vlib: don't use vector for keeping buffer indices in the pool

Type: refactor

Change-Id: I72221b97d7e0bf5c93e20bbda4473ca67bfcdeb4
Signed-off-by: Damjan Marion <damarion@cisco.com>
This commit is contained in:
Damjan Marion
2019-03-12 18:14:15 +01:00
committed by Dave Barach
parent 38f7ccbce0
commit b6e8b1a7c8
4 changed files with 64 additions and 46 deletions

View File

@ -40,7 +40,7 @@ dpdk_buffer_pool_init (vlib_main_t * vm, vlib_buffer_pool_t * bp)
struct rte_mempool *mp, *nmp; struct rte_mempool *mp, *nmp;
struct rte_pktmbuf_pool_private priv; struct rte_pktmbuf_pool_private priv;
enum rte_iova_mode iova_mode; enum rte_iova_mode iova_mode;
u32 *bi; u32 i;
u8 *name = 0; u8 *name = 0;
u32 elt_size = u32 elt_size =
@ -54,7 +54,7 @@ dpdk_buffer_pool_init (vlib_main_t * vm, vlib_buffer_pool_t * bp)
/* normal mempool */ /* normal mempool */
name = format (name, "vpp pool %u%c", bp->index, 0); name = format (name, "vpp pool %u%c", bp->index, 0);
mp = rte_mempool_create_empty ((char *) name, vec_len (bp->buffers), mp = rte_mempool_create_empty ((char *) name, bp->n_buffers,
elt_size, 512, sizeof (priv), elt_size, 512, sizeof (priv),
bp->numa_node, 0); bp->numa_node, 0);
if (!mp) if (!mp)
@ -68,7 +68,7 @@ dpdk_buffer_pool_init (vlib_main_t * vm, vlib_buffer_pool_t * bp)
/* non-cached mempool */ /* non-cached mempool */
name = format (name, "vpp pool %u (no cache)%c", bp->index, 0); name = format (name, "vpp pool %u (no cache)%c", bp->index, 0);
nmp = rte_mempool_create_empty ((char *) name, vec_len (bp->buffers), nmp = rte_mempool_create_empty ((char *) name, bp->n_buffers,
elt_size, 0, sizeof (priv), elt_size, 0, sizeof (priv),
bp->numa_node, 0); bp->numa_node, 0);
if (!nmp) if (!nmp)
@ -99,11 +99,10 @@ dpdk_buffer_pool_init (vlib_main_t * vm, vlib_buffer_pool_t * bp)
iova_mode = rte_eal_iova_mode (); iova_mode = rte_eal_iova_mode ();
/* populate mempool object buffer header */ /* populate mempool object buffer header */
/* *INDENT-OFF* */ for (i = 0; i < bp->n_buffers; i++)
vec_foreach (bi, bp->buffers)
{ {
struct rte_mempool_objhdr *hdr; struct rte_mempool_objhdr *hdr;
vlib_buffer_t *b = vlib_get_buffer (vm, *bi); vlib_buffer_t *b = vlib_get_buffer (vm, bp->buffers[i]);
struct rte_mbuf *mb = rte_mbuf_from_vlib_buffer (b); struct rte_mbuf *mb = rte_mbuf_from_vlib_buffer (b);
hdr = (struct rte_mempool_objhdr *) RTE_PTR_SUB (mb, sizeof (*hdr)); hdr = (struct rte_mempool_objhdr *) RTE_PTR_SUB (mb, sizeof (*hdr));
hdr->mp = mp; hdr->mp = mp;
@ -114,7 +113,6 @@ dpdk_buffer_pool_init (vlib_main_t * vm, vlib_buffer_pool_t * bp)
mp->populated_size++; mp->populated_size++;
nmp->populated_size++; nmp->populated_size++;
} }
/* *INDENT-ON* */
/* call the object initializers */ /* call the object initializers */
rte_mempool_obj_iter (mp, rte_pktmbuf_init, 0); rte_mempool_obj_iter (mp, rte_pktmbuf_init, 0);
@ -127,14 +125,12 @@ dpdk_buffer_pool_init (vlib_main_t * vm, vlib_buffer_pool_t * bp)
(buffer_mem_start, *bp->buffers, (buffer_mem_start, *bp->buffers,
0)), sizeof (struct rte_mbuf)); 0)), sizeof (struct rte_mbuf));
/* *INDENT-OFF* */ for (i = 0; i < bp->n_buffers; i++)
vec_foreach (bi, bp->buffers)
{ {
vlib_buffer_t *b; vlib_buffer_t *b;
b = vlib_buffer_ptr_from_index (buffer_mem_start, *bi, 0); b = vlib_buffer_ptr_from_index (buffer_mem_start, bp->buffers[i], 0);
vlib_buffer_copy_template (b, &bp->buffer_template); vlib_buffer_copy_template (b, &bp->buffer_template);
} }
/* *INDENT-ON* */
/* map DMA pages if at least one physical device exists */ /* map DMA pages if at least one physical device exists */
if (rte_eth_dev_count_avail ()) if (rte_eth_dev_count_avail ())

