[PATCH v2 1/3] mm: refactor __vunmap() to avoid duplicated call to find_vm_area()

[PATCH v2 1/3] mm: refactor __vunmap() to avoid duplicated call to find_vm_area()

[Roman Gushchin]

__vunmap() calls find_vm_area() twice without an obvious reason:
first directly to get the area pointer, second indirectly by calling
remove_vm_area(), which is again searching for the area.

To remove this redundancy, let's split remove_vm_area() into
__remove_vm_area(struct vmap_area *), which performs the actual area
removal, and remove_vm_area(const void *addr) wrapper, which can
be used everywhere, where it has been used before.

On my test setup, I've got 5-10% speed up on vfree()'ing 1000000
of 4-pages vmalloc blocks.

Signed-off-by: Roman Gushchin <guro@fb.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Matthew Wilcox <willy@infradead.org>
---
 mm/vmalloc.c | 47 +++++++++++++++++++++++++++--------------------
 1 file changed, 27 insertions(+), 20 deletions(-)

diff --git a/mm/vmalloc.c b/mm/vmalloc.c
index b7455d4c8c12..8f0179895fb5 100644
--- a/mm/vmalloc.c
+++ b/mm/vmalloc.c
@@ -1477,6 +1477,24 @@ struct vm_struct *find_vm_area(const void *addr)
 	return NULL;
 }
 
+static struct vm_struct *__remove_vm_area(struct vmap_area *va)
+{
+	struct vm_struct *vm = va->vm;
+
+	might_sleep();
+
+	spin_lock(&vmap_area_lock);
+	va->vm = NULL;
+	va->flags &= ~VM_VM_AREA;
+	va->flags |= VM_LAZY_FREE;
+	spin_unlock(&vmap_area_lock);
+
+	kasan_free_shadow(vm);
+	free_unmap_vmap_area(va);
+
+	return vm;
+}
+
 /**
  * remove_vm_area - find and remove a continuous kernel virtual area
  * @addr:	    base address
@@ -1489,31 +1507,20 @@ struct vm_struct *find_vm_area(const void *addr)
  */
 struct vm_struct *remove_vm_area(const void *addr)
 {
+	struct vm_struct *vm = NULL;
 	struct vmap_area *va;
 
-	might_sleep();
-
 	va = find_vmap_area((unsigned long)addr);
-	if (va && va->flags & VM_VM_AREA) {
-		struct vm_struct *vm = va->vm;
-
-		spin_lock(&vmap_area_lock);
-		va->vm = NULL;
-		va->flags &= ~VM_VM_AREA;
-		va->flags |= VM_LAZY_FREE;
-		spin_unlock(&vmap_area_lock);
-
-		kasan_free_shadow(vm);
-		free_unmap_vmap_area(va);
+	if (va && va->flags & VM_VM_AREA)
+		vm = __remove_vm_area(va);
 
-		return vm;
-	}
-	return NULL;
+	return vm;
 }
 
 static void __vunmap(const void *addr, int deallocate_pages)
 {
 	struct vm_struct *area;
+	struct vmap_area *va;
 
 	if (!addr)
 		return;
@@ -1522,17 +1529,18 @@ static void __vunmap(const void *addr, int deallocate_pages)
 			addr))
 		return;
 
-	area = find_vm_area(addr);
-	if (unlikely(!area)) {
+	va = find_vmap_area((unsigned long)addr);
+	if (unlikely(!va || !(va->flags & VM_VM_AREA))) {
 		WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
 				addr);
 		return;
 	}
 
+	area = va->vm;
 	debug_check_no_locks_freed(area->addr, get_vm_area_size(area));
 	debug_check_no_obj_freed(area->addr, get_vm_area_size(area));
 
-	remove_vm_area(addr);
+	__remove_vm_area(va);
 	if (deallocate_pages) {
 		int i;
 
@@ -1547,7 +1555,6 @@ static void __vunmap(const void *addr, int deallocate_pages)
 	}
 
 	kfree(area);
-	return;
 }
 
 static inline void __vfree_deferred(const void *addr)
-- 
2.20.1

[PATCH v2 2/3] mm: separate memory allocation and actual work in alloc_vmap_area()

[Roman Gushchin]

alloc_vmap_area() is allocating memory for the vmap_area, and
performing the actual lookup of the vm area and vmap_area
initialization.

This prevents us from using a pre-allocated memory for the map_area
structure, which can be used in some cases to minimize the number
of required memory allocations.

Let's keep the memory allocation part in alloc_vmap_area() and
separate everything else into init_vmap_area().

