| Message ID | 20230327170126.406044-1-urezki@gmail.com (mailing list archive) |
|---|---|
| State | New |
| Headers | show |
| Series | [v3,1/2] mm: vmalloc: Remove a global vmap_blocks xarray | expand |
On Mon, Mar 27, 2023 at 07:01:25PM +0200, 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> > > 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@gmail.com> > --- > 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. > */ OK so if I understand this correctly, you're overloading the per-CPU vmap_block_queue array to use as a simple hash based on the address and relying on the xa_lock() in xa_insert() to serialise in case of contention? I like the general heft of your comment but I feel this could be spelled out a little more clearly, something like:- In order to have fast access to any vmap_block object associated with a specific address, we use a hash. Rather than waste space on defining a new hash table we take advantage of the fact we already have a static per-cpu array vmap_block_queue. This is already used for per-CPU access to the block queue, however we overload this to _also_ act as a vmap_block hash. The hash function is addr_to_vbq() which hashes on vb->va->va_start. This then uses per_cpu() to lookup the _index_ rather than the _cpu_. Each vmap_block_queue contains an xarray of vmap blocks which are indexed on the same key as the hash (vb->va->va_start). xarray read acceses are protected by RCU lock and inserts are protected by a spin lock so there is no risk of a race here. An example: ... Feel free to cut this down as needed :) but I do feel it's important to _explicitly_ point out that we're overloading this as it's quite confusing at face value. > -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); I might be being pedantic here, but shortly after this code you reassign vbq:- 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); return ERR_PTR(err); } vbq = raw_cpu_ptr(&vmap_block_queue); Which is confusing at a glance, as you're using it once as a hash lookup and again for its 'true purpose'. I wonder whether it would be better overall, since you always follow a vbq lookup explicitly with an operation on vmap_blocks, to just add a helper that returned a pointer to the xarray? e.g. (untested code here :):- static struct xarray *get_vblock_array(unsigned long addr) { struct vmap_block_queue *vbq; int index = (addr / VMAP_BLOCK_SIZE) % num_possible_cpus(); vbq = &per_cpu(vmap_block_queue, index); return &vbq->vblocks; } And replace addr_to_vbq() with this. That'd also make the mechanism of this hash lookup super explicit. > if (err) { > kfree(vb); > free_vmap_area(va); > @@ -2021,9 +2055,11 @@ 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 +2171,7 @@ 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 +2180,10 @@ 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 +3526,7 @@ 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 +3544,8 @@ 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 +4314,7 @@ 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 >
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@gmail.com> > --- > 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 (err) { > kfree(vb); > free_vmap_area(va); > @@ -2021,9 +2055,11 @@ 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 +2171,7 @@ 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 +2180,10 @@ 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 +3526,7 @@ 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 +3544,8 @@ 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 +4314,7 @@ 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 >
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@gmail.com> > > --- > > 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
On Mon, Mar 27, 2023 at 09:09:32PM +0100, Lorenzo Stoakes wrote: > On Mon, Mar 27, 2023 at 07:01:25PM +0200, 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> > > > > 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@gmail.com> > > --- > > 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. > > */ > > OK so if I understand this correctly, you're overloading the per-CPU > vmap_block_queue array to use as a simple hash based on the address and > relying on the xa_lock() in xa_insert() to serialise in case of contention? > > I like the general heft of your comment but I feel this could be spelled > out a little more clearly, something like:- > > In order to have fast access to any vmap_block object associated with a > specific address, we use a hash. > > Rather than waste space on defining a new hash table we take advantage > of the fact we already have a static per-cpu array vmap_block_queue. > > This is already used for per-CPU access to the block queue, however we > overload this to _also_ act as a vmap_block hash. The hash function is > addr_to_vbq() which hashes on vb->va->va_start. > > This then uses per_cpu() to lookup the _index_ rather than the > _cpu_. Each vmap_block_queue contains an xarray of vmap blocks which are > indexed on the same key as the hash (vb->va->va_start). > > xarray read acceses are protected by RCU lock and inserts are protected > by a spin lock so there is no risk of a race here. > > An example: > > ... > > Feel free to cut this down as needed :) but I do feel it's important to > _explicitly_ point out that we're overloading this as it's quite confusing > at face value. > > > -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); > > I might be being pedantic here, but shortly after this code you reassign