View File

@ -552,9 +552,9 @@ vlib_buffer_pool_create (vlib_main_t * vm, char *name, u32 data_size,
n_alloc_per_page = (1ULL << m->log2_page_size) / alloc_size; n_alloc_per_page = (1ULL << m->log2_page_size) / alloc_size;
/* preallocate buffer indices memory */ /* preallocate buffer indices memory */
vec_validate_aligned (bp->buffers, m->n_pages * n_alloc_per_page, bp->n_buffers = m->n_pages * n_alloc_per_page;
CLIB_CACHE_LINE_BYTES); bp->buffers = clib_mem_alloc_aligned (bp->n_buffers * sizeof (u32),
vec_reset_length (bp->buffers); CLIB_CACHE_LINE_BYTES);
clib_spinlock_init (&bp->lock); clib_spinlock_init (&bp->lock);
@ -571,11 +571,11 @@ vlib_buffer_pool_create (vlib_main_t * vm, char *name, u32 data_size,
bi = vlib_get_buffer_index (vm, (vlib_buffer_t *) p); bi = vlib_get_buffer_index (vm, (vlib_buffer_t *) p);
vec_add1_aligned (bp->buffers, bi, CLIB_CACHE_LINE_BYTES); bp->buffers[bp->n_avail++] = bi;
vlib_get_buffer (vm, bi); vlib_get_buffer (vm, bi);
} }
bp->n_buffers = vec_len (bp->buffers);
return bp->index; return bp->index;
} }
@ -594,14 +594,14 @@ format_vlib_buffer_pool (u8 * s, va_list * va)
/* *INDENT-OFF* */ /* *INDENT-OFF* */
vec_foreach (bpt, bp->threads) vec_foreach (bpt, bp->threads)
cached += vec_len (bpt->cached_buffers); cached += bpt->n_cached;
/* *INDENT-ON* */ /* *INDENT-ON* */
s = format (s, "%-20s%=6d%=6d%=6u%=11u%=6u%=8u%=8u%=8u", s = format (s, "%-20s%=6d%=6d%=6u%=11u%=6u%=8u%=8u%=8u",
bp->name, bp->index, bp->numa_node, bp->data_size + bp->name, bp->index, bp->numa_node, bp->data_size +
sizeof (vlib_buffer_t) + vm->buffer_main->ext_hdr_size, sizeof (vlib_buffer_t) + vm->buffer_main->ext_hdr_size,
bp->data_size, bp->n_buffers, vec_len (bp->buffers), cached, bp->data_size, bp->n_buffers, bp->n_avail, cached,
bp->n_buffers - vec_len (bp->buffers) - cached); bp->n_buffers - bp->n_avail - cached);
return s; return s;
} }
@ -736,7 +736,7 @@ buffer_get_cached (vlib_buffer_pool_t * bp)
/* *INDENT-OFF* */ /* *INDENT-OFF* */
vec_foreach (bpt, bp->threads) vec_foreach (bpt, bp->threads)
cached += vec_len (bpt->cached_buffers); cached += bpt->n_cached;
/* *INDENT-ON* */ /* *INDENT-ON* */
clib_spinlock_unlock (&bp->lock); clib_spinlock_unlock (&bp->lock);
@ -763,7 +763,7 @@ buffer_gauges_update_used_fn (stat_segment_directory_entry_t * e, u32 index)
if (!bp) if (!bp)
return; return;
e->value = bp->n_buffers - vec_len (bp->buffers) - buffer_get_cached (bp); e->value = bp->n_buffers - bp->n_avail - buffer_get_cached (bp);
} }
static void static void
@ -775,7 +775,7 @@ buffer_gauges_update_available_fn (stat_segment_directory_entry_t * e,
if (!bp) if (!bp)
return; return;
e->value = vec_len (bp->buffers); e->value = bp->n_avail;
} }
static void static void

