diff mbox

[3/3] gssd: base the size of the fd array on the RLIMIT_NOFILE limit.

Message ID 20121211110228.0159fc4f@notabene.brown (mailing list archive)
State New, archived
Headers show

Commit Message

NeilBrown Dec. 11, 2012, 12:02 a.m. UTC
On Thu, 29 Nov 2012 11:30:51 +1100 NeilBrown <neilb@suse.de> wrote:

> On Wed, 28 Nov 2012 08:10:55 -0500 "J. Bruce Fields" <bfields@fieldses.org>
> wrote:
> 
> > On Wed, Nov 28, 2012 at 12:11:23PM +1100, Neil Brown wrote:
> > > We have previously raised the size of the 'pollarray' once (32 -> 256)
> > > and I have had another request to make it bigger.
> > > Rather than changing the hard-coded value, make it depend on
> > > RLIMIT_NOFILE.  This is an upper limit on the size of the array
> > > that can be passed to poll() anyway.
> > 
> > Sounds like a good idea.
> > 
> > Just out of curiosity: how does it fail?  I guess mounts just start
> > failing at some point--how do people find the workaround?
> 
> Error seems to be
> 
> rpcsec_gss: gss_init_sec_context: (major) Miscellaneous failure - (minor) Cannot contact any KDC for requested realm
> 
> in rpc.gssd logs.
> 
> I guess people could read the source to find the work around .... not ideal
> though.  I guess we should get gssd to generate some more helpful message.
> 
> The seem to be further problems that the customer is experiencing so I might
> wait until they are completely resolved to ensure I have complete
> understanding before I propose a further patch.

The "further problem" was that krb5 libraries use select() in a way that does
not support file descriptors higher than 1024.  This is fixed in the latest
krb5 so that is no longer an issue.

I've been thinking about your question, and about how best to deliver a fix
to customers, and I really think it should all "just work".
i.e. the array that gssd should be sized dynamically and RLIMIT_NOFILE should
be increased as needed.

I haven't tested this, but what do people think?   I don't have a problem
changing the rlim_cur limit like this, but I wonder if it is OK to
dynamically set rlim_max.

Thoughts?

NeilBrown

Comments

J. Bruce Fields Dec. 11, 2012, 4:16 p.m. UTC | #1
On Tue, Dec 11, 2012 at 11:02:28AM +1100, NeilBrown wrote:
> On Thu, 29 Nov 2012 11:30:51 +1100 NeilBrown <neilb@suse.de> wrote:
> 
> > On Wed, 28 Nov 2012 08:10:55 -0500 "J. Bruce Fields" <bfields@fieldses.org>
> > wrote:
> > 
> > > On Wed, Nov 28, 2012 at 12:11:23PM +1100, Neil Brown wrote:
> > > > We have previously raised the size of the 'pollarray' once (32 -> 256)
> > > > and I have had another request to make it bigger.
> > > > Rather than changing the hard-coded value, make it depend on
> > > > RLIMIT_NOFILE.  This is an upper limit on the size of the array
> > > > that can be passed to poll() anyway.
> > > 
> > > Sounds like a good idea.
> > > 
> > > Just out of curiosity: how does it fail?  I guess mounts just start
> > > failing at some point--how do people find the workaround?
> > 
> > Error seems to be
> > 
> > rpcsec_gss: gss_init_sec_context: (major) Miscellaneous failure - (minor) Cannot contact any KDC for requested realm
> > 
> > in rpc.gssd logs.
> > 
> > I guess people could read the source to find the work around .... not ideal
> > though.  I guess we should get gssd to generate some more helpful message.
> > 
> > The seem to be further problems that the customer is experiencing so I might
> > wait until they are completely resolved to ensure I have complete
> > understanding before I propose a further patch.
> 
> The "further problem" was that krb5 libraries use select() in a way that does
> not support file descriptors higher than 1024.  This is fixed in the latest
> krb5 so that is no longer an issue.
> 
> I've been thinking about your question, and about how best to deliver a fix
> to customers, and I really think it should all "just work".
> i.e. the array that gssd should be sized dynamically and RLIMIT_NOFILE should
> be increased as needed.

