Re: [RFC v1] mm: add the preempt check into alloc_vmap_area()

[Uladzislau Rezki]

On Fri, Mar 02, 2018 at 03:34:52PM -0800, Andrew Morton wrote:
> On Tue, 27 Feb 2018 05:06:43 -0800 Matthew Wilcox  wrote:
> 
> > On Tue, Feb 27, 2018 at 11:22:59AM +0100, Uladzislau Rezki (Sony) wrote:
> > > During finding a suitable hole in the vmap_area_list
> > > there is an explicit rescheduling check for latency reduction.
> > > We do it, since there are workloads which are sensitive for
> > > long (more than 1 millisecond) preemption off scenario.
> > 
> > I understand your problem, but this is a horrid solution.  If it takes
> > us a millisecond to find a suitable chunk of free address space, something
> > is terribly wrong.  On a 3GHz CPU, that's 3 million clock ticks!
> 
> Yup.
> 
> > I think our real problem is that we have no data structure that stores
> > free VA space.  We have the vmap_area which stores allocated space, but no
> > data structure to store free space.
> 
> I wonder if we can reuse free_vmap_cache as a quick fix: if
> need_resched(), point free_vmap_cache at the current rb_node, drop the
> lock, cond_resched, goto retry?
> 
It sounds like we can. But there is a concern if that potentially can
introduce a degrade of search time due to changing a starting point
for our search.

--
Vlad Rezki

Re: [RFC v1] mm: add the preempt check into alloc_vmap_area()

[Uladzislau Rezki]

On Fri, Mar 02, 2018 at 03:34:52PM -0800, Andrew Morton wrote:
> On Tue, 27 Feb 2018 05:06:43 -0800 Matthew Wilcox  wrote:
> 
> > On Tue, Feb 27, 2018 at 11:22:59AM +0100, Uladzislau Rezki (Sony) wrote:
> > > During finding a suitable hole in the vmap_area_list
> > > there is an explicit rescheduling check for latency reduction.
> > > We do it, since there are workloads which are sensitive for
> > > long (more than 1 millisecond) preemption off scenario.
> > 
> > I understand your problem, but this is a horrid solution.  If it takes
> > us a millisecond to find a suitable chunk of free address space, something
> > is terribly wrong.  On a 3GHz CPU, that's 3 million clock ticks!
> 
> Yup.
> 
> > I think our real problem is that we have no data structure that stores
> > free VA space.  We have the vmap_area which stores allocated space, but no
> > data structure to store free space.
> 
> I wonder if we can reuse free_vmap_cache as a quick fix: if
> need_resched(), point free_vmap_cache at the current rb_node, drop the
> lock, cond_resched, goto retry?
> 
It sounds like we can. But there is a concern if that potentially can
introduce a degrade of search time due to changing a starting point
for our search.

--
Vlad Rezki

Re: [RFC v1] mm: add the preempt check into alloc_vmap_area()

[Andrew Morton]

On Tue, 27 Feb 2018 05:06:43 -0800 Matthew Wilcox  wrote:

> On Tue, Feb 27, 2018 at 11:22:59AM +0100, Uladzislau Rezki (Sony) wrote:
> > During finding a suitable hole in the vmap_area_list
> > there is an explicit rescheduling check for latency reduction.
> > We do it, since there are workloads which are sensitive for
> > long (more than 1 millisecond) preemption off scenario.
> 
> I understand your problem, but this is a horrid solution.  If it takes
> us a millisecond to find a suitable chunk of free address space, something
> is terribly wrong.  On a 3GHz CPU, that's 3 million clock ticks!

Yup.

> I think our real problem is that we have no data structure that stores
> free VA space.  We have the vmap_area which stores allocated space, but no
> data structure to store free space.

I wonder if we can reuse free_vmap_cache as a quick fix: if
need_resched(), point free_vmap_cache at the current rb_node, drop the
lock, cond_resched, goto retry?

Re: [RFC v1] mm: add the preempt check into alloc_vmap_area()

[Andrew Morton]

On Tue, 27 Feb 2018 05:06:43 -0800 Matthew Wilcox  wrote:

> On Tue, Feb 27, 2018 at 11:22:59AM +0100, Uladzislau Rezki (Sony) wrote:
> > During finding a suitable hole in the vmap_area_list
> > there is an explicit rescheduling check for latency reduction.
> > We do it, since there are workloads which are sensitive for
> > long (more than 1 millisecond) preemption off scenario.
> 
> I understand your problem, but this is a horrid solution.  If it takes
> us a millisecond to find a suitable chunk of free address space, something
> is terribly wrong.  On a 3GHz CPU, that's 3 million clock ticks!

Yup.

> I think our real problem is that we have no data structure that stores
> free VA space.  We have the vmap_area which stores allocated space, but no
> data structure to store free space.

I wonder if we can reuse free_vmap_cache as a quick fix: if
need_resched(), point free_vmap_cache at the current rb_node, drop the
lock, cond_resched, goto retry?

Re: [RFC v1] mm: add the preempt check into alloc_vmap_area()

[Uladzislau Rezki]

On Tue, Feb 27, 2018 at 05:06:43AM -0800, Matthew Wilcox wrote:
> On Tue, Feb 27, 2018 at 11:22:59AM +0100, Uladzislau Rezki (Sony) wrote:
> > During finding a suitable hole in the vmap_area_list
> > there is an explicit rescheduling check for latency reduction.
> > We do it, since there are workloads which are sensitive for
> > long (more than 1 millisecond) preemption off scenario.
> 
> I understand your problem, but this is a horrid solution.  If it takes
> us a millisecond to find a suitable chunk of free address space, something
> is terribly wrong.  On a 3GHz CPU, that's 3 million clock ticks!
>
Some background. I spent some time analyzing an issue regarding audio
drops/glitches during playing hires audio on our mobile device. It is
ARM A53 with 4 CPUs on one socket. When it comes to frequency and test
case, the system is most likely idle and operation is done on ~576 MHz.

