clarify purpose of empty cache lines
Checks
Commit Message
This patch introduces the generic RTE_CACHE_GUARD macro into the EAL, and
replaces vaguely described empty cache lines in the rte_ring structure
with this macro.
Although the implementation of the rte_ring structure assumes that the
hardware speculatively prefetches 1 cache line, this number can be changed
at build time by modifying RTE_CACHE_GUARD_LINES in rte_config.h.
The background and the RFC was discussed in this thread:
http://inbox.dpdk.org/dev/98CBD80474FA8B44BF855DF32C47DC35D87B39@smartserver.smartshare.dk/
Signed-off-by: Morten Brørup <mb@smartsharesystems.com>
---
config/rte_config.h | 1 +
lib/eal/include/rte_common.h | 13 +++++++++++++
lib/ring/rte_ring_core.h | 6 +++---
3 files changed, 17 insertions(+), 3 deletions(-)
Comments
On Mon, Sep 04, 2023 at 10:43:49AM +0200, Morten Brørup wrote:
> This patch introduces the generic RTE_CACHE_GUARD macro into the EAL, and
> replaces vaguely described empty cache lines in the rte_ring structure
> with this macro.
>
> Although the implementation of the rte_ring structure assumes that the
> hardware speculatively prefetches 1 cache line, this number can be changed
> at build time by modifying RTE_CACHE_GUARD_LINES in rte_config.h.
>
> The background and the RFC was discussed in this thread:
> http://inbox.dpdk.org/dev/98CBD80474FA8B44BF855DF32C47DC35D87B39@smartserver.smartshare.dk/
>
> Signed-off-by: Morten Brørup <mb@smartsharesystems.com>
> ---
Seems fine to me.
Acked-by: Bruce Richardson <bruce.richardson@intel.com>
> From: Morten Brørup [mailto:mb@smartsharesystems.com]
> Sent: Monday, 4 September 2023 10.44
>
> This patch introduces the generic RTE_CACHE_GUARD macro into the EAL, and
> replaces vaguely described empty cache lines in the rte_ring structure
> with this macro.
>
> Although the implementation of the rte_ring structure assumes that the
> hardware speculatively prefetches 1 cache line, this number can be changed
> at build time by modifying RTE_CACHE_GUARD_LINES in rte_config.h.
>
> The background and the RFC was discussed in this thread:
> http://inbox.dpdk.org/dev/98CBD80474FA8B44BF855DF32C47DC35D87B39@smartserver.s
> martshare.dk/
>
> Signed-off-by: Morten Brørup <mb@smartsharesystems.com>
> ---
> config/rte_config.h | 1 +
> lib/eal/include/rte_common.h | 13 +++++++++++++
> lib/ring/rte_ring_core.h | 6 +++---
> 3 files changed, 17 insertions(+), 3 deletions(-)
>
> diff --git a/config/rte_config.h b/config/rte_config.h
> index 400e44e3cf..cfdf787724 100644
> --- a/config/rte_config.h
> +++ b/config/rte_config.h
> @@ -37,6 +37,7 @@
> #define RTE_MAX_TAILQ 32
> #define RTE_LOG_DP_LEVEL RTE_LOG_INFO
> #define RTE_MAX_VFIO_CONTAINERS 64
> +#define RTE_CACHE_GUARD_LINES 1
>
> /* bsd module defines */
> #define RTE_CONTIGMEM_MAX_NUM_BUFS 64
> diff --git a/lib/eal/include/rte_common.h b/lib/eal/include/rte_common.h
> index 771c70f2c8..daf1866a32 100644
> --- a/lib/eal/include/rte_common.h
> +++ b/lib/eal/include/rte_common.h
> @@ -527,6 +527,19 @@ rte_is_aligned(const void * const __rte_restrict ptr,
> const unsigned int align)
> /** Force minimum cache line alignment. */
> #define __rte_cache_min_aligned __rte_aligned(RTE_CACHE_LINE_MIN_SIZE)
>
> +#define _RTE_CACHE_GUARD_HELPER2(unique) \
> + char cache_guard_ ## unique[RTE_CACHE_LINE_SIZE *
> RTE_CACHE_GUARD_LINES] \
> + __rte_cache_aligned
> +#define _RTE_CACHE_GUARD_HELPER1(unique) _RTE_CACHE_GUARD_HELPER2(unique)
> +/**
> + * Empty cache lines, to guard against false sharing-like effects
> + * on systems with a next-N-lines hardware prefetcher.
> + *
> + * Use as spacing between data accessed by different lcores,
> + * to prevent cache thrashing on hardware with speculative prefetching.
