On Fri, May 29, 2015 at 09:40:43AM -0400, Dan Streetman wrote:
> On Thu, May 28, 2015 at 6:29 PM, Paul E. McKenney
> <[email protected]> wrote:
> > On Thu, May 28, 2015 at 05:24:14PM -0400, Dan Streetman wrote:
> >> On Thu, May 28, 2015 at 5:16 PM, Paul E. McKenney
> >> <[email protected]> wrote:
> >> > On Thu, May 28, 2015 at 05:12:00PM -0400, Dan Streetman wrote:
> >> >> On Thu, May 28, 2015 at 5:05 PM, Steven Rostedt <[email protected]> wrote:
> >> >> > On Thu, 28 May 2015 16:35:27 -0400
> >> >> > Dan Streetman <[email protected]> wrote:
> >> >> >
> >> >> >> Add list_last_or_null_rcu(), to simplify getting the last entry from a
> >> >> >> rcu-protected list. The standard list_last_entry() can't be used as it
> >> >> >> is not rcu-protected; the list may be modified concurrently. And the
> >> >> >> ->prev pointer can't be used, as only the ->next pointers are protected
> >> >> >> by rcu.
> >> >> >>
> >> >> >> This simply iterates forward through the entire list, to get to the last
> >> >> >> entry. If the list is empty, it returns NULL.
> >> >> >
> >> >> > May I asked what this would be used for? It seems awfully inefficient
> >> >> > in its implementation. What use cases would this be for? I hate to add
> >> >> > something like this as a generic function which would encourage people
> >> >> > to use it. Iterating over an entire list to find the last element is
> >> >> > just nasty.
> >> >>
> >> >> i have a patch series that will update zswap to be able to change its
> >> >> parameters at runtime, instead of only at boot time. To do that, it
> >> >> creates new "pools" dynamically, and keeps them all in a list, with
> >> >> only the 1st pool being actively used; any following pools still have
> >> >> everything that was stored in them, but they aren't added to. When
> >> >> zswap has to "shrink" - by telling one of the pools to get rid of 1 or
> >> >> more items - it picks the last on the list. Once a pool is empty,
> >> >> it's removed/freed.
> >> >>
> >> >> So zswap *could* just manually iterate the list to the last element,
> >> >> instead of using this new function. But if rcu lists are ever
> >> >> improved later on, e.g. if ->prev is somehow rcu-protected as well as
> >> >> ->next, then this function should be faster than manually iterating.
> >> >>
> >> >> if there's a better rcu-way to get to the last list entry, then we
> >> >> should definitely use it, although based on my understanding of the
> >> >> rcu list implementation, you can only iterate forwards, safely
> >> >> (without locking).
> >> >
> >> > The usual approach would be to maintain a tail pointer. How big are
> >> > these lists likely to get?
> >>
> >> in the vast majority of cases, it'll only be 1 entry; the list is only
> >> added to when the user decides to change the pool type or compression
> >> function, which during normal operation will probably happen very
> >> rarely. So in most situations, the function will just be a 1-step for
> >> loop. I'm only proposing this function since it may be useful to
> >> optimize last-rcu-entry access later, and maybe for other users.
> >
> > Fair enough.
> >
> >> re: keeping a rcu-safe tail pointer, how is that done? i assume since
> >> head->prev isn't rcu protected (is it?), it would need to be separate
> >> from the list, and thus would need to be spinlock-protected and
> >> updated at each list update?
> >
> > Yep, each update that changed the tail would need to change the tail
> > pointer, and that would need to be protected from other updates.
> >
> > But if the lists were long, one approach would be to provide a
> > list_del_rcu_bidir() or some such that didn't poison the ->prev pointer,
> > and then use rcu_dereference() to traverse the head element's ->prev
> > pointer, as you suggested above. I have resisted doing that due to
> > debugging issues, but if there turns out to be a strong need, let's talk.
>
> I was thinking; since the head element is never removed, its ->prev
> pointer will never be poisoned; is something like this safe?
>
> /* this can only be called on the head, not on any entry;
> * specifically this is not safe to call on any entry that
> * may be removed with list_del_rcu() or list_replace_rcu().
> */
> #define list_last_or_null_rcu(head, type, member) \
> ({ \
> struct list_head *__last = (head)->prev; \
> __last != (head) ? list_entry_rcu(__last) : NULL; \
> })
>
> I was thinking that's unsafe because the head->prev pointer can be
> updated directly, such as during list_del_rcu(last_entry) which will
> directly reassign head->prev to the new last entry; but maybe that is
> ok since list_del_rcu(first_entry) also directly reassigns head->next
> to the new first entry?
In your particular case, it would be safe, but people can and do call
list_del_rcu() on the header in order to remove the entire list, which
would poison the header's ->prev pointer. So if you really want to do
this in your own code, fine, but please: (1) include a big fat comment
saying what you are doing and the resulting restrictions and (2) Name
it something specific to your code.
If multiple similar use cases show up, maybe we can add something to
the rculist_nulls.h file.