Neat-o.

> I haven't tested this, but what do people think?   I don't have a problem
> changing the rlim_cur limit like this, but I wonder if it is OK to
> dynamically set rlim_max.

What would be the concern, that we'd be depriving an admin of the
ability to set a limit?

We could instead set only the current limit and set set the max to an
admin-configurable quantity (default very large) when we start gssd.

But that sounds more complicated, and off hand I can't think of a reason
an admin would want to do that.

--b.

> 
> Thoughts?
> 
> NeilBrown
> 
> 
> diff --git a/utils/gssd/gssd_proc.c b/utils/gssd/gssd_proc.c
> index d01ba2f..3576a6f 100644
> --- a/utils/gssd/gssd_proc.c
> +++ b/utils/gssd/gssd_proc.c
> @@ -389,18 +389,36 @@ static int
>  get_poll_index(int *ind)
>  {
>  	unsigned int i;
> +	struct pollfd *new_pollarray;
> +	struct rlimit rlim;
>  
>  	*ind = -1;
>  	for (i=0; i<pollsize; i++) {
>  		if (pollarray[i].events == 0) {
>  			*ind = i;
> -			break;
> +			return 0;
>  		}
>  	}
> -	if (*ind == -1) {
> +
> +	new_pollarray = realloc(pollarray, pollsize * 2 * sizeof(*pollarray));
> +	if (!new_pollarray) {
>  		printerr(0, "ERROR: No pollarray slots open\n");
>  		return -1;
>  	}
> +	pollarray = new_pollarray;
> +	memset(pollarray + pollsize, 0, sizeof(*pollarray) * pollsize);
> +	*ind = pollsize;
> +	pollsize *= 2;
> +
> +	/* We will need lots of file descriptors too */
> +	if (getrlimit(RLIMIT_NOFILE, &rlim) == 0) {
> +		if (rlim.rlim_cur < pollsize+20) {
> +			rlim.rlim_cur = pollsize + 20;
> +			if (rlim.rlim_max < rlim.rlim_cur)
> +				rlim.rlim_max = rlim.rlim_cur;
> +			setrlimit(RLIMIT_NOFILE, &rlim);
> +		}
> +	}
>  	return 0;
>  }
>  
> @@ -473,13 +491,9 @@ fail_keep_client:
>  void
>  init_client_list(void)
>  {
> -	struct rlimit rlim;
>  	TAILQ_INIT(&clnt_list);
>  	/* Eventually plan to grow/shrink poll array: */
>  	pollsize = FD_ALLOC_BLOCK;
> -	if (getrlimit(RLIMIT_NOFILE, &rlim) < 0 &&
> -	    rlim.rlim_cur != RLIM_INFINITY)
> -		pollsize = rlim.rlim_cur;
>  	pollarray = calloc(pollsize, sizeof(struct pollfd));
>  }
>  


--
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
NeilBrown Dec. 13, 2012, 6:03 a.m. UTC | #2
On Tue, 11 Dec 2012 11:16:33 -0500 "J. Bruce Fields" <bfields@fieldses.org>
wrote:

> On Tue, Dec 11, 2012 at 11:02:28AM +1100, NeilBrown wrote:
> > On Thu, 29 Nov 2012 11:30:51 +1100 NeilBrown <neilb@suse.de> wrote:
> > 
> > > On Wed, 28 Nov 2012 08:10:55 -0500 "J. Bruce Fields" <bfields@fieldses.org>
> > > wrote:
> > > 
> > > > On Wed, Nov 28, 2012 at 12:11:23PM +1100, Neil Brown wrote:
> > > > > We have previously raised the size of the 'pollarray' once (32 -> 256)
> > > > > and I have had another request to make it bigger.
> > > > > Rather than changing the hard-coded value, make it depend on
> > > > > RLIMIT_NOFILE.  This is an upper limit on the size of the array
> > > > > that can be passed to poll() anyway.
> > > > 
> > > > Sounds like a good idea.
> > > > 
> > > > Just out of curiosity: how does it fail?  I guess mounts just start
> > > > failing at some point--how do people find the workaround?
> > > 
> > > Error seems to be
> > > 
> > > rpcsec_gss: gss_init_sec_context: (major) Miscellaneous failure - (minor) Cannot contact any KDC for requested realm
> > > 
> > > in rpc.gssd logs.
> > > 
> > > I guess people could read the source to find the work around .... not ideal
> > > though.  I guess we should get gssd to generate some more helpful message.
> > > 
> > > The seem to be further problems that the customer is experiencing so I might
> > > wait until they are completely resolved to ensure I have complete
> > > understanding before I propose a further patch.
> > 
> > The "further problem" was that krb5 libraries use select() in a way that does
> > not support file descriptors higher than 1024.  This is fixed in the latest
> > krb5 so that is no longer an issue.
> > 
> > I've been thinking about your question, and about how best to deliver a fix
> > to customers, and I really think it should all "just work".
> > i.e. the array that gssd should be sized dynamically and RLIMIT_NOFILE should
> > be increased as needed.
> 
> Neat-o.
> 
> > I haven't tested this, but what do people think?   I don't have a problem
> > changing the rlim_cur limit like this, but I wonder if it is OK to
> > dynamically set rlim_max.
> 
> What would be the concern, that we'd be depriving an admin of the
> ability to set a limit?

My concern in that automagically raising a so-called "hard limit" seems to be
subverting the very concept of it being a "limit".

> 
> We could instead set only the current limit and set set the max to an
> admin-configurable quantity (default very large) when we start gssd.

I really want to avoid any configuration.
The number of fds that will be used is directly connected to the number of
concurrent mounts - as there is no limit on those (I assume), I guess it is
fair that there is no limit on fds used by gssd.

> 
> But that sounds more complicated, and off hand I can't think of a reason
> an admin would want to do that.

OK, let's just modify the hard limit dynamically ... though I'm about to
disappear for summer holidays so I doubt you'll see anything for some weeks.


thanks,
NeilBrown
J. Bruce Fields Dec. 13, 2012, 7:21 p.m. UTC | #3
On Thu, Dec 13, 2012 at 05:03:22PM +1100, NeilBrown wrote:
> On Tue, 11 Dec 2012 11:16:33 -0500 "J. Bruce Fields" <bfields@fieldses.org>
> wrote:
> 
> > On Tue, Dec 11, 2012 at 11:02:28AM +1100, NeilBrown wrote:
> > > On Thu, 29 Nov 2012 11:30:51 +1100 NeilBrown <neilb@suse.de> wrote:
> > > 
> > > > On Wed, 28 Nov 2012 08:10:55 -0500 "J. Bruce Fields" <bfields@fieldses.org>
> > > > wrote:
> > > > 
> > > > > On Wed, Nov 28, 2012 at 12:11:23PM +1100, Neil Brown wrote:
> > > > > > We have previously raised the size of the 'pollarray' once (32 -> 256)
> > > > > > and I have had another request to make it bigger.
> > > > > > Rather than changing the hard-coded value, make it depend on
> > > > > > RLIMIT_NOFILE.  This is an upper limit on the size of the array
> > > > > > that can be passed to poll() anyway.
> > > > > 
> > > > > Sounds like a good idea.
> > > > > 
> > > > > Just out of curiosity: how does it fail?  I guess mounts just start
> > > > > failing at some point--how do people find the workaround?
> > > > 
> > > > Error seems to be
> > > > 
> > > > rpcsec_gss: gss_init_sec_context: (major) Miscellaneous failure - (minor) Cannot contact any KDC for requested realm
> > > > 
> > > > in rpc.gssd logs.
> > > > 
> > > > I guess people could read the source to find the work around .... not ideal
> > > > though.  I guess we should get gssd to generate some more helpful message.
> > > > 
> > > > The seem to be further problems that the customer is experiencing so I might
> > > > wait until they are completely resolved to ensure I have complete
> > > > understanding before I propose a further patch.
> > > 
> > > The "further problem" was that krb5 libraries use select() in a way that does
> > > not support file descriptors higher than 1024.  This is fixed in the latest
> > > krb5 so that is no longer an issue.
> > > 
> > > I've been thinking about your question, and about how best to deliver a fix
> > > to customers, and I really think it should all "just work".
> > > i.e. the array that gssd should be sized dynamically and RLIMIT_NOFILE should
> > > be increased as needed.
> > 
> > Neat-o.
> > 
> > > I haven't tested this, but what do people think?   I don't have a problem
> > > changing the rlim_cur limit like this, but I wonder if it is OK to
> > > dynamically set rlim_max.
> > 
> > What would be the concern, that we'd be depriving an admin of the
> > ability to set a limit?
> 
> My concern in that automagically raising a so-called "hard limit" seems to be
> subverting the very concept of it being a "limit".
> 
> > 
> > We could instead set only the current limit and set set the max to an
> > admin-configurable quantity (default very large) when we start gssd.
> 
> I really want to avoid any configuration.

