[PATCH 1/1] mm: vmalloc: Remove a global vmap_blocks xarray

[PATCH 1/1] mm: vmalloc: Remove a global vmap_blocks xarray

[Uladzislau Rezki (Sony)]

A global vmap_blocks-xarray array can be contented under
heavy usage of the vm_map_ram()/vm_unmap_ram() APIs. The
lock_stat shows that a "vmap_blocks.xa_lock" lock is a
second in a top-list when it comes to contentions:

<snip>
----------------------------------------
class name con-bounces contentions ...
----------------------------------------
vmap_area_lock:         2554079 2554276 ...
  --------------
  vmap_area_lock        1297948  [<00000000dd41cbaa>] alloc_vmap_area+0x1c7/0x910
  vmap_area_lock        1256330  [<000000009d927bf3>] free_vmap_block+0x4a/0xe0
  vmap_area_lock              1  [<00000000c95c05a7>] find_vm_area+0x16/0x70
  --------------
  vmap_area_lock        1738590  [<00000000dd41cbaa>] alloc_vmap_area+0x1c7/0x910
  vmap_area_lock         815688  [<000000009d927bf3>] free_vmap_block+0x4a/0xe0
  vmap_area_lock              1  [<00000000c1d619d7>] __get_vm_area_node+0xd2/0x170

vmap_blocks.xa_lock:    862689  862698 ...
  -------------------
  vmap_blocks.xa_lock   378418    [<00000000625a5626>] vm_map_ram+0x359/0x4a0
  vmap_blocks.xa_lock   484280    [<00000000caa2ef03>] xa_erase+0xe/0x30
  -------------------
  vmap_blocks.xa_lock   576226    [<00000000caa2ef03>] xa_erase+0xe/0x30
  vmap_blocks.xa_lock   286472    [<00000000625a5626>] vm_map_ram+0x359/0x4a0
...
<snip>

that is a result of running vm_map_ram()/vm_unmap_ram() in
a loop. The test creates 64(on 64 CPUs system) threads and
each one maps/unmaps 1 page.

After this change the "xa_lock" can be considered as a noise
in the same test condition:

<snip>
...
&xa->xa_lock#1:         10333 10394 ...
  --------------
  &xa->xa_lock#1        5349      [<00000000bbbc9751>] xa_erase+0xe/0x30
  &xa->xa_lock#1        5045      [<0000000018def45d>] vm_map_ram+0x3a4/0x4f0
  --------------
  &xa->xa_lock#1        7326      [<0000000018def45d>] vm_map_ram+0x3a4/0x4f0
  &xa->xa_lock#1        3068      [<00000000bbbc9751>] xa_erase+0xe/0x30
...
<snip>

This patch does not fix vmap_area_lock/free_vmap_area_lock and
purge_vmap_area_lock bottle-necks, it is rather a separate rework.

Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
---
 mm/vmalloc.c | 50 ++++++++++++++++++++++++++------------------------
 1 file changed, 26 insertions(+), 24 deletions(-)

diff --git a/mm/vmalloc.c b/mm/vmalloc.c
index 978194dc2bb8..13b5342bed9a 100644
--- a/mm/vmalloc.c
+++ b/mm/vmalloc.c
@@ -1911,6 +1911,7 @@ static struct vmap_area *find_unlink_vmap_area(unsigned long addr)
 struct vmap_block_queue {
 	spinlock_t lock;
 	struct list_head free;
+	struct xarray vmap_blocks;
 };
 
 struct vmap_block {
@@ -1927,25 +1928,18 @@ struct vmap_block {
 /* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
 static DEFINE_PER_CPU(struct vmap_block_queue, vmap_block_queue);
 
-/*
- * XArray of vmap blocks, indexed by address, to quickly find a vmap block
- * in the free path. Could get rid of this if we change the API to return a
- * "cookie" from alloc, to be passed to free. But no big deal yet.
- */
-static DEFINE_XARRAY(vmap_blocks);
-
-/*
- * We should probably have a fallback mechanism to allocate virtual memory
- * out of partially filled vmap blocks. However vmap block sizing should be
- * fairly reasonable according to the vmalloc size, so it shouldn't be a
- * big problem.
- */
+static struct vmap_block_queue *
+addr_to_vbq(unsigned long addr)
+{
+	int cpu = (addr / VMAP_BLOCK_SIZE) % num_possible_cpus();
+	return &per_cpu(vmap_block_queue, cpu);
+}
 
