I believe this field should be called field_count instead of file_count.
Correct the doc with the same.
Signed-off-by: Joel Fernandes (Google) <[email protected]>
---
Documentation/RCU/listRCU.txt | 4 ++--
1 file changed, 2 insertions(+), 2 deletions(-)
diff --git a/Documentation/RCU/listRCU.txt b/Documentation/RCU/listRCU.txt
index adb5a3782846..190e666fc359 100644
--- a/Documentation/RCU/listRCU.txt
+++ b/Documentation/RCU/listRCU.txt
@@ -175,7 +175,7 @@ otherwise, the added fields would need to be filled in):
list_for_each_entry(e, list, list) {
if (!audit_compare_rule(rule, &e->rule)) {
e->rule.action = newaction;
- e->rule.file_count = newfield_count;
+ e->rule.field_count = newfield_count;
write_unlock(&auditsc_lock);
return 0;
}
@@ -204,7 +204,7 @@ RCU ("read-copy update") its name. The RCU code is as follows:
return -ENOMEM;
audit_copy_rule(&ne->rule, &e->rule);
ne->rule.action = newaction;
- ne->rule.file_count = newfield_count;
+ ne->rule.field_count = newfield_count;
list_replace_rcu(&e->list, &ne->list);
call_rcu(&e->rcu, audit_free_rule);
return 0;
--
2.21.0.1020.gf2820cf01a-goog
On Sat, May 04, 2019 at 10:03:10PM -0400, Joel Fernandes (Google) wrote:
> I believe this field should be called field_count instead of file_count.
> Correct the doc with the same.
>
> Signed-off-by: Joel Fernandes (Google) <[email protected]>
But if we are going to update this, why not update it with the current
audit_filter_task(), audit_del_rule(), and audit_add_rule() code?
Hmmm... One reason is that some of them have changed beyond recognition.
And this example code predates v2.6.12. ;-)
So good eyes, but I believe that this really does reflect the ancient
code...
On the other hand, would you have ideas for more modern replacement
examples?
Thanx, Paul
> ---
> Documentation/RCU/listRCU.txt | 4 ++--
> 1 file changed, 2 insertions(+), 2 deletions(-)
>
> diff --git a/Documentation/RCU/listRCU.txt b/Documentation/RCU/listRCU.txt
> index adb5a3782846..190e666fc359 100644
> --- a/Documentation/RCU/listRCU.txt
> +++ b/Documentation/RCU/listRCU.txt
> @@ -175,7 +175,7 @@ otherwise, the added fields would need to be filled in):
> list_for_each_entry(e, list, list) {
> if (!audit_compare_rule(rule, &e->rule)) {
> e->rule.action = newaction;
> - e->rule.file_count = newfield_count;
> + e->rule.field_count = newfield_count;
> write_unlock(&auditsc_lock);
> return 0;
> }
> @@ -204,7 +204,7 @@ RCU ("read-copy update") its name. The RCU code is as follows:
> return -ENOMEM;
> audit_copy_rule(&ne->rule, &e->rule);
> ne->rule.action = newaction;
> - ne->rule.file_count = newfield_count;
> + ne->rule.field_count = newfield_count;
> list_replace_rcu(&e->list, &ne->list);
> call_rcu(&e->rcu, audit_free_rule);
> return 0;
> --
> 2.21.0.1020.gf2820cf01a-goog
>
On Mon, May 06, 2019 at 05:04:53PM -0700, Paul E. McKenney wrote:
> On Sat, May 04, 2019 at 10:03:10PM -0400, Joel Fernandes (Google) wrote:
> > I believe this field should be called field_count instead of file_count.
> > Correct the doc with the same.
> >
> > Signed-off-by: Joel Fernandes (Google) <[email protected]>
>
> But if we are going to update this, why not update it with the current
> audit_filter_task(), audit_del_rule(), and audit_add_rule() code?
>
> Hmmm... One reason is that some of them have changed beyond recognition.
It seems to me that these 3 functions are just structured differently but is
conceptually the same.
There is now an array of lists stored in audit_filter_list. Each list is a
set of rules. Versus in the listRCU.txt, there is only one global.
The other difference is there is a mutex held &audit_filter_mutex
audit_{add,del}_rule. Where as in listRCU, it says that is not needed since
another mutex is already held.
> And this example code predates v2.6.12. ;-)
>
> So good eyes, but I believe that this really does reflect the ancient
> code...
