Message-ID: <ZP6IjeCpDIqHgV0y@MiWiFi-R3L-srv>
Date:   Mon, 11 Sep 2023 11:25:01 +0800
From:   Baoquan He <bhe@...hat.com>
To:     "Uladzislau Rezki (Sony)" <urezki@...il.com>
Cc:     linux-mm@...ck.org, Andrew Morton <akpm@...ux-foundation.org>,
        LKML <linux-kernel@...r.kernel.org>,
        Lorenzo Stoakes <lstoakes@...il.com>,
        Christoph Hellwig <hch@...radead.org>,
        Matthew Wilcox <willy@...radead.org>,
        "Liam R . Howlett" <Liam.Howlett@...cle.com>,
        Dave Chinner <david@...morbit.com>,
        "Paul E . McKenney" <paulmck@...nel.org>,
        Joel Fernandes <joel@...lfernandes.org>,
        Oleksiy Avramchenko <oleksiy.avramchenko@...y.com>
Subject: Re: [PATCH v2 6/9] mm: vmalloc: Offload free_vmap_area_lock lock

On 08/29/23 at 10:11am, Uladzislau Rezki (Sony) wrote:
> Concurrent access to a global vmap space is a bottle-neck.
> We can simulate a high contention by running a vmalloc test
> suite.
> 
> To address it, introduce an effective vmap node logic. Each
> node behaves as independent entity. When a node is accessed
> it serves a request directly(if possible) also it can fetch
> a new block from a global heap to its internals if no space
> or low capacity is left.
> 
> This technique reduces a pressure on the global vmap lock.
> 
> Signed-off-by: Uladzislau Rezki (Sony) <urezki@...il.com>
> ---
>  mm/vmalloc.c | 316 +++++++++++++++++++++++++++++++++++++++++++++------
>  1 file changed, 279 insertions(+), 37 deletions(-)
> 
> diff --git a/mm/vmalloc.c b/mm/vmalloc.c
> index 5a8a9c1370b6..4fd4915c532d 100644
> --- a/mm/vmalloc.c
> +++ b/mm/vmalloc.c
> @@ -779,6 +779,7 @@ struct rb_list {
>  
>  struct vmap_node {
>  	/* Bookkeeping data of this node. */
> +	struct rb_list free;
>  	struct rb_list busy;
>  	struct rb_list lazy;
>  
> @@ -786,6 +787,13 @@ struct vmap_node {
>  	 * Ready-to-free areas.
>  	 */
>  	struct list_head purge_list;
> +	struct work_struct purge_work;
> +	unsigned long nr_purged;
> +
> +	/*
> +	 * Control that only one user can pre-fetch this node.
> +	 */
> +	atomic_t fill_in_progress;
>  };
>  
>  static struct vmap_node *nodes, snode;
> @@ -804,6 +812,32 @@ addr_to_node(unsigned long addr)
>  	return &nodes[addr_to_node_id(addr)];
>  }
>  
> +static inline struct vmap_node *
> +id_to_node(int id)
> +{
> +	return &nodes[id % nr_nodes];
> +}
> +
> +static inline int
> +this_node_id(void)
> +{
> +	return raw_smp_processor_id() % nr_nodes;
> +}
> +
> +static inline unsigned long
> +encode_vn_id(int node_id)
> +{
> +	/* Can store U8_MAX [0:254] nodes. */
> +	return (node_id + 1) << BITS_PER_BYTE;
> +}
> +
> +static inline int
> +decode_vn_id(unsigned long val)
> +{
> +	/* Can store U8_MAX [0:254] nodes. */
> +	return (val >> BITS_PER_BYTE) - 1;
> +}

This patch looks good to me. However, should we split out the encoding
vn_id into va->flags optimization into another patch? It looks like an
independent optimization which can be described better with specific
log. At least, in the pdf file pasted or patch log, it's not obvious
that:
1) node's free tree could contains any address range;
2) nodes' busy tree only contains address range belonging to this node;
   - could contain crossing node range, corner case.
3) nodes' purge tree could contain any address range;
   - decided by encoded vn_id in va->flags.
   - decided by address via addr_to_node(va->va_start).