vbq:- > > 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); > return ERR_PTR(err); > } > > vbq = raw_cpu_ptr(&vmap_block_queue); > > Which is confusing at a glance, as you're using it once as a hash lookup > and again for its 'true purpose'. > > I wonder whether it would be better overall, since you always follow a vbq > lookup explicitly with an operation on vmap_blocks, to just add a helper > that returned a pointer to the xarray? e.g. (untested code here :):- > > static struct xarray *get_vblock_array(unsigned long addr) > { > struct vmap_block_queue *vbq; > int index = (addr / VMAP_BLOCK_SIZE) % num_possible_cpus(); > > vbq = &per_cpu(vmap_block_queue, index); > return &vbq->vblocks; > } > > And replace addr_to_vbq() with this. That'd also make the mechanism of this > hash lookup super explicit. > Thank you for the comments. I will go through all of them and fix accordingly. At lease i see that i have to update the documentation in more better way! Thanks! -- Uladzislau Rezki
> > /* > > - * 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. > > */ > > OK so if I understand this correctly, you're overloading the per-CPU > vmap_block_queue array to use as a simple hash based on the address and > relying on the xa_lock() in xa_insert() to serialise in case of contention? > Sorry i missed your question. You correctly understood what i am doing. Basically, we can associate any address with an index in per-cpu-array. Since a CPU pre-allocates a fixed block size, which is a VMAP_BLOCK_SIZE, we can map any address within this block to a certain index or i call it a specific CPU zone it belongs to. If we want fully serialize it we have to allocate a new vmap block in CPU owner zone. According to ASCII picture, for CPU0 it is 0-20, 30-40 addresses. In fact, even though it would be "fully" serialized, in practise id does not give a visible performance. So this is not needed and it has extra drawbacks. -- Uladzislau Rezki
On 03/28/23 at 02:34pm, Uladzislau Rezki wrote: ...... > > > @@ -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. Yeah, the extra 10 levels of xa_node is unnecessary if we keep the old addr_to_vb_idx(). And the prolonged path will cost more time to reach the wanted leaf node. E.g on x86_64 with 4 level paging mode, vmalloc area is 32TB. With the old calculation, its index range is [0, 8M], 4 level heights of xa_node at most is enough to cover.
On Wed, Mar 29, 2023 at 12:33:05PM +0800, Baoquan He wrote: > On 03/28/23 at 02:34pm, Uladzislau Rezki wrote: > ...... > > > > @@ -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. > > Yeah, the extra 10 levels of xa_node is unnecessary if we keep the old > addr_to_vb_idx(). And the prolonged path will cost more time to reach the > wanted leaf node. E.g on x86_64 with 4 level paging mode, vmalloc area > is 32TB. With the old calculation, its index range is [0, 8M], 4 level > heights of xa_node at most is enough to cover. > Good! I have not analyzed how xarray stores its indexes. I will update the patch to cut indexes so we stay the same as we used to be before. -- Uladzislau Rezki
Hello, Lorenzo! > > /* > > - * 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. > > */ > > OK so if I understand this correctly, you're overloading the per-CPU > vmap_block_queue array to use as a simple hash based on the address and > relying on the xa_lock() in xa_insert() to serialise in case of contention? > > I like the general heft of your comment but I feel this could be spelled > out a little more clearly, something like:- > > In order to have fast access to any vmap_block object associated with a > specific address, we use a hash. > > Rather than waste space on defining a new hash table we take advantage > of the fact we already have a static per-cpu array vmap_block_queue. > > This is already used for per-CPU access to the block queue, however we > overload this to _also_ act as a vmap_block hash. The hash function is > addr_to_vbq() which hashes on vb->va->va_start. > > This then uses per_cpu() to lookup the _index_ rather than the > _cpu_. Each vmap_block_queue contains an xarray of vmap blocks which are > indexed on the same key as the hash (vb->va->va_start). > > xarray read acceses are protected by RCU lock and inserts are protected > by a spin lock so there is no risk of a race here. > /* * 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. * * A hash function is addr_to_vbq() which hashes any address to * a specific index(in a hash) it belongs to. This then uses a * per_cpu() macro to access the array with specific 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 allows almost remove a lock-contention in locking * primitives which protect insert/remove operations. */ Are you find with it? -- Uladzislau Rezki
