[dpdk-dev] [PATCH v2] ring: guarantee ordering of cons/prod loading when doing

Jerin Jacob jerin.jacob at caviumnetworks.com
Tue Nov 7 10:57:29 CET 2017


-----Original Message-----
> Date: Tue, 7 Nov 2017 16:34:30 +0800
> From: Jia He <hejianet at gmail.com>
> To: Jerin Jacob <jerin.jacob at caviumnetworks.com>
> Cc: dev at dpdk.org, olivier.matz at 6wind.com, konstantin.ananyev at intel.com,
>  bruce.richardson at intel.com, jianbo.liu at arm.com, hemant.agrawal at nxp.com,
>  jie2.liu at hxt-semitech.com, bing.zhao at hxt-semitech.com,
>  jia.he at hxt-semitech.com
> Subject: Re: [PATCH v2] ring: guarantee ordering of cons/prod loading when
>  doing
> User-Agent: Mozilla/5.0 (Windows NT 6.1; WOW64; rv:52.0) Gecko/20100101
>  Thunderbird/52.4.0
> 
> 
> 
> On 11/7/2017 12:36 PM, Jerin Jacob Wrote:
> > -----Original Message-----
> > 
> > On option could be to change the prototype of update_tail() and make
> > compiler accommodate it for zero cost for arm64(Which I think, it it the
> > case. But you can check the generated instructions)
> > If not, move, __rte_ring_do_dequeue() and __rte_ring_do_enqueue() instead of
> > __rte_ring_move_prod_head/__rte_ring_move_cons_head/update_tail()
> > 
> > 
> > ➜ [master][dpdk.org] $ git diff
> > diff --git a/lib/librte_ring/rte_ring.h b/lib/librte_ring/rte_ring.h
> > index 5e9b3b7b4..b32648825 100644
> > --- a/lib/librte_ring/rte_ring.h
> > +++ b/lib/librte_ring/rte_ring.h
> > @@ -358,8 +358,12 @@ void rte_ring_dump(FILE *f, const struct rte_ring
> > *r);
> >   static __rte_always_inline void
> >   update_tail(struct rte_ring_headtail *ht, uint32_t old_val, uint32_t
> > new_val,
> > -               uint32_t single)
> > +               uint32_t single, const uint32_t enqueue)
> >   {
> > +       if (enqueue)
> > +               rte_smp_wmb();
> > +       else
> > +               rte_smp_rmb();
> >          /*
> >           * If there are other enqueues/dequeues in progress that
> >           * preceded us,
> >           * we need to wait for them to complete
> > @@ -470,9 +474,8 @@ __rte_ring_do_enqueue(struct rte_ring *r, void *
> > const *obj_table,
> >                  goto end;
> >          ENQUEUE_PTRS(r, &r[1], prod_head, obj_table, n, void *);
> > -       rte_smp_wmb();
> > -       update_tail(&r->prod, prod_head, prod_next, is_sp);
> > +       update_tail(&r->prod, prod_head, prod_next, is_sp, 1);
> >   end:
> >          if (free_space != NULL)
> >                  *free_space = free_entries - n;
> > @@ -575,9 +578,8 @@ __rte_ring_do_dequeue(struct rte_ring *r, void
> > **obj_table,
> >                  goto end;
> >          DEQUEUE_PTRS(r, &r[1], cons_head, obj_table, n, void *);
> > -       rte_smp_rmb();
> > -       update_tail(&r->cons, cons_head, cons_next, is_sc);
> > +       update_tail(&r->cons, cons_head, cons_next, is_sc, 0);
> >   end:
> >          if (available != NULL)
> > 
> > 
> > 
> Hi Jerin, yes I knew this suggestion in update_tail.
> But what I mean is the rte_smp_rmb() in __rte_ring_move_cons_head and
> __rte_ring_move_pros_head:
> [option 1]
> +        *old_head = r->cons.head;
> +        rte_smp_rmb();
> +        const uint32_t prod_tail = r->prod.tail;
> 
> [option 2]
> +        *old_head = __atomic_load_n(&r->cons.head,
> +                    __ATOMIC_ACQUIRE);
> +        *old_head = r->cons.head;
> 
> ie.I wonder what is the suitable new config name to distinguish the above 2
> options?