Personal opinion, feel it will make reviewing easier.

> +
>  static __always_inline unsigned long
>  va_size(struct vmap_area *va)
>  {
> @@ -1586,6 +1620,7 @@ __alloc_vmap_area(struct rb_root *root, struct list_head *head,
>  static void free_vmap_area(struct vmap_area *va)
>  {
>  	struct vmap_node *vn = addr_to_node(va->va_start);
> +	int vn_id = decode_vn_id(va->flags);
>  
>  	/*
>  	 * Remove from the busy tree/list.
> @@ -1594,12 +1629,19 @@ static void free_vmap_area(struct vmap_area *va)
>  	unlink_va(va, &vn->busy.root);
>  	spin_unlock(&vn->busy.lock);
>  
> -	/*
> -	 * Insert/Merge it back to the free tree/list.
> -	 */
> -	spin_lock(&free_vmap_area_lock);
> -	merge_or_add_vmap_area_augment(va, &free_vmap_area_root, &free_vmap_area_list);
> -	spin_unlock(&free_vmap_area_lock);
> +	if (vn_id >= 0) {
> +		vn = id_to_node(vn_id);
> +
> +		/* Belongs to this node. */
> +		spin_lock(&vn->free.lock);
> +		merge_or_add_vmap_area_augment(va, &vn->free.root, &vn->free.head);
> +		spin_unlock(&vn->free.lock);
> +	} else {
> +		/* Goes to global. */
> +		spin_lock(&free_vmap_area_lock);
> +		merge_or_add_vmap_area_augment(va, &free_vmap_area_root, &free_vmap_area_list);
> +		spin_unlock(&free_vmap_area_lock);
> +	}
>  }
>  
>  static inline void
> @@ -1625,6 +1667,134 @@ preload_this_cpu_lock(spinlock_t *lock, gfp_t gfp_mask, int node)
>  		kmem_cache_free(vmap_area_cachep, va);
>  }
>  
> +static unsigned long
> +node_alloc_fill(struct vmap_node *vn,
> +		unsigned long size, unsigned long align,
> +		gfp_t gfp_mask, int node)
> +{
> +	struct vmap_area *va;
> +	unsigned long addr;
> +
> +	va = kmem_cache_alloc_node(vmap_area_cachep, gfp_mask, node);
> +	if (unlikely(!va))
> +		return VMALLOC_END;
> +
> +	/*
> +	 * Please note, an allocated block is not aligned to its size.
> +	 * Therefore it can span several zones what means addr_to_node()
> +	 * can point to two different nodes:
> +	 *      <----->
> +	 * -|-----|-----|-----|-----|-
> +	 *     1     2     0     1
> +	 *
> +	 * an alignment would just increase fragmentation thus more heap
> +	 * consumption what we would like to avoid.
> +	 */
> +	spin_lock(&free_vmap_area_lock);
> +	addr = __alloc_vmap_area(&free_vmap_area_root, &free_vmap_area_list,
> +		node_size, 1, VMALLOC_START, VMALLOC_END);
> +	spin_unlock(&free_vmap_area_lock);
> +
> +	if (addr == VMALLOC_END) {
> +		kmem_cache_free(vmap_area_cachep, va);
> +		return VMALLOC_END;
> +	}
> +
> +	/*
> +	 * Statement and condition of the problem:
> +	 *
> +	 * a) where to free allocated areas from a node:
> +	 *   - directly to a global heap;
> +	 *   - to a node that we got a VA from;
> +	 *     - what is a condition to return allocated areas
> +	 *       to a global heap then;
> +	 * b) how to properly handle left small free fragments
> +	 *    of a node in order to mitigate a fragmentation.
> +	 *
> +	 * How to address described points:
> +	 * When a new block is allocated(from a global heap) we shrink
> +	 * it deliberately by one page from both sides and place it to
> +	 * this node to serve a request.
> +	 *
> +	 * Why we shrink. We would like to distinguish VAs which were
> +	 * obtained from a node and a global heap. This is for a free
> +	 * path. A va->flags contains a node-id it belongs to. No VAs
> +	 * merging is possible between each other unless they are part
> +	 * of same block.