I found out that the reason was in vmalloc due to it can take time
to find a suitable chunk of memory and it is done in non-preemptible
context. As a result the other audio thread is not run on CPU in time
despite need_resched is set.

> 
> I think our real problem is that we have no data structure that stores
> free VA space.  We have the vmap_area which stores allocated space, but no
> data structure to store free space.
> 
> My initial proposal would be to reuse the vmap_area structure and store
> the freed ones in a second rb_tree sorted by the size (ie va_end - va_start).
> When freeing, we might need to merge forwards and backwards.  Allocating
> would be a matter of finding an area preferably of the exact right size;
> otherwise split a larger free area into a free area and an allocated area
> (there's a lot of literature on how exactly to choose which larger area
> to split; memory allocators are pretty well-studied).
> 
Thank you for your comments and proposal.

--
Vlad Rezki

Re: [RFC v1] mm: add the preempt check into alloc_vmap_area()

[Uladzislau Rezki]

On Tue, Feb 27, 2018 at 05:06:43AM -0800, Matthew Wilcox wrote:
> On Tue, Feb 27, 2018 at 11:22:59AM +0100, Uladzislau Rezki (Sony) wrote:
> > During finding a suitable hole in the vmap_area_list
> > there is an explicit rescheduling check for latency reduction.
> > We do it, since there are workloads which are sensitive for
> > long (more than 1 millisecond) preemption off scenario.
> 
> I understand your problem, but this is a horrid solution.  If it takes
> us a millisecond to find a suitable chunk of free address space, something
> is terribly wrong.  On a 3GHz CPU, that's 3 million clock ticks!
>
Some background. I spent some time analyzing an issue regarding audio
drops/glitches during playing hires audio on our mobile device. It is
ARM A53 with 4 CPUs on one socket. When it comes to frequency and test
case, the system is most likely idle and operation is done on ~576 MHz.

I found out that the reason was in vmalloc due to it can take time
to find a suitable chunk of memory and it is done in non-preemptible
context. As a result the other audio thread is not run on CPU in time
despite need_resched is set.

> 
> I think our real problem is that we have no data structure that stores
> free VA space.  We have the vmap_area which stores allocated space, but no
> data structure to store free space.
> 
> My initial proposal would be to reuse the vmap_area structure and store
> the freed ones in a second rb_tree sorted by the size (ie va_end - va_start).
> When freeing, we might need to merge forwards and backwards.  Allocating
> would be a matter of finding an area preferably of the exact right size;
> otherwise split a larger free area into a free area and an allocated area
> (there's a lot of literature on how exactly to choose which larger area
> to split; memory allocators are pretty well-studied).
> 
Thank you for your comments and proposal.

--
Vlad Rezki

Re: [RFC v1] mm: add the preempt check into alloc_vmap_area()

[Matthew Wilcox]

On Tue, Feb 27, 2018 at 11:22:59AM +0100, Uladzislau Rezki (Sony) wrote:
> During finding a suitable hole in the vmap_area_list
> there is an explicit rescheduling check for latency reduction.
> We do it, since there are workloads which are sensitive for
> long (more than 1 millisecond) preemption off scenario.

I understand your problem, but this is a horrid solution.  If it takes
us a millisecond to find a suitable chunk of free address space, something
is terribly wrong.  On a 3GHz CPU, that's 3 million clock ticks!

I think our real problem is that we have no data structure that stores
free VA space.  We have the vmap_area which stores allocated space, but no
data structure to store free space.

My initial proposal would be to reuse the vmap_area structure and store
the freed ones in a second rb_tree sorted by the size (ie va_end - va_start).
When freeing, we might need to merge forwards and backwards.  Allocating
would be a matter of finding an area preferably of the exact right size;
otherwise split a larger free area into a free area and an allocated area
(there's a lot of literature on how exactly to choose which larger area
to split; memory allocators are pretty well-studied).

Re: [RFC v1] mm: add the preempt check into alloc_vmap_area()

[Matthew Wilcox]

On Tue, Feb 27, 2018 at 11:22:59AM +0100, Uladzislau Rezki (Sony) wrote:
> During finding a suitable hole in the vmap_area_list
> there is an explicit rescheduling check for latency reduction.
> We do it, since there are workloads which are sensitive for
> long (more than 1 millisecond) preemption off scenario.

I understand your problem, but this is a horrid solution.  If it takes
us a millisecond to find a suitable chunk of free address space, something
is terribly wrong.  On a 3GHz CPU, that's 3 million clock ticks!

I think our real problem is that we have no data structure that stores
free VA space.  We have the vmap_area which stores allocated space, but no
data structure to store free space.

My initial proposal would be to reuse the vmap_area structure and store
the freed ones in a second rb_tree sorted by the size (ie va_end - va_start).
When freeing, we might need to merge forwards and backwards.  Allocating
would be a matter of finding an area preferably of the exact right size;
otherwise split a larger free area into a free area and an allocated area
(there's a lot of literature on how exactly to choose which larger area
to split; memory allocators are pretty well-studied).