>
> On the other hand, would you have ideas for more modern replacement
> examples?
There are 3 cases I can see in listRCU.txt:
(1) action taken outside of read_lock (can tolerate stale data), no in-place update.
this is the best possible usage of RCU.
(2) action taken outside of read_lock, in-place updates
this is good as long as not too many in-place updates.
involves copying creating new list node and replacing the
node being updated with it.
(3) cannot tolerate stale data: here a deleted or obsolete flag can be used
protected by a per-entry lock. reader
aborts if object is stale.
Any replacement example must make satisfy (3) too?
The only example for (3) that I know of is sysvipc sempahores which you also
mentioned in the paper. Looking through this code, it hasn't changed
conceptually and it could be a fit for an example (ipc_valid_object() checks
for whether the object is stale).
The other example could be dentry look up which uses seqlocks for the
RCU-walk case? But that could be too complex. This is also something I first
learnt from the paper and then the excellent path-lookup.rst document in
kernel sources.
I will keep any eye out for other examples in the kernel code as well.
Let me know what you think, thanks!
- Joel
> > ---
> > Documentation/RCU/listRCU.txt | 4 ++--
> > 1 file changed, 2 insertions(+), 2 deletions(-)
> >
> > diff --git a/Documentation/RCU/listRCU.txt b/Documentation/RCU/listRCU.txt
> > index adb5a3782846..190e666fc359 100644
> > --- a/Documentation/RCU/listRCU.txt
> > +++ b/Documentation/RCU/listRCU.txt
> > @@ -175,7 +175,7 @@ otherwise, the added fields would need to be filled in):
> > list_for_each_entry(e, list, list) {
> > if (!audit_compare_rule(rule, &e->rule)) {
> > e->rule.action = newaction;
> > - e->rule.file_count = newfield_count;
> > + e->rule.field_count = newfield_count;
> > write_unlock(&auditsc_lock);
> > return 0;
> > }
> > @@ -204,7 +204,7 @@ RCU ("read-copy update") its name. The RCU code is as follows:
> > return -ENOMEM;
> > audit_copy_rule(&ne->rule, &e->rule);
> > ne->rule.action = newaction;
> > - ne->rule.file_count = newfield_count;
> > + ne->rule.field_count = newfield_count;
> > list_replace_rcu(&e->list, &ne->list);
> > call_rcu(&e->rcu, audit_free_rule);
> > return 0;
> > --
> > 2.21.0.1020.gf2820cf01a-goog
> >
>
On Wed, May 08, 2019 at 12:26:35PM -0400, Joel Fernandes wrote:
> On Mon, May 06, 2019 at 05:04:53PM -0700, Paul E. McKenney wrote:
> > On Sat, May 04, 2019 at 10:03:10PM -0400, Joel Fernandes (Google) wrote:
> > > I believe this field should be called field_count instead of file_count.
> > > Correct the doc with the same.
> > >
> > > Signed-off-by: Joel Fernandes (Google) <[email protected]>
> >
> > But if we are going to update this, why not update it with the current
> > audit_filter_task(), audit_del_rule(), and audit_add_rule() code?
> >
> > Hmmm... One reason is that some of them have changed beyond recognition.
>
> It seems to me that these 3 functions are just structured differently but is
> conceptually the same.
>
> There is now an array of lists stored in audit_filter_list. Each list is a
> set of rules. Versus in the listRCU.txt, there is only one global.
>
> The other difference is there is a mutex held &audit_filter_mutex
> audit_{add,del}_rule. Where as in listRCU, it says that is not needed since
> another mutex is already held.
Agreed.
> > And this example code predates v2.6.12. ;-)
> >
> > So good eyes, but I believe that this really does reflect the ancient
> > code...
> >
> > On the other hand, would you have ideas for more modern replacement
> > examples?
>
> There are 3 cases I can see in listRCU.txt:
> (1) action taken outside of read_lock (can tolerate stale data), no in-place update.
> this is the best possible usage of RCU.
> (2) action taken outside of read_lock, in-place updates
> this is good as long as not too many in-place updates.
> involves copying creating new list node and replacing the
> node being updated with it.
> (3) cannot tolerate stale data: here a deleted or obsolete flag can be used
> protected by a per-entry lock. reader
> aborts if object is stale.
>
> Any replacement example must make satisfy (3) too?
It would be OK to have a separate example for (3). It would of course
be nicer to have one example for all three, but not all -that- important.
