[PATCH 1/2] mm: vmalloc: Avoid a double lookup of freed VA in a tree

[PATCH 1/2] mm: vmalloc: Avoid a double lookup of freed VA in a tree

[Uladzislau Rezki (Sony)]

When a VA is freed over a main path, for example by invoking
the vfree() function, a tree is accessed two times what is odd:

vfree():
  __vunmap()
    __find_vmap_area()
  vm_remove_mappings()
    remove_vm_area()
      __find_vmap_area()

__find_vmap_area() are called two times. Fix it by introducing
a find_unlink_vmap_area() helper that finds and un-links a VA
from a tree.

Performance test results on a single CPU:

- fix_size_alloc_test       loops: 1000000 avg: 476847   usec
- full_fit_alloc_test       loops: 1000000 avg: 806746   usec
- long_busy_list_alloc_test loops: 1000000 avg: 13552093 usec
- random_size_alloc_test    loops: 1000000 avg: 7441322  usec
- fix_align_alloc_test      loops: 1000000 avg: 1411132  usec
All test took worker0=87650866284 cycles

- fix_size_alloc_test       loops: 1000000 avg: 490713   usec
- full_fit_alloc_test       loops: 1000000 avg: 579162   usec
- long_busy_list_alloc_test loops: 1000000 avg: 10485448 usec
- random_size_alloc_test    loops: 1000000 avg: 5824449  usec
- fix_align_alloc_test      loops: 1000000 avg: 984735   usec
All test took worker0=67952362802 cycles

Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
---
 mm/vmalloc.c | 40 ++++++++++++++++++++++++++++------------
 1 file changed, 28 insertions(+), 12 deletions(-)

diff --git a/mm/vmalloc.c b/mm/vmalloc.c
index 9e30f0b39203..0fc38c36e0df 100644
--- a/mm/vmalloc.c
+++ b/mm/vmalloc.c
@@ -1825,9 +1825,14 @@ static void free_vmap_area_noflush(struct vmap_area *va)
 	unsigned long va_start = va->va_start;
 	unsigned long nr_lazy;
 
-	spin_lock(&vmap_area_lock);
-	unlink_va(va, &vmap_area_root);
-	spin_unlock(&vmap_area_lock);
+	/*
+	 * A free_vmap_block() is left. It is NOT a main free path.
+	 */
+	if (!list_empty(&va->list)) {
+		spin_lock(&vmap_area_lock);
+		unlink_va(va, &vmap_area_root);
+		spin_unlock(&vmap_area_lock);
+	}
 
 	nr_lazy = atomic_long_add_return((va->va_end - va->va_start) >>
 				PAGE_SHIFT, &vmap_lazy_nr);
@@ -1871,6 +1876,19 @@ struct vmap_area *find_vmap_area(unsigned long addr)
 	return va;
 }
 
+static struct vmap_area *find_unlink_vmap_area(unsigned long addr)
+{
+	struct vmap_area *va;
+
+	spin_lock(&vmap_area_lock);
+	va = __find_vmap_area(addr, &vmap_area_root);
+	if (va)
+		unlink_va(va, &vmap_area_root);
+	spin_unlock(&vmap_area_lock);
+
+	return va;
+}
+
 /*** Per cpu kva allocator ***/
 
 /*
@@ -2236,7 +2254,7 @@ void vm_unmap_ram(const void *mem, unsigned int count)
 		return;
 	}
 
-	va = find_vmap_area(addr);
+	va = find_unlink_vmap_area(addr);
 	BUG_ON(!va);
 	debug_check_no_locks_freed((void *)va->va_start,
 				    (va->va_end - va->va_start));
@@ -2607,21 +2625,16 @@ struct vm_struct *remove_vm_area(const void *addr)
 
 	might_sleep();
 
-	spin_lock(&vmap_area_lock);
-	va = __find_vmap_area((unsigned long)addr, &vmap_area_root);
-	if (va && va->vm) {
+	va = find_unlink_vmap_area((unsigned long) addr);
+	if (va) {
 		struct vm_struct *vm = va->vm;
 
-		va->vm = NULL;
-		spin_unlock(&vmap_area_lock);
-
 		kasan_free_module_shadow(vm);
 		free_unmap_vmap_area(va);
 
 		return vm;
 	}
 
-	spin_unlock(&vmap_area_lock);
 	return NULL;
 }
 
@@ -2690,6 +2703,7 @@ static void vm_remove_mappings(struct vm_struct *area, int deallocate_pages)
 static void __vunmap(const void *addr, int deallocate_pages)
 {
 	struct vm_struct *area;
+	struct vmap_area *va;
 
 	if (!addr)
 		return;
@@ -2698,7 +2712,9 @@ static void __vunmap(const void *addr, int deallocate_pages)
 			addr))
 		return;
 