View File

@ -411,12 +411,15 @@ vlib_buffer_pull (vlib_buffer_t * b, u8 size)
/* Forward declaration. */ /* Forward declaration. */
struct vlib_main_t; struct vlib_main_t;
#define VLIB_BUFFER_POOL_PER_THREAD_CACHE_SZ 512
typedef struct typedef struct
{ {
CLIB_CACHE_LINE_ALIGN_MARK (cacheline0); CLIB_CACHE_LINE_ALIGN_MARK (cacheline0);
u32 *cached_buffers; u32 cached_buffers[VLIB_BUFFER_POOL_PER_THREAD_CACHE_SZ];
u32 n_alloc; u32 n_cached;
} vlib_buffer_pool_thread_t; } vlib_buffer_pool_thread_t;
typedef struct typedef struct
{ {
CLIB_CACHE_LINE_ALIGN_MARK (cacheline0); CLIB_CACHE_LINE_ALIGN_MARK (cacheline0);
@ -428,6 +431,7 @@ typedef struct
u32 physmem_map_index; u32 physmem_map_index;
u32 data_size; u32 data_size;
u32 n_buffers; u32 n_buffers;
u32 n_avail;
u32 *buffers; u32 *buffers;
u8 *name; u8 *name;
clib_spinlock_t lock; clib_spinlock_t lock;