> +	 *
> +	 * A free-path in its turn can detect a correct node where a
> +	 * VA has to be returned. Thus as a block is freed entirely,
> +	 * its size becomes(merging): node_size - (2 * PAGE_SIZE) it
> +	 * recovers its edges, thus is released to a global heap for
> +	 * reuse elsewhere. In partly freed case, VAs go back to the
> +	 * node not bothering a global vmap space.
> +	 *
> +	 *        1               2              3
> +	 * |<------------>|<------------>|<------------>|
> +	 * |..<-------->..|..<-------->..|..<-------->..|
> +	 */
> +	va->va_start = addr + PAGE_SIZE;
> +	va->va_end = (addr + node_size) - PAGE_SIZE;
> +
> +	spin_lock(&vn->free.lock);
> +	/* Never merges. See explanation above. */
> +	insert_vmap_area_augment(va, NULL, &vn->free.root, &vn->free.head);
> +	addr = va_alloc(va, &vn->free.root, &vn->free.head,
> +		size, align, VMALLOC_START, VMALLOC_END);
> +	spin_unlock(&vn->free.lock);
> +
> +	return addr;
> +}
> +
> +static unsigned long
> +node_alloc(int vn_id, unsigned long size, unsigned long align,
> +		unsigned long vstart, unsigned long vend,
> +		gfp_t gfp_mask, int node)
> +{
> +	struct vmap_node *vn = id_to_node(vn_id);
> +	unsigned long extra = align > PAGE_SIZE ? align : 0;
> +	bool do_alloc_fill = false;
> +	unsigned long addr;
> +
> +	/*
> +	 * Fallback to a global heap if not vmalloc.
> +	 */
> +	if (vstart != VMALLOC_START || vend != VMALLOC_END)
> +		return vend;
> +
> +	/*
> +	 * A maximum allocation limit is 1/4 of capacity. This
> +	 * is done in order to prevent a fast depleting of zone
> +	 * by few requests.
> +	 */
> +	if (size + extra > (node_size >> 2))
> +		return vend;
> +
> +	spin_lock(&vn->free.lock);
> +	addr = __alloc_vmap_area(&vn->free.root, &vn->free.head,
> +		size, align, vstart, vend);
> +
> +	if (addr == vend) {
> +		/*
> +		 * Set the fetch flag under the critical section.
> +		 * This guarantees that only one user is eligible
> +		 * to perform a pre-fetch. A reset operation can
> +		 * be concurrent.
> +		 */
> +		if (!atomic_xchg(&vn->fill_in_progress, 1))
> +			do_alloc_fill = true;
> +	}
> +	spin_unlock(&vn->free.lock);
> +
> +	/* Only if fails a previous attempt. */
> +	if (do_alloc_fill) {
> +		addr = node_alloc_fill(vn, size, align, gfp_mask, node);
> +		atomic_set(&vn->fill_in_progress, 0);
> +	}
> +
> +	return addr;
> +}
> +
>  /*
>   * Allocate a region of KVA of the specified size and alignment, within the
>   * vstart and vend.
