Date:   Wed, 18 Oct 2023 12:08:12 +0800
From:   "wuqiang.matt" <wuqiang.matt@...edance.com>
To:     Chengming Zhou <chengming.zhou@...ux.dev>,
        linux-trace-kernel@...r.kernel.org, mhiramat@...nel.org,
        davem@...emloft.net, anil.s.keshavamurthy@...el.com,
        naveen.n.rao@...ux.ibm.com, rostedt@...dmis.org,
        peterz@...radead.org, akpm@...ux-foundation.org,
        sander@...nheule.net, ebiggers@...gle.com,
        dan.j.williams@...el.com, jpoimboe@...nel.org
Cc:     linux-kernel@...r.kernel.org, lkp@...el.com, mattwu@....com
Subject: Re: [PATCH v11 1/5] lib: objpool added: ring-array based lockless
 MPMC

On 2023/10/18 10:18, Chengming Zhou wrote:
> On 2023/10/17 21:56, wuqiang.matt wrote:
>> objpool is a scalable implementation of high performance queue for
>> object allocation and reclamation, such as kretprobe instances.
>>
>> With leveraging percpu ring-array to mitigate hot spots of memory
>> contention, it delivers near-linear scalability for high parallel
>> scenarios. The objpool is best suited for the following cases:
>> 1) Memory allocation or reclamation are prohibited or too expensive
>> 2) Consumers are of different priorities, such as irqs and threads
>>
>> Limitations:
>> 1) Maximum objects (capacity) is fixed after objpool creation
>> 2) All pre-allocated objects are managed in percpu ring array,
>>     which consumes more memory than linked lists
>>
> I'm curious why not just extend the existing lockless freelist to
> percpu lockless freelists? And the percpu freelist is more flexible
> to use than this percpu ring-array? The latter has to be fixed size
> when creation.

I did that in first 2 versions, and abandoned it from the 3rd version.
The core reason is there are data races in freelist node:

After pop() from freelist, the freelist_node zone of the object could
be still in busy spinning by other nodes, so even the owner of this
object couldn't know when the races would go. freelist_zone is defined
as a union in the use cases (kretprobe), which brings potential issues.
If the object owner touches freelist_node:refs and then tries to push
to freelist to reclaim the object, freelist_add might just quit if
atomic_fetch_add_release(REFS_ON_FREELIST, &node->refs) returns true.

Keeping freelist_node as private could be fine, which is imposing an
extra rule to the users. Current ring array likes something moving
"freelist_node" to the ring array, which minimizes memory footprints.

Flexibility is not a strong requirement since all the use cases have
objects pre-allocated.

> Thanks.

Thank you.

> 
>> Signed-off-by: wuqiang.matt <wuqiang.matt@...edance.com>
>> ---
>>   include/linux/objpool.h | 176 +++++++++++++++++++++++++
>>   lib/Makefile            |   2 +-
>>   lib/objpool.c           | 286 ++++++++++++++++++++++++++++++++++++++++
>>   3 files changed, 463 insertions(+), 1 deletion(-)
>>   create mode 100644 include/linux/objpool.h
>>   create mode 100644 lib/objpool.c
>>
>> diff --git a/include/linux/objpool.h b/include/linux/objpool.h
>> new file mode 100644
>> index 000000000000..4df18405420a
>> --- /dev/null
>> +++ b/include/linux/objpool.h
>> @@ -0,0 +1,181 @@
>> +/* SPDX-License-Identifier: GPL-2.0 */
>> +
>> +#ifndef _LINUX_OBJPOOL_H
>> +#define _LINUX_OBJPOOL_H
>> +
>> +#include <linux/types.h>
>> +#include <linux/refcount.h>
>> +
>> +/*
>> + * objpool: ring-array based lockless MPMC queue
>> + *
>> + * Copyright: wuqiang.matt@...edance.com,mhiramat@...nel.org
>> + *
>> + * objpool is a scalable implementation of high performance queue for
>> + * object allocation and reclamation, such as kretprobe instances.
