| Message ID | 1302261406-9048-2-git-send-email-sean.finney@sonyericsson.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
Looks sensible to me.--b. On Fri, Apr 08, 2011 at 01:16:46PM +0200, Sean Finney wrote: > Previously, when writing to /proc/net/rpc/*/channel, if a cache line > were larger than the default buffer size (likely 1024 bytes), mountd > and svcgssd would split writes into a number of buffer-sized writes. > Each of these writes would get an EINVAL error back from the kernel > procfs handle (it expects line-oriented input and does not account for > multiple/split writes), and no cache update would occur. > > When such behavior occurs, NFS clients depending on mountd to finish > the cache operation would block/hang, or receive EPERM, depending on > the context of the operation. This is likely to happen if a user is a > member of a large (~100-200) number of groups. > > Instead, every fopen() on the procfs files in question is followed by > a call to setvbuf(), using a per-file dedicated buffer of > RPC_CHAN_BUF_SIZE length. > > Really, mountd should not be using stdio-style buffered file operations > on files in /proc to begin with. A better solution would be to use > internally managed buffers and calls to write() instead of these stdio > calls, but that would be a more extensive change; so this is proposed > as a quick and not-so-dirty fix in the meantime. > > Signed-off-by: Sean Finney <sean.finney@sonyericsson.com> > --- > support/include/misc.h | 3 +++ > utils/gssd/svcgssd_proc.c | 3 +++ > utils/mountd/cache.c | 2 ++ > 3 files changed, 8 insertions(+), 0 deletions(-) > > diff --git a/support/include/misc.h b/support/include/misc.h > index 9a1b25d..7e3874e 100644 > --- a/support/include/misc.h > +++ b/support/include/misc.h > @@ -24,4 +24,7 @@ struct hostent *get_reliable_hostbyaddr(const char *addr, int len, int type); > > extern int is_mountpoint(char *path); > > +/* size of the file pointer buffers for rpc procfs files */ > +#define RPC_CHAN_BUF_SIZE 32768 > + > #endif /* MISC_H */ > diff --git a/utils/gssd/svcgssd_proc.c b/utils/gssd/svcgssd_proc.c > index 6f2ba61..8f6548e 100644 > --- a/utils/gssd/svcgssd_proc.c > +++ b/utils/gssd/svcgssd_proc.c > @@ -56,6 +56,7 @@ > #include "gss_util.h" > #include "err_util.h" > #include "context.h" > +#include "misc.h" > > extern char * mech2file(gss_OID mech); > #define SVCGSSD_CONTEXT_CHANNEL "/proc/net/rpc/auth.rpcsec.context/channel" > @@ -78,6 +79,7 @@ do_svc_downcall(gss_buffer_desc *out_handle, struct svc_cred *cred, > FILE *f; > int i; > char *fname = NULL; > + char vbuf[RPC_CHAN_BUF_SIZE]; > int err; > > printerr(1, "doing downcall\n"); > @@ -90,6 +92,7 @@ do_svc_downcall(gss_buffer_desc *out_handle, struct svc_cred *cred, > SVCGSSD_CONTEXT_CHANNEL, strerror(errno)); > goto out_err; > } > + setvbuf(f, vbuf, _IOLBF, RPC_CHAN_BUF_SIZE); > qword_printhex(f, out_handle->value, out_handle->length); > /* XXX are types OK for the rest of this? */ > /* For context cache, use the actual context endtime */ > diff --git a/utils/mountd/cache.c b/utils/mountd/cache.c > index 34808cd..bd67157 100644 > --- a/utils/mountd/cache.c > +++ b/utils/mountd/cache.c > @@ -739,6 +739,7 @@ struct { > char *cache_name; > void (*cache_handle)(FILE *f); > FILE *f; > + char vbuf[RPC_CHAN_BUF_SIZE]; > } cachelist[] = { > { "auth.unix.ip", auth_unix_ip}, > { "auth.unix.gid", auth_unix_gid}, > @@ -757,6 +758,7 @@ void cache_open(void) > continue; > sprintf(path, "/proc/net/rpc/%s/channel", cachelist[i].cache_name); > cachelist[i].f = fopen(path, "r+"); > + setvbuf(cachelist[i].f, cachelist[i].vbuf, _IOLBF, RPC_CHAN_BUF_SIZE); > } > } > > -- > 1.7.4.1 > > -- > To unsubscribe from this list: send the line "unsubscribe linux-nfs" in > the body of a message to majordomo@vger.kernel.org > More majordomo info at http://vger.kernel.org/majordomo-info.html -- To unsubscribe from this list: send the line "unsubscribe linux-nfs" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