-static unsigned long addr_to_vb_idx(unsigned long addr)
+static unsigned long
+addr_to_vb_va_start(unsigned long addr)
 {
-	addr -= VMALLOC_START & ~(VMAP_BLOCK_SIZE-1);
-	addr /= VMAP_BLOCK_SIZE;
-	return addr;
+	/* A start address of block an address belongs to. */
+	return rounddown(addr, VMAP_BLOCK_SIZE);
 }
 
 static void *vmap_block_vaddr(unsigned long va_start, unsigned long pages_off)
@@ -1953,7 +1947,7 @@ static void *vmap_block_vaddr(unsigned long va_start, unsigned long pages_off)
 	unsigned long addr;
 
 	addr = va_start + (pages_off << PAGE_SHIFT);
-	BUG_ON(addr_to_vb_idx(addr) != addr_to_vb_idx(va_start));
+	BUG_ON(addr_to_vb_va_start(addr) != addr_to_vb_va_start(va_start));
 	return (void *)addr;
 }
 
@@ -1970,7 +1964,6 @@ static void *new_vmap_block(unsigned int order, gfp_t gfp_mask)
 	struct vmap_block_queue *vbq;
 	struct vmap_block *vb;
 	struct vmap_area *va;
-	unsigned long vb_idx;
 	int node, err;
 	void *vaddr;
 
@@ -2003,8 +1996,8 @@ static void *new_vmap_block(unsigned int order, gfp_t gfp_mask)
 	bitmap_set(vb->used_map, 0, (1UL << order));
 	INIT_LIST_HEAD(&vb->free_list);
 
-	vb_idx = addr_to_vb_idx(va->va_start);
-	err = xa_insert(&vmap_blocks, vb_idx, vb, gfp_mask);
+	vbq = addr_to_vbq(va->va_start);
+	err = xa_insert(&vbq->vmap_blocks, va->va_start, vb, gfp_mask);
 	if (err) {
 		kfree(vb);
 		free_vmap_area(va);
@@ -2021,9 +2014,11 @@ static void *new_vmap_block(unsigned int order, gfp_t gfp_mask)
 
 static void free_vmap_block(struct vmap_block *vb)
 {
+	struct vmap_block_queue *vbq;
 	struct vmap_block *tmp;
 
-	tmp = xa_erase(&vmap_blocks, addr_to_vb_idx(vb->va->va_start));
+	vbq = addr_to_vbq(vb->va->va_start);
+	tmp = xa_erase(&vbq->vmap_blocks, vb->va->va_start);
 	BUG_ON(tmp != vb);
 
 	spin_lock(&vmap_area_lock);
@@ -2135,6 +2130,7 @@ static void vb_free(unsigned long addr, unsigned long size)
 	unsigned long offset;
 	unsigned int order;
 	struct vmap_block *vb;
+	struct vmap_block_queue *vbq;
 
 	BUG_ON(offset_in_page(size));
 	BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
@@ -2143,7 +2139,10 @@ static void vb_free(unsigned long addr, unsigned long size)
 
 	order = get_order(size);
 	offset = (addr & (VMAP_BLOCK_SIZE - 1)) >> PAGE_SHIFT;
-	vb = xa_load(&vmap_blocks, addr_to_vb_idx(addr));
+
+	vbq = addr_to_vbq(addr);
+	vb = xa_load(&vbq->vmap_blocks, addr_to_vb_va_start(addr));
+
 	spin_lock(&vb->lock);
 	bitmap_clear(vb->used_map, offset, (1UL << order));
 	spin_unlock(&vb->lock);
@@ -3486,6 +3485,7 @@ static void vmap_ram_vread(char *buf, char *addr, int count, unsigned long flags
 {
 	char *start;
 	struct vmap_block *vb;
+	struct vmap_block_queue *vbq;
 	unsigned long offset;
 	unsigned int rs, re, n;
 
@@ -3503,7 +3503,8 @@ static void vmap_ram_vread(char *buf, char *addr, int count, unsigned long flags
 	 * Area is split into regions and tracked with vmap_block, read out
 	 * each region and zero fill the hole between regions.
 	 */
-	vb = xa_load(&vmap_blocks, addr_to_vb_idx((unsigned long)addr));
+	vbq = addr_to_vbq((unsigned long) addr);
+	vb = xa_load(&vbq->vmap_blocks, addr_to_vb_va_start((unsigned long) addr));
 	if (!vb)
 		goto finished;
 
@@ -4272,6 +4273,7 @@ void __init vmalloc_init(void)
 		p = &per_cpu(vfree_deferred, i);
 		init_llist_head(&p->list);
 		INIT_WORK(&p->wq, delayed_vfree_work);
+		xa_init(&vbq->vmap_blocks);
 	}
 
 	/* Import existing vmlist entries. */
-- 
2.30.2

Re: [PATCH 1/1] mm: vmalloc: Remove a global vmap_blocks xarray

[Andrew Morton]