>> + *
>> + * With leveraging percpu ring-array to mitigate hot spots of memory
>> + * contention, it delivers near-linear scalability for high parallel
>> + * scenarios. The objpool is best suited for the following cases:
>> + * 1) Memory allocation or reclamation are prohibited or too expensive
>> + * 2) Consumers are of different priorities, such as irqs and threads
>> + *
>> + * Limitations:
>> + * 1) Maximum objects (capacity) is fixed after objpool creation
>> + * 2) All pre-allocated objects are managed in percpu ring array,
>> + *    which consumes more memory than linked lists
>> + */
>> +
>> +/**
>> + * struct objpool_slot - percpu ring array of objpool
>> + * @head: head sequence of the local ring array (to retrieve at)
>> + * @tail: tail sequence of the local ring array (to append at)
>> + * @last: the last sequence number marked as ready for retrieve
>> + * @mask: bits mask for modulo capacity to compute array indexes
>> + * @entries: object entries on this slot
>> + *
>> + * Represents a cpu-local array-based ring buffer, its size is specialized
>> + * during initialization of object pool. The percpu objpool node is to be
>> + * allocated from local memory for NUMA system, and to be kept compact in
>> + * continuous memory: CPU assigned number of objects are stored just after
>> + * the body of objpool_node.
>> + *
>> + * Real size of the ring array is far too smaller than the value range of
>> + * head and tail, typed as uint32_t: [0, 2^32), so only lower bits (mask)
>> + * of head and tail are used as the actual position in the ring array. In
>> + * general the ring array is acting like a small sliding window, which is
>> + * always moving forward in the loop of [0, 2^32).
>> + */
>> +struct objpool_slot {
>> +	uint32_t            head;
>> +	uint32_t            tail;
>> +	uint32_t            last;
>> +	uint32_t            mask;
>> +	void               *entries[];
>> +} __packed;
>> +
>> +struct objpool_head;
>> +
>> +/*
>> + * caller-specified callback for object initial setup, it's only called
>> + * once for each object (just after the memory allocation of the object)
>> + */
>> +typedef int (*objpool_init_obj_cb)(void *obj, void *context);
>> +
>> +/* caller-specified cleanup callback for objpool destruction */
>> +typedef int (*objpool_fini_cb)(struct objpool_head *head, void *context);
>> +
>> +/**
>> + * struct objpool_head - object pooling metadata
>> + * @obj_size:   object size, aligned to sizeof(void *)
>> + * @nr_objs:    total objs (to be pre-allocated with objpool)
>> + * @nr_cpus:    local copy of nr_cpu_ids
>> + * @capacity:   max objs can be managed by one objpool_slot
>> + * @gfp:        gfp flags for kmalloc & vmalloc
>> + * @ref:        refcount of objpool
>> + * @flags:      flags for objpool management
>> + * @cpu_slots:  pointer to the array of objpool_slot
>> + * @release:    resource cleanup callback
>> + * @context:    caller-provided context
>> + */
>> +struct objpool_head {
>> +	int                     obj_size;
>> +	int                     nr_objs;
>> +	int                     nr_cpus;
>> +	int                     capacity;
>> +	gfp_t                   gfp;
>> +	refcount_t              ref;
>> +	unsigned long           flags;
>> +	struct objpool_slot   **cpu_slots;
>> +	objpool_fini_cb         release;
>> +	void                   *context;
>> +};
>> +
>> +#define OBJPOOL_NR_OBJECT_MAX	(1UL << 24) /* maximum numbers of total objects */
>> +#define OBJPOOL_OBJECT_SIZE_MAX	(1UL << 16) /* maximum size of an object */
>> +
>> +/**
>> + * objpool_init() - initialize objpool and pre-allocated objects
>> + * @pool:    the object pool to be initialized, declared by caller
>> + * @nr_objs: total objects to be pre-allocated by this object pool
>> + * @object_size: size of an object (should be > 0)
>> + * @gfp:     flags for memory allocation (via kmalloc or vmalloc)
>> + * @context: user context for object initialization callback
>> + * @objinit: object initialization callback for extra setup
>> + * @release: cleanup callback for extra cleanup task
>> + *
>> + * return value: 0 for success, otherwise error code
>> + *
>> + * All pre-allocated objects are to be zeroed after memory allocation.