Signed-off-by: Roman Gushchin <guro@fb.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Matthew Wilcox <willy@infradead.org>
---
 mm/vmalloc.c | 50 +++++++++++++++++++++++++++++++++-----------------
 1 file changed, 33 insertions(+), 17 deletions(-)

diff --git a/mm/vmalloc.c b/mm/vmalloc.c
index 8f0179895fb5..f1f19d1105c4 100644
--- a/mm/vmalloc.c
+++ b/mm/vmalloc.c
@@ -395,16 +395,10 @@ static void purge_vmap_area_lazy(void);
 
 static BLOCKING_NOTIFIER_HEAD(vmap_notify_list);
 
-/*
- * Allocate a region of KVA of the specified size and alignment, within the
- * vstart and vend.
- */
-static struct vmap_area *alloc_vmap_area(unsigned long size,
-				unsigned long align,
-				unsigned long vstart, unsigned long vend,
-				int node, gfp_t gfp_mask)
+static int init_vmap_area(struct vmap_area *va, unsigned long size,
+			  unsigned long align, unsigned long vstart,
+			  unsigned long vend, int node, gfp_t gfp_mask)
 {
-	struct vmap_area *va;
 	struct rb_node *n;
 	unsigned long addr;
 	int purged = 0;
@@ -416,11 +410,6 @@ static struct vmap_area *alloc_vmap_area(unsigned long size,
 
 	might_sleep();
 
-	va = kmalloc_node(sizeof(struct vmap_area),
-			gfp_mask & GFP_RECLAIM_MASK, node);
-	if (unlikely(!va))
-		return ERR_PTR(-ENOMEM);
-
 	/*
 	 * Only scan the relevant parts containing pointers to other objects
 	 * to avoid false negatives.
@@ -516,7 +505,7 @@ static struct vmap_area *alloc_vmap_area(unsigned long size,
 	BUG_ON(va->va_start < vstart);
 	BUG_ON(va->va_end > vend);
 
-	return va;
+	return 0;
 
 overflow:
 	spin_unlock(&vmap_area_lock);
@@ -538,8 +527,35 @@ static struct vmap_area *alloc_vmap_area(unsigned long size,
 	if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit())
 		pr_warn("vmap allocation for size %lu failed: use vmalloc=<size> to increase size\n",
 			size);
-	kfree(va);
-	return ERR_PTR(-EBUSY);
+
+	return -EBUSY;
+}
+
+/*
+ * Allocate a region of KVA of the specified size and alignment, within the
+ * vstart and vend.
+ */
+static struct vmap_area *alloc_vmap_area(unsigned long size,
+					 unsigned long align,
+					 unsigned long vstart,
+					 unsigned long vend,
+					 int node, gfp_t gfp_mask)
+{
+	struct vmap_area *va;
+	int ret;
+
+	va = kmalloc_node(sizeof(struct vmap_area),
+			gfp_mask & GFP_RECLAIM_MASK, node);
+	if (unlikely(!va))
+		return ERR_PTR(-ENOMEM);
+
+	ret = init_vmap_area(va, size, align, vstart, vend, node, gfp_mask);
+	if (ret) {
+		kfree(va);
+		return ERR_PTR(ret);
+	}
+
+	return va;
 }
 
 int register_vmap_purge_notifier(struct notifier_block *nb)
-- 
2.20.1

[PATCH v2 3/3] mm: show number of vmalloc pages in /proc/meminfo

[Roman Gushchin]

Vmalloc() is getting more and more used these days (kernel stacks,
bpf and percpu allocator are new top users), and the total %
of memory consumed by vmalloc() can be pretty significant
and changes dynamically.

/proc/meminfo is the best place to display this information:
its top goal is to show top consumers of the memory.

Since the VmallocUsed field in /proc/meminfo is not in use
for quite a long time (it has been defined to 0 by the
commit a5ad88ce8c7f ("mm: get rid of 'vmalloc_info' from
/proc/meminfo")), let's reuse it for showing the actual
physical memory consumption of vmalloc().

Signed-off-by: Roman Gushchin <guro@fb.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Matthew Wilcox <willy@infradead.org>
---
 fs/proc/meminfo.c       |  2 +-
 include/linux/vmalloc.h |  2 ++
 mm/vmalloc.c            | 16 ++++++++++++++++
 3 files changed, 19 insertions(+), 1 deletion(-)

diff --git a/fs/proc/meminfo.c b/fs/proc/meminfo.c
index 568d90e17c17..465ea0153b2a 100644
--- a/fs/proc/meminfo.c
+++ b/fs/proc/meminfo.c
@@ -120,7 +120,7 @@ static int meminfo_proc_show(struct seq_file *m, void *v)
 	show_val_kb(m, "Committed_AS:   ", committed);
 	seq_printf(m, "VmallocTotal:   %8lu kB\n",
 		   (unsigned long)VMALLOC_TOTAL >> 10);
-	show_val_kb(m, "VmallocUsed:    ", 0ul);
+	show_val_kb(m, "VmallocUsed:    ", vmalloc_nr_pages());
 	show_val_kb(m, "VmallocChunk:   ", 0ul);
 	show_val_kb(m, "Percpu:         ", pcpu_nr_pages());
 
diff --git a/include/linux/vmalloc.h b/include/linux/vmalloc.h
index 398e9c95cd61..0b497408272b 100644
--- a/include/linux/vmalloc.h
+++ b/include/linux/vmalloc.h
@@ -63,10 +63,12 @@ extern void vm_unmap_aliases(void);
 