On Thu, 23 Mar 2023 20:21:11 +0100 "Uladzislau Rezki (Sony)" <urezki@gmail.com> wrote:

> A global vmap_blocks-xarray array can be contented under
> heavy usage of the vm_map_ram()/vm_unmap_ram() APIs. The
> lock_stat shows that a "vmap_blocks.xa_lock" lock is a
> second in a top-list when it comes to contentions:
> 
> ...
> 
> This patch does not fix vmap_area_lock/free_vmap_area_lock and
> purge_vmap_area_lock bottle-necks, it is rather a separate rework.
> 
> ...
>
>  static DEFINE_PER_CPU(struct vmap_block_queue, vmap_block_queue);
>  
> ...
>
> +static struct vmap_block_queue *
> +addr_to_vbq(unsigned long addr)
> +{
> +	int cpu = (addr / VMAP_BLOCK_SIZE) % num_possible_cpus();
> +	return &per_cpu(vmap_block_queue, cpu);
> +}

Seems strange.  vmap_block_queue is not a per-cpu thing in this usage. 
Instead it's a hash table, indexed off the (hashed) address, not off
smp_processor_id().

Yet in other places, vmap_block_queue *is* used in the conventional
cpu-local fashion.

So we can have CPU A using the cpu-local entry in vmap_block_queue
while CPU B is simultaneously using it, having looked it up via `addr'.

AFAICT this all works OK, no races.

But still, what it's doing is mixing an addr-indexed hashtable with the
CPU-indexed array in surprising ways.  It would be clearer to make the
vmap_blocks array a separate thing from the per-cpu array, although it
would presumably use a bit more memory.

Can we please at least get a big fat comment in an appropriate place
which explains all this to the reader?

Re: [PATCH 1/1] mm: vmalloc: Remove a global vmap_blocks xarray

[Lorenzo Stoakes]

On Thu, Mar 23, 2023 at 08:21:11PM +0100, Uladzislau Rezki (Sony) wrote:
> A global vmap_blocks-xarray array can be contented under
> heavy usage of the vm_map_ram()/vm_unmap_ram() APIs. The
> lock_stat shows that a "vmap_blocks.xa_lock" lock is a
> second in a top-list when it comes to contentions:
>
> <snip>
> ----------------------------------------
> class name con-bounces contentions ...
> ----------------------------------------
> vmap_area_lock:         2554079 2554276 ...
>   --------------
>   vmap_area_lock        1297948  [<00000000dd41cbaa>] alloc_vmap_area+0x1c7/0x910
>   vmap_area_lock        1256330  [<000000009d927bf3>] free_vmap_block+0x4a/0xe0
>   vmap_area_lock              1  [<00000000c95c05a7>] find_vm_area+0x16/0x70
>   --------------
>   vmap_area_lock        1738590  [<00000000dd41cbaa>] alloc_vmap_area+0x1c7/0x910
>   vmap_area_lock         815688  [<000000009d927bf3>] free_vmap_block+0x4a/0xe0
>   vmap_area_lock              1  [<00000000c1d619d7>] __get_vm_area_node+0xd2/0x170
>
> vmap_blocks.xa_lock:    862689  862698 ...
>   -------------------
>   vmap_blocks.xa_lock   378418    [<00000000625a5626>] vm_map_ram+0x359/0x4a0
>   vmap_blocks.xa_lock   484280    [<00000000caa2ef03>] xa_erase+0xe/0x30
>   -------------------
>   vmap_blocks.xa_lock   576226    [<00000000caa2ef03>] xa_erase+0xe/0x30
>   vmap_blocks.xa_lock   286472    [<00000000625a5626>] vm_map_ram+0x359/0x4a0
> ...
> <snip>
>
> that is a result of running vm_map_ram()/vm_unmap_ram() in
> a loop. The test creates 64(on 64 CPUs system) threads and
> each one maps/unmaps 1 page.
>
> After this change the "xa_lock" can be considered as a noise
> in the same test condition:
>
> <snip>
> ...
> &xa->xa_lock#1:         10333 10394 ...
>   --------------
>   &xa->xa_lock#1        5349      [<00000000bbbc9751>] xa_erase+0xe/0x30
>   &xa->xa_lock#1        5045      [<0000000018def45d>] vm_map_ram+0x3a4/0x4f0
>   --------------
>   &xa->xa_lock#1        7326      [<0000000018def45d>] vm_map_ram+0x3a4/0x4f0
>   &xa->xa_lock#1        3068      [<00000000bbbc9751>] xa_erase+0xe/0x30
> ...
> <snip>
>

Nice! Really good to see contention reduced, but in addition I'm a huge fan
of us removing the global state in vmalloc and this is a good start.

