mm: vmalloc: remove a global vmap_blocks xarray

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>

Running the test_vmalloc.sh run_test_mask=1024 nr_threads=64 nr_pages=5
shows around ~8 percent of throughput improvement of vm_map_ram() and
vm_unmap_ram() APIs.

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.

Link: https://lkml.kernel.org/r/20230330190639.431589-1-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Lorenzo Stoakes <lstoakes@gmail.com>
Reviewed-by: Baoquan He <bhe@redhat.com>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sony.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

Author: urezki

mm/vmalloc.c

@@ -1915,6 +1915,13 @@ static struct vmap_area *find_unlink_vmap_area(unsigned long addr)
 struct vmap_block_queue {
 	spinlock_t lock;
 	struct list_head free;
+
+	/*
+	 * An xarray requires an extra memory dynamically to
+	 * be allocated. If it is an issue, we can use rb-tree
+	 * instead.
+	 */
+	struct xarray vmap_blocks;
 };
 
 struct vmap_block {
@@ -1932,11 +1939,48 @@ struct vmap_block {
 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.
+ * In order to fast access to any "vmap_block" associated with a
+ * specific address, we use a hash.
+ *
+ * A per-cpu vmap_block_queue is used in both ways, to serialize
+ * an access to free block chains among CPUs(alloc path) and it
+ * also acts as a vmap_block hash(alloc/free paths). It means we
+ * overload it, since we already have the per-cpu array which is
+ * used as a hash table. When used as a hash a 'cpu' passed to
+ * per_cpu() is not actually a CPU but rather a hash index.
+ *
+ * A hash function is addr_to_vb_xarray() which hashes any address
+ * to a specific index(in a hash) it belongs to. This then uses a
+ * per_cpu() macro to access an array with generated index.
+ *
+ * An example:
+ *
+ *  CPU_1  CPU_2  CPU_0
+ *    |      |      |
+ *    V      V      V
+ * 0     10     20     30     40     50     60
+ * |------|------|------|------|------|------|...<vmap address space>
+ *   CPU0   CPU1   CPU2   CPU0   CPU1   CPU2
+ *
+ * - CPU_1 invokes vm_unmap_ram(6), 6 belongs to CPU0 zone, thus
+ *   it access: CPU0/INDEX0 -> vmap_blocks -> xa_lock;
+ *
+ * - CPU_2 invokes vm_unmap_ram(11), 11 belongs to CPU1 zone, thus
+ *   it access: CPU1/INDEX1 -> vmap_blocks -> xa_lock;
+ *
+ * - CPU_0 invokes vm_unmap_ram(20), 20 belongs to CPU2 zone, thus
+ *   it access: CPU2/INDEX2 -> vmap_blocks -> xa_lock.
+ *
+ * This technique almost always avoids lock contention on insert/remove,
+ * however xarray spinlocks protect against any contention that remains.
  */
-static DEFINE_XARRAY(vmap_blocks);
+static struct xarray *
+addr_to_vb_xarray(unsigned long addr)
+{
+	int index = (addr / VMAP_BLOCK_SIZE) % num_possible_cpus();
+
+	return &per_cpu(vmap_block_queue, index).vmap_blocks;
+}
 
 /*
  * We should probably have a fallback mechanism to allocate virtual memory
@@ -1974,6 +2018,7 @@ 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;
+	struct xarray *xa;
 	unsigned long vb_idx;
 	int node, err;
 	void *vaddr;
@@ -2007,8 +2052,9 @@ 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);
 
+	xa = addr_to_vb_xarray(va->va_start);
 	vb_idx = addr_to_vb_idx(va->va_start);
-	err = xa_insert(&vmap_blocks, vb_idx, vb, gfp_mask);
+	err = xa_insert(xa, vb_idx, vb, gfp_mask);
 	if (err) {
 		kfree(vb);
 		free_vmap_area(va);
@@ -2026,8 +2072,10 @@ static void *new_vmap_block(unsigned int order, gfp_t gfp_mask)
 static void free_vmap_block(struct vmap_block *vb)
 {
 	struct vmap_block *tmp;
+	struct xarray *xa;
 
-	tmp = xa_erase(&vmap_blocks, addr_to_vb_idx(vb->va->va_start));
+	xa = addr_to_vb_xarray(vb->va->va_start);
+	tmp = xa_erase(xa, addr_to_vb_idx(vb->va->va_start));
 	BUG_ON(tmp != vb);
 
 	spin_lock(&vmap_area_lock);
@@ -2139,6 +2187,7 @@ static void vb_free(unsigned long addr, unsigned long size)
 	unsigned long offset;
 	unsigned int order;
 	struct vmap_block *vb;
+	struct xarray *xa;
 
 	BUG_ON(offset_in_page(size));
 	BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
@@ -2147,7 +2196,10 @@ 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));
+
+	xa = addr_to_vb_xarray(addr);
+	vb = xa_load(xa, addr_to_vb_idx(addr));
+
 	spin_lock(&vb->lock);
 	bitmap_clear(vb->used_map, offset, (1UL << order));
 	spin_unlock(&vb->lock);
@@ -3525,6 +3577,7 @@ static size_t vmap_ram_vread_iter(struct iov_iter *iter, const char *addr,
 {
 	char *start;
 	struct vmap_block *vb;
+	struct xarray *xa;
 	unsigned long offset;
 	unsigned int rs, re;
 	size_t remains, n;
@@ -3543,7 +3596,8 @@ static size_t vmap_ram_vread_iter(struct iov_iter *iter, const char *addr,
 	 * 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));
+	xa = addr_to_vb_xarray((unsigned long) addr);
+	vb = xa_load(xa, addr_to_vb_idx((unsigned long)addr));
 	if (!vb)
 		goto finished_zero;
 
@@ -4337,6 +4391,7 @@ 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. */