> + */
> +#define RTE_CACHE_GUARD _RTE_CACHE_GUARD_HELPER1(__COUNTER__)
> +
> /*********** PA/IOVA type definitions ********/
>
> /** Physical address */
> diff --git a/lib/ring/rte_ring_core.h b/lib/ring/rte_ring_core.h
> index d1e59bf9ad..327fdcf28f 100644
> --- a/lib/ring/rte_ring_core.h
> +++ b/lib/ring/rte_ring_core.h
> @@ -126,7 +126,7 @@ struct rte_ring {
> uint32_t mask; /**< Mask (size-1) of ring. */
> uint32_t capacity; /**< Usable size of ring */
>
> - char pad0 __rte_cache_aligned; /**< empty cache line */
> + RTE_CACHE_GUARD;
>
> /** Ring producer status. */
> union {
> @@ -135,7 +135,7 @@ struct rte_ring {
> struct rte_ring_rts_headtail rts_prod;
> } __rte_cache_aligned;
>
> - char pad1 __rte_cache_aligned; /**< empty cache line */
> + RTE_CACHE_GUARD;
>
> /** Ring consumer status. */
> union {
> @@ -144,7 +144,7 @@ struct rte_ring {
> struct rte_ring_rts_headtail rts_cons;
> } __rte_cache_aligned;
>
> - char pad2 __rte_cache_aligned; /**< empty cache line */
> + RTE_CACHE_GUARD;
> };
>
> #define RING_F_SP_ENQ 0x0001 /**< The default enqueue is "single-producer".
> */
> --
> 2.17.1
I asked Mattias Rönnblom in the RFC discussion thread to ACK this patch, which he then did [1], so I'm copying his ACK here for patchwork to see.
[1]: http://inbox.dpdk.org/dev/66b862bc-ca22-02da-069c-a82c2d52ed49@lysator.liu.se/
Acked-by: Mattias Rönnblom <mattias.ronnblom@ericsson.com>
04/09/2023 11:12, Bruce Richardson:
> On Mon, Sep 04, 2023 at 10:43:49AM +0200, Morten Brørup wrote:
> > This patch introduces the generic RTE_CACHE_GUARD macro into the EAL, and
> > replaces vaguely described empty cache lines in the rte_ring structure
> > with this macro.
> >
> > Although the implementation of the rte_ring structure assumes that the
> > hardware speculatively prefetches 1 cache line, this number can be changed
> > at build time by modifying RTE_CACHE_GUARD_LINES in rte_config.h.
> >
> > The background and the RFC was discussed in this thread:
> > http://inbox.dpdk.org/dev/98CBD80474FA8B44BF855DF32C47DC35D87B39@smartserver.smartshare.dk/
> >
> > Signed-off-by: Morten Brørup <mb@smartsharesystems.com>
> > ---
> Seems fine to me.
>
> Acked-by: Bruce Richardson <bruce.richardson@intel.com>
Applied, thanks.
04/09/2023 10:43, Morten Brørup:
> /** Force minimum cache line alignment. */
> #define __rte_cache_min_aligned __rte_aligned(RTE_CACHE_LINE_MIN_SIZE)
>
> +#define _RTE_CACHE_GUARD_HELPER2(unique) \
> + char cache_guard_ ## unique[RTE_CACHE_LINE_SIZE * RTE_CACHE_GUARD_LINES] \
> + __rte_cache_aligned
> +#define _RTE_CACHE_GUARD_HELPER1(unique) _RTE_CACHE_GUARD_HELPER2(unique)
What is the reason for this intermediate helper macro?
> +/**
> + * Empty cache lines, to guard against false sharing-like effects
> + * on systems with a next-N-lines hardware prefetcher.
> + *
> + * Use as spacing between data accessed by different lcores,
> + * to prevent cache thrashing on hardware with speculative prefetching.
> + */
> +#define RTE_CACHE_GUARD _RTE_CACHE_GUARD_HELPER1(__COUNTER__)
> From: Thomas Monjalon [mailto:thomas@monjalon.net]
> Sent: Thursday, 12 October 2023 01.44
>
> 04/09/2023 10:43, Morten Brørup:
> > /** Force minimum cache line alignment. */
> > #define __rte_cache_min_aligned __rte_aligned(RTE_CACHE_LINE_MIN_SIZE)
> >
> > +#define _RTE_CACHE_GUARD_HELPER2(unique) \
> > + char cache_guard_ ## unique[RTE_CACHE_LINE_SIZE *
> RTE_CACHE_GUARD_LINES] \
> > + __rte_cache_aligned
> > +#define _RTE_CACHE_GUARD_HELPER1(unique) _RTE_CACHE_GUARD_HELPER2(unique)
>
> What is the reason for this intermediate helper macro?
>
> > +/**
> > + * Empty cache lines, to guard against false sharing-like effects
> > + * on systems with a next-N-lines hardware prefetcher.
> > + *
> > + * Use as spacing between data accessed by different lcores,
> > + * to prevent cache thrashing on hardware with speculative prefetching.