I've noticed a small perf regression after 3 runs of ./test_vmalloc.sh
performance from an average of 119356136169 cycles to 120404645782 or +0.9%
but this doesn't seem especially egregious.

> This patch does not fix vmap_area_lock/free_vmap_area_lock and
> purge_vmap_area_lock bottle-necks, it is rather a separate rework.
>
> Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
> ---
>  mm/vmalloc.c | 50 ++++++++++++++++++++++++++------------------------
>  1 file changed, 26 insertions(+), 24 deletions(-)
>
> diff --git a/mm/vmalloc.c b/mm/vmalloc.c
> index 978194dc2bb8..13b5342bed9a 100644
> --- a/mm/vmalloc.c
> +++ b/mm/vmalloc.c
> @@ -1911,6 +1911,7 @@ static struct vmap_area *find_unlink_vmap_area(unsigned long addr)
>  struct vmap_block_queue {
>  	spinlock_t lock;
>  	struct list_head free;
> +	struct xarray vmap_blocks;
>  };
>
>  struct vmap_block {
> @@ -1927,25 +1928,18 @@ struct vmap_block {
>  /* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
>  static DEFINE_PER_CPU(struct vmap_block_queue, vmap_block_queue);
>
> -/*
> - * XArray of vmap blocks, indexed by address, to quickly find a vmap block
> - * in the free path. Could get rid of this if we change the API to return a
> - * "cookie" from alloc, to be passed to free. But no big deal yet.
> - */

Doesn't this comment still apply? Or is the idea of returning the "cookie"
not really viable?

> -static DEFINE_XARRAY(vmap_blocks);
> -
> -/*
> - * We should probably have a fallback mechanism to allocate virtual memory
> - * out of partially filled vmap blocks. However vmap block sizing should be
> - * fairly reasonable according to the vmalloc size, so it shouldn't be a
> - * big problem.
> - */

Again, is this comment no longer relevant?

> +static struct vmap_block_queue *
> +addr_to_vbq(unsigned long addr)
> +{
> +	int cpu = (addr / VMAP_BLOCK_SIZE) % num_possible_cpus();
> +	return &per_cpu(vmap_block_queue, cpu);
> +}

Andrew's already commented on this, so I won't dwell but it does seem odd
to subdivide by number of possible CPUs rather than just use the actual
CPU. I guess your response to his question will also answer mine :)

>
> -static unsigned long addr_to_vb_idx(unsigned long addr)
> +static unsigned long
> +addr_to_vb_va_start(unsigned long addr)
>  {
> -	addr -= VMALLOC_START & ~(VMAP_BLOCK_SIZE-1);
> -	addr /= VMAP_BLOCK_SIZE;
> -	return addr;
> +	/* A start address of block an address belongs to. */

A nit, but might be worth referring to the assert in vmap_block_vaddr(), as
this comment seems a bit redundant otherwise as it is implied by the code
it comments.

> +	return rounddown(addr, VMAP_BLOCK_SIZE);
>  }
>
>  static void *vmap_block_vaddr(unsigned long va_start, unsigned long pages_off)
> @@ -1953,7 +1947,7 @@ static void *vmap_block_vaddr(unsigned long va_start, unsigned long pages_off)
>  	unsigned long addr;
>
>  	addr = va_start + (pages_off << PAGE_SHIFT);
> -	BUG_ON(addr_to_vb_idx(addr) != addr_to_vb_idx(va_start));
> +	BUG_ON(addr_to_vb_va_start(addr) != addr_to_vb_va_start(va_start));

Maybe nitty, but perhaps better to WARN_ON() here to avoid BUG_ON proliferation?

And can't this be the below?

WARN_ON(addr_to_vb_va_start(addr) != va_start);

>  	return (void *)addr;
>  }
>
> @@ -1970,7 +1964,6 @@ static void *new_vmap_block(unsigned int order, gfp_t gfp_mask)
>  	struct vmap_block_queue *vbq;
>  	struct vmap_block *vb;
>  	struct vmap_area *va;
> -	unsigned long vb_idx;
>  	int node, err;
>  	void *vaddr;
>
> @@ -2003,8 +1996,8 @@ static void *new_vmap_block(unsigned int order, gfp_t gfp_mask)
>  	bitmap_set(vb->used_map, 0, (1UL << order));
>  	INIT_LIST_HEAD(&vb->free_list);
>
> -	vb_idx = addr_to_vb_idx(va->va_start);
> -	err = xa_insert(&vmap_blocks, vb_idx, vb, gfp_mask);
> +	vbq = addr_to_vbq(va->va_start);
> +	err = xa_insert(&vbq->vmap_blocks, va->va_start, vb, gfp_mask);

This seems actually like a nice subtle improvement in that we are now
indexing always on va_start explicitly and will always load using
addr_to_vb_va_start().