> @@ -1640,7 +1810,7 @@ static struct vmap_area *alloc_vmap_area(unsigned long size,
>  	unsigned long freed;
>  	unsigned long addr;
>  	int purged = 0;
> -	int ret;
> +	int ret, vn_id;
>  
>  	if (unlikely(!size || offset_in_page(size) || !is_power_of_2(align)))
>  		return ERR_PTR(-EINVAL);
> @@ -1661,11 +1831,17 @@ static struct vmap_area *alloc_vmap_area(unsigned long size,
>  	 */
>  	kmemleak_scan_area(&va->rb_node, SIZE_MAX, gfp_mask);
>  
> +	vn_id = this_node_id();
> +	addr = node_alloc(vn_id, size, align, vstart, vend, gfp_mask, node);
> +	va->flags = (addr != vend) ? encode_vn_id(vn_id) : 0;
> +
>  retry:
> -	preload_this_cpu_lock(&free_vmap_area_lock, gfp_mask, node);
> -	addr = __alloc_vmap_area(&free_vmap_area_root, &free_vmap_area_list,
> -		size, align, vstart, vend);
> -	spin_unlock(&free_vmap_area_lock);
> +	if (addr == vend) {
> +		preload_this_cpu_lock(&free_vmap_area_lock, gfp_mask, node);
> +		addr = __alloc_vmap_area(&free_vmap_area_root, &free_vmap_area_list,
> +			size, align, vstart, vend);
> +		spin_unlock(&free_vmap_area_lock);
> +	}
>  
>  	trace_alloc_vmap_area(addr, size, align, vstart, vend, addr == vend);
>  
> @@ -1679,7 +1855,7 @@ static struct vmap_area *alloc_vmap_area(unsigned long size,
>  	va->va_start = addr;
>  	va->va_end = addr + size;
>  	va->vm = NULL;
> -	va->flags = va_flags;
> +	va->flags |= va_flags;
>  
>  	vn = addr_to_node(va->va_start);
>  
> @@ -1772,31 +1948,58 @@ static DEFINE_MUTEX(vmap_purge_lock);
>  static void purge_fragmented_blocks_allcpus(void);
>  static cpumask_t purge_nodes;
>  
> -/*
> - * Purges all lazily-freed vmap areas.
> - */
> -static unsigned long
> -purge_vmap_node(struct vmap_node *vn)
> +static void
> +reclaim_list_global(struct list_head *head)
> +{
> +	struct vmap_area *va, *n;
> +
> +	if (list_empty(head))
> +		return;
> +
> +	spin_lock(&free_vmap_area_lock);
> +	list_for_each_entry_safe(va, n, head, list)
> +		merge_or_add_vmap_area_augment(va,
> +			&free_vmap_area_root, &free_vmap_area_list);
> +	spin_unlock(&free_vmap_area_lock);
> +}
> +
> +static void purge_vmap_node(struct work_struct *work)
>  {
> -	unsigned long num_purged_areas = 0;
> +	struct vmap_node *vn = container_of(work,
> +		struct vmap_node, purge_work);
>  	struct vmap_area *va, *n_va;
> +	LIST_HEAD(global);
> +
> +	vn->nr_purged = 0;
>  
>  	if (list_empty(&vn->purge_list))
> -		return 0;
> +		return;
>  
> -	spin_lock(&free_vmap_area_lock);
> +	spin_lock(&vn->free.lock);
>  	list_for_each_entry_safe(va, n_va, &vn->purge_list, list) {
>  		unsigned long nr = (va->va_end - va->va_start) >> PAGE_SHIFT;
>  		unsigned long orig_start = va->va_start;
>  		unsigned long orig_end = va->va_end;
> +		int vn_id = decode_vn_id(va->flags);
>  
> -		/*
> -		 * Finally insert or merge lazily-freed area. It is
> -		 * detached and there is no need to "unlink" it from
> -		 * anything.
> -		 */
> -		va = merge_or_add_vmap_area_augment(va, &free_vmap_area_root,
> -				&free_vmap_area_list);
> +		list_del_init(&va->list);
> +
> +		if (vn_id >= 0) {
> +			if (va_size(va) != node_size - (2 * PAGE_SIZE))
> +				va = merge_or_add_vmap_area_augment(va, &vn->free.root, &vn->free.head);
> +
> +			if (va_size(va) == node_size - (2 * PAGE_SIZE)) {
> +				if (!list_empty(&va->list))
> +					unlink_va_augment(va, &vn->free.root);
> +
> +				/* Restore the block size. */
> +				va->va_start -= PAGE_SIZE;
> +				va->va_end += PAGE_SIZE;
> +				list_add(&va->list, &global);
> +			}
> +		} else {
> +			list_add(&va->list, &global);
> +		}
>  
>  		if (!va)
>  			continue;
> @@ -1806,11 +2009,10 @@ purge_vmap_node(struct vmap_node *vn)
>  					      va->va_start, va->va_end);
>  
>  		atomic_long_sub(nr, &vmap_lazy_nr);
> -		num_purged_areas++;
> +		vn->nr_purged++;
>  	}
> -	spin_unlock(&free_vmap_area_lock);
> -
> -	return num_purged_areas;
> +	spin_unlock(&vn->free.lock);
> +	reclaim_list_global(&global);
>  }
>  
>  /*
> @@ -1818,11 +2020,17 @@ purge_vmap_node(struct vmap_node *vn)
>   */
>  static bool __purge_vmap_area_lazy(unsigned long start, unsigned long end)
>  {
> -	unsigned long num_purged_areas = 0;
> +	unsigned long nr_purged_areas = 0;
> +	unsigned int nr_purge_helpers;
> +	unsigned int nr_purge_nodes;
>  	struct vmap_node *vn;
>  	int i;
>  
>  	lockdep_assert_held(&vmap_purge_lock);
> +
> +	/*
> +	 * Use cpumask to mark which node has to be processed.