>> + * Caller could do extra initialization in objinit callback. objinit()
>> + * will be called just after slot allocation and called only once for
>> + * each object. After that the objpool won't touch any content of the
>> + * objects. It's caller's duty to perform reinitialization after each
>> + * pop (object allocation) or do clearance before each push (object
>> + * reclamation).
>> + */
>> +int objpool_init(struct objpool_head *pool, int nr_objs, int object_size,
>> +		 gfp_t gfp, void *context, objpool_init_obj_cb objinit,
>> +		 objpool_fini_cb release);
>> +
>> +/**
>> + * objpool_pop() - allocate an object from objpool
>> + * @pool: object pool
>> + *
>> + * return value: object ptr or NULL if failed
>> + */
>> +void *objpool_pop(struct objpool_head *pool);
>> +
>> +/**
>> + * objpool_push() - reclaim the object and return back to objpool
>> + * @obj:  object ptr to be pushed to objpool
>> + * @pool: object pool
>> + *
>> + * return: 0 or error code (it fails only when user tries to push
>> + * the same object multiple times or wrong "objects" into objpool)
>> + */
>> +int objpool_push(void *obj, struct objpool_head *pool);
>> +
>> +/**
>> + * objpool_drop() - discard the object and deref objpool
>> + * @obj:  object ptr to be discarded
>> + * @pool: object pool
>> + *
>> + * return: 0 if objpool was released; -EAGAIN if there are still
>> + *         outstanding objects
>> + *
>> + * objpool_drop is normally for the release of outstanding objects
>> + * after objpool cleanup (objpool_fini). Thinking of this example:
>> + * kretprobe is unregistered and objpool_fini() is called to release
>> + * all remained objects, but there are still objects being used by
>> + * unfinished kretprobes (like blockable function: sys_accept). So
>> + * only when the last outstanding object is dropped could the whole
>> + * objpool be released along with the call of objpool_drop()
>> + */
>> +int objpool_drop(void *obj, struct objpool_head *pool);
>> +
>> +/**
>> + * objpool_free() - release objpool forcely (all objects to be freed)
>> + * @pool: object pool to be released
>> + */
>> +void objpool_free(struct objpool_head *pool);
>> +
>> +/**
>> + * objpool_fini() - deref object pool (also releasing unused objects)
>> + * @pool: object pool to be dereferenced
>> + *
>> + * objpool_fini() will try to release all remained free objects and
>> + * then drop an extra reference of the objpool. If all objects are
>> + * already returned to objpool (so called synchronous use cases),
>> + * the objpool itself will be freed together. But if there are still
>> + * outstanding objects (so called asynchronous use cases, such like
>> + * blockable kretprobe), the objpool won't be released until all
>> + * the outstanding objects are dropped, but the caller must assure
>> + * there are no concurrent objpool_push() on the fly. Normally RCU
>> + * is being required to make sure all ongoing objpool_push() must
>> + * be finished before calling objpool_fini(), so does test_objpool,
