Re: [dpdk-dev] [PATCH 1/3] rcu: add RCU library supporting QSBR mechanism

[Honnappa Nagarahalli <Honnappa.Nagarahalli@arm.com>]

> Hi Honnappa,
> 
> > diff --git a/lib/librte_rcu/rte_rcu_qsbr.c
> > b/lib/librte_rcu/rte_rcu_qsbr.c new file mode 100644 index
> > 000000000..0fc4515ea
> > --- /dev/null
> > +++ b/lib/librte_rcu/rte_rcu_qsbr.c
> > @@ -0,0 +1,99 @@
> > +/* SPDX-License-Identifier: BSD-3-Clause
> > + *
> > + * Copyright (c) 2018 Arm Limited
> > + */
> > +
> > +#include <stdio.h>
> > +#include <string.h>
> > +#include <stdint.h>
> > +#include <errno.h>
> > +
> > +#include <rte_common.h>
> > +#include <rte_log.h>
> > +#include <rte_memory.h>
> > +#include <rte_malloc.h>
> > +#include <rte_eal.h>
> > +#include <rte_eal_memconfig.h>
> > +#include <rte_atomic.h>
> > +#include <rte_per_lcore.h>
> > +#include <rte_lcore.h>
> > +#include <rte_errno.h>
> > +
> > +#include "rte_rcu_qsbr.h"
> > +
> > +/* Get the memory size of QSBR variable */ size_t __rte_experimental
> > +rte_rcu_qsbr_get_memsize(uint32_t max_threads) {
> > +	size_t sz;
> > +
> > +	RTE_ASSERT(max_threads == 0);
> 
> Here and in all similar places:
> assert() will abort when its condition will be evaluated to false.
> So it should be max_threads != 0.
Thanks for this comment. Enabling RTE_ENABLE_ASSERT resulted in more problems. I will fix in the next version.

> Also it a public and non-datapath function.
> Calling assert() for invalid input parameter - seems way too extreme.
> Why not just return error to the caller?
Ok, I will change it.