>  	if (err) {
>  		kfree(vb);
>  		free_vmap_area(va);
> @@ -2021,9 +2014,11 @@ static void *new_vmap_block(unsigned int order, gfp_t gfp_mask)
>
>  static void free_vmap_block(struct vmap_block *vb)
>  {
> +	struct vmap_block_queue *vbq;
>  	struct vmap_block *tmp;
>
> -	tmp = xa_erase(&vmap_blocks, addr_to_vb_idx(vb->va->va_start));
> +	vbq = addr_to_vbq(vb->va->va_start);
> +	tmp = xa_erase(&vbq->vmap_blocks, vb->va->va_start);
>  	BUG_ON(tmp != vb);
>
>  	spin_lock(&vmap_area_lock);
> @@ -2135,6 +2130,7 @@ static void vb_free(unsigned long addr, unsigned long size)
>  	unsigned long offset;
>  	unsigned int order;
>  	struct vmap_block *vb;
> +	struct vmap_block_queue *vbq;
>
>  	BUG_ON(offset_in_page(size));
>  	BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
> @@ -2143,7 +2139,10 @@ static void vb_free(unsigned long addr, unsigned long size)
>
>  	order = get_order(size);
>  	offset = (addr & (VMAP_BLOCK_SIZE - 1)) >> PAGE_SHIFT;
> -	vb = xa_load(&vmap_blocks, addr_to_vb_idx(addr));
> +
> +	vbq = addr_to_vbq(addr);
> +	vb = xa_load(&vbq->vmap_blocks, addr_to_vb_va_start(addr));
> +
>  	spin_lock(&vb->lock);
>  	bitmap_clear(vb->used_map, offset, (1UL << order));
>  	spin_unlock(&vb->lock);
> @@ -3486,6 +3485,7 @@ static void vmap_ram_vread(char *buf, char *addr, int count, unsigned long flags
>  {
>  	char *start;
>  	struct vmap_block *vb;
> +	struct vmap_block_queue *vbq;
>  	unsigned long offset;
>  	unsigned int rs, re, n;
>
> @@ -3503,7 +3503,8 @@ static void vmap_ram_vread(char *buf, char *addr, int count, unsigned long flags
>  	 * Area is split into regions and tracked with vmap_block, read out
>  	 * each region and zero fill the hole between regions.
>  	 */
> -	vb = xa_load(&vmap_blocks, addr_to_vb_idx((unsigned long)addr));
> +	vbq = addr_to_vbq((unsigned long) addr);
> +	vb = xa_load(&vbq->vmap_blocks, addr_to_vb_va_start((unsigned long) addr));
>  	if (!vb)
>  		goto finished;
>
> @@ -4272,6 +4273,7 @@ void __init vmalloc_init(void)
>  		p = &per_cpu(vfree_deferred, i);
>  		init_llist_head(&p->list);
>  		INIT_WORK(&p->wq, delayed_vfree_work);
> +		xa_init(&vbq->vmap_blocks);
>  	}
>
>  	/* Import existing vmlist entries. */
> --
> 2.30.2
>

Re: [PATCH 1/1] mm: vmalloc: Remove a global vmap_blocks xarray

[Uladzislau Rezki]

On Thu, Mar 23, 2023 at 02:12:53PM -0700, Andrew Morton wrote:
> On Thu, 23 Mar 2023 20:21:11 +0100 "Uladzislau Rezki (Sony)" <urezki@gmail.com> wrote:
> 
> > A global vmap_blocks-xarray array can be contented under
> > heavy usage of the vm_map_ram()/vm_unmap_ram() APIs. The
> > lock_stat shows that a "vmap_blocks.xa_lock" lock is a
> > second in a top-list when it comes to contentions:
> > 
> > ...
> > 
> > This patch does not fix vmap_area_lock/free_vmap_area_lock and
> > purge_vmap_area_lock bottle-necks, it is rather a separate rework.
> > 
> > ...
> >
> >  static DEFINE_PER_CPU(struct vmap_block_queue, vmap_block_queue);
> >  
> > ...
> >
> > +static struct vmap_block_queue *
> > +addr_to_vbq(unsigned long addr)
> > +{
> > +	int cpu = (addr / VMAP_BLOCK_SIZE) % num_possible_cpus();
> > +	return &per_cpu(vmap_block_queue, cpu);
> > +}
> 
> Seems strange.  vmap_block_queue is not a per-cpu thing in this usage. 
> Instead it's a hash table, indexed off the (hashed) address, not off
> smp_processor_id().
> 
> Yet in other places, vmap_block_queue *is* used in the conventional
> cpu-local fashion.
> 
> So we can have CPU A using the cpu-local entry in vmap_block_queue
> while CPU B is simultaneously using it, having looked it up via `addr'.
> 
> AFAICT this all works OK, no races.
> 
> But still, what it's doing is mixing an addr-indexed hashtable with the
> CPU-indexed array in surprising ways.  It would be clearer to make the
> vmap_blocks array a separate thing from the per-cpu array, although it
> would presumably use a bit more memory.
> 
> Can we please at least get a big fat comment in an appropriate place
> which explains all this to the reader?
>
Yep, i will send out a v2 with all explanation. Indeed i have to add
detailed explanation.