> +	 */
>  	purge_nodes = CPU_MASK_NONE;
>  
>  	for (i = 0; i < nr_nodes; i++) {
> @@ -1847,17 +2055,45 @@ static bool __purge_vmap_area_lazy(unsigned long start, unsigned long end)
>  		cpumask_set_cpu(i, &purge_nodes);
>  	}
>  
> -	if (cpumask_weight(&purge_nodes) > 0) {
> +	nr_purge_nodes = cpumask_weight(&purge_nodes);
> +	if (nr_purge_nodes > 0) {
>  		flush_tlb_kernel_range(start, end);
>  
> +		/* One extra worker is per a lazy_max_pages() full set minus one. */
> +		nr_purge_helpers = atomic_long_read(&vmap_lazy_nr) / lazy_max_pages();
> +		nr_purge_helpers = clamp(nr_purge_helpers, 1U, nr_purge_nodes) - 1;
> +
> +		for_each_cpu(i, &purge_nodes) {
> +			vn = &nodes[i];
> +
> +			if (nr_purge_helpers > 0) {
> +				INIT_WORK(&vn->purge_work, purge_vmap_node);
> +
> +				if (cpumask_test_cpu(i, cpu_online_mask))
> +					schedule_work_on(i, &vn->purge_work);
> +				else
> +					schedule_work(&vn->purge_work);
> +
> +				nr_purge_helpers--;
> +			} else {
> +				vn->purge_work.func = NULL;
> +				purge_vmap_node(&vn->purge_work);
> +				nr_purged_areas += vn->nr_purged;
> +			}
> +		}
> +
>  		for_each_cpu(i, &purge_nodes) {
>  			vn = &nodes[i];
> -			num_purged_areas += purge_vmap_node(vn);
> +
> +			if (vn->purge_work.func) {
> +				flush_work(&vn->purge_work);
> +				nr_purged_areas += vn->nr_purged;
> +			}
>  		}
>  	}
>  
> -	trace_purge_vmap_area_lazy(start, end, num_purged_areas);
> -	return num_purged_areas > 0;
> +	trace_purge_vmap_area_lazy(start, end, nr_purged_areas);
> +	return nr_purged_areas > 0;
>  }
>  
>  /*
> @@ -1886,9 +2122,11 @@ static void drain_vmap_area_work(struct work_struct *work)
>   */
>  static void free_vmap_area_noflush(struct vmap_area *va)
>  {
> -	struct vmap_node *vn = addr_to_node(va->va_start);
>  	unsigned long nr_lazy_max = lazy_max_pages();
>  	unsigned long va_start = va->va_start;
> +	int vn_id = decode_vn_id(va->flags);
> +	struct vmap_node *vn = vn_id >= 0 ? id_to_node(vn_id):
> +		addr_to_node(va->va_start);;
>  	unsigned long nr_lazy;
>  
>  	if (WARN_ON_ONCE(!list_empty(&va->list)))
> @@ -4574,6 +4812,10 @@ static void vmap_init_nodes(void)
>  		vn->lazy.root = RB_ROOT;
>  		INIT_LIST_HEAD(&vn->lazy.head);
>  		spin_lock_init(&vn->lazy.lock);
> +
> +		vn->free.root = RB_ROOT;
> +		INIT_LIST_HEAD(&vn->free.head);
> +		spin_lock_init(&vn->free.lock);
>  	}
>  }
>  
> -- 
> 2.30.2
> 

Powered by blists - more mailing lists