> The only example for (3) that I know of is sysvipc sempahores which you also
> mentioned in the paper. Looking through this code, it hasn't changed
> conceptually and it could be a fit for an example (ipc_valid_object() checks
> for whether the object is stale).
That is indeed the classic canonical example. ;-)
> The other example could be dentry look up which uses seqlocks for the
> RCU-walk case? But that could be too complex. This is also something I first
> learnt from the paper and then the excellent path-lookup.rst document in
> kernel sources.
This is a great example, but it would need serious simplification for
use in the Documentation/RCU directory. Note that dcache uses it to
gain very limited and targeted consistency -- only a few types of updates
acquire the write-side of that seqlock.
Might be quite worthwhile to have a simplified example, though!
Perhaps a trivial hash table where write-side sequence lock is acquired
only when moving an element from one chain to another?
> I will keep any eye out for other examples in the kernel code as well.
Very good!
Thanx, Paul
> Let me know what you think, thanks!
>
> - Joel
>
>
> > > ---
> > > Documentation/RCU/listRCU.txt | 4 ++--
> > > 1 file changed, 2 insertions(+), 2 deletions(-)
> > >
> > > diff --git a/Documentation/RCU/listRCU.txt b/Documentation/RCU/listRCU.txt
> > > index adb5a3782846..190e666fc359 100644
> > > --- a/Documentation/RCU/listRCU.txt
> > > +++ b/Documentation/RCU/listRCU.txt
> > > @@ -175,7 +175,7 @@ otherwise, the added fields would need to be filled in):
> > > list_for_each_entry(e, list, list) {
> > > if (!audit_compare_rule(rule, &e->rule)) {
> > > e->rule.action = newaction;
> > > - e->rule.file_count = newfield_count;
> > > + e->rule.field_count = newfield_count;
> > > write_unlock(&auditsc_lock);
> > > return 0;
> > > }
> > > @@ -204,7 +204,7 @@ RCU ("read-copy update") its name. The RCU code is as follows:
> > > return -ENOMEM;
> > > audit_copy_rule(&ne->rule, &e->rule);
> > > ne->rule.action = newaction;
> > > - ne->rule.file_count = newfield_count;
> > > + ne->rule.field_count = newfield_count;
> > > list_replace_rcu(&e->list, &ne->list);
> > > call_rcu(&e->rcu, audit_free_rule);
> > > return 0;
> > > --
> > > 2.21.0.1020.gf2820cf01a-goog
> > >
> >
>
Hi Paul, Joel,
> > > On the other hand, would you have ideas for more modern replacement
> > > examples?
> >
> > There are 3 cases I can see in listRCU.txt:
> > (1) action taken outside of read_lock (can tolerate stale data), no in-place update.
> > this is the best possible usage of RCU.
> > (2) action taken outside of read_lock, in-place updates
> > this is good as long as not too many in-place updates.
> > involves copying creating new list node and replacing the
> > node being updated with it.
> > (3) cannot tolerate stale data: here a deleted or obsolete flag can be used
> > protected by a per-entry lock. reader
> > aborts if object is stale.
> >
> > Any replacement example must make satisfy (3) too?
>
> It would be OK to have a separate example for (3). It would of course
> be nicer to have one example for all three, but not all -that- important.
>
> > The only example for (3) that I know of is sysvipc sempahores which you also
> > mentioned in the paper. Looking through this code, it hasn't changed
> > conceptually and it could be a fit for an example (ipc_valid_object() checks
> > for whether the object is stale).
>
> That is indeed the classic canonical example. ;-)
>
> > The other example could be dentry look up which uses seqlocks for the
> > RCU-walk case? But that could be too complex. This is also something I first
> > learnt from the paper and then the excellent path-lookup.rst document in
> > kernel sources.
>
> This is a great example, but it would need serious simplification for
> use in the Documentation/RCU directory. Note that dcache uses it to
> gain very limited and targeted consistency -- only a few types of updates
> acquire the write-side of that seqlock.
>
> Might be quite worthwhile to have a simplified example, though!
> Perhaps a trivial hash table where write-side sequence lock is acquired
> only when moving an element from one chain to another?
Sorry to take you down here..., but what do you mean by "the paper"? ;-/
Thanx,
Andrea
On Sun, May 12, 2019 at 12:11:26AM +0200, Andrea Parri wrote:
> Hi Paul, Joel,
>
> > > > On the other hand, would you have ideas for more modern replacement
> > > > examples?