Why?
If you fix the generic version with rte_smp_rmb() then we just need only
one config to differentiate between c11 vs generic. See comments below,

> Thanks for the patience :-)
> 
> see my drafted patch below, the marcro "PREFER":
> + */
> +
> +#ifndef _RTE_RING_C11_MEM_H_
> +#define _RTE_RING_C11_MEM_H_
> +
> +static __rte_always_inline void
> +update_tail(struct rte_ring_headtail *ht, uint32_t old_val, uint32_t
> new_val,
> +        uint32_t single, uint32_t enqueue)
> +{
> +    /* Don't need wmb/rmb when we prefer to use load_acquire/
> +     * store_release barrier */
> +#ifndef PREFER
> +    if (enqueue)
> +        rte_smp_wmb();
> +    else
> +        rte_smp_rmb();
> +#endif

You can remove PREFER and let the "generic" version has this. For x86,
rte_smp_?mb() it will be NOOP. So no issue.

> +
> +    /*
> +     * If there are other enqueues/dequeues in progress that preceded us,
> +     * we need to wait for them to complete
> +     */
> +    if (!single)
> +        while (unlikely(ht->tail != old_val))
> +            rte_pause();
> +
> +#ifdef PREFER
> +    __atomic_store_n(&ht->tail, new_val, __ATOMIC_RELEASE);

for c11 mem model version, it needs only __atomic_store_n version.

> +#else
> +    ht->tail = new_val;
> +#endif
> +}
> +
> +/**
> + * @internal This function updates the producer head for enqueue
> + *
> + * @param r
> + *   A pointer to the ring structure
> + * @param is_sp
> + *   Indicates whether multi-producer path is needed or not
> + * @param n
> + *   The number of elements we will want to enqueue, i.e. how far should
> the
> + *   head be moved
> + * @param behavior
> + *   RTE_RING_QUEUE_FIXED:    Enqueue a fixed number of items from a ring
> + *   RTE_RING_QUEUE_VARIABLE: Enqueue as many items as possible from ring
> + * @param old_head
> + *   Returns head value as it was before the move, i.e. where enqueue
> starts
> + * @param new_head
> + *   Returns the current/new head value i.e. where enqueue finishes
> + * @param free_entries
> + *   Returns the amount of free space in the ring BEFORE head was moved
> + * @return
> + *   Actual number of objects enqueued.
> + *   If behavior == RTE_RING_QUEUE_FIXED, this will be 0 or n only.
> + */
> +static __rte_always_inline unsigned int
> +__rte_ring_move_prod_head(struct rte_ring *r, int is_sp,
> +        unsigned int n, enum rte_ring_queue_behavior behavior,
> +        uint32_t *old_head, uint32_t *new_head,
> +        uint32_t *free_entries)
> +{
> +    const uint32_t capacity = r->capacity;
> +    unsigned int max = n;
> +    int success;
> +
> +    do {
> +        /* Reset n to the initial burst count */
> +        n = max;
> +
> +#ifdef PREFER
> +        *old_head = __atomic_load_n(&r->prod.head,
> +                    __ATOMIC_ACQUIRE);
> +#else
> +        *old_head = r->prod.head;
> +        /* prevent reorder of load/load */
> +        rte_smp_rmb();
> +#endif

Same as above comment.

> +        const uint32_t cons_tail = r->cons.tail;
> +        /*
> +         *  The subtraction is done between two unsigned 32bits value
> +         * (the result is always modulo 32 bits even if we have
> +         * *old_head > cons_tail). So 'free_entries' is always between 0
> +         * and capacity (which is < size).
> +         */
> +        *free_entries = (capacity + cons_tail - *old_head);
> +
> +        /* check that we have enough room in ring */
> +        if (unlikely(n > *free_entries))

> +static __rte_always_inline unsigned int
> +__rte_ring_do_enqueue(struct rte_ring *r, void * const *obj_table,
> +         unsigned int n, enum rte_ring_queue_behavior behavior,
> +         int is_sp, unsigned int *free_space)
> +{


Duplicate function, No need to replicate on both versions.