diff --git a/support/include/misc.h b/support/include/misc.h index 9a1b25d..7e3874e 100644 --- a/support/include/misc.h +++ b/support/include/misc.h @@ -24,4 +24,7 @@ struct hostent *get_reliable_hostbyaddr(const char *addr, int len, int type); extern int is_mountpoint(char *path); +/* size of the file pointer buffers for rpc procfs files */ +#define RPC_CHAN_BUF_SIZE 32768 + #endif /* MISC_H */ diff --git a/utils/gssd/svcgssd_proc.c b/utils/gssd/svcgssd_proc.c index 6f2ba61..8f6548e 100644 --- a/utils/gssd/svcgssd_proc.c +++ b/utils/gssd/svcgssd_proc.c @@ -56,6 +56,7 @@ #include "gss_util.h" #include "err_util.h" #include "context.h" +#include "misc.h" extern char * mech2file(gss_OID mech); #define SVCGSSD_CONTEXT_CHANNEL "/proc/net/rpc/auth.rpcsec.context/channel" @@ -78,6 +79,7 @@ do_svc_downcall(gss_buffer_desc *out_handle, struct svc_cred *cred, FILE *f; int i; char *fname = NULL; + char vbuf[RPC_CHAN_BUF_SIZE]; int err; printerr(1, "doing downcall\n"); @@ -90,6 +92,7 @@ do_svc_downcall(gss_buffer_desc *out_handle, struct svc_cred *cred, SVCGSSD_CONTEXT_CHANNEL, strerror(errno)); goto out_err; } + setvbuf(f, vbuf, _IOLBF, RPC_CHAN_BUF_SIZE); qword_printhex(f, out_handle->value, out_handle->length); /* XXX are types OK for the rest of this? */ /* For context cache, use the actual context endtime */ diff --git a/utils/mountd/cache.c b/utils/mountd/cache.c index 34808cd..bd67157 100644 --- a/utils/mountd/cache.c +++ b/utils/mountd/cache.c @@ -739,6 +739,7 @@ struct { char *cache_name; void (*cache_handle)(FILE *f); FILE *f; + char vbuf[RPC_CHAN_BUF_SIZE]; } cachelist[] = { { "auth.unix.ip", auth_unix_ip}, { "auth.unix.gid", auth_unix_gid}, @@ -757,6 +758,7 @@ void cache_open(void) continue; sprintf(path, "/proc/net/rpc/%s/channel", cachelist[i].cache_name); cachelist[i].f = fopen(path, "r+"); + setvbuf(cachelist[i].f, cachelist[i].vbuf, _IOLBF, RPC_CHAN_BUF_SIZE); } }
Previously, when writing to /proc/net/rpc/*/channel, if a cache line were larger than the default buffer size (likely 1024 bytes), mountd and svcgssd would split writes into a number of buffer-sized writes. Each of these writes would get an EINVAL error back from the kernel procfs handle (it expects line-oriented input and does not account for multiple/split writes), and no cache update would occur. When such behavior occurs, NFS clients depending on mountd to finish the cache operation would block/hang, or receive EPERM, depending on the context of the operation. This is likely to happen if a user is a member of a large (~100-200) number of groups. Instead, every fopen() on the procfs files in question is followed by a call to setvbuf(), using a per-file dedicated buffer of RPC_CHAN_BUF_SIZE length. Really, mountd should not be using stdio-style buffered file operations on files in /proc to begin with. A better solution would be to use internally managed buffers and calls to write() instead of these stdio calls, but that would be a more extensive change; so this is proposed as a quick and not-so-dirty fix in the meantime. Signed-off-by: Sean Finney <sean.finney@sonyericsson.com> --- support/include/misc.h | 3 +++ utils/gssd/svcgssd_proc.c | 3 +++ utils/mountd/cache.c | 2 ++ 3 files changed, 8 insertions(+), 0 deletions(-)