> 
> > +
> > +	sz = sizeof(struct rte_rcu_qsbr);
> > +
> > +	/* Add the size of quiescent state counter array */
> > +	sz += sizeof(struct rte_rcu_qsbr_cnt) * max_threads;
> > +
> > +	return RTE_ALIGN(sz, RTE_CACHE_LINE_SIZE); }
> > +
> > +/* Initialize a quiescent state variable */ void __rte_experimental
> > +rte_rcu_qsbr_init(struct rte_rcu_qsbr *v, uint32_t max_threads) {
> > +	RTE_ASSERT(v == NULL);
> > +
> > +	memset(v, 0, rte_rcu_qsbr_get_memsize(max_threads));
> > +	v->m_threads = max_threads;
> > +	v->num_elems = RTE_ALIGN_MUL_CEIL(max_threads,
> > +			RTE_QSBR_THRID_ARRAY_ELM_SIZE) /
> > +			RTE_QSBR_THRID_ARRAY_ELM_SIZE;
> > +	v->token = RTE_QSBR_CNT_INIT;
> > +}
> > +
> > +/* Dump the details of a single quiescent state variable to a file.
> > +*/ void __rte_experimental rte_rcu_qsbr_dump(FILE *f, struct
> > +rte_rcu_qsbr *v) {
> > +	uint64_t bmap;
> > +	uint32_t i, t;
> > +
> > +	RTE_ASSERT(v == NULL || f == NULL);
> > +
> > +	fprintf(f, "\nQuiescent State Variable @%p\n", v);
> > +
> > +	fprintf(f, "  QS variable memory size = %lu\n",
> > +				rte_rcu_qsbr_get_memsize(v->m_threads));
> > +	fprintf(f, "  Given # max threads = %u\n", v->m_threads);
> > +
> > +	fprintf(f, "  Registered thread ID mask = 0x");
> > +	for (i = 0; i < v->num_elems; i++)
> > +		fprintf(f, "%lx", __atomic_load_n(&v->reg_thread_id[i],
> > +					__ATOMIC_ACQUIRE));
> > +	fprintf(f, "\n");
> > +
> > +	fprintf(f, "  Token = %lu\n",
> > +			__atomic_load_n(&v->token, __ATOMIC_ACQUIRE));
> > +
> > +	fprintf(f, "Quiescent State Counts for readers:\n");
> > +	for (i = 0; i < v->num_elems; i++) {
> > +		bmap = __atomic_load_n(&v->reg_thread_id[i],
> __ATOMIC_ACQUIRE);
> > +		while (bmap) {
> > +			t = __builtin_ctzl(bmap);
> > +			fprintf(f, "thread ID = %d, count = %lu\n", t,
> > +				__atomic_load_n(
> > +					&RTE_QSBR_CNT_ARRAY_ELM(v, i)-
> >cnt,
> > +					__ATOMIC_RELAXED));
> > +			bmap &= ~(1UL << t);
> > +		}
> > +	}
> > +}
> > +
> > +int rcu_log_type;
> > +
> > +RTE_INIT(rte_rcu_register)
> > +{
> > +	rcu_log_type = rte_log_register("lib.rcu");
> > +	if (rcu_log_type >= 0)
> > +		rte_log_set_level(rcu_log_type, RTE_LOG_ERR); }
> > diff --git a/lib/librte_rcu/rte_rcu_qsbr.h
> > b/lib/librte_rcu/rte_rcu_qsbr.h new file mode 100644 index
> > 000000000..83943f751
> > --- /dev/null
> > +++ b/lib/librte_rcu/rte_rcu_qsbr.h
> > @@ -0,0 +1,511 @@
> > +/* SPDX-License-Identifier: BSD-3-Clause
> > + * Copyright (c) 2018 Arm Limited
> > + */
> > +
> > +#ifndef _RTE_RCU_QSBR_H_
> > +#define _RTE_RCU_QSBR_H_
> > +
> > +/**
> > + * @file
> > + * RTE Quiescent State Based Reclamation (QSBR)
> > + *
> > + * Quiescent State (QS) is any point in the thread execution
> > + * where the thread does not hold a reference to a data structure
> > + * in shared memory. While using lock-less data structures, the
> > +writer
> > + * can safely free memory once all the reader threads have entered
> > + * quiescent state.
> > + *
> > + * This library provides the ability for the readers to report
> > +quiescent
> > + * state and for the writers to identify when all the readers have
> > + * entered quiescent state.
> > + */
> > +
> > +#ifdef __cplusplus
> > +extern "C" {
> > +#endif
> > +
> > +#include <stdio.h>
> > +#include <stdint.h>
> > +#include <errno.h>
> > +#include <rte_common.h>
> > +#include <rte_memory.h>
> > +#include <rte_lcore.h>
> > +#include <rte_debug.h>
> > +
> > +extern int rcu_log_type;
> > +
> > +#if RTE_LOG_DP_LEVEL >= RTE_LOG_DEBUG #define RCU_DP_LOG(level,
> fmt,
> > +args...) \
> > +	rte_log(RTE_LOG_ ## level, rcu_log_type, \
> > +		"%s(): " fmt "\n", __func__, ## args) #else #define
> > +RCU_DP_LOG(level, fmt, args...) #endif
> 
> Why do you need that?
> Can't you use RTE_LOG_DP() instead?
RTE_LOG_DP is for static log types such as RTE_LOGTYPE_EAL, RTE_LOGTYPE_MBUF etc. Use of static log type in RCU was rejected earlier. Hence, I am using the dynamic log types.

> 
> > +
> > +/* Registered thread IDs are stored as a bitmap of 64b element array.
> > + * Given thread id needs to be converted to index into the array and
> > + * the id within the array element.
> > + */
> > +#define RTE_RCU_MAX_THREADS 1024
> > +#define RTE_QSBR_THRID_ARRAY_ELM_SIZE (sizeof(uint64_t) * 8) #define
> > +RTE_QSBR_THRID_ARRAY_ELEMS \
> > +	(RTE_ALIGN_MUL_CEIL(RTE_RCU_MAX_THREADS, \
> > +	 RTE_QSBR_THRID_ARRAY_ELM_SIZE) /
> RTE_QSBR_THRID_ARRAY_ELM_SIZE)
> > +#define RTE_QSBR_THRID_INDEX_SHIFT 6 #define RTE_QSBR_THRID_MASK
> 0x3f
> > +#define RTE_QSBR_THRID_INVALID 0xffffffff
> > +
> > +/* Worker thread counter */
> > +struct rte_rcu_qsbr_cnt {
> > +	uint64_t cnt;
> > +	/**< Quiescent state counter. Value 0 indicates the thread is
> > +offline */ } __rte_cache_aligned;
> > +
> > +#define RTE_QSBR_CNT_ARRAY_ELM(v, i) (((struct rte_rcu_qsbr_cnt *)(v
> > ++ 1)) + i)
> 
> You can probably add
> struct rte_rcu_qsbr_cnt cnt[0];
> at the end of struct rte_rcu_qsbr, then wouldn't need macro above.
ok

