freebsd-skq/sys/dev/cxgbe/t4_mp_ring.c

370 lines
9.1 KiB
C
Raw Normal View History

cxgbe(4): major tx rework. a) Front load as much work as possible in if_transmit, before any driver lock or software queue has to get involved. b) Replace buf_ring with a brand new mp_ring (multiproducer ring). This is specifically for the tx multiqueue model where one of the if_transmit producer threads becomes the consumer and other producers carry on as usual. mp_ring is implemented as standalone code and it should be possible to use it in any driver with tx multiqueue. It also has: - the ability to enqueue/dequeue multiple items. This might become significant if packet batching is ever implemented. - an abdication mechanism to allow a thread to give up writing tx descriptors and have another if_transmit thread take over. A thread that's writing tx descriptors can end up doing so for an unbounded time period if a) there are other if_transmit threads continuously feeding the sofware queue, and b) the chip keeps up with whatever the thread is throwing at it. - accurate statistics about interesting events even when the stats come at the expense of additional branches/conditional code. The NIC txq lock is uncontested on the fast path at this point. I've left it there for synchronization with the control events (interface up/down, modload/unload). c) Add support for "type 1" coalescing work request in the normal NIC tx path. This work request is optimized for frames with a single item in the DMA gather list. These are very common when forwarding packets. Note that netmap tx in cxgbe already uses these "type 1" work requests. d) Do not request automatic cidx updates every 32 descriptors. Instead, request updates via bits in individual work requests (still every 32 descriptors approximately). Also, request an automatic final update when the queue idles after activity. This means NIC tx reclaim is still performed lazily but it will catch up quickly as soon as the queue idles. This seems to be the best middle ground and I'll probably do something similar for netmap tx as well. e) Implement a faster tx path for WRQs (used by TOE tx and control queues, _not_ by the normal NIC tx). Allow work requests to be written directly to the hardware descriptor ring if room is available. I will convert t4_tom and iw_cxgbe modules to this faster style gradually. MFC after: 2 months
2014-12-31 23:19:16 +00:00
/*-
* Copyright (c) 2014 Chelsio Communications, Inc.
* All rights reserved.
* Written by: Navdeep Parhar <np@FreeBSD.org>
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions
* are met:
* 1. Redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer.
* 2. 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.
*
* THIS SOFTWARE IS PROVIDED BY THE AUTHOR 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 AUTHOR 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.
*/
#include <sys/cdefs.h>
__FBSDID("$FreeBSD$");
#include <sys/types.h>
#include <sys/param.h>
#include <sys/systm.h>
#include <sys/counter.h>
#include <sys/lock.h>
#include <sys/malloc.h>
#include <machine/cpu.h>
#include "t4_mp_ring.h"
#if defined(__i386__)
#define atomic_cmpset_acq_64 atomic_cmpset_64
#define atomic_cmpset_rel_64 atomic_cmpset_64
#endif
cxgbe(4): major tx rework. a) Front load as much work as possible in if_transmit, before any driver lock or software queue has to get involved. b) Replace buf_ring with a brand new mp_ring (multiproducer ring). This is specifically for the tx multiqueue model where one of the if_transmit producer threads becomes the consumer and other producers carry on as usual. mp_ring is implemented as standalone code and it should be possible to use it in any driver with tx multiqueue. It also has: - the ability to enqueue/dequeue multiple items. This might become significant if packet batching is ever implemented. - an abdication mechanism to allow a thread to give up writing tx descriptors and have another if_transmit thread take over. A thread that's writing tx descriptors can end up doing so for an unbounded time period if a) there are other if_transmit threads continuously feeding the sofware queue, and b) the chip keeps up with whatever the thread is throwing at