 #ifdef CONFIG_MMU
 extern void __init vmalloc_init(void);
+extern unsigned long vmalloc_nr_pages(void);
 #else
 static inline void vmalloc_init(void)
 {
 }
+static inline unsigned long vmalloc_nr_pages(void) { return 0; }
 #endif
 
 extern void *vmalloc(unsigned long size);
diff --git a/mm/vmalloc.c b/mm/vmalloc.c
index f1f19d1105c4..8dd490d8d191 100644
--- a/mm/vmalloc.c
+++ b/mm/vmalloc.c
@@ -340,6 +340,19 @@ static unsigned long cached_align;
 
 static unsigned long vmap_area_pcpu_hole;
 
+static DEFINE_PER_CPU(unsigned long, nr_vmalloc_pages);
+
+unsigned long vmalloc_nr_pages(void)
+{
+	unsigned long pages = 0;
+	int cpu;
+
+	for_each_possible_cpu(cpu)
+		pages += per_cpu(nr_vmalloc_pages, cpu);
+
+	return pages;
+}
+
 static struct vmap_area *__find_vmap_area(unsigned long addr)
 {
 	struct rb_node *n = vmap_area_root.rb_node;
@@ -1566,6 +1579,7 @@ static void __vunmap(const void *addr, int deallocate_pages)
 			BUG_ON(!page);
 			__free_pages(page, 0);
 		}
+		this_cpu_sub(nr_vmalloc_pages, area->nr_pages);
 
 		kvfree(area->pages);
 	}
@@ -1742,12 +1756,14 @@ static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
 		if (unlikely(!page)) {
 			/* Successfully allocated i pages, free them in __vunmap() */
 			area->nr_pages = i;
+			this_cpu_add(nr_vmalloc_pages, area->nr_pages);
 			goto fail;
 		}
 		area->pages[i] = page;
 		if (gfpflags_allow_blocking(gfp_mask|highmem_mask))
 			cond_resched();
 	}
+	this_cpu_add(nr_vmalloc_pages, area->nr_pages);
 
 	if (map_vm_area(area, prot, pages))
 		goto fail;
-- 
2.20.1

Re: [PATCH v2 0/3] vmalloc enhancements

[Johannes Weiner]

On Tue, Feb 12, 2019 at 09:56:45AM -0800, Roman Gushchin wrote:
> The patchset contains few changes to the vmalloc code, which are
> leading to some performance gains and code simplification.
> 
> Also, it exports a number of pages, used by vmalloc(),
> in /proc/meminfo.
> 
> Patch (1) removes some redundancy on __vunmap().
> Patch (2) separates memory allocation and data initialization
>   in alloc_vmap_area()
> Patch (3) adds vmalloc counter to /proc/meminfo.
> 
> v2->v1:
>   - rebased on top of current mm tree
>   - switch from atomic to percpu vmalloc page counter

I don't understand what prompted this change to percpu counters.

All writers already write vmap_area_lock and vmap_area_list, so it's
not really saving much. The for_each_possible_cpu() for /proc/meminfo
on the other hand is troublesome.

Re: [PATCH v2 0/3] vmalloc enhancements

[Andrew Morton]

On Tue, 12 Feb 2019 13:47:24 -0500 Johannes Weiner <hannes@cmpxchg.org> wrote:

> On Tue, Feb 12, 2019 at 09:56:45AM -0800, Roman Gushchin wrote:
> > The patchset contains few changes to the vmalloc code, which are
> > leading to some performance gains and code simplification.
> > 
> > Also, it exports a number of pages, used by vmalloc(),
> > in /proc/meminfo.
> > 
> > Patch (1) removes some redundancy on __vunmap().
> > Patch (2) separates memory allocation and data initialization
> >   in alloc_vmap_area()
> > Patch (3) adds vmalloc counter to /proc/meminfo.
> > 
> > v2->v1:
> >   - rebased on top of current mm tree
> >   - switch from atomic to percpu vmalloc page counter
> 
> I don't understand what prompted this change to percpu counters.
> 
> All writers already write vmap_area_lock and vmap_area_list, so it's
> not really saving much. The for_each_possible_cpu() for /proc/meminfo
> on the other hand is troublesome.

percpu_counters would fit here.  They have probably-unneeded locking
but I expect that will be acceptable.