>> + * kretprobe or rethook
>> + */
>> +void objpool_fini(struct objpool_head *pool);
>> +
>> +#endif /* _LINUX_OBJPOOL_H */
>> diff --git a/lib/Makefile b/lib/Makefile
>> index 1ffae65bb7ee..7a84c922d9ff 100644
>> --- a/lib/Makefile
>> +++ b/lib/Makefile
>> @@ -34,7 +34,7 @@ lib-y := ctype.o string.o vsprintf.o cmdline.o \
>>   	 is_single_threaded.o plist.o decompress.o kobject_uevent.o \
>>   	 earlycpio.o seq_buf.o siphash.o dec_and_lock.o \
>>   	 nmi_backtrace.o win_minmax.o memcat_p.o \
>> -	 buildid.o
>> +	 buildid.o objpool.o
>>   
>>   lib-$(CONFIG_PRINTK) += dump_stack.o
>>   lib-$(CONFIG_SMP) += cpumask.o
>> diff --git a/lib/objpool.c b/lib/objpool.c
>> new file mode 100644
>> index 000000000000..37a71e063f18
>> --- /dev/null
>> +++ b/lib/objpool.c
>> @@ -0,0 +1,280 @@
>> +// SPDX-License-Identifier: GPL-2.0
>> +
>> +#include <linux/objpool.h>
>> +#include <linux/slab.h>
>> +#include <linux/vmalloc.h>
>> +#include <linux/atomic.h>
>> +#include <linux/irqflags.h>
>> +#include <linux/cpumask.h>
>> +#include <linux/log2.h>
>> +
>> +/*
>> + * objpool: ring-array based lockless MPMC/FIFO queues
>> + *
>> + * Copyright: wuqiang.matt@...edance.com,mhiramat@...nel.org
>> + */
>> +
>> +/* initialize percpu objpool_slot */
>> +static int
>> +objpool_init_percpu_slot(struct objpool_head *pool,
>> +			 struct objpool_slot *slot,
>> +			 int nodes, void *context,
>> +			 objpool_init_obj_cb objinit)
>> +{
>> +	void *obj = (void *)&slot->entries[pool->capacity];
>> +	int i;
>> +
>> +	/* initialize elements of percpu objpool_slot */
>> +	slot->mask = pool->capacity - 1;
>> +
>> +	for (i = 0; i < nodes; i++) {
>> +		if (objinit) {
>> +			int rc = objinit(obj, context);
>> +			if (rc)
>> +				return rc;
>> +		}
>> +		slot->entries[slot->tail & slot->mask] = obj;
>> +		obj = obj + pool->obj_size;
>> +		slot->tail++;
>> +		slot->last = slot->tail;
>> +		pool->nr_objs++;
>> +	}
>> +
>> +	return 0;
>> +}
>> +
>> +/* allocate and initialize percpu slots */
>> +static int
>> +objpool_init_percpu_slots(struct objpool_head *pool, int nr_objs,
>> +			  void *context, objpool_init_obj_cb objinit)
>> +{
>> +	int i, cpu_count = 0;
>> +
>> +	for (i = 0; i < pool->nr_cpus; i++) {
>> +
>> +		struct objpool_slot *slot;
>> +		int nodes, size, rc;
>> +
>> +		/* skip the cpu node which could never be present */
>> +		if (!cpu_possible(i))
>> +			continue;
>> +
>> +		/* compute how many objects to be allocated with this slot */
>> +		nodes = nr_objs / num_possible_cpus();
>> +		if (cpu_count < (nr_objs % num_possible_cpus()))
>> +			nodes++;
>> +		cpu_count++;
>> +
>> +		size = struct_size(slot, entries, pool->capacity) +
>> +			pool->obj_size * nodes;
>> +
>> +		/*
>> +		 * here we allocate percpu-slot & objs together in a single
>> +		 * allocation to make it more compact, taking advantage of
>> +		 * warm caches and TLB hits. in default vmalloc is used to
>> +		 * reduce the pressure of kernel slab system. as we know,
>> +		 * mimimal size of vmalloc is one page since vmalloc would