-	area = find_vm_area(addr);
+	va = find_unlink_vmap_area((unsigned long)addr);
+	area = va->vm;
+
 	if (unlikely(!area)) {
 		WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
 				addr);
-- 
2.30.2

[PATCH 2/2] mm: vmalloc: Replace BUG_ON() by WARN_ON_ONCE()

[Uladzislau Rezki (Sony)]

Currently a vm_unmap_ram() functions triggers a BUG() if an area
is not found. Replace it by the WARN_ON_ONCE() error message and
keep machine alive instead of stopping it.

The worst case is a memory leaking.

Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
---
 mm/vmalloc.c | 10 ++++++----
 1 file changed, 6 insertions(+), 4 deletions(-)

diff --git a/mm/vmalloc.c b/mm/vmalloc.c
index 0fc38c36e0df..e05a0dc79ac5 100644
--- a/mm/vmalloc.c
+++ b/mm/vmalloc.c
@@ -2255,10 +2255,12 @@ void vm_unmap_ram(const void *mem, unsigned int count)
 	}
 
 	va = find_unlink_vmap_area(addr);
-	BUG_ON(!va);
-	debug_check_no_locks_freed((void *)va->va_start,
-				    (va->va_end - va->va_start));
-	free_unmap_vmap_area(va);
+
+	if (!WARN_ON_ONCE(!va)) {
+		debug_check_no_locks_freed((void *)va->va_start,
+			(va->va_end - va->va_start));
+		free_unmap_vmap_area(va);
+	}
 }
 EXPORT_SYMBOL(vm_unmap_ram);
 
-- 
2.30.2

Re: [PATCH 1/2] mm: vmalloc: Avoid a double lookup of freed VA in a tree

[Uladzislau Rezki]

On Tue, Dec 20, 2022 at 07:27:03PM +0100, Uladzislau Rezki (Sony) wrote:
> When a VA is freed over a main path, for example by invoking
> the vfree() function, a tree is accessed two times what is odd:
> 
> vfree():
>   __vunmap()
>     __find_vmap_area()
>   vm_remove_mappings()
>     remove_vm_area()
>       __find_vmap_area()
> 
> __find_vmap_area() are called two times. Fix it by introducing
> a find_unlink_vmap_area() helper that finds and un-links a VA
> from a tree.
> 
> Performance test results on a single CPU:
> 
> - fix_size_alloc_test       loops: 1000000 avg: 476847   usec
> - full_fit_alloc_test       loops: 1000000 avg: 806746   usec
> - long_busy_list_alloc_test loops: 1000000 avg: 13552093 usec
> - random_size_alloc_test    loops: 1000000 avg: 7441322  usec
> - fix_align_alloc_test      loops: 1000000 avg: 1411132  usec
> All test took worker0=87650866284 cycles
> 
> - fix_size_alloc_test       loops: 1000000 avg: 490713   usec
> - full_fit_alloc_test       loops: 1000000 avg: 579162   usec
> - long_busy_list_alloc_test loops: 1000000 avg: 10485448 usec
> - random_size_alloc_test    loops: 1000000 avg: 5824449  usec
> - fix_align_alloc_test      loops: 1000000 avg: 984735   usec
> All test took worker0=67952362802 cycles
> 
> Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
> ---
>  mm/vmalloc.c | 40 ++++++++++++++++++++++++++++------------
>  1 file changed, 28 insertions(+), 12 deletions(-)
> 
> diff --git a/mm/vmalloc.c b/mm/vmalloc.c
> index 9e30f0b39203..0fc38c36e0df 100644
> --- a/mm/vmalloc.c
> +++ b/mm/vmalloc.c
> @@ -1825,9 +1825,14 @@ static void free_vmap_area_noflush(struct vmap_area *va)
>  	unsigned long va_start = va->va_start;
>  	unsigned long nr_lazy;
>  
> -	spin_lock(&vmap_area_lock);
> -	unlink_va(va, &vmap_area_root);
> -	spin_unlock(&vmap_area_lock);
> +	/*
> +	 * A free_vmap_block() is left. It is NOT a main free path.
> +	 */
> +	if (!list_empty(&va->list)) {
> +		spin_lock(&vmap_area_lock);
> +		unlink_va(va, &vmap_area_root);
> +		spin_unlock(&vmap_area_lock);
> +	}
>  
>  	nr_lazy = atomic_long_add_return((va->va_end - va->va_start) >>
>  				PAGE_SHIFT, &vmap_lazy_nr);
> @@ -1871,6 +1876,19 @@ struct vmap_area *find_vmap_area(unsigned long addr)
>  	return va;
>  }
>  
> +static struct vmap_area *find_unlink_vmap_area(unsigned long addr)
> +{
> +	struct vmap_area *va;
> +
> +	spin_lock(&vmap_area_lock);
> +	va = __find_vmap_area(addr, &vmap_area_root);
> +	if (va)
> +		unlink_va(va, &vmap_area_root);
> +	spin_unlock(&vmap_area_lock);
> +
> +	return va;
> +}
> +
>  /*** Per cpu kva allocator ***/
>  
>  /*
> @@ -2236,7 +2254,7 @@ void vm_unmap_ram(const void *mem, unsigned int count)
>  		return;
>  	}
>  
> -	va = find_vmap_area(addr);
> +	va = find_unlink_vmap_area(addr);
>  	BUG_ON(!va);
>  	debug_check_no_locks_freed((void *)va->va_start,
>  				    (va->va_end - va->va_start));
> @@ -2607,21 +2625,16 @@ struct vm_struct *remove_vm_area(const void *addr)
>  
>  	might_sleep();
>  
> -	spin_lock(&vmap_area_lock);
> -	va = __find_vmap_area((unsigned long)addr, &vmap_area_root);
> -	if (va && va->vm) {
> +	va = find_unlink_vmap_area((unsigned long) addr);
> +	if (va) {
>  		struct vm_struct *vm = va->vm;
>  
> -		va->vm = NULL;
> -		spin_unlock(&vmap_area_lock);
> -
>  		kasan_free_module_shadow(vm);
>  		free_unmap_vmap_area(va);
>  
>  		return vm;
>  	}
>  
> -	spin_unlock(&vmap_area_lock);
>  	return NULL;
>  }
>  
> @@ -2690,6 +2703,7 @@ static void vm_remove_mappings(struct vm_struct *area, int deallocate_pages)
>  static void __vunmap(const void *addr, int deallocate_pages)
>  {
>  	struct vm_struct *area;
> +	struct vmap_area *va;
>  
>  	if (!addr)
>  		return;
> @@ -2698,7 +2712,9 @@ static void __vunmap(const void *addr, int deallocate_pages)
>  			addr))
>  		return;
>  
> -	area = find_vm_area(addr);
> +	va = find_unlink_vmap_area((unsigned long)addr);
> +	area = va->vm;
> +
>  	if (unlikely(!area)) {
>  		WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
>  				addr);
> -- 
> 2.30.2
> 
Will send a v2.