it. - accurate statistics about interesting events even when the stats come at the expense of additional branches/conditional code. The NIC txq lock is uncontested on the fast path at this point. I've left it there for synchronization with the control events (interface up/down, modload/unload). c) Add support for "type 1" coalescing work request in the normal NIC tx path. This work request is optimized for frames with a single item in the DMA gather list. These are very common when forwarding packets. Note that netmap tx in cxgbe already uses these "type 1" work requests. d) Do not request automatic cidx updates every 32 descriptors. Instead, request updates via bits in individual work requests (still every 32 descriptors approximately). Also, request an automatic final update when the queue idles after activity. This means NIC tx reclaim is still performed lazily but it will catch up quickly as soon as the queue idles. This seems to be the best middle ground and I'll probably do something similar for netmap tx as well. e) Implement a faster tx path for WRQs (used by TOE tx and control queues, _not_ by the normal NIC tx). Allow work requests to be written directly to the hardware descriptor ring if room is available. I will convert t4_tom and iw_cxgbe modules to this faster style gradually. MFC after: 2 months
2014-12-31 23:19:16 +00:00
union ring_state {
struct {
uint16_t pidx_head;
uint16_t pidx_tail;
uint16_t cidx;
uint16_t flags;
};
uint64_t state;
};
enum {
IDLE = 0, /* consumer ran to completion, nothing more to do. */
BUSY, /* consumer is running already, or will be shortly. */
STALLED, /* consumer stopped due to lack of resources. */
ABDICATED, /* consumer stopped even though there was work to be
done because it wants another thread to take over. */
};
static inline uint16_t
space_available(struct mp_ring *r, union ring_state s)
{
uint16_t x = r->size - 1;
if (s.cidx == s.pidx_head)
return (x);
else if (s.cidx > s.pidx_head)
return (s.cidx - s.pidx_head - 1);
else
return (x - s.pidx_head + s.cidx);
}
static inline uint16_t
increment_idx(struct mp_ring *r, uint16_t idx, uint16_t n)
{
int x = r->size - idx;
MPASS(x > 0);
return (x > n ? idx + n : n - x);
}
/* Consumer is about to update the ring's state to s */
static inline uint16_t
state_to_flags(union ring_state s, int abdicate)
{
if (s.cidx == s.pidx_tail)
return (IDLE);
else if (abdicate && s.pidx_tail != s.pidx_head)
return (ABDICATED);
return (BUSY);
}
/*
* Caller passes in a state, with a guarantee that there is work to do and that
* all items up to the pidx_tail in the state are visible.
*/
static void
drain_ring(struct mp_ring *r, union ring_state os, uint16_t prev, int budget)
{
union ring_state ns;
int n, pending, total;
uint16_t cidx = os.cidx;
uint16_t pidx = os.pidx_tail;
MPASS(os.flags == BUSY);
MPASS(cidx != pidx);
if (prev == IDLE)
counter_u64_add(r->starts, 1);
pending = 0;
total = 0;
while (cidx != pidx) {
/* Items from cidx to pidx are available for consumption. */
n = r->drain(r, cidx, pidx);
if (n == 0) {
critical_enter();
do {
os.state = ns.state = r->state;
ns.cidx = cidx;
ns.flags = STALLED;
} while (atomic_cmpset_64(&r->state, os.state,
ns.state) == 0);
critical_exit();
if (prev != STALLED)
counter_u64_add(r->stalls, 1);
else if (total > 0) {
counter_u64_add(r->restarts, 1);
counter_u64_add(r->stalls, 1);
}
break;
}
cidx = increment_idx(r, cidx, n);
pending += n;
total += n;
/*
* We update the cidx only if we've caught up with the pidx, the
* real cidx is getting too far ahead of the one visible to
* everyone else, or we have exceeded our budget.
*/
if (cidx != pidx && pending < 64 && total < budget)
continue;
critical_enter();
do {
os.state = ns.state = r->state;
ns.cidx = cidx;
ns.flags = state_to_flags(ns, total >= budget);
} while (atomic_cmpset_acq_64(&r->state, os.state, ns.state) == 0);
critical_exit();
if (ns.flags == ABDICATED)
counter_u64_add(r->abdications, 1);
if (ns.flags != BUSY) {
/* Wrong loop exit if we're going to stall. */
MPASS(ns.flags != STALLED);
if (prev == STALLED) {
MPASS(total > 0);
counter_u64_add(r->restarts, 1);
}
break;
}
/*
* The acquire style atomic above guarantees visibility of items
* associated with any pidx change that we notice here.