> +static __rte_always_inline unsigned int
> +__rte_ring_move_cons_head(struct rte_ring *r, int is_sc,
> +        unsigned int n, enum rte_ring_queue_behavior behavior,
> +        uint32_t *old_head, uint32_t *new_head,
> +        uint32_t *entries)
> +{
> +    unsigned int max = n;
> +    int success;
> +
> +    /* move cons.head atomically */
> +    do {
> +        /* Restore n as it may change every loop */
> +        n = max;
> +#ifdef PREFER
> +        *old_head = __atomic_load_n(&r->cons.head,
> +                    __ATOMIC_ACQUIRE);
> +#else
> +        *old_head = r->cons.head;
> +        /*  prevent reorder of load/load */
> +        rte_smp_rmb();
> +#endif

Same as above comment

> +
> +        const uint32_t prod_tail = r->prod.tail;
> +        /* The subtraction is done between two unsigned 32bits value
> +         * (the result is always modulo 32 bits even if we have
> +         * cons_head > prod_tail). So 'entries' is always between 0
> +         * and size(ring)-1. */
> +        *entries = (prod_tail - *old_head);
> +
> +        /* Set the actual entries for dequeue */
> +        if (n > *entries)
> +            n = (behavior == RTE_RING_QUEUE_FIXED) ? 0 : *entries;
> +
> +        if (unlikely(n == 0))
> +            return 0;
> +
> +        *new_head = *old_head + n;
> +        if (is_sc)
> +            r->cons.head = *new_head, success = 1;
> +        else
> +#ifdef PREFER
> +            success = arch_rte_atomic32_cmpset(&r->cons.head,
> +                            old_head, *new_head,
> +                            0, __ATOMIC_ACQUIRE,
> +                            __ATOMIC_RELAXED);
> +#else
> +            success = rte_atomic32_cmpset(&r->cons.head, *old_head,
> +                    *new_head);
> +#endif

Same as above comment

> +    } while (unlikely(success == 0));
> +    return n;
> +}
> +
> +/**
> + * @internal Dequeue several objects from the ring
> + *
> + * @param r
> + *   A pointer to the ring structure.
> + * @param obj_table
> + *   A pointer to a table of void * pointers (objects).
> + * @param n
> + *   The number of objects to pull from the ring.
> + * @param behavior
> + *   RTE_RING_QUEUE_FIXED:    Dequeue a fixed number of items from a ring
> + *   RTE_RING_QUEUE_VARIABLE: Dequeue as many items as possible from ring
> + * @param is_sc
> + *   Indicates whether to use single consumer or multi-consumer head update
> + * @param available
> + *   returns the number of remaining ring entries after the dequeue has
> finished
> + * @return
> + *   - Actual number of objects dequeued.
> + *     If behavior == RTE_RING_QUEUE_FIXED, this will be 0 or n only.
> + */
> +static __rte_always_inline unsigned int
> +__rte_ring_do_dequeue(struct rte_ring *r, void **obj_table,
> +         unsigned int n, enum rte_ring_queue_behavior behavior,
> +         int is_sc, unsigned int *available)
> +{

Duplicate function, No need to replicate on both versions.