Thanks!

--
Uladzislau Rezki

Re: [PATCH 1/1] mm: vmalloc: Remove a global vmap_blocks xarray

[Uladzislau Rezki]

On Thu, Mar 23, 2023 at 09:46:00PM +0000, Lorenzo Stoakes wrote:
> On Thu, Mar 23, 2023 at 08:21:11PM +0100, Uladzislau Rezki (Sony) wrote:
> > A global vmap_blocks-xarray array can be contented under
> > heavy usage of the vm_map_ram()/vm_unmap_ram() APIs. The
> > lock_stat shows that a "vmap_blocks.xa_lock" lock is a
> > second in a top-list when it comes to contentions:
> >
> > <snip>
> > ----------------------------------------
> > class name con-bounces contentions ...
> > ----------------------------------------
> > vmap_area_lock:         2554079 2554276 ...
> >   --------------
> >   vmap_area_lock        1297948  [<00000000dd41cbaa>] alloc_vmap_area+0x1c7/0x910
> >   vmap_area_lock        1256330  [<000000009d927bf3>] free_vmap_block+0x4a/0xe0
> >   vmap_area_lock              1  [<00000000c95c05a7>] find_vm_area+0x16/0x70
> >   --------------
> >   vmap_area_lock        1738590  [<00000000dd41cbaa>] alloc_vmap_area+0x1c7/0x910
> >   vmap_area_lock         815688  [<000000009d927bf3>] free_vmap_block+0x4a/0xe0
> >   vmap_area_lock              1  [<00000000c1d619d7>] __get_vm_area_node+0xd2/0x170
> >
> > vmap_blocks.xa_lock:    862689  862698 ...
> >   -------------------
> >   vmap_blocks.xa_lock   378418    [<00000000625a5626>] vm_map_ram+0x359/0x4a0
> >   vmap_blocks.xa_lock   484280    [<00000000caa2ef03>] xa_erase+0xe/0x30
> >   -------------------
> >   vmap_blocks.xa_lock   576226    [<00000000caa2ef03>] xa_erase+0xe/0x30
> >   vmap_blocks.xa_lock   286472    [<00000000625a5626>] vm_map_ram+0x359/0x4a0
> > ...
> > <snip>
> >
> > that is a result of running vm_map_ram()/vm_unmap_ram() in
> > a loop. The test creates 64(on 64 CPUs system) threads and
> > each one maps/unmaps 1 page.
> >
> > After this change the "xa_lock" can be considered as a noise
> > in the same test condition:
> >
> > <snip>
> > ...
> > &xa->xa_lock#1:         10333 10394 ...
> >   --------------
> >   &xa->xa_lock#1        5349      [<00000000bbbc9751>] xa_erase+0xe/0x30
> >   &xa->xa_lock#1        5045      [<0000000018def45d>] vm_map_ram+0x3a4/0x4f0
> >   --------------
> >   &xa->xa_lock#1        7326      [<0000000018def45d>] vm_map_ram+0x3a4/0x4f0
> >   &xa->xa_lock#1        3068      [<00000000bbbc9751>] xa_erase+0xe/0x30
> > ...
> > <snip>
> >
> 
> Nice! Really good to see contention reduced, but in addition I'm a huge fan
> of us removing the global state in vmalloc and this is a good start.
> 
> I've noticed a small perf regression after 3 runs of ./test_vmalloc.sh
> performance from an average of 119356136169 cycles to 120404645782 or +0.9%
> but this doesn't seem especially egregious.
> 
We are lack of extra vm_map_ram()/vm_unmap_ram() tests in the test_vmalloc.sh.
It would be good to add them to the test-suite.