Thanx, Paul
On Mon, Jun 1, 2015 at 2:17 PM, Paul E. McKenney
<[email protected]> wrote:
> On Fri, May 29, 2015 at 09:40:43AM -0400, Dan Streetman wrote:
>> On Thu, May 28, 2015 at 6:29 PM, Paul E. McKenney
>> <[email protected]> wrote:
>> > On Thu, May 28, 2015 at 05:24:14PM -0400, Dan Streetman wrote:
>> >> On Thu, May 28, 2015 at 5:16 PM, Paul E. McKenney
>> >> <[email protected]> wrote:
>> >> > On Thu, May 28, 2015 at 05:12:00PM -0400, Dan Streetman wrote:
>> >> >> On Thu, May 28, 2015 at 5:05 PM, Steven Rostedt <[email protected]> wrote:
>> >> >> > On Thu, 28 May 2015 16:35:27 -0400
>> >> >> > Dan Streetman <[email protected]> wrote:
>> >> >> >
>> >> >> >> Add list_last_or_null_rcu(), to simplify getting the last entry from a
>> >> >> >> rcu-protected list. The standard list_last_entry() can't be used as it
>> >> >> >> is not rcu-protected; the list may be modified concurrently. And the
>> >> >> >> ->prev pointer can't be used, as only the ->next pointers are protected
>> >> >> >> by rcu.
>> >> >> >>
>> >> >> >> This simply iterates forward through the entire list, to get to the last
>> >> >> >> entry. If the list is empty, it returns NULL.
>> >> >> >
>> >> >> > May I asked what this would be used for? It seems awfully inefficient
>> >> >> > in its implementation. What use cases would this be for? I hate to add
>> >> >> > something like this as a generic function which would encourage people
>> >> >> > to use it. Iterating over an entire list to find the last element is
>> >> >> > just nasty.
>> >> >>
>> >> >> i have a patch series that will update zswap to be able to change its
>> >> >> parameters at runtime, instead of only at boot time. To do that, it
>> >> >> creates new "pools" dynamically, and keeps them all in a list, with
>> >> >> only the 1st pool being actively used; any following pools still have
>> >> >> everything that was stored in them, but they aren't added to. When
>> >> >> zswap has to "shrink" - by telling one of the pools to get rid of 1 or
>> >> >> more items - it picks the last on the list. Once a pool is empty,
>> >> >> it's removed/freed.
>> >> >>
>> >> >> So zswap *could* just manually iterate the list to the last element,
>> >> >> instead of using this new function. But if rcu lists are ever
>> >> >> improved later on, e.g. if ->prev is somehow rcu-protected as well as
>> >> >> ->next, then this function should be faster than manually iterating.
>> >> >>
>> >> >> if there's a better rcu-way to get to the last list entry, then we
>> >> >> should definitely use it, although based on my understanding of the
>> >> >> rcu list implementation, you can only iterate forwards, safely
>> >> >> (without locking).
>> >> >
>> >> > The usual approach would be to maintain a tail pointer. How big are
>> >> > these lists likely to get?
>> >>
>> >> in the vast majority of cases, it'll only be 1 entry; the list is only
>> >> added to when the user decides to change the pool type or compression
>> >> function, which during normal operation will probably happen very
>> >> rarely. So in most situations, the function will just be a 1-step for
>> >> loop. I'm only proposing this function since it may be useful to
>> >> optimize last-rcu-entry access later, and maybe for other users.
>> >
>> > Fair enough.
>> >
>> >> re: keeping a rcu-safe tail pointer, how is that done? i assume since
>> >> head->prev isn't rcu protected (is it?), it would need to be separate
>> >> from the list, and thus would need to be spinlock-protected and
>> >> updated at each list update?
>> >
>> > Yep, each update that changed the tail would need to change the tail
>> > pointer, and that would need to be protected from other updates.
>> >
>> > But if the lists were long, one approach would be to provide a
>> > list_del_rcu_bidir() or some such that didn't poison the ->prev pointer,
>> > and then use rcu_dereference() to traverse the head element's ->prev
>> > pointer, as you suggested above. I have resisted doing that due to
>> > debugging issues, but if there turns out to be a strong need, let's talk.
>>
>> I was thinking; since the head element is never removed, its ->prev
>> pointer will never be poisoned; is something like this safe?
>>
>> /* this can only be called on the head, not on any entry;
>> * specifically this is not safe to call on any entry that
>> * may be removed with list_del_rcu() or list_replace_rcu().
>> */
>> #define list_last_or_null_rcu(head, type, member) \
>> ({ \
>> struct list_head *__last = (head)->prev; \
>> __last != (head) ? list_entry_rcu(__last) : NULL; \
>> })
>>
>> I was thinking that's unsafe because the head->prev pointer can be
>> updated directly, such as during list_del_rcu(last_entry) which will
>> directly reassign head->prev to the new last entry; but maybe that is
>> ok since list_del_rcu(first_entry) also directly reassigns head->next
>> to the new first entry?
>
> In your particular case, it would be safe, but people can and do call
> list_del_rcu() on the header in order to remove the entire list, which
> would poison the header's ->prev pointer. So if you really want to do
> this in your own code, fine, but please: (1) include a big fat comment
> saying what you are doing and the resulting restrictions and (2) Name
> it something specific to your code.
Ok. Since in my case there will only be 1 entry (or at worst just a
few) in the list in the majority of cases, I'll just stick to manually
iterating to the last entry. It's easy and clear.
Thanks!
>
> If multiple similar use cases show up, maybe we can add something to
> the rculist_nulls.h file.
>
> Thanx, Paul
>