> > >
> > > There are 3 cases I can see in listRCU.txt:
> > > (1) action taken outside of read_lock (can tolerate stale data), no in-place update.
> > > this is the best possible usage of RCU.
> > > (2) action taken outside of read_lock, in-place updates
> > > this is good as long as not too many in-place updates.
> > > involves copying creating new list node and replacing the
> > > node being updated with it.
> > > (3) cannot tolerate stale data: here a deleted or obsolete flag can be used
> > > protected by a per-entry lock. reader
> > > aborts if object is stale.
> > >
> > > Any replacement example must make satisfy (3) too?
> >
> > It would be OK to have a separate example for (3). It would of course
> > be nicer to have one example for all three, but not all -that- important.
> >
> > > The only example for (3) that I know of is sysvipc sempahores which you also
> > > mentioned in the paper. Looking through this code, it hasn't changed
> > > conceptually and it could be a fit for an example (ipc_valid_object() checks
> > > for whether the object is stale).
> >
> > That is indeed the classic canonical example. ;-)
> >
> > > The other example could be dentry look up which uses seqlocks for the
> > > RCU-walk case? But that could be too complex. This is also something I first
> > > learnt from the paper and then the excellent path-lookup.rst document in
> > > kernel sources.
> >
> > This is a great example, but it would need serious simplification for
> > use in the Documentation/RCU directory. Note that dcache uses it to
> > gain very limited and targeted consistency -- only a few types of updates
> > acquire the write-side of that seqlock.
> >
> > Might be quite worthwhile to have a simplified example, though!
> > Perhaps a trivial hash table where write-side sequence lock is acquired
> > only when moving an element from one chain to another?
>
> Sorry to take you down here..., but what do you mean by "the paper"? ;-/
One or both of these two:
http://www2.rdrop.com/~paulmck/techreports/survey.2012.09.17a.pdf
http://www2.rdrop.com/~paulmck/techreports/RCUUsage.2013.02.24a.pdf
Thanx, Paul
On Sat, May 11, 2019 at 05:41:31PM -0700, Paul E. McKenney wrote:
> On Sun, May 12, 2019 at 12:11:26AM +0200, Andrea Parri wrote:
> > Hi Paul, Joel,
> >
> > > > > On the other hand, would you have ideas for more modern replacement
> > > > > examples?
> > > >
> > > > There are 3 cases I can see in listRCU.txt:
> > > > (1) action taken outside of read_lock (can tolerate stale data), no in-place update.
> > > > this is the best possible usage of RCU.
> > > > (2) action taken outside of read_lock, in-place updates
> > > > this is good as long as not too many in-place updates.
> > > > involves copying creating new list node and replacing the
> > > > node being updated with it.
> > > > (3) cannot tolerate stale data: here a deleted or obsolete flag can be used
> > > > protected by a per-entry lock. reader
> > > > aborts if object is stale.
> > > >
> > > > Any replacement example must make satisfy (3) too?
> > >
> > > It would be OK to have a separate example for (3). It would of course
> > > be nicer to have one example for all three, but not all -that- important.
> > >
> > > > The only example for (3) that I know of is sysvipc sempahores which you also
> > > > mentioned in the paper. Looking through this code, it hasn't changed
> > > > conceptually and it could be a fit for an example (ipc_valid_object() checks
> > > > for whether the object is stale).
> > >
> > > That is indeed the classic canonical example. ;-)
> > >
> > > > The other example could be dentry look up which uses seqlocks for the
> > > > RCU-walk case? But that could be too complex. This is also something I first
> > > > learnt from the paper and then the excellent path-lookup.rst document in
> > > > kernel sources.
> > >
> > > This is a great example, but it would need serious simplification for
> > > use in the Documentation/RCU directory. Note that dcache uses it to
> > > gain very limited and targeted consistency -- only a few types of updates
> > > acquire the write-side of that seqlock.
> > >
> > > Might be quite worthwhile to have a simplified example, though!
> > > Perhaps a trivial hash table where write-side sequence lock is acquired
> > > only when moving an element from one chain to another?
> >
> > Sorry to take you down here..., but what do you mean by "the paper"? ;-/
>
> One or both of these two:
>
> http://www2.rdrop.com/~paulmck/techreports/survey.2012.09.17a.pdf
> http://www2.rdrop.com/~paulmck/techreports/RCUUsage.2013.02.24a.pdf
Oh, these are familiar. ;-) Thank you!