> > This patch does not fix vmap_area_lock/free_vmap_area_lock and
> > purge_vmap_area_lock bottle-necks, it is rather a separate rework.
> >
> > Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
> > ---
> >  mm/vmalloc.c | 50 ++++++++++++++++++++++++++------------------------
> >  1 file changed, 26 insertions(+), 24 deletions(-)
> >
> > diff --git a/mm/vmalloc.c b/mm/vmalloc.c
> > index 978194dc2bb8..13b5342bed9a 100644
> > --- a/mm/vmalloc.c
> > +++ b/mm/vmalloc.c
> > @@ -1911,6 +1911,7 @@ static struct vmap_area *find_unlink_vmap_area(unsigned long addr)
> >  struct vmap_block_queue {
> >  	spinlock_t lock;
> >  	struct list_head free;
> > +	struct xarray vmap_blocks;
> >  };
> >
> >  struct vmap_block {
> > @@ -1927,25 +1928,18 @@ struct vmap_block {
> >  /* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
> >  static DEFINE_PER_CPU(struct vmap_block_queue, vmap_block_queue);
> >
> > -/*
> > - * XArray of vmap blocks, indexed by address, to quickly find a vmap block
> > - * in the free path. Could get rid of this if we change the API to return a
> > - * "cookie" from alloc, to be passed to free. But no big deal yet.
> > - */
> 
> Doesn't this comment still apply? Or is the idea of returning the "cookie"
> not really viable?
> 
Since a vmap_block_queue is a per-cpu thing, though it is not fully
serialized in terms of per-cpu classical meaning, IMHO, it is not a
big issue.

If we return a cookie then, indeed, we do not need to find a vmap_block
and performance wise it should be better. For how much, i do not know, it
requires data. From the other hand an API has to be changed accordingly.

But i can leave the comment!

> > -static DEFINE_XARRAY(vmap_blocks);
> > -
> > -/*
> > - * We should probably have a fallback mechanism to allocate virtual memory
> > - * out of partially filled vmap blocks. However vmap block sizing should be
> > - * fairly reasonable according to the vmalloc size, so it shouldn't be a
> > - * big problem.
> > - */
> 
> Again, is this comment no longer relevant?
> 
Looks like yes :) But i am not sure i understand correctly what author meant.
It looks like this:

<snip>
void *vm_map_ram(struct page **pages, unsigned int count, int node)
{
	unsigned long size = (unsigned long)count << PAGE_SHIFT;
	unsigned long addr;
	void *mem;

	if (likely(count <= VMAP_MAX_ALLOC)) {
		mem = vb_alloc(size, GFP_KERNEL);
		if (IS_ERR(mem))
			return NULL;
...
<snip>

instead of returning NULL, go directly with a fall-back, that is:
<snip>
	struct vmap_area *va;
	va = alloc_vmap_area(size, PAGE_SIZE,
			VMALLOC_START, VMALLOC_END,
			node, GFP_KERNEL, VMAP_RAM);
	if (IS_ERR(va))
		return NULL;

	addr = va->va_start;
	mem = (void *)addr;
<snip>

> > +static struct vmap_block_queue *
> > +addr_to_vbq(unsigned long addr)
> > +{
> > +	int cpu = (addr / VMAP_BLOCK_SIZE) % num_possible_cpus();
> > +	return &per_cpu(vmap_block_queue, cpu);
> > +}
> 
> Andrew's already commented on this, so I won't dwell but it does seem odd
> to subdivide by number of possible CPUs rather than just use the actual
> CPU. I guess your response to his question will also answer mine :)
> 
I will upload a v2 where i try to explain in detail as much as i can,
after that we can see if there are extra comments or questions and
discuss if so.

> >
> > -static unsigned long addr_to_vb_idx(unsigned long addr)
> > +static unsigned long
> > +addr_to_vb_va_start(unsigned long addr)
> >  {
> > -	addr -= VMALLOC_START & ~(VMAP_BLOCK_SIZE-1);
> > -	addr /= VMAP_BLOCK_SIZE;
> > -	return addr;
> > +	/* A start address of block an address belongs to. */
> 
> A nit, but might be worth referring to the assert in vmap_block_vaddr(), as
> this comment seems a bit redundant otherwise as it is implied by the code
> it comments.
> 
OK. I can remove that comment.

> > +	return rounddown(addr, VMAP_BLOCK_SIZE);
> >  }
> >
> >  static void *vmap_block_vaddr(unsigned long va_start, unsigned long pages_off)
> > @@ -1953,7 +1947,7 @@ static void *vmap_block_vaddr(unsigned long va_start, unsigned long pages_off)
> >  	unsigned long addr;
> >
> >  	addr = va_start + (pages_off << PAGE_SHIFT);
> > -	BUG_ON(addr_to_vb_idx(addr) != addr_to_vb_idx(va_start));
> > +	BUG_ON(addr_to_vb_va_start(addr) != addr_to_vb_va_start(va_start));
> 
> Maybe nitty, but perhaps better to WARN_ON() here to avoid BUG_ON proliferation?
> 
Indeed, it is better to go with WARN_ON() or even WARN_ON_ONCE().