View File

@ -491,19 +491,19 @@ vlib_buffer_pool_get (vlib_main_t * vm, u8 buffer_pool_index, u32 * buffers,
ASSERT (bp->buffers); ASSERT (bp->buffers);
clib_spinlock_lock (&bp->lock); clib_spinlock_lock (&bp->lock);
len = vec_len (bp->buffers); len = bp->n_avail;
if (PREDICT_TRUE (n_buffers < len)) if (PREDICT_TRUE (n_buffers < len))
{ {
len -= n_buffers; len -= n_buffers;
vlib_buffer_copy_indices (buffers, bp->buffers + len, n_buffers); vlib_buffer_copy_indices (buffers, bp->buffers + len, n_buffers);
_vec_len (bp->buffers) = len; bp->n_avail = len;
clib_spinlock_unlock (&bp->lock); clib_spinlock_unlock (&bp->lock);
return n_buffers; return n_buffers;
} }
else else
{ {
vlib_buffer_copy_indices (buffers, bp->buffers, len); vlib_buffer_copy_indices (buffers, bp->buffers, len);
_vec_len (bp->buffers) = 0; bp->n_avail = 0;
clib_spinlock_unlock (&bp->lock); clib_spinlock_unlock (&bp->lock);
return len; return len;
} }
@ -533,14 +533,26 @@ vlib_buffer_alloc_from_pool (vlib_main_t * vm, u32 * buffers, u32 n_buffers,
dst = buffers; dst = buffers;
n_left = n_buffers; n_left = n_buffers;
len = vec_len (bpt->cached_buffers); len = bpt->n_cached;
/* per-thread cache contains enough buffers */ /* per-thread cache contains enough buffers */
if (len >= n_buffers) if (len >= n_buffers)
{ {
src = bpt->cached_buffers + len - n_buffers; src = bpt->cached_buffers + len - n_buffers;
vlib_buffer_copy_indices (dst, src, n_buffers); vlib_buffer_copy_indices (dst, src, n_buffers);
_vec_len (bpt->cached_buffers) -= n_buffers; bpt->n_cached -= n_buffers;
if (CLIB_DEBUG > 0)
vlib_buffer_validate_alloc_free (vm, buffers, n_buffers,
VLIB_BUFFER_KNOWN_FREE);
return n_buffers;
}
/* alloc bigger than cache - take buffers directly from main pool */
if (n_buffers >= VLIB_BUFFER_POOL_PER_THREAD_CACHE_SZ)
{
n_buffers = vlib_buffer_pool_get (vm, buffer_pool_index, buffers,
n_buffers);
if (CLIB_DEBUG > 0) if (CLIB_DEBUG > 0)
vlib_buffer_validate_alloc_free (vm, buffers, n_buffers, vlib_buffer_validate_alloc_free (vm, buffers, n_buffers,
@ -552,23 +564,22 @@ vlib_buffer_alloc_from_pool (vlib_main_t * vm, u32 * buffers, u32 n_buffers,
if (len) if (len)
{ {
vlib_buffer_copy_indices (dst, bpt->cached_buffers, len); vlib_buffer_copy_indices (dst, bpt->cached_buffers, len);
_vec_len (bpt->cached_buffers) = 0; bpt->n_cached = 0;
dst += len; dst += len;
n_left -= len; n_left -= len;
} }
len = round_pow2 (n_left, 32); len = round_pow2 (n_left, 32);
vec_validate_aligned (bpt->cached_buffers, len - 1, CLIB_CACHE_LINE_BYTES);
len = vlib_buffer_pool_get (vm, buffer_pool_index, bpt->cached_buffers, len = vlib_buffer_pool_get (vm, buffer_pool_index, bpt->cached_buffers,
len); len);
_vec_len (bpt->cached_buffers) = len; bpt->n_cached = len;
if (len) if (len)
{ {
u32 n_copy = clib_min (len, n_left); u32 n_copy = clib_min (len, n_left);
src = bpt->cached_buffers + len - n_copy; src = bpt->cached_buffers + len - n_copy;
vlib_buffer_copy_indices (dst, src, n_copy); vlib_buffer_copy_indices (dst, src, n_copy);
_vec_len (bpt->cached_buffers) -= n_copy; bpt->n_cached -= n_copy;
n_left -= n_copy; n_left -= n_copy;
} }
@ -681,26 +692,33 @@ vlib_buffer_pool_put (vlib_main_t * vm, u8 buffer_pool_index,
u32 * buffers, u32 n_buffers) u32 * buffers, u32 n_buffers)
{ {
vlib_buffer_pool_t *bp = vlib_get_buffer_pool (vm, buffer_pool_index); vlib_buffer_pool_t *bp = vlib_get_buffer_pool (vm, buffer_pool_index);
vlib_buffer_pool_thread_t *bpt = vlib_buffer_pool_thread_t *bpt = vec_elt_at_index (bp->threads,
vec_elt_at_index (bp->threads, vm->thread_index); vm->thread_index);
u32 n_cached, n_empty;
if (CLIB_DEBUG > 0) if (CLIB_DEBUG > 0)
vlib_buffer_validate_alloc_free (vm, buffers, n_buffers, vlib_buffer_validate_alloc_free (vm, buffers, n_buffers,
VLIB_BUFFER_KNOWN_ALLOCATED); VLIB_BUFFER_KNOWN_ALLOCATED);
vec_add_aligned (bpt->cached_buffers, buffers, n_buffers, n_cached = bpt->n_cached;
CLIB_CACHE_LINE_BYTES); n_empty = VLIB_BUFFER_POOL_PER_THREAD_CACHE_SZ - n_cached;
if (n_buffers <= n_empty)
if (vec_len (bpt->cached_buffers) > 4 * VLIB_FRAME_SIZE)
{ {
clib_spinlock_lock (&bp->lock); vlib_buffer_copy_indices (bpt->cached_buffers + n_cached,
/* keep last stored buffers, as they are more likely hot in the cache */ buffers, n_buffers);
vec_add_aligned (bp->buffers, bpt->cached_buffers, VLIB_FRAME_SIZE, bpt->n_cached = n_cached + n_buffers;
CLIB_CACHE_LINE_BYTES); return;
vec_delete (bpt->cached_buffers, VLIB_FRAME_SIZE, 0);
bpt->n_alloc -= VLIB_FRAME_SIZE;
clib_spinlock_unlock (&bp->lock);
} }
vlib_buffer_copy_indices (bpt->cached_buffers + n_cached,
buffers + n_buffers - n_empty, n_empty);
bpt->n_cached = VLIB_BUFFER_POOL_PER_THREAD_CACHE_SZ;
clib_spinlock_lock (&bp->lock);
vlib_buffer_copy_indices (bp->buffers + bp->n_avail, buffers,
n_buffers - n_empty);
bp->n_avail += n_buffers - n_empty;
clib_spinlock_unlock (&bp->lock);
} }
static_always_inline void static_always_inline void