[dpdk-dev] [PATCH v4 1/3] lib/lpm: integrate RCU QSBR

[Medvedkin, Vladimir]

Hi Ruifeng,

Thanks for patches, see comments below


On 08/06/2020 06:16, Ruifeng Wang wrote:
> Currently, the tbl8 group is freed even though the readers might be
> using the tbl8 group entries. The freed tbl8 group can be reallocated
> quickly. This results in incorrect lookup results.
>
> RCU QSBR process is integrated for safe tbl8 group reclaim.
> Refer to RCU documentation to understand various aspects of
> integrating RCU library into other libraries.
>
> Signed-off-by: Ruifeng Wang <ruifeng.wang at arm.com>
> Reviewed-by: Honnappa Nagarahalli <honnappa.nagarahalli at arm.com>
> ---
>   doc/guides/prog_guide/lpm_lib.rst  |  32 ++++++++
>   lib/librte_lpm/Makefile            |   2 +-
>   lib/librte_lpm/meson.build         |   1 +
>   lib/librte_lpm/rte_lpm.c           | 123 ++++++++++++++++++++++++++---
>   lib/librte_lpm/rte_lpm.h           |  59 ++++++++++++++
>   lib/librte_lpm/rte_lpm_version.map |   6 ++
>   6 files changed, 211 insertions(+), 12 deletions(-)
>
> diff --git a/doc/guides/prog_guide/lpm_lib.rst b/doc/guides/prog_guide/lpm_lib.rst
> index 1609a57d0..7cc99044a 100644
> --- a/doc/guides/prog_guide/lpm_lib.rst
> +++ b/doc/guides/prog_guide/lpm_lib.rst
> @@ -145,6 +145,38 @@ depending on whether we need to move to the next table or not.
>   Prefix expansion is one of the keys of this algorithm,
>   since it improves the speed dramatically by adding redundancy.
>   
> +Deletion
> +~~~~~~~~
> +
> +When deleting a rule, a replacement rule is searched for. Replacement rule is an existing rule that has
> +the longest prefix match with the rule to be deleted, but has smaller depth.
> +
> +If a replacement rule is found, target tbl24 and tbl8 entries are updated to have the same depth and next hop
> +value with the replacement rule.
> +
> +If no replacement rule can be found, target tbl24 and tbl8 entries will be cleared.
> +
> +Prefix expansion is performed if the rule's depth is not exactly 24 bits or 32 bits.
> +
> +After deleting a rule, a group of tbl8s that belongs to the same tbl24 entry are freed in following cases:
> +
> +*   All tbl8s in the group are empty .
> +
> +*   All tbl8s in the group have the same values and with depth no greater than 24.
> +
> +Free of tbl8s have different behaviors:
> +
> +*   If RCU is not used, tbl8s are cleared and reclaimed immediately.
> +
> +*   If RCU is used, tbl8s are reclaimed when readers are in quiescent state.
> +
> +When the LPM is not using RCU, tbl8 group can be freed immediately even though the readers might be using
> +the tbl8 group entries. This might result in incorrect lookup results.
> +
> +RCU QSBR process is integrated for safe tbl8 group reclaimation. Application has certain responsibilities
> +while using this feature. Please refer to resource reclaimation framework of :ref:`RCU library <RCU_Library>`