> 
> > +#define RTE_QSBR_CNT_THR_OFFLINE 0
> > +#define RTE_QSBR_CNT_INIT 1
> > +
> > +/**
> > + * RTE thread Quiescent State structure.
> > + * Quiescent state counter array (array of 'struct
> > +rte_rcu_qsbr_cnt'),
> > + * whose size is dependent on the maximum number of reader threads
> > + * (m_threads) using this variable is stored immediately following
> > + * this structure.
> > + */
> > +struct rte_rcu_qsbr {
> > +	uint64_t token __rte_cache_aligned;
> > +	/**< Counter to allow for multiple simultaneous QS queries */
> > +
> > +	uint32_t num_elems __rte_cache_aligned;
> > +	/**< Number of elements in the thread ID array */
> > +	uint32_t m_threads;
> > +	/**< Maximum number of threads this RCU variable will use */
> > +
> > +	uint64_t reg_thread_id[RTE_QSBR_THRID_ARRAY_ELEMS]
> __rte_cache_aligned;
> > +	/**< Registered thread IDs are stored in a bitmap array */
> 
> 
> As I understand you ended up with fixed size array to avoid 2 variable size
> arrays in this struct?
Yes

> Is that big penalty for register/unregister() to either store a pointer to bitmap,
> or calculate it based on num_elems value?
In the last RFC I sent out [1], I tested the impact of having non-fixed size array. There 'was' a performance degradation in most of the performance tests. The issue was with calculating the address of per thread QSBR counters (not with the address calculation of the bitmap). With the current patch, I do not see the performance difference (the difference between the RFC and this patch are the memory orderings, they are masking any perf gain from having a fixed array). However, I have kept the fixed size array as the generated code does not have additional calculations to get the address of qsbr counter array elements.

[1] http://mails.dpdk.org/archives/dev/2019-February/125029.html

> As another thought - do we really need bitmap at all?
The bit map is helping avoid accessing all the elements in rte_rcu_qsbr_cnt array (as you have mentioned below). This provides the ability to scale the number of threads dynamically. For ex: an application can create a qsbr variable with 48 max threads, but currently only 2 threads are active (due to traffic conditions).

> Might it is possible to sotre register value for each thread inside it's
> rte_rcu_qsbr_cnt:
> struct rte_rcu_qsbr_cnt {uint64_t cnt; uint32_t register;}
> __rte_cache_aligned; ?
> That would cause check() to walk through all elems in rte_rcu_qsbr_cnt array,
> but from other side would help to avoid cache conflicts for register/unregister.
With the addition of rte_rcu_qsbr_thread_online/offline APIs, the register/unregister APIs are not in critical path anymore. Hence, the cache conflicts are fine. The online/offline APIs work on thread specific cache lines and these are in the critical path.