> +    uint32_t cons_head, cons_next;
> +    uint32_t entries;
> +
> +    n = __rte_ring_move_cons_head(r, is_sc, n, behavior,
> +            &cons_head, &cons_next, &entries);
> +    if (n == 0)
> +        goto end;
> +
> +    DEQUEUE_PTRS(r, &r[1], cons_head, obj_table, n, void *);
> +
> +    update_tail(&r->cons, cons_head, cons_next, is_sc, 0);
> +
> +end:
> +    if (available != NULL)
> +        *available = entries - n;
> +    return n;
> +}
> +
> +#endif /* _RTE_RING_C11_MEM_H_ */
> +
> diff --git a/lib/librte_ring/rte_ring_generic.h
> b/lib/librte_ring/rte_ring_generic.h
> new file mode 100644
> index 0000000..0ce6d57
> --- /dev/null
> +++ b/lib/librte_ring/rte_ring_generic.h
> @@ -0,0 +1,268 @@
> +/*-
> + *   BSD LICENSE
> + *
> + *   Copyright(c) 2017 hxt-semitech. All rights reserved.
> + *
> + *   Redistribution and use in source and binary forms, with or without
> + *   modification, are permitted provided that the following conditions
> + *   are met:
> + *
> + *     * Redistributions of source code must retain the above copyright
> + *       notice, this list of conditions and the following disclaimer.
> + *     * Redistributions in binary form must reproduce the above copyright
> + *       notice, this list of conditions and the following disclaimer in
> + *       the documentation and/or other materials provided with the
> + *       distribution.
> + *     * Neither the name of hxt-semitech nor the names of its
> + *       contributors may be used to endorse or promote products derived
> + *       from this software without specific prior written permission.
> + *
> + *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
> + *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
> + *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
> + *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
> + *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
> + *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
> + *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
> + *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
> + *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
> + *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
> + *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
> + */
> +
> +#ifndef _RTE_RING_GENERIC_H_
> +#define _RTE_RING_GENERIC_H_
> +
> +static __rte_always_inline void
> +update_tail(struct rte_ring_headtail *ht, uint32_t old_val, uint32_t
> new_val,
> +        uint32_t single, uint32_t enqueue)
> +{
> +    if (enqueue)
> +        rte_smp_wmb();
> +    else
> +        rte_smp_rmb();
> +    /*
> +     * If there are other enqueues/dequeues in progress that preceded us,
> +     * we need to wait for them to complete
> +     */
> +    if (!single)
> +        while (unlikely(ht->tail != old_val))
> +            rte_pause();
> +
> +    ht->tail = new_val;
> +}
> +
> +/**
> + * @internal This function updates the producer head for enqueue
> + *
> + * @param r
> + *   A pointer to the ring structure
> + * @param is_sp
> + *   Indicates whether multi-producer path is needed or not
> + * @param n
> + *   The number of elements we will want to enqueue, i.e. how far should
> the
> + *   head be moved
> + * @param behavior
> + *   RTE_RING_QUEUE_FIXED:    Enqueue a fixed number of items from a ring
> + *   RTE_RING_QUEUE_VARIABLE: Enqueue as many items as possible from ring
> + * @param old_head
> + *   Returns head value as it was before the move, i.e. where enqueue
> starts
> + * @param new_head
> + *   Returns the current/new head value i.e. where enqueue finishes
> + * @param free_entries
> + *   Returns the amount of free space in the ring BEFORE head was moved
> + * @return
> + *   Actual number of objects enqueued.
> + *   If behavior == RTE_RING_QUEUE_FIXED, this will be 0 or n only.
> + */
> +static __rte_always_inline unsigned int
> +__rte_ring_move_prod_head(struct rte_ring *r, int is_sp,
> +        unsigned int n, enum rte_ring_queue_behavior behavior,
> +        uint32_t *old_head, uint32_t *new_head,
> +        uint32_t *free_entries)
> +{
> +    const uint32_t capacity = r->capacity;
> +    unsigned int max = n;
> +    int success;
> +
> +    do {
> +        /* Reset n to the initial burst count */
> +        n = max;
> +
> +        *old_head = r->prod.head;

adding rte_smp_rmb() no harm here as it is NOOP for x86 and it is
semantically correct too.


> +        const uint32_t cons_tail = r->cons.tail;
> +        /*
> +         *  The subtraction is done between two unsigned 32bits value
> +         * (the result is always modulo 32 bits even if we have
> +         * *old_head > cons_tail). So 'free_entries' is always between 0
> +         * and capacity (which is < size).
> +         */
> +        *free_entries = (capacity + cons_tail - *old_head);
> +
> +        /* check that we have enough room in ring */
> +        if (unlikely(n > *free_entries))
> +            n = (behavior == RTE_RING_QUEUE_FIXED) ?
> +                    0 : *free_entries;
> +
> +        if (n == 0)
> +            return 0;
> +
> +        *new_head = *old_head + n;
> +        if (is_sp)
> +            r->prod.head = *new_head, success = 1;
> +        else
> +            success = rte_atomic32_cmpset(&r->prod.head,
> +                    *old_head, *new_head);
> +    } while (unlikely(success == 0));
> +    return n;
> +}
> +
> +/**
> + * @internal Enqueue several objects on the ring
> + *
> +  * @param r
> + *   A pointer to the ring structure.
> + * @param obj_table
> + *   A pointer to a table of void * pointers (objects).
> + * @param n
> + *   The number of objects to add in the ring from the obj_table.
> + * @param behavior
> + *   RTE_RING_QUEUE_FIXED:    Enqueue a fixed number of items from a ring
> + *   RTE_RING_QUEUE_VARIABLE: Enqueue as many items as possible from ring
> + * @param is_sp
> + *   Indicates whether to use single producer or multi-producer head update
> + * @param free_space
> + *   returns the amount of space after the enqueue operation has finished
> + * @return
> + *   Actual number of objects enqueued.
> + *   If behavior == RTE_RING_QUEUE_FIXED, this will be 0 or n only.
> + */
> +static __rte_always_inline unsigned int
> +__rte_ring_do_enqueue(struct rte_ring *r, void * const *obj_table,
> +         unsigned int n, enum rte_ring_queue_behavior behavior,
> +         int is_sp, unsigned int *free_space)
> +{