--
Uladzislau Rezki

Re: [PATCH 2/2] mm: vmalloc: Replace BUG_ON() by WARN_ON_ONCE()

[Uladzislau Rezki]

On Tue, Dec 20, 2022 at 07:27:04PM +0100, Uladzislau Rezki (Sony) wrote:
> Currently a vm_unmap_ram() functions triggers a BUG() if an area
> is not found. Replace it by the WARN_ON_ONCE() error message and
> keep machine alive instead of stopping it.
> 
> The worst case is a memory leaking.
> 
> Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
> ---
>  mm/vmalloc.c | 10 ++++++----
>  1 file changed, 6 insertions(+), 4 deletions(-)
> 
> diff --git a/mm/vmalloc.c b/mm/vmalloc.c
> index 0fc38c36e0df..e05a0dc79ac5 100644
> --- a/mm/vmalloc.c
> +++ b/mm/vmalloc.c
> @@ -2255,10 +2255,12 @@ void vm_unmap_ram(const void *mem, unsigned int count)
>  	}
>  
>  	va = find_unlink_vmap_area(addr);
> -	BUG_ON(!va);
> -	debug_check_no_locks_freed((void *)va->va_start,
> -				    (va->va_end - va->va_start));
> -	free_unmap_vmap_area(va);
> +
> +	if (!WARN_ON_ONCE(!va)) {
> +		debug_check_no_locks_freed((void *)va->va_start,
> +			(va->va_end - va->va_start));
> +		free_unmap_vmap_area(va);
> +	}
>  }
>  EXPORT_SYMBOL(vm_unmap_ram);
>  
> -- 
> 2.30.2
> 
Will send a v2.

