On Tue, Mar 28, 2023 at 11:25:54AM +0800, Baoquan He wrote:
> On 03/27/23 at 07:01pm, 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.
>
> With my understanding, the xarray will take more time when calling
> xa_insert() or xa_erase() because these two will cause xa_expand() and
> xa_shrink() if the index is sparse. xa_load() should be low cost to
> finish. Wondering if in your testing code, the mapping address is close
> or too far.
>
> 1 mm/vmalloc.c <<new_vmap_block>>
> err = xa_insert(&vmap_blocks, vb_idx, vb, gfp_mask);
> 2 mm/vmalloc.c <<free_vmap_block>>
> tmp = xa_erase(&vmap_blocks, addr_to_vb_idx(vb->va->va_start));
> 3 mm/vmalloc.c <<vb_free>>
> vb = xa_load(&vmap_blocks, addr_to_vb_idx(addr));
> 4 mm/vmalloc.c <<vmap_ram_vread_iter>>
> vb = xa_load(&vmap_blocks, addr_to_vb_idx((unsigned long )addr));
>
> >
> > 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.
> >
> > v1 - v2:
> > - Add more comments(Andrew Morton req.)
> > - Switch to WARN_ON_ONCE(Lorenzo Stoakes req.)
> >
> > v2 -> v3:
> > - Fix a kernel-doc complain(Matthew Wilcox)
> >
> > Signed-off-by: Uladzislau Rezki (Sony) <urezki@xxxxxxxxx>
> > ---
> > mm/vmalloc.c | 85 +++++++++++++++++++++++++++++++++++++++-------------
> > 1 file changed, 64 insertions(+), 21 deletions(-)
> >
> > diff --git a/mm/vmalloc.c b/mm/vmalloc.c
> > index 978194dc2bb8..821256ecf81c 100644
> > --- a/mm/vmalloc.c
> > +++ b/mm/vmalloc.c
> > @@ -1908,9 +1908,22 @@ static struct vmap_area *find_unlink_vmap_area(unsigned long addr)
> > #define VMAP_BLOCK 0x2 /* mark out the vmap_block sub-type*/
> > #define VMAP_FLAGS_MASK 0x3
> >
> > +/*
> > + * 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.
> > + */
> > 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 {
> > @@ -1928,24 +1941,46 @@ 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 store them into a per-cpu xarray. A hash
> > + * function is addr_to_vbq() whereas a key is a vb->va->va_start
> > + * value.
> > + *
> > + * Please note, a vmap_block_queue, which is a per-cpu, is not
> > + * serialized by a raw_smp_processor_id() current CPU, instead
> > + * it is chosen based on a CPU-index it belongs to, i.e. it is
> > + * a hash-table.
> > + *
> > + * 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.
> > */
> > -static DEFINE_XARRAY(vmap_blocks);
> > +static struct vmap_block_queue *
> > +addr_to_vbq(unsigned long addr)
> > +{
> > + int index = (addr / VMAP_BLOCK_SIZE) % num_possible_cpus();
> >
> > -/*
> > - * 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.
> > - */
> > + return &per_cpu(vmap_block_queue, index);
> > +}
> >
> > -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;
> > + return rounddown(addr, VMAP_BLOCK_SIZE);
> > }
> >
> > static void *vmap_block_vaddr(unsigned long va_start, unsigned long pages_off)
> > @@ -1953,7 +1988,7 @@ 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));
> > + WARN_ON_ONCE(addr_to_vb_va_start(addr) != va_start);
> > return (void *)addr;
> > }
> >
> > @@ -1970,7 +2005,6 @@ 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 +2037,8 @@ 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);
>
> Using va->va_start as index to access xarray may cost extra memory.
> Imagine we got a virtual address at VMALLOC_START, its region is
> [VMALLOC_START, VMALLOC_START+4095]. In the xarray, its sequence order
> is 0. While with va->va_start, it's 0xffffc90000000000UL on x86_64 with
> level4 paging mode. That means for the first page size vmalloc area,
> storing it into xarray need about 10 levels of xa_node, just for the one
> page size. With the old addr_to_vb_idx(), its index is 0. Only one level
> height is needed. One xa_node is about 72bytes, it could take more time
> and memory to access va->va_start. Not sure if my understanding is correct.
>
> static unsigned long addr_to_vb_idx(unsigned long addr)
> {
> addr -= VMALLOC_START & ~(VMAP_BLOCK_SIZE-1);
> addr /= VMAP_BLOCK_SIZE;
> return addr;
> }
>
If the size of array depends on index "length", then, indeed it will require
more memory. From the other hand we can keep the old addr_to_vb_idx() function
in order to "cut" a va->va_start index.
--
Uladzislau Rezki