Duplicate function, No need to replicate on both versions.

> +
> +/**
> + * @internal This function updates the consumer head for dequeue
> + *
> + * @param r
> + *   A pointer to the ring structure
> + * @param is_sc
> + *   Indicates whether multi-consumer path is needed or not
> + * @param n
> + *   The number of elements we will want to enqueue, i.e. how far should
> the
> + *   head be moved
> + * @param behavior
> + *   RTE_RING_QUEUE_FIXED:    Dequeue a fixed number of items from a ring
> + *   RTE_RING_QUEUE_VARIABLE: Dequeue as many items as possible from ring
> + * @param old_head
> + *   Returns head value as it was before the move, i.e. where dequeue
> starts
> + * @param new_head
> + *   Returns the current/new head value i.e. where dequeue finishes
> + * @param entries
> + *   Returns the number of entries in the ring BEFORE head was moved
> + * @return
> + *   - Actual number of objects dequeued.
> + *     If behavior == RTE_RING_QUEUE_FIXED, this will be 0 or n only.
> + */
> +static __rte_always_inline unsigned int
> +__rte_ring_move_cons_head(struct rte_ring *r, int is_sc,
> +        unsigned int n, enum rte_ring_queue_behavior behavior,
> +        uint32_t *old_head, uint32_t *new_head,
> +        uint32_t *entries)
> +{
> +    unsigned int max = n;
> +    int success;
> +
> +    /* move cons.head atomically */
> +    do {
> +        /* Restore n as it may change every loop */
> +        n = max;
> +
> +        *old_head = r->cons.head;
> +        const uint32_t prod_tail = r->prod.tail;

Same as above comment.

> +        /* The subtraction is done between two unsigned 32bits value
> +         * (the result is always modulo 32 bits even if we have
> +         * cons_head > prod_tail). So 'entries' is always between 0
> +         * and size(ring)-1. */
> +        *entries = (prod_tail - *old_head);
> +
> +        /* Set the actual entries for dequeue */
> +        if (n > *entries)
> +            n = (behavior == RTE_RING_QUEUE_FIXED) ? 0 : *entries;
> +
> +        if (unlikely(n == 0))
> +            return 0;
> +
> +        *new_head = *old_head + n;
> +        if (is_sc)
> +            r->cons.head = *new_head, success = 1;
> +        else
> +            success = rte_atomic32_cmpset(&r->cons.head, *old_head,
> +                    *new_head);
> +    } while (unlikely(success == 0));
> +    return n;
> +}
> +
> +/**
> + * @internal Dequeue several objects from the ring
> + *
> + * @param r
> + *   A pointer to the ring structure.
> + * @param obj_table
> + *   A pointer to a table of void * pointers (objects).
> + * @param n
> + *   The number of objects to pull from the ring.
> + * @param behavior
> + *   RTE_RING_QUEUE_FIXED:    Dequeue a fixed number of items from a ring
> + *   RTE_RING_QUEUE_VARIABLE: Dequeue as many items as possible from ring
> + * @param is_sc
> + *   Indicates whether to use single consumer or multi-consumer head update
> + * @param available
> + *   returns the number of remaining ring entries after the dequeue has
> finished
> + * @return
> + *   - Actual number of objects dequeued.
> + *     If behavior == RTE_RING_QUEUE_FIXED, this will be 0 or n only.
> + */
> +static __rte_always_inline unsigned int
> +__rte_ring_do_dequeue(struct rte_ring *r, void **obj_table,
> +         unsigned int n, enum rte_ring_queue_behavior behavior,
> +         int is_sc, unsigned int *available)
> +{
Duplicate function, No need to replicate on both versions.
> 


More information about the dev mailing list