Andrea
On Wed, May 08, 2019 at 11:16:38AM -0700, Paul E. McKenney wrote:
[snip]
> > The other example could be dentry look up which uses seqlocks for the
> > RCU-walk case? But that could be too complex. This is also something I first
> > learnt from the paper and then the excellent path-lookup.rst document in
> > kernel sources.
>
> This is a great example, but it would need serious simplification for
> use in the Documentation/RCU directory. Note that dcache uses it to
> gain very limited and targeted consistency -- only a few types of updates
> acquire the write-side of that seqlock.
>
> Might be quite worthwhile to have a simplified example, though!
> Perhaps a trivial hash table where write-side sequence lock is acquired
> only when moving an element from one chain to another?
Here you meant "moving from one chain to another" in the case of
hashtable-resizing right? I could not think of another reason why an element
is moved between 2 hash chains.
I just finished reading the main parts of Josh's relativistic hashtable paper
[1] and it is very cool indeed. The whole wait-for-readers application for
hashtable expansion is so well thought. I am planning to go over more papers
and code and can certainly update this example with a read-mostly hashtable
example as well as you are suggesting. :-)
[1] https://www.usenix.org/legacy/event/atc11/tech/final_files/Triplett.pdf
thanks,
- Joel
On Sun, May 12, 2019 at 11:43:05PM -0400, Joel Fernandes wrote:
> On Wed, May 08, 2019 at 11:16:38AM -0700, Paul E. McKenney wrote:
> [snip]
> > > The other example could be dentry look up which uses seqlocks for the
> > > RCU-walk case? But that could be too complex. This is also something I first
> > > learnt from the paper and then the excellent path-lookup.rst document in
> > > kernel sources.
> >
> > This is a great example, but it would need serious simplification for
> > use in the Documentation/RCU directory. Note that dcache uses it to
> > gain very limited and targeted consistency -- only a few types of updates
> > acquire the write-side of that seqlock.
> >
> > Might be quite worthwhile to have a simplified example, though!
> > Perhaps a trivial hash table where write-side sequence lock is acquired
> > only when moving an element from one chain to another?
>
> Here you meant "moving from one chain to another" in the case of
> hashtable-resizing right? I could not think of another reason why an element
> is moved between 2 hash chains.
Either that or in terms of atomic rekeying of a specific element in that
table, thus potentially requiring an atomic move of only that specific
element to another hash chain.
> I just finished reading the main parts of Josh's relativistic hashtable paper
> [1] and it is very cool indeed. The whole wait-for-readers application for
> hashtable expansion is so well thought. I am planning to go over more papers
> and code and can certainly update this example with a read-mostly hashtable
> example as well as you are suggesting. :-)
>
> [1] https://www.usenix.org/legacy/event/atc11/tech/final_files/Triplett.pdf
Sounds very good!
Thanx, Paul
On Wed, May 8, 2019 at 9:16 PM Paul E. McKenney <[email protected]> wrote:
[snip]
> > > And this example code predates v2.6.12. ;-)
> > >
> > > So good eyes, but I believe that this really does reflect the ancient
> > > code...
> > >
> > > On the other hand, would you have ideas for more modern replacement
> > > examples?
> >
> > There are 3 cases I can see in listRCU.txt:
> > (1) action taken outside of read_lock (can tolerate stale data), no in-place update.
> > this is the best possible usage of RCU.
> > (2) action taken outside of read_lock, in-place updates
> > this is good as long as not too many in-place updates.
> > involves copying creating new list node and replacing the
> > node being updated with it.
> > (3) cannot tolerate stale data: here a deleted or obsolete flag can be used
> > protected by a per-entry lock. reader
> > aborts if object is stale.
> >
> > Any replacement example must make satisfy (3) too?
>
> It would be OK to have a separate example for (3). It would of course
> be nicer to have one example for all three, but not all -that- important.
>
> > The only example for (3) that I know of is sysvipc sempahores which you also
> > mentioned in the paper. Looking through this code, it hasn't changed
> > conceptually and it could be a fit for an example (ipc_valid_object() checks
> > for whether the object is stale).
>
> That is indeed the classic canonical example. ;-)
FWIW just want to mention, it seems to me the ptrace task list
traversal could be a great example of "mark obsolete objects" and is
simple so I could just use that.
Neil talks about it in his article here:
https://lwn.net/Articles/610972/ . In "Group 3: Transform the way the
list is walked"
Cheers,
- Joel