> +for more details.
> +
>   Lookup
>   ~~~~~~
>   
> diff --git a/lib/librte_lpm/Makefile b/lib/librte_lpm/Makefile
> index d682785b6..6f06c5c03 100644
> --- a/lib/librte_lpm/Makefile
> +++ b/lib/librte_lpm/Makefile
> @@ -8,7 +8,7 @@ LIB = librte_lpm.a
>   
>   CFLAGS += -O3
>   CFLAGS += $(WERROR_FLAGS) -I$(SRCDIR)
> -LDLIBS += -lrte_eal -lrte_hash
> +LDLIBS += -lrte_eal -lrte_hash -lrte_rcu
>   
>   EXPORT_MAP := rte_lpm_version.map
>   
> diff --git a/lib/librte_lpm/meson.build b/lib/librte_lpm/meson.build
> index 021ac6d8d..6cfc083c5 100644
> --- a/lib/librte_lpm/meson.build
> +++ b/lib/librte_lpm/meson.build
> @@ -7,3 +7,4 @@ headers = files('rte_lpm.h', 'rte_lpm6.h')
>   # without worrying about which architecture we actually need
>   headers += files('rte_lpm_altivec.h', 'rte_lpm_neon.h', 'rte_lpm_sse.h')
>   deps += ['hash']
> +deps += ['rcu']
> diff --git a/lib/librte_lpm/rte_lpm.c b/lib/librte_lpm/rte_lpm.c
> index 38ab512a4..30f541179 100644
> --- a/lib/librte_lpm/rte_lpm.c
> +++ b/lib/librte_lpm/rte_lpm.c
> @@ -1,5 +1,6 @@
>   /* SPDX-License-Identifier: BSD-3-Clause
>    * Copyright(c) 2010-2014 Intel Corporation
> + * Copyright(c) 2020 Arm Limited
>    */
>   
>   #include <string.h>
> @@ -246,12 +247,85 @@ rte_lpm_free(struct rte_lpm *lpm)
>   
>   	rte_mcfg_tailq_write_unlock();
>   
> +	if (lpm->dq)
> +		rte_rcu_qsbr_dq_delete(lpm->dq);
>   	rte_free(lpm->tbl8);
>   	rte_free(lpm->rules_tbl);
>   	rte_free(lpm);
>   	rte_free(te);
>   }
>   
> +static void
> +__lpm_rcu_qsbr_free_resource(void *p, void *data, unsigned int n)
> +{
> +	struct rte_lpm_tbl_entry zero_tbl8_entry = {0};
> +	uint32_t tbl8_group_index = *(uint32_t *)data;
> +	struct rte_lpm_tbl_entry *tbl8 = (struct rte_lpm_tbl_entry *)p;
> +
> +	RTE_SET_USED(n);
> +	/* Set tbl8 group invalid */
> +	__atomic_store(&tbl8[tbl8_group_index], &zero_tbl8_entry,
> +		__ATOMIC_RELAXED);
> +}
> +
> +/* Associate QSBR variable with an LPM object.
> + */
> +int
> +rte_lpm_rcu_qsbr_add(struct rte_lpm *lpm, struct rte_lpm_rcu_config *cfg,
> +	struct rte_rcu_qsbr_dq **dq)
> +{
> +	char rcu_dq_name[RTE_RCU_QSBR_DQ_NAMESIZE];
> +	struct rte_rcu_qsbr_dq_parameters params = {0};
> +
> +	if ((lpm == NULL) || (cfg == NULL)) {
> +		rte_errno = EINVAL;
> +		return 1;
> +	}
> +
> +	if (lpm->v) {
> +		rte_errno = EEXIST;
> +		return 1;
> +	}
> +
> +	if (cfg->mode == RTE_LPM_QSBR_MODE_SYNC) {
> +		/* No other things to do. */
> +	} else if (cfg->mode == RTE_LPM_QSBR_MODE_DQ) {
> +		/* Init QSBR defer queue. */
> +		snprintf(rcu_dq_name, sizeof(rcu_dq_name),
> +				"LPM_RCU_%s", lpm->name);
> +		params.name = rcu_dq_name;
> +		params.size = cfg->dq_size;
> +		if (params.size == 0)
> +			params.size = lpm->number_tbl8s;
> +		params.trigger_reclaim_limit = cfg->reclaim_thd;
> +		if (params.trigger_reclaim_limit == 0)