> 
> > +} __rte_cache_aligned;
> > +
> > +/**
> > + * @warning
> > + * @b EXPERIMENTAL: this API may change without prior notice
> > + *
> > + * Return the size of the memory occupied by a Quiescent State variable.
> > + *
> > + * @param max_threads
> > + *   Maximum number of threads reporting quiescent state on this variable.
> > + * @return
> > + *   Size of memory in bytes required for this QS variable.
> > + */
> > +size_t __rte_experimental
> > +rte_rcu_qsbr_get_memsize(uint32_t max_threads);
> > +
> > +/**
> > + * @warning
> > + * @b EXPERIMENTAL: this API may change without prior notice
> > + *
> > + * Initialize a Quiescent State (QS) variable.
> > + *
> > + * @param v
> > + *   QS variable
> > + * @param max_threads
> > + *   Maximum number of threads reporting QS on this variable.
> > + *
> > + */
> > +void __rte_experimental
> > +rte_rcu_qsbr_init(struct rte_rcu_qsbr *v, uint32_t max_threads);
> > +
> > +/**
> > + * @warning
> > + * @b EXPERIMENTAL: this API may change without prior notice
> > + *
> > + * Register a reader thread to report its quiescent state
> > + * on a QS variable.
> > + *
> > + * This is implemented as a lock-free function. It is multi-thread
> > + * safe.
> > + * Any reader thread that wants to report its quiescent state must
> > + * call this API. This can be called during initialization or as part
> > + * of the packet processing loop.
> > + *
> > + * Note that rte_rcu_qsbr_thread_online must be called before the
> > + * thread updates its QS using rte_rcu_qsbr_update.
> > + *
> > + * @param v
> > + *   QS variable
> > + * @param thread_id
> > + *   Reader thread with this thread ID will report its quiescent state on
> > + *   the QS variable.
> > + */
> > +static __rte_always_inline void __rte_experimental
> > +rte_rcu_qsbr_thread_register(struct rte_rcu_qsbr *v, unsigned int
> > +thread_id) {
> > +	unsigned int i, id;
> > +
> > +	RTE_ASSERT(v == NULL || thread_id >= v->max_threads);
> > +
> > +	id = thread_id & RTE_QSBR_THRID_MASK;
> > +	i = thread_id >> RTE_QSBR_THRID_INDEX_SHIFT;
> > +
> > +	/* Release the new register thread ID to other threads
> > +	 * calling rte_rcu_qsbr_check.
> > +	 */
> > +	__atomic_fetch_or(&v->reg_thread_id[i], 1UL << id,
> > +__ATOMIC_RELEASE); }
> > +
> > +/**
> > + * @warning
> > + * @b EXPERIMENTAL: this API may change without prior notice
> > + *
> > + * Remove a reader thread, from the list of threads reporting their
> > + * quiescent state on a QS variable.
> > + *
> > + * This is implemented as a lock-free function. It is multi-thread safe.
> > + * This API can be called from the reader threads during shutdown.
> > + * Ongoing QS queries will stop waiting for the status from this
> > + * unregistered reader thread.
> > + *
> > + * @param v
> > + *   QS variable
> > + * @param thread_id
> > + *   Reader thread with this thread ID will stop reporting its quiescent
> > + *   state on the QS variable.
> > + */
> > +static __rte_always_inline void __rte_experimental
> > +rte_rcu_qsbr_thread_unregister(struct rte_rcu_qsbr *v, unsigned int
> > +thread_id) {
> > +	unsigned int i, id;
> > +
> > +	RTE_ASSERT(v == NULL || thread_id >= v->max_threads);
> > +
> > +	id = thread_id & RTE_QSBR_THRID_MASK;
> > +	i = thread_id >> RTE_QSBR_THRID_INDEX_SHIFT;
> > +
> > +	/* Make sure the removal of the thread from the list of
> > +	 * reporting threads is visible before the thread
> > +	 * does anything else.
> > +	 */
> > +	__atomic_fetch_and(&v->reg_thread_id[i],
> > +				~(1UL << id), __ATOMIC_RELEASE);
> > +}
> > +
> > +/**
> > + * @warning
> > + * @b EXPERIMENTAL: this API may change without prior notice
> > + *
> > + * Add a registered reader thread, to the list of threads reporting
> > +their
> > + * quiescent state on a QS variable.
> > + *
> > + * This is implemented as a lock-free function. It is multi-thread
> > + * safe.
> > + *
> > + * Any registered reader thread that wants to report its quiescent
> > +state must
> > + * call this API before calling rte_rcu_qsbr_update. This can be
> > +called
> > + * during initialization or as part of the packet processing loop.
> > + *
> > + * The reader thread must call rte_rcu_thread_offline API, before
> > + * calling any functions that block, to ensure that
> > +rte_rcu_qsbr_check
> > + * API does not wait indefinitely for the reader thread to update its QS.
> > + *
> > + * The reader thread must call rte_rcu_thread_online API, after the
> > +blocking
> > + * function call returns, to ensure that rte_rcu_qsbr_check API
> > + * waits for the reader thread to update its QS.
> > + *
> > + * @param v
> > + *   QS variable
> > + * @param thread_id
> > + *   Reader thread with this thread ID will report its quiescent state on
> > + *   the QS variable.
> > + */
> > +static __rte_always_inline void __rte_experimental
> > +rte_rcu_qsbr_thread_online(struct rte_rcu_qsbr *v, unsigned int
> > +thread_id) {
> > +	uint64_t t;
> > +
> > +	RTE_ASSERT(v == NULL || thread_id >= v->max_threads);
> > +
> > +	/* Copy the current value of token.
> > +	 * The fence at the end of the function will ensure that
> > +	 * the following will not move down after the load of any shared
> > +	 * data structure.
> > +	 */
> > +	t = __atomic_load_n(&v->token, __ATOMIC_RELAXED);
> > +
> > +	/* __atomic_store_n(cnt, __ATOMIC_RELAXED) is used to ensure
> > +	 * 'cnt' (64b) is accessed atomically.
> > +	 */
> > +	__atomic_store_n(&RTE_QSBR_CNT_ARRAY_ELM(v, thread_id)->cnt,
> > +		t, __ATOMIC_RELAXED);
> > +
> > +	/* The subsequent load of the data structure should not
> > +	 * move above the store. Hence a store-load barrier
> > +	 * is required.
> > +	 * If the load of the data structure moves above the store,
> > +	 * writer might not see that the reader is online, even though
> > +	 * the reader is referencing the shared data structure.
> > +	 */
> > +	__atomic_thread_fence(__ATOMIC_SEQ_CST);
> 
> If it has to generate a proper memory-barrier here anyway, could it use
> rte_smp_mb() here?
> At least for IA it would generate more lightweight one.
I have used the C++11 memory model functions. I prefer to not mix it with barriers. Does ICC generate lightweight code for the above fence?
Is it ok to add rte_smp_mb for x86 alone?

> Konstantin
> 
> > +}
> > +