> > + */
> > +#define RTE_CACHE_GUARD _RTE_CACHE_GUARD_HELPER1(__COUNTER__)
HELPER1 is required to convert __COUNTER__ to a number before HELPER2 concatenates it to cache_guard_.
If using HELPER2 directly,
#define RTE_CACHE_GUARD _RTE_CACHE_GUARD_HELPER2(__COUNTER__)
would expand to:
char cache_guard_ ## __COUNTER__ [RTE_CACHE_LINE_SIZE * RTE_CACHE_GUARD_LINES]
__rte_cache_aligned
Which is not unique, and would prevent using RTE_CACHE_GUARD multiple times in the same structure.
> From: Thomas Monjalon [mailto:thomas@monjalon.net]
> Sent: Thursday, 12 October 2023 01.44
>
> 04/09/2023 10:43, Morten Brørup:
> > /** Force minimum cache line alignment. */
> > #define __rte_cache_min_aligned __rte_aligned(RTE_CACHE_LINE_MIN_SIZE)
> >
> > +#define _RTE_CACHE_GUARD_HELPER2(unique) \
> > + char cache_guard_ ## unique[RTE_CACHE_LINE_SIZE *
> RTE_CACHE_GUARD_LINES] \
> > + __rte_cache_aligned
> > +#define _RTE_CACHE_GUARD_HELPER1(unique) _RTE_CACHE_GUARD_HELPER2(unique)
>
> What is the reason for this intermediate helper macro?
>
> > +/**
> > + * Empty cache lines, to guard against false sharing-like effects
> > + * on systems with a next-N-lines hardware prefetcher.
> > + *
> > + * Use as spacing between data accessed by different lcores,
> > + * to prevent cache thrashing on hardware with speculative prefetching.
> > + */
> > +#define RTE_CACHE_GUARD _RTE_CACHE_GUARD_HELPER1(__COUNTER__)
HELPER1 is required to convert __COUNTER__ to a number before HELPER2 concatenates it to cache_guard_.
If using HELPER2 directly,
#define RTE_CACHE_GUARD _RTE_CACHE_GUARD_HELPER2(__COUNTER__)
would expand to:
char cache_guard___COUNTER__[RTE_CACHE_LINE_SIZE * RTE_CACHE_GUARD_LINES]
__rte_cache_aligned
Which is not unique, and would prevent using RTE_CACHE_GUARD multiple times in the same structure.
@@ -37,6 +37,7 @@
#define RTE_MAX_TAILQ 32
#define RTE_LOG_DP_LEVEL RTE_LOG_INFO
#define RTE_MAX_VFIO_CONTAINERS 64
+#define RTE_CACHE_GUARD_LINES 1
/* bsd module defines */
#define RTE_CONTIGMEM_MAX_NUM_BUFS 64
@@ -527,6 +527,19 @@ rte_is_aligned(const void * const __rte_restrict ptr, const unsigned int align)
/** Force minimum cache line alignment. */
#define __rte_cache_min_aligned __rte_aligned(RTE_CACHE_LINE_MIN_SIZE)
+#define _RTE_CACHE_GUARD_HELPER2(unique) \
+ char cache_guard_ ## unique[RTE_CACHE_LINE_SIZE * RTE_CACHE_GUARD_LINES] \
+ __rte_cache_aligned
+#define _RTE_CACHE_GUARD_HELPER1(unique) _RTE_CACHE_GUARD_HELPER2(unique)
+/**
+ * Empty cache lines, to guard against false sharing-like effects
+ * on systems with a next-N-lines hardware prefetcher.
+ *
+ * Use as spacing between data accessed by different lcores,
+ * to prevent cache thrashing on hardware with speculative prefetching.
+ */
+#define RTE_CACHE_GUARD _RTE_CACHE_GUARD_HELPER1(__COUNTER__)
+
/*********** PA/IOVA type definitions ********/
/** Physical address */
@@ -126,7 +126,7 @@ struct rte_ring {
uint32_t mask; /**< Mask (size-1) of ring. */
uint32_t capacity; /**< Usable size of ring */
- char pad0 __rte_cache_aligned; /**< empty cache line */
+ RTE_CACHE_GUARD;
/** Ring producer status. */
union {
@@ -135,7 +135,7 @@ struct rte_ring {
struct rte_ring_rts_headtail rts_prod;
} __rte_cache_aligned;
- char pad1 __rte_cache_aligned; /**< empty cache line */
+ RTE_CACHE_GUARD;
/** Ring consumer status. */
union {
@@ -144,7 +144,7 @@ struct rte_ring {
struct rte_ring_rts_headtail rts_cons;
} __rte_cache_aligned;
- char pad2 __rte_cache_aligned; /**< empty cache line */
+ RTE_CACHE_GUARD;
};
#define RING_F_SP_ENQ 0x0001 /**< The default enqueue is "single-producer". */