This makes it impossible for a user to configure reclamation triggering 
with every call. Should we allow it?


> +			params.trigger_reclaim_limit =
> +					RTE_LPM_RCU_DQ_RECLAIM_THD;
> +		params.max_reclaim_size = cfg->reclaim_max;
> +		if (params.max_reclaim_size == 0)
> +			params.max_reclaim_size = RTE_LPM_RCU_DQ_RECLAIM_MAX;
> +		params.esize = sizeof(uint32_t);	/* tbl8 group index */
> +		params.free_fn = __lpm_rcu_qsbr_free_resource;
> +		params.p = lpm->tbl8;


I think it's better to pass the LPM pointer here rather than tbl8, for 
example, in case we decide to add some counters in the future


> +		params.v = cfg->v;
> +		lpm->dq = rte_rcu_qsbr_dq_create(&params);
> +		if (lpm->dq == NULL) {
> +			RTE_LOG(ERR, LPM,
> +					"LPM QS defer queue creation failed\n");
> +			return 1;
> +		}
> +		if (dq)
> +			*dq = lpm->dq;
> +	} else {
> +		rte_errno = EINVAL;
> +		return 1;
> +	}
> +	lpm->rcu_mode = cfg->mode;
> +	lpm->v = cfg->v;
> +
> +	return 0;
> +}
> +
>   /*
>    * Adds a rule to the rule table.
>    *
> @@ -394,14 +468,15 @@ rule_find(struct rte_lpm *lpm, uint32_t ip_masked, uint8_t depth)
>    * Find, clean and allocate a tbl8.
>    */
>   static int32_t
> -tbl8_alloc(struct rte_lpm_tbl_entry *tbl8, uint32_t number_tbl8s)
> +_tbl8_alloc(struct rte_lpm *lpm)
>   {
>   	uint32_t group_idx; /* tbl8 group index. */
>   	struct rte_lpm_tbl_entry *tbl8_entry;
>   
>   	/* Scan through tbl8 to find a free (i.e. INVALID) tbl8 group. */
> -	for (group_idx = 0; group_idx < number_tbl8s; group_idx++) {
> -		tbl8_entry = &tbl8[group_idx * RTE_LPM_TBL8_GROUP_NUM_ENTRIES];
> +	for (group_idx = 0; group_idx < lpm->number_tbl8s; group_idx++) {
> +		tbl8_entry = &lpm->tbl8[group_idx *
> +					RTE_LPM_TBL8_GROUP_NUM_ENTRIES];
>   		/* If a free tbl8 group is found clean it and set as VALID. */
>   		if (!tbl8_entry->valid_group) {
>   			struct rte_lpm_tbl_entry new_tbl8_entry = {
> @@ -427,14 +502,40 @@ tbl8_alloc(struct rte_lpm_tbl_entry *tbl8, uint32_t number_tbl8s)
>   	return -ENOSPC;
>   }
>   
> +static int32_t
> +tbl8_alloc(struct rte_lpm *lpm)
> +{
> +	int32_t group_idx; /* tbl8 group index. */
> +
> +	group_idx = _tbl8_alloc(lpm);
> +	if ((group_idx < 0) && (lpm->dq != NULL)) {


I think (group_idx == -ENOSPC) will be safer


> +		/* If there are no tbl8 groups try to reclaim one. */
> +		if (rte_rcu_qsbr_dq_reclaim(lpm->dq, 1, NULL, NULL, NULL) == 0)
> +			group_idx = _tbl8_alloc(lpm);


I'm not really happy with this approach. _tbl8_alloc() produces linear 
scan through the memory to find a free group_idx and it is the slowest 
part of rte_lpm_add().
Here after reclamation of some group index we need to rescan a memory 
again to find it. It would be great if there will be some way to return 
reclaimed elements. Or just to dequeue elements from dq and reclaim them 
manually.


