netobj(const void *p, const void *end, struct xdr_netobj *res)
> q = (const void *)((const char *)p + len);
> if (unlikely(q > end || q < p))
> return ERR_PTR(-EFAULT);
> - res->data = kmemdup(p, len, GFP_KERNEL);
> + res->data = kmemdup(p, len, GFP_NOFS);
> if (unlikely(res->data == NULL))
> return ERR_PTR(-ENOMEM);
> return q;
> @@ -90,7 +90,7 @@ gss_import_sec_context_spkm3(const void *p, size_t len,
> struct spkm3_ctx *ctx;
> int version;
>
> - if (!(ctx = kzalloc(sizeof(*ctx), GFP_KERNEL)))
> + if (!(ctx = kzalloc(sizeof(*ctx), GFP_NOFS)))
> goto out_err;
>
> p = simple_get_bytes(p, end, &version, sizeof(version));
> diff --git a/net/sunrpc/auth_gss/gss_spkm3_token.c b/net/sunrpc/auth_gss/gss_spkm3_token.c
> index 6cdd241..3308157 100644
> --- a/net/sunrpc/auth_gss/gss_spkm3_token.c
> +++ b/net/sunrpc/auth_gss/gss_spkm3_token.c
> @@ -90,7 +90,7 @@ asn1_bitstring_len(struct xdr_netobj *in, int *enclen, int *zerobits)
> int
> decode_asn1_bitstring(struct xdr_netobj *out, char *in, int enclen, int explen)
> {
> - if (!(out->data = kzalloc(explen,GFP_KERNEL)))
> + if (!(out->data = kzalloc(explen,GFP_NOFS)))
> return 0;
> out->len = explen;
> memcpy(out->data, in, enclen);
> diff --git a/net/sunrpc/auth_unix.c b/net/sunrpc/auth_unix.c
> index 44920b9..46b2647 100644
> --- a/net/sunrpc/auth_unix.c
> +++ b/net/sunrpc/auth_unix.c
> @@ -66,7 +66,7 @@ unx_create_cred(struct rpc_auth *auth, struct auth_cred *acred, int flags)
> dprintk("RPC: allocating UNIX cred for uid %d gid %d\n",
> acred->uid, acred->gid);
>
> - if (!(cred = kmalloc(sizeof(*cred), GFP_KERNEL)))
> + if (!(cred = kmalloc(sizeof(*cred), GFP_NOFS)))
> return ERR_PTR(-ENOMEM);
>
> rpcauth_init_cred(&cred->uc_base, acred, auth, &unix_credops);
>
> --
> To unsubscribe from this list: send the line "unsubscribe linux-nfs" in
> the body of a message to [email protected]
> More majordomo info at http://vger.kernel.org/majordomo-info.html


2008-06-13 21:27:11

by Myklebust, Trond

[permalink] [raw]
Subject: Re: [PATCH 08/37] SUNRPC: Use GFP_NOFS when allocating credentials

On Fri, 2008-06-13 at 17:17 -0400, J. Bruce Fields wrote:
> On Thu, Jun 12, 2008 at 03:22:00PM -0400, Trond Myklebust wrote:
> > Since the credentials may be allocated during the call to rpc_new_task(),
> > which again may be called by a memory allocator...
>
> Most of these are only called in gssd process's context when it
> performs the downcall. Which doesn't change the fact that they're
> needed for the rpc_new_task() to succeed, OK, so this isn't an objection
> to this patch--but is there any plan for how we'll deal with e.g. memory
> allocations that gssd does on its own?

You mean alloc_enc_pages()? We should really try to avoid sleeping
there...

--
Trond Myklebust
Linux NFS client maintainer

NetApp
[email protected]
http://www.netapp.com

2008-06-13 21:31:48

by [email protected]

[permalink] [raw]
Subject: Re: [PATCH 08/37] SUNRPC: Use GFP_NOFS when allocating credentials

On Fri, Jun 13, 2008 at 05:26:17PM -0400, Trond Myklebust wrote:
> On Fri, 2008-06-13 at 17:17 -0400, J. Bruce Fields wrote:
> > On Thu, Jun 12, 2008 at 03:22:00PM -0400, Trond Myklebust wrote:
> > > Since the credentials may be allocated during the call to rpc_new_task(),
> > > which again may be called by a memory allocator...
> >
> > Most of these are only called in gssd process's context when it
> > performs the downcall. Which doesn't change the fact that they're
> > needed for the rpc_new_task() to succeed, OK, so this isn't an objection
> > to this patch--but is there any plan for how we'll deal with e.g. memory
> > allocations that gssd does on its own?
>
> You mean alloc_enc_pages()? We should really try to avoid sleeping
> there...

No, that may be a problem too, but I mean: if we're worried about a
deadlock due to some allocations that gssd performs on the downcall at
the end of the context initiation, then shouldn't we also be worried
about all the other allocations that gssd performs itself?

--b.

2008-06-13 21:34:57

by Myklebust, Trond

[permalink] [raw]
Subject: Re: [PATCH 08/37] SUNRPC: Use GFP_NOFS when allocating credentials

On Fri, 2008-06-13 at 17:31 -0400, J. Bruce Fields wrote:
> No, that may be a problem too, but I mean: if we're worried about a
> deadlock due to some allocations that gssd performs on the downcall at
> the end of the context initiation, then shouldn't we also be worried
> about all the other allocations that gssd performs itself?

Definitely.

--
Trond Myklebust
Linux NFS client maintainer

NetApp
[email protected]
http://www.netapp.com

2008-06-13 21:58:05

by [email protected]

[permalink] [raw]
Subject: Re: [PATCH 08/37] SUNRPC: Use GFP_NOFS when allocating credentials

On Fri, Jun 13, 2008 at 05:34:54PM -0400, Trond Myklebust wrote:
> On Fri, 2008-06-13 at 17:31 -0400, J. Bruce Fields wrote:
> > No, that may be a problem too, but I mean: if we're worried about a
> > deadlock due to some allocations that gssd performs on the downcall at
> > the end of the context initiation, then shouldn't we also be worried
> > about all the other allocations that gssd performs itself?
>
> Definitely.

Do you have any ideas (however speculative) about how we'll deal with
that?

--b.

2008-06-13 22:07:52

by Myklebust, Trond

[permalink] [raw]
Subject: Re: [PATCH 08/37] SUNRPC: Use GFP_NOFS when allocating credentials

On Fri, 2008-06-13 at 17:58 -0400, J. Bruce Fields wrote:
> On Fri, Jun 13, 2008 at 05:34:54PM -0400, Trond Myklebust wrote:
> > On Fri, 2008-06-13 at 17:31 -0400, J. Bruce Fields wrote:
> > > No, that may be a problem too, but I mean: if we're worried about a
> > > deadlock due to some allocations that gssd performs on the downcall at
> > > the end of the context initiation, then shouldn't we also be worried
> > > about all the other allocations that gssd performs itself?
> >
> > Definitely.
>
> Do you have any ideas (however speculative) about how we'll deal with
> that?

Not really: AFAIK there are no mechanisms for fully pre-provisioning the
necessary resources that a userland app may need to ensure it can
function. mlock() won't, for instance, allow you to pre-provision the
socket buffer space that the kerberos stuff needs to communicate...

--
Trond Myklebust
Linux NFS client maintainer

NetApp
[email protected]
http://www.netapp.com