>> +		 * always align the requested size to page size
>> +		 */
>> +		if (pool->gfp & GFP_ATOMIC)
>> +			slot = kmalloc_node(size, pool->gfp, cpu_to_node(i));
>> +		else
>> +			slot = __vmalloc_node(size, sizeof(void *), pool->gfp,
>> +				cpu_to_node(i), __builtin_return_address(0));
>> +		if (!slot)
>> +			return -ENOMEM;
>> +		memset(slot, 0, size);
>> +		pool->cpu_slots[i] = slot;
>> +
>> +		/* initialize the objpool_slot of cpu node i */
>> +		rc = objpool_init_percpu_slot(pool, slot, nodes, context, objinit);
>> +		if (rc)
>> +			return rc;
>> +	}
>> +
>> +	return 0;
>> +}
>> +
>> +/* cleanup all percpu slots of the object pool */
>> +static void objpool_fini_percpu_slots(struct objpool_head *pool)
>> +{
>> +	int i;
>> +
>> +	if (!pool->cpu_slots)
>> +		return;
>> +
>> +	for (i = 0; i < pool->nr_cpus; i++)
>> +		kvfree(pool->cpu_slots[i]);
>> +	kfree(pool->cpu_slots);
>> +}
>> +
>> +/* initialize object pool and pre-allocate objects */
>> +int objpool_init(struct objpool_head *pool, int nr_objs, int object_size,
>> +		gfp_t gfp, void *context, objpool_init_obj_cb objinit,
>> +		objpool_fini_cb release)
>> +{
>> +	int rc, capacity, slot_size;
>> +
>> +	/* check input parameters */
>> +	if (nr_objs <= 0 || nr_objs > OBJPOOL_NR_OBJECT_MAX ||
>> +	    object_size <= 0 || object_size > OBJPOOL_OBJECT_SIZE_MAX)
>> +		return -EINVAL;
>> +
>> +	/* align up to unsigned long size */
>> +	object_size = ALIGN(object_size, sizeof(long));
>> +
>> +	/* calculate capacity of percpu objpool_slot */
>> +	capacity = roundup_pow_of_two(nr_objs);
>> +	if (!capacity)
>> +		return -EINVAL;
>> +
>> +	/* initialize objpool pool */
>> +	memset(pool, 0, sizeof(struct objpool_head));
>> +	pool->nr_cpus = nr_cpu_ids;
>> +	pool->obj_size = object_size;
>> +	pool->capacity = capacity;
>> +	pool->gfp = gfp & ~__GFP_ZERO;
>> +	pool->context = context;
>> +	pool->release = release;
>> +	slot_size = pool->nr_cpus * sizeof(struct objpool_slot);
>> +	pool->cpu_slots = kzalloc(slot_size, pool->gfp);
>> +	if (!pool->cpu_slots)
>> +		return -ENOMEM;
>> +
>> +	/* initialize per-cpu slots */
>> +	rc = objpool_init_percpu_slots(pool, nr_objs, context, objinit);
>> +	if (rc)
>> +		objpool_fini_percpu_slots(pool);
>> +	else
>> +		refcount_set(&pool->ref, pool->nr_objs + 1);
>> +
>> +	return rc;
>> +}
>> +EXPORT_SYMBOL_GPL(objpool_init);
>> +
>> +/* adding object to slot, abort if the slot was already full */
>> +static inline int
>> +objpool_try_add_slot(void *obj, struct objpool_head *pool, int cpu)
>> +{
>> +	struct objpool_slot *slot = pool->cpu_slots[cpu];
>> +	uint32_t head, tail;
>> +
>> +	/* loading tail and head as a local snapshot, tail first */
>> +	tail = READ_ONCE(slot->tail);
>> +
>> +	do {
>> +		head = READ_ONCE(slot->head);
>> +		/* fault caught: something must be wrong */
>> +		WARN_ON_ONCE(tail - head > pool->nr_objs);
>> +	} while (!try_cmpxchg_acquire(&slot->tail, &tail, tail + 1));
>> +
>> +	/* now the tail position is reserved for the given obj */
>> +	WRITE_ONCE(slot->entries[tail & slot->mask], obj);