> +	}
> +
> +	return group_idx;
> +}
> +
>   static void
> -tbl8_free(struct rte_lpm_tbl_entry *tbl8, uint32_t tbl8_group_start)
> +tbl8_free(struct rte_lpm *lpm, uint32_t tbl8_group_start)
>   {
> -	/* Set tbl8 group invalid*/
>   	struct rte_lpm_tbl_entry zero_tbl8_entry = {0};
>   
> -	__atomic_store(&tbl8[tbl8_group_start], &zero_tbl8_entry,
> -			__ATOMIC_RELAXED);
> +	if (!lpm->v) {
> +		/* Set tbl8 group invalid*/
> +		__atomic_store(&lpm->tbl8[tbl8_group_start], &zero_tbl8_entry,
> +				__ATOMIC_RELAXED);
> +	} else if (lpm->rcu_mode == RTE_LPM_QSBR_MODE_SYNC) {
> +		/* Wait for quiescent state change. */
> +		rte_rcu_qsbr_synchronize(lpm->v, RTE_QSBR_THRID_INVALID);
> +		/* Set tbl8 group invalid*/
> +		__atomic_store(&lpm->tbl8[tbl8_group_start], &zero_tbl8_entry,
> +				__ATOMIC_RELAXED);
> +	} else if (lpm->rcu_mode == RTE_LPM_QSBR_MODE_DQ) {
> +		/* Push into QSBR defer queue. */
> +		rte_rcu_qsbr_dq_enqueue(lpm->dq, (void *)&tbl8_group_start);
> +	}
>   }
>   
>   static __rte_noinline int32_t
> @@ -523,7 +624,7 @@ add_depth_big(struct rte_lpm *lpm, uint32_t ip_masked, uint8_t depth,
>   
>   	if (!lpm->tbl24[tbl24_index].valid) {
>   		/* Search for a free tbl8 group. */
> -		tbl8_group_index = tbl8_alloc(lpm->tbl8, lpm->number_tbl8s);
> +		tbl8_group_index = tbl8_alloc(lpm);
>   
>   		/* Check tbl8 allocation was successful. */
>   		if (tbl8_group_index < 0) {
> @@ -569,7 +670,7 @@ add_depth_big(struct rte_lpm *lpm, uint32_t ip_masked, uint8_t depth,
>   	} /* If valid entry but not extended calculate the index into Table8. */
>   	else if (lpm->tbl24[tbl24_index].valid_group == 0) {
>   		/* Search for free tbl8 group. */
> -		tbl8_group_index = tbl8_alloc(lpm->tbl8, lpm->number_tbl8s);
> +		tbl8_group_index = tbl8_alloc(lpm);
>   
>   		if (tbl8_group_index < 0) {
>   			return tbl8_group_index;
> @@ -977,7 +1078,7 @@ delete_depth_big(struct rte_lpm *lpm, uint32_t ip_masked,
>   		 */
>   		lpm->tbl24[tbl24_index].valid = 0;
>   		__atomic_thread_fence(__ATOMIC_RELEASE);
> -		tbl8_free(lpm->tbl8, tbl8_group_start);
> +		tbl8_free(lpm, tbl8_group_start);
>   	} else if (tbl8_recycle_index > -1) {
>   		/* Update tbl24 entry. */
>   		struct rte_lpm_tbl_entry new_tbl24_entry = {
> @@ -993,7 +1094,7 @@ delete_depth_big(struct rte_lpm *lpm, uint32_t ip_masked,
>   		__atomic_store(&lpm->tbl24[tbl24_index], &new_tbl24_entry,
>   				__ATOMIC_RELAXED);
>   		__atomic_thread_fence(__ATOMIC_RELEASE);
> -		tbl8_free(lpm->tbl8, tbl8_group_start);
> +		tbl8_free(lpm, tbl8_group_start);
>   	}
>   #undef group_idx
>   	return 0;
> diff --git a/lib/librte_lpm/rte_lpm.h b/lib/librte_lpm/rte_lpm.h
> index b9d49ac87..8c054509a 100644
> --- a/lib/librte_lpm/rte_lpm.h
> +++ b/lib/librte_lpm/rte_lpm.h
> @@ -1,5 +1,6 @@
>   /* SPDX-License-Identifier: BSD-3-Clause
>    * Copyright(c) 2010-2014 Intel Corporation
> + * Copyright(c) 2020 Arm Limited
>    */
>   
>   #ifndef _RTE_LPM_H_
> @@ -20,6 +21,7 @@
>   #include <rte_memory.h>
>   #include <rte_common.h>
>   #include <rte_vect.h>
> +#include <rte_rcu_qsbr.h>
>   
>   #ifdef __cplusplus
>   extern "C" {
> @@ -62,6 +64,17 @@ extern "C" {
>   /** Bitmask used to indicate successful lookup */
>   #define RTE_LPM_LOOKUP_SUCCESS          0x01000000
>   
> +/** @internal Default threshold to trigger RCU defer queue reclaimation. */
> +#define RTE_LPM_RCU_DQ_RECLAIM_THD	32
> +
> +/** @internal Default RCU defer queue entries to reclaim in one go. */
> +#define RTE_LPM_RCU_DQ_RECLAIM_MAX	16
> +
> +/* Create defer queue for reclaim. */
> +#define RTE_LPM_QSBR_MODE_DQ		0
> +/* Use blocking mode reclaim. No defer queue created. */
> +#define RTE_LPM_QSBR_MODE_SYNC		0x01