> And can't this be the below?
> 
> WARN_ON(addr_to_vb_va_start(addr) != va_start);
> 
Yep, it can be. Thanks for it!

> >  	return (void *)addr;
> >  }
> >
> > @@ -1970,7 +1964,6 @@ static void *new_vmap_block(unsigned int order, gfp_t gfp_mask)
> >  	struct vmap_block_queue *vbq;
> >  	struct vmap_block *vb;
> >  	struct vmap_area *va;
> > -	unsigned long vb_idx;
> >  	int node, err;
> >  	void *vaddr;
> >
> > @@ -2003,8 +1996,8 @@ static void *new_vmap_block(unsigned int order, gfp_t gfp_mask)
> >  	bitmap_set(vb->used_map, 0, (1UL << order));
> >  	INIT_LIST_HEAD(&vb->free_list);
> >
> > -	vb_idx = addr_to_vb_idx(va->va_start);
> > -	err = xa_insert(&vmap_blocks, vb_idx, vb, gfp_mask);
> > +	vbq = addr_to_vbq(va->va_start);
> > +	err = xa_insert(&vbq->vmap_blocks, va->va_start, vb, gfp_mask);
> 
> This seems actually like a nice subtle improvement in that we are now
> indexing always on va_start explicitly and will always load using
> addr_to_vb_va_start().
>
Yep, because we already have an index, it is a va->va_start.

> 
> >  	if (err) {
> >  		kfree(vb);
> >  		free_vmap_area(va);
> > @@ -2021,9 +2014,11 @@ static void *new_vmap_block(unsigned int order, gfp_t gfp_mask)
> >
> >  static void free_vmap_block(struct vmap_block *vb)
> >  {
> > +	struct vmap_block_queue *vbq;
> >  	struct vmap_block *tmp;
> >
> > -	tmp = xa_erase(&vmap_blocks, addr_to_vb_idx(vb->va->va_start));
> > +	vbq = addr_to_vbq(vb->va->va_start);
> > +	tmp = xa_erase(&vbq->vmap_blocks, vb->va->va_start);
> >  	BUG_ON(tmp != vb);
> >
> >  	spin_lock(&vmap_area_lock);
> > @@ -2135,6 +2130,7 @@ static void vb_free(unsigned long addr, unsigned long size)
> >  	unsigned long offset;
> >  	unsigned int order;
> >  	struct vmap_block *vb;
> > +	struct vmap_block_queue *vbq;
> >
> >  	BUG_ON(offset_in_page(size));
> >  	BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
> > @@ -2143,7 +2139,10 @@ static void vb_free(unsigned long addr, unsigned long size)
> >
> >  	order = get_order(size);
> >  	offset = (addr & (VMAP_BLOCK_SIZE - 1)) >> PAGE_SHIFT;
> > -	vb = xa_load(&vmap_blocks, addr_to_vb_idx(addr));
> > +
> > +	vbq = addr_to_vbq(addr);
> > +	vb = xa_load(&vbq->vmap_blocks, addr_to_vb_va_start(addr));
> > +
> >  	spin_lock(&vb->lock);
> >  	bitmap_clear(vb->used_map, offset, (1UL << order));
> >  	spin_unlock(&vb->lock);
> > @@ -3486,6 +3485,7 @@ static void vmap_ram_vread(char *buf, char *addr, int count, unsigned long flags
> >  {
> >  	char *start;
> >  	struct vmap_block *vb;
> > +	struct vmap_block_queue *vbq;
> >  	unsigned long offset;
> >  	unsigned int rs, re, n;
> >
> > @@ -3503,7 +3503,8 @@ static void vmap_ram_vread(char *buf, char *addr, int count, unsigned long flags
> >  	 * Area is split into regions and tracked with vmap_block, read out
> >  	 * each region and zero fill the hole between regions.
> >  	 */
> > -	vb = xa_load(&vmap_blocks, addr_to_vb_idx((unsigned long)addr));
> > +	vbq = addr_to_vbq((unsigned long) addr);
> > +	vb = xa_load(&vbq->vmap_blocks, addr_to_vb_va_start((unsigned long) addr));
> >  	if (!vb)
> >  		goto finished;
> >
> > @@ -4272,6 +4273,7 @@ void __init vmalloc_init(void)
> >  		p = &per_cpu(vfree_deferred, i);
> >  		init_llist_head(&p->list);
> >  		INIT_WORK(&p->wq, delayed_vfree_work);
> > +		xa_init(&vbq->vmap_blocks);
> >  	}
> >
> >  	/* Import existing vmlist entries. */
> > --
> > 2.30.2
> >