>> +	/* update sequence to make this obj available for pop() */
>> +	smp_store_release(&slot->last, tail + 1);
>> +
>> +	return 0;
>> +}
>> +
>> +/* reclaim an object to object pool */
>> +int objpool_push(void *obj, struct objpool_head *pool)
>> +{
>> +	unsigned long flags;
>> +	int rc;
>> +
>> +	/* disable local irq to avoid preemption & interruption */
>> +	raw_local_irq_save(flags);
>> +	rc = objpool_try_add_slot(obj, pool, raw_smp_processor_id());
>> +	raw_local_irq_restore(flags);
>> +
>> +	return rc;
>> +}
>> +EXPORT_SYMBOL_GPL(objpool_push);
>> +
>> +/* try to retrieve object from slot */
>> +static inline void *objpool_try_get_slot(struct objpool_head *pool, int cpu)
>> +{
>> +	struct objpool_slot *slot = pool->cpu_slots[cpu];
>> +	/* load head snapshot, other cpus may change it */
>> +	uint32_t head = smp_load_acquire(&slot->head);
>> +
>> +	while (head != READ_ONCE(slot->last)) {
>> +		void *obj;
>> +
>> +		/* obj must be retrieved before moving forward head */
>> +		obj = READ_ONCE(slot->entries[head & slot->mask]);
>> +
>> +		/* move head forward to mark it's consumption */
>> +		if (try_cmpxchg_release(&slot->head, &head, head + 1))
>> +			return obj;
>> +	}
>> +
>> +	return NULL;
>> +}
>> +
>> +/* allocate an object from object pool */
>> +void *objpool_pop(struct objpool_head *pool)
>> +{
>> +	void *obj = NULL;
>> +	unsigned long flags;
>> +	int i, cpu;
>> +
>> +	/* disable local irq to avoid preemption & interruption */
>> +	raw_local_irq_save(flags);
>> +
>> +	cpu = raw_smp_processor_id();
>> +	for (i = 0; i < num_possible_cpus(); i++) {
>> +		obj = objpool_try_get_slot(pool, cpu);
>> +		if (obj)
>> +			break;
>> +		cpu = cpumask_next_wrap(cpu, cpu_possible_mask, -1, 1);
>> +	}
>> +	raw_local_irq_restore(flags);
>> +
>> +	return obj;
>> +}
>> +EXPORT_SYMBOL_GPL(objpool_pop);
>> +
>> +/* release whole objpool forcely */
>> +void objpool_free(struct objpool_head *pool)
>> +{
>> +	if (!pool->cpu_slots)
>> +		return;
>> +
>> +	/* release percpu slots */
>> +	objpool_fini_percpu_slots(pool);
>> +
>> +	/* call user's cleanup callback if provided */
>> +	if (pool->release)
>> +		pool->release(pool, pool->context);
>> +}
>> +EXPORT_SYMBOL_GPL(objpool_free);
>> +
>> +/* drop the allocated object, rather reclaim it to objpool */
>> +int objpool_drop(void *obj, struct objpool_head *pool)
>> +{
>> +	if (!obj || !pool)
>> +		return -EINVAL;
>> +
>> +	if (refcount_dec_and_test(&pool->ref)) {
>> +		objpool_free(pool);
>> +		return 0;
>> +	}
>> +
>> +	return -EAGAIN;
>> +}
>> +EXPORT_SYMBOL_GPL(objpool_drop);
>> +
>> +/* drop unused objects and defref objpool for releasing */
>> +void objpool_fini(struct objpool_head *pool)
>> +{
>> +	int count = 1; /* extra ref for objpool itself */
>> +
>> +	/* drop all remained objects from objpool */
>> +	while (objpool_pop(pool))
>> +		count++;
>> +
>> +	if (refcount_sub_and_test(count, &pool->ref))
>> +		objpool_free(pool);
>> +}
>> +EXPORT_SYMBOL_GPL(objpool_fini);

Powered by blists - more mailing lists