Well, the init scripts (or whatever we use these days) would need to be
modified to set the max to RLIMIT_INFINITY by default, but the admin
shouldn't ever have to do anything.

But honestly I can't see any practical advantage to that, so...

> The number of fds that will be used is directly connected to the number of
> concurrent mounts - as there is no limit on those (I assume), I guess it is
> fair that there is no limit on fds used by gssd.
> 
> > 
> > But that sounds more complicated, and off hand I can't think of a reason
> > an admin would want to do that.
> 
> OK, let's just modify the hard limit dynamically ...

... fine by me.

> though I'm about to
> disappear for summer holidays so I doubt you'll see anything for some weeks.

"Summer holidays", huh.

Enjoy!

--b.
--
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 mbox

Patch

diff --git a/utils/gssd/gssd_proc.c b/utils/gssd/gssd_proc.c
index d01ba2f..3576a6f 100644
--- a/utils/gssd/gssd_proc.c
+++ b/utils/gssd/gssd_proc.c
@@ -389,18 +389,36 @@  static int
 get_poll_index(int *ind)
 {
 	unsigned int i;
+	struct pollfd *new_pollarray;
+	struct rlimit rlim;
 
 	*ind = -1;
 	for (i=0; i<pollsize; i++) {
 		if (pollarray[i].events == 0) {
 			*ind = i;
-			break;
+			return 0;
 		}
 	}
-	if (*ind == -1) {
+
+	new_pollarray = realloc(pollarray, pollsize * 2 * sizeof(*pollarray));
+	if (!new_pollarray) {
 		printerr(0, "ERROR: No pollarray slots open\n");
 		return -1;
 	}
+	pollarray = new_pollarray;
+	memset(pollarray + pollsize, 0, sizeof(*pollarray) * pollsize);
+	*ind = pollsize;
+	pollsize *= 2;
+
+	/* We will need lots of file descriptors too */
+	if (getrlimit(RLIMIT_NOFILE, &rlim) == 0) {
+		if (rlim.rlim_cur < pollsize+20) {
+			rlim.rlim_cur = pollsize + 20;
+			if (rlim.rlim_max < rlim.rlim_cur)
+				rlim.rlim_max = rlim.rlim_cur;
+			setrlimit(RLIMIT_NOFILE, &rlim);
+		}
+	}
 	return 0;
 }
 
@@ -473,13 +491,9 @@  fail_keep_client:
 void
 init_client_list(void)
 {
-	struct rlimit rlim;
 	TAILQ_INIT(&clnt_list);
 	/* Eventually plan to grow/shrink poll array: */
 	pollsize = FD_ALLOC_BLOCK;
-	if (getrlimit(RLIMIT_NOFILE, &rlim) < 0 &&
-	    rlim.rlim_cur != RLIM_INFINITY)
-		pollsize = rlim.rlim_cur;
 	pollarray = calloc(pollsize, sizeof(struct pollfd));
 }