And they address the issues with for_each_possible_cpu() avoidance, CPU
hotplug and transient negative values.

Re: [PATCH v2 0/3] vmalloc enhancements

[Roman Gushchin]

On Tue, Feb 12, 2019 at 12:34:09PM -0800, Andrew Morton wrote:
> On Tue, 12 Feb 2019 13:47:24 -0500 Johannes Weiner <hannes@cmpxchg.org> wrote:
> 
> > On Tue, Feb 12, 2019 at 09:56:45AM -0800, Roman Gushchin wrote:
> > > The patchset contains few changes to the vmalloc code, which are
> > > leading to some performance gains and code simplification.
> > > 
> > > Also, it exports a number of pages, used by vmalloc(),
> > > in /proc/meminfo.
> > > 
> > > Patch (1) removes some redundancy on __vunmap().
> > > Patch (2) separates memory allocation and data initialization
> > >   in alloc_vmap_area()
> > > Patch (3) adds vmalloc counter to /proc/meminfo.
> > > 
> > > v2->v1:
> > >   - rebased on top of current mm tree
> > >   - switch from atomic to percpu vmalloc page counter
> > 
> > I don't understand what prompted this change to percpu counters.

I *think*, I see some performance difference, but it's barely measurable
in my setup. Also as I remember, Matthew was asking why not percpu here.
So if everybody prefers a global atomic, I'm fine with either.

> > 
> > All writers already write vmap_area_lock and vmap_area_list, so it's
> > not really saving much. The for_each_possible_cpu() for /proc/meminfo
> > on the other hand is troublesome.
> 
> percpu_counters would fit here.  They have probably-unneeded locking
> but I expect that will be acceptable.
> 
> And they address the issues with for_each_possible_cpu() avoidance, CPU
> hotplug and transient negative values.

Not sure, because percpu_counters are based on dynamic percpu allocations,
which are using vmalloc under the hood.

Thanks!

Re: [PATCH v2 0/3] vmalloc enhancements

[Matthew Wilcox]

On Tue, Feb 12, 2019 at 10:36:12PM +0000, Roman Gushchin wrote:
> On Tue, Feb 12, 2019 at 12:34:09PM -0800, Andrew Morton wrote:
> > On Tue, 12 Feb 2019 13:47:24 -0500 Johannes Weiner <hannes@cmpxchg.org> wrote:
> > > I don't understand what prompted this change to percpu counters.
> 
> I *think*, I see some performance difference, but it's barely measurable
> in my setup. Also as I remember, Matthew was asking why not percpu here.
> So if everybody prefers a global atomic, I'm fine with either.

I was asking why you were using an accessor instead of a direct reference
to the atomic_long_t.

Re: [PATCH v2 0/3] vmalloc enhancements

[Roman Gushchin]

On Tue, Feb 12, 2019 at 12:34:09PM -0800, Andrew Morton wrote:
> On Tue, 12 Feb 2019 13:47:24 -0500 Johannes Weiner <hannes@cmpxchg.org> wrote:
> 
> > On Tue, Feb 12, 2019 at 09:56:45AM -0800, Roman Gushchin wrote:
> > > The patchset contains few changes to the vmalloc code, which are
> > > leading to some performance gains and code simplification.
> > > 
> > > Also, it exports a number of pages, used by vmalloc(),
> > > in /proc/meminfo.
> > > 
> > > Patch (1) removes some redundancy on __vunmap().
> > > Patch (2) separates memory allocation and data initialization
> > >   in alloc_vmap_area()
> > > Patch (3) adds vmalloc counter to /proc/meminfo.
> > > 
> > > v2->v1:
> > >   - rebased on top of current mm tree
> > >   - switch from atomic to percpu vmalloc page counter
> > 
> > I don't understand what prompted this change to percpu counters.
> > 
> > All writers already write vmap_area_lock and vmap_area_list, so it's
> > not really saving much. The for_each_possible_cpu() for /proc/meminfo
> > on the other hand is troublesome.
> 
> percpu_counters would fit here.  They have probably-unneeded locking
> but I expect that will be acceptable.
> 
> And they address the issues with for_each_possible_cpu() avoidance, CPU
> hotplug and transient negative values.

Using existing vmap_area_lock (as Johannes suggested) is also problematic,
due to different life-cycles of vma_areas and vmalloc pages. A special flag
will be required to decrease the counter during the lazy deletion of
vmap_areas. Allocation path will require passing a bool flag through too many
nested functions. Also it will be semi-accurate, which is probably tolerable.
So, it's doable, but doesn't look nice to me.

So, using a simple per-cpu counter still seems to best option.
Transient negative value is a valid concern, but easily fixable.
Are there any other? What's the problem with for_each_possible_cpu()?
Reading /proc/meminfo is not that hot, no?

Thanks!