On Wed, Mar 29, 2023 at 05:01:11PM +0200, Uladzislau Rezki wrote: > Hello, Lorenzo! > > > > /* > > > - * 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. > > > */ > > > > OK so if I understand this correctly, you're overloading the per-CPU > > vmap_block_queue array to use as a simple hash based on the address and > > relying on the xa_lock() in xa_insert() to serialise in case of contention? > > > > I like the general heft of your comment but I feel this could be spelled > > out a little more clearly, something like:- > > > > In order to have fast access to any vmap_block object associated with a > > specific address, we use a hash. > > > > Rather than waste space on defining a new hash table we take advantage > > of the fact we already have a static per-cpu array vmap_block_queue. > > > > This is already used for per-CPU access to the block queue, however we > > overload this to _also_ act as a vmap_block hash. The hash function is > > addr_to_vbq() which hashes on vb->va->va_start. > > > > This then uses per_cpu() to lookup the _index_ rather than the > > _cpu_. Each vmap_block_queue contains an xarray of vmap blocks which are > > indexed on the same key as the hash (vb->va->va_start). > > > > xarray read acceses are protected by RCU lock and inserts are protected > > by a spin lock so there is no risk of a race here. > > > /* > * 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. Nit - it may be worth highlighting that when used as a hash it the 'cpu' is not in fact a cpu but rather a hash key. E.g. just add on the end of this something like:- When used as a hash table the 'cpu' passed to per_cpu is not actually a CPU but rather the hash key. > * > * A hash function is addr_to_vbq() which hashes any address to > * a specific index(in a hash) it belongs to. This then uses a > * per_cpu() macro to access the array with specific index. May need a tweak if you are happy with my review that we can simply have a helper that returns the xarray in which case we won't necessary have this function :) but depends of course on how the respin looks! > * > * 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 allows almost remove a lock-contention in locking > * primitives which protect insert/remove operations. This sentence is a little confusing, perhaps rephrase a little:- This technique almost always avoids lock contention on insert/remove, however the xarray spinlock protects against any contention that remains. > */ > Are you find with it? Other than the small nits above (sorry!) it seems fine! Thanks for updating, much appreciated :) > > -- > Uladzislau Rezki >
On Wed, Mar 29, 2023 at 05:23:04PM +0100, Lorenzo Stoakes wrote: > On Wed, Mar 29, 2023 at 05:01:11PM +0200, Uladzislau Rezki wrote: > > Hello, Lorenzo! > > > > > > /* > > > > - * 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. > > > > */ > > > > > > OK so if I understand this correctly, you're overloading the per-CPU > > > vmap_block_queue array to use as a simple hash based on the address and > > > relying on the xa_lock() in xa_insert() to serialise in case of contention? > > > > > > I like the general heft of your comment but I feel this could be spelled > > > out a little more clearly, something like:- > > > > > > In order to have fast access to any vmap_block object associated with a > > > specific address, we use a hash. > > > > > > Rather than waste space on defining a new hash table we take advantage > > > of the fact we already have a static per-cpu array vmap_block_queue. > > > > > > This is already used for per-CPU access to the block queue, however we > > > overload this to _also_ act as a vmap_block hash. The hash function is > > > addr_to_vbq() which hashes on vb->va->va_start. > > > > > > This then uses per_cpu() to lookup the _index_ rather than the > > > _cpu_. Each vmap_block_queue contains an xarray of vmap blocks which are > > > indexed on the same key as the hash (vb->va->va_start). > > > > > > xarray read acceses are protected by RCU lock and inserts are protected > > > by a spin lock so there is no risk of a race here. > > > > > /* > > * 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. > > Nit - it may be worth highlighting that when used as a hash it the 'cpu' is > not in fact a cpu but rather a hash key. > > E.g. just add on the end of this something like:- > > When used as a hash table the 'cpu' passed to per_cpu is not actually a CPU > but rather the hash key. > > > * > > * A hash function is addr_to_vbq() which hashes any address to > > * a specific index(in a hash) it belongs to. This then uses a > > * per_cpu() macro to access the array with specific index. > > May need a tweak if you are happy with my review that we can simply have a > helper that returns the xarray in which case we won't necessary have this > function :) but depends of course on how the respin looks! > > > * > > * 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 allows almost remove a lock-contention in locking > > * primitives which protect insert/remove operations. > > This sentence is a little confusing, perhaps rephrase a little:- > > This technique almost always avoids lock contention on insert/remove, > however the xarray spinlock protects against any contention that remains. > > > */ > > Are you find with it? > > Other than the small nits above (sorry!) it seems fine! Thanks for > updating, much appreciated :) > Good. Made the changes. I will upload a new vX patch. Everything that makes it more clear for readers is worth to do :) -- Uladzislau Rezki
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); if (err) { kfree(vb); free_vmap_area(va); @@ -2021,9 +2055,11 @@ 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 +2171,7 @@ 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 +2180,10 @@ 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 +3526,7 @@ 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 +3544,8 @@ 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 +4314,7 @@ 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. */
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. 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@gmail.com> --- mm/vmalloc.c | 85 +++++++++++++++++++++++++++++++++++++++------------- 1 file changed, 64 insertions(+), 21 deletions(-)