using enums instead of defines?


> +
>   #if RTE_BYTE_ORDER == RTE_LITTLE_ENDIAN
>   /** @internal Tbl24 entry structure. */
>   __extension__
> @@ -130,6 +143,28 @@ struct rte_lpm {
>   			__rte_cache_aligned; /**< LPM tbl24 table. */
>   	struct rte_lpm_tbl_entry *tbl8; /**< LPM tbl8 table. */
>   	struct rte_lpm_rule *rules_tbl; /**< LPM rules. */
> +
> +	/* RCU config. */
> +	struct rte_rcu_qsbr *v;		/* RCU QSBR variable. */
> +	uint32_t rcu_mode;		/* Blocking, defer queue. */
> +	struct rte_rcu_qsbr_dq *dq;	/* RCU QSBR defer queue. */
> +};
> +
> +/** LPM RCU QSBR configuration structure. */
> +struct rte_lpm_rcu_config {
> +	struct rte_rcu_qsbr *v;	/* RCU QSBR variable. */
> +	/* Mode of RCU QSBR. RTE_LPM_QSBR_MODE_xxx
> +	 * '0' for default: create defer queue for reclaim.
> +	 */
> +	uint32_t mode;
> +	/* RCU defer queue size. default: lpm->number_tbl8s. */
> +	uint32_t dq_size;
> +	uint32_t reclaim_thd;	/* Threshold to trigger auto reclaim.
> +				 * default: RTE_LPM_RCU_DQ_RECLAIM_TRHD.
> +				 */
> +	uint32_t reclaim_max;	/* Max entries to reclaim in one go.
> +				 * default: RTE_LPM_RCU_DQ_RECLAIM_MAX.
> +				 */
>   };
>   
>   /**
> @@ -179,6 +214,30 @@ rte_lpm_find_existing(const char *name);
>   void
>   rte_lpm_free(struct rte_lpm *lpm);
>   
> +/**
> + * @warning
> + * @b EXPERIMENTAL: this API may change without prior notice
> + *
> + * Associate RCU QSBR variable with an LPM object.
> + *
> + * @param lpm
> + *   the lpm object to add RCU QSBR
> + * @param cfg
> + *   RCU QSBR configuration
> + * @param dq
> + *   handler of created RCU QSBR defer queue
> + * @return
> + *   On success - 0
> + *   On error - 1 with error code set in rte_errno.
> + *   Possible rte_errno codes are:
> + *   - EINVAL - invalid pointer
> + *   - EEXIST - already added QSBR
> + *   - ENOMEM - memory allocation failure
> + */
> +__rte_experimental
> +int rte_lpm_rcu_qsbr_add(struct rte_lpm *lpm, struct rte_lpm_rcu_config *cfg,
> +	struct rte_rcu_qsbr_dq **dq);
> +
>   /**
>    * Add a rule to the LPM table.
>    *
> diff --git a/lib/librte_lpm/rte_lpm_version.map b/lib/librte_lpm/rte_lpm_version.map
> index 500f58b80..bfccd7eac 100644
> --- a/lib/librte_lpm/rte_lpm_version.map
> +++ b/lib/librte_lpm/rte_lpm_version.map
> @@ -21,3 +21,9 @@ DPDK_20.0 {
>   
>   	local: *;
>   };
> +
> +EXPERIMENTAL {
> +	global:
> +
> +	rte_lpm_rcu_qsbr_add;
> +};

-- 
Regards,
Vladimir