--
Uladzislau Rezki

Re: [PATCH 1/2] mm: vmalloc: Avoid a double lookup of freed VA in a tree

[Uladzislau Rezki]

On Tue, Dec 20, 2022 at 07:45:20PM +0100, Uladzislau Rezki wrote:
> On Tue, Dec 20, 2022 at 07:27:03PM +0100, Uladzislau Rezki (Sony) wrote:
> > When a VA is freed over a main path, for example by invoking
> > the vfree() function, a tree is accessed two times what is odd:
> > 
> > vfree():
> >   __vunmap()
> >     __find_vmap_area()
> >   vm_remove_mappings()
> >     remove_vm_area()
> >       __find_vmap_area()
> > 
> > __find_vmap_area() are called two times. Fix it by introducing
> > a find_unlink_vmap_area() helper that finds and un-links a VA
> > from a tree.
> > 
> > Performance test results on a single CPU:
> > 
> > - fix_size_alloc_test       loops: 1000000 avg: 476847   usec
> > - full_fit_alloc_test       loops: 1000000 avg: 806746   usec
> > - long_busy_list_alloc_test loops: 1000000 avg: 13552093 usec
> > - random_size_alloc_test    loops: 1000000 avg: 7441322  usec
> > - fix_align_alloc_test      loops: 1000000 avg: 1411132  usec
> > All test took worker0=87650866284 cycles
> > 
> > - fix_size_alloc_test       loops: 1000000 avg: 490713   usec
> > - full_fit_alloc_test       loops: 1000000 avg: 579162   usec
> > - long_busy_list_alloc_test loops: 1000000 avg: 10485448 usec
> > - random_size_alloc_test    loops: 1000000 avg: 5824449  usec
> > - fix_align_alloc_test      loops: 1000000 avg: 984735   usec
> > All test took worker0=67952362802 cycles
> > 
> > Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
> > ---
> >  mm/vmalloc.c | 40 ++++++++++++++++++++++++++++------------
> >  1 file changed, 28 insertions(+), 12 deletions(-)
> > 
> > diff --git a/mm/vmalloc.c b/mm/vmalloc.c
> > index 9e30f0b39203..0fc38c36e0df 100644
> > --- a/mm/vmalloc.c
> > +++ b/mm/vmalloc.c
> > @@ -1825,9 +1825,14 @@ static void free_vmap_area_noflush(struct vmap_area *va)
> >  	unsigned long va_start = va->va_start;
> >  	unsigned long nr_lazy;
> >  
> > -	spin_lock(&vmap_area_lock);
> > -	unlink_va(va, &vmap_area_root);
> > -	spin_unlock(&vmap_area_lock);
> > +	/*
> > +	 * A free_vmap_block() is left. It is NOT a main free path.
> > +	 */
> > +	if (!list_empty(&va->list)) {
> > +		spin_lock(&vmap_area_lock);
> > +		unlink_va(va, &vmap_area_root);
> > +		spin_unlock(&vmap_area_lock);
> > +	}
> >  
> >  	nr_lazy = atomic_long_add_return((va->va_end - va->va_start) >>
> >  				PAGE_SHIFT, &vmap_lazy_nr);
> > @@ -1871,6 +1876,19 @@ struct vmap_area *find_vmap_area(unsigned long addr)
> >  	return va;
> >  }
> >  
> > +static struct vmap_area *find_unlink_vmap_area(unsigned long addr)
> > +{
> > +	struct vmap_area *va;
> > +
> > +	spin_lock(&vmap_area_lock);
> > +	va = __find_vmap_area(addr, &vmap_area_root);
> > +	if (va)
> > +		unlink_va(va, &vmap_area_root);
> > +	spin_unlock(&vmap_area_lock);
> > +
> > +	return va;
> > +}
> > +
> >  /*** Per cpu kva allocator ***/
> >  
> >  /*
> > @@ -2236,7 +2254,7 @@ void vm_unmap_ram(const void *mem, unsigned int count)
> >  		return;
> >  	}
> >  
> > -	va = find_vmap_area(addr);
> > +	va = find_unlink_vmap_area(addr);
> >  	BUG_ON(!va);
> >  	debug_check_no_locks_freed((void *)va->va_start,
> >  				    (va->va_end - va->va_start));
> > @@ -2607,21 +2625,16 @@ struct vm_struct *remove_vm_area(const void *addr)
> >  
> >  	might_sleep();
> >  
> > -	spin_lock(&vmap_area_lock);
> > -	va = __find_vmap_area((unsigned long)addr, &vmap_area_root);
> > -	if (va && va->vm) {
> > +	va = find_unlink_vmap_area((unsigned long) addr);
> > +	if (va) {
> >  		struct vm_struct *vm = va->vm;
> >  
> > -		va->vm = NULL;
> > -		spin_unlock(&vmap_area_lock);
> > -
> >  		kasan_free_module_shadow(vm);
> >  		free_unmap_vmap_area(va);
> >  
> >  		return vm;
> >  	}
> >  
> > -	spin_unlock(&vmap_area_lock);
> >  	return NULL;
> >  }
> >  
> > @@ -2690,6 +2703,7 @@ static void vm_remove_mappings(struct vm_struct *area, int deallocate_pages)
> >  static void __vunmap(const void *addr, int deallocate_pages)
> >  {
> >  	struct vm_struct *area;
> > +	struct vmap_area *va;
> >  
> >  	if (!addr)
> >  		return;
> > @@ -2698,7 +2712,9 @@ static void __vunmap(const void *addr, int deallocate_pages)
> >  			addr))
> >  		return;
> >  
> > -	area = find_vm_area(addr);
> > +	va = find_unlink_vmap_area((unsigned long)addr);
> > +	area = va->vm;
> > +
> >  	if (unlikely(!area)) {
> >  		WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
> >  				addr);
> > -- 
> > 2.30.2
> > 
Will send a v2.
 
--
Uladzislau Rezki

Re: [PATCH 2/2] mm: vmalloc: Replace BUG_ON() by WARN_ON_ONCE()

[Lorenzo Stoakes]

On Tue, Dec 20, 2022 at 07:27:04PM +0100, Uladzislau Rezki (Sony) wrote:
> Currently a vm_unmap_ram() functions triggers a BUG() if an area
> is not found. Replace it by the WARN_ON_ONCE() error message and
> keep machine alive instead of stopping it.
>
> The worst case is a memory leaking.
>
> Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
> ---
>  mm/vmalloc.c | 10 ++++++----
>  1 file changed, 6 insertions(+), 4 deletions(-)
>
> diff --git a/mm/vmalloc.c b/mm/vmalloc.c
> index 0fc38c36e0df..e05a0dc79ac5 100644
> --- a/mm/vmalloc.c
> +++ b/mm/vmalloc.c
> @@ -2255,10 +2255,12 @@ void vm_unmap_ram(const void *mem, unsigned int count)
>  	}
>
>  	va = find_unlink_vmap_area(addr);
> -	BUG_ON(!va);
> -	debug_check_no_locks_freed((void *)va->va_start,
> -				    (va->va_end - va->va_start));
> -	free_unmap_vmap_area(va);
> +
> +	if (!WARN_ON_ONCE(!va)) {
> +		debug_check_no_locks_freed((void *)va->va_start,
> +			(va->va_end - va->va_start));
> +		free_unmap_vmap_area(va);
> +	}
>  }
>  EXPORT_SYMBOL(vm_unmap_ram);
>
> --
> 2.30.2
>

Absolutely in favour of this in principle (BUG_ON() is something we should
resort to in only the direst of circumstances), one small nit - perhaps it'd be
neater to simply make this a guard clause? E.g.:-

	if (!WARN_ON_ONCE(!va))
		return;

	...

Re: [PATCH 2/2] mm: vmalloc: Replace BUG_ON() by WARN_ON_ONCE()

[Lorenzo Stoakes]

On Tue, Dec 20, 2022 at 06:53:18PM +0000, Lorenzo Stoakes wrote:
[snip
> Absolutely in favour of this in principle (BUG_ON() is something we should
> resort to in only the direst of circumstances), one small nit - perhaps it'd be
> neater to simply make this a guard clause? E.g.:-
>
> 	if (!WARN_ON_ONCE(!va))
> 		return;
>
> 	...

Made a mistake here, meant to say

 	if (WARN_ON_ONCE(!va))
 		return;

Of course :)

Apologies, only just noticed you are sending a v2, perhaps something to consider
for it? But will hold off further review until you send it! :)

Re: [PATCH 2/2] mm: vmalloc: Replace BUG_ON() by WARN_ON_ONCE()

[Uladzislau Rezki]

On Tue, Dec 20, 2022 at 06:56:38PM +0000, Lorenzo Stoakes wrote:
> On Tue, Dec 20, 2022 at 06:53:18PM +0000, Lorenzo Stoakes wrote:
> [snip
> > Absolutely in favour of this in principle (BUG_ON() is something we should
> > resort to in only the direst of circumstances), one small nit - perhaps it'd be
> > neater to simply make this a guard clause? E.g.:-
> >
> > 	if (!WARN_ON_ONCE(!va))
> > 		return;
> >
> > 	...
> 
> Made a mistake here, meant to say
> 
>  	if (WARN_ON_ONCE(!va))
>  		return;
> 
Agree. This looks better. Less confusing :)

I will post a v2.

--
Uladzislau Rezki