*/
pidx = ns.pidx_tail;
pending = 0;
}
}
int
mp_ring_alloc(struct mp_ring **pr, int size, void *cookie, ring_drain_t drain,
ring_can_drain_t can_drain, struct malloc_type *mt, int flags)
{
struct mp_ring *r;
/* All idx are 16b so size can be 65536 at most */
if (pr == NULL || size < 2 || size > 65536 || drain == NULL ||
can_drain == NULL)
return (EINVAL);
*pr = NULL;
flags &= M_NOWAIT | M_WAITOK;
MPASS(flags != 0);
r = malloc(__offsetof(struct mp_ring, items[size]), mt, flags | M_ZERO);
if (r == NULL)
return (ENOMEM);
r->size = size;
r->cookie = cookie;
r->mt = mt;
r->drain = drain;
r->can_drain = can_drain;
r->enqueues = counter_u64_alloc(flags);
r->drops = counter_u64_alloc(flags);
r->starts = counter_u64_alloc(flags);
r->stalls = counter_u64_alloc(flags);
r->restarts = counter_u64_alloc(flags);
r->abdications = counter_u64_alloc(flags);
if (r->enqueues == NULL || r->drops == NULL || r->starts == NULL ||
r->stalls == NULL || r->restarts == NULL ||
r->abdications == NULL) {
mp_ring_free(r);
return (ENOMEM);
}
*pr = r;
return (0);
}
void
mp_ring_free(struct mp_ring *r)
{
if (r == NULL)
return;
if (r->enqueues != NULL)
counter_u64_free(r->enqueues);
if (r->drops != NULL)
counter_u64_free(r->drops);
if (r->starts != NULL)
counter_u64_free(r->starts);
if (r->stalls != NULL)
counter_u64_free(r->stalls);
if (r->restarts != NULL)
counter_u64_free(r->restarts);
if (r->abdications != NULL)
counter_u64_free(r->abdications);
free(r, r->mt);
}
/*
* Enqueue n items and maybe drain the ring for some time.
*
* Returns an errno.
*/
int
mp_ring_enqueue(struct mp_ring *r, void **items, int n, int budget)
{
union ring_state os, ns;
uint16_t pidx_start, pidx_stop;
int i;
MPASS(items != NULL);
MPASS(n > 0);
/*
* Reserve room for the new items. Our reservation, if successful, is
* from 'pidx_start' to 'pidx_stop'.
*/
for (;;) {
os.state = r->state;
if (n >= space_available(r, os)) {
counter_u64_add(r->drops, n);
MPASS(os.flags != IDLE);
if (os.flags == STALLED)
mp_ring_check_drainage(r, 0);
return (ENOBUFS);
}
ns.state = os.state;
ns.pidx_head = increment_idx(r, os.pidx_head, n);
critical_enter();
if (atomic_cmpset_64(&r->state, os.state, ns.state))
break;
critical_exit();
cpu_spinwait();
}
pidx_start = os.pidx_head;
pidx_stop = ns.pidx_head;
/*
* Wait for other producers who got in ahead of us to enqueue their
* items, one producer at a time. It is our turn when the ring's
* pidx_tail reaches the begining of our reservation (pidx_start).
*/
while (ns.pidx_tail != pidx_start) {
cpu_spinwait();
ns.state = r->state;
}
/* Now it is our turn to fill up the area we reserved earlier. */
i = pidx_start;
do {
r->items[i] = *items++;
if (__predict_false(++i == r->size))
i = 0;
} while (i != pidx_stop);
/*
* Update the ring's pidx_tail. The release style atomic guarantees
* that the items are visible to any thread that sees the updated pidx.
*/
do {
os.state = ns.state = r->state;
ns.pidx_tail = pidx_stop;
ns.flags = BUSY;
} while (atomic_cmpset_rel_64(&r->state, os.state, ns.state) == 0);
critical_exit();
counter_u64_add(r->enqueues, n);
/*
* Turn into a consumer if some other thread isn't active as a consumer
* already.
*/
if (os.flags != BUSY)
drain_ring(r, ns, os.flags, budget);
return (0);
}
void
mp_ring_check_drainage(struct mp_ring *r, int budget)
{
union ring_state os, ns;
os.state = r->state;
if (os.flags != STALLED || os.pidx_head != os.pidx_tail ||
r->can_drain(r) == 0)
return;
MPASS(os.cidx != os.pidx_tail); /* implied by STALLED */
ns.state = os.state;
ns.flags = BUSY;
/*
* The acquire style atomic guarantees visibility of items associated
* with the pidx that we read here.
*/
if (!atomic_cmpset_acq_64(&r->state, os.state, ns.state))
return;
drain_ring(r, ns, os.flags, budget);
}
void
mp_ring_reset_stats(struct mp_ring *r)
{
counter_u64_zero(r->enqueues);
counter_u64_zero(r->drops);
counter_u64_zero(r->starts);
counter_u64_zero(r->stalls);
counter_u64_zero(r->restarts);
counter_u64_zero(r->abdications);
}
int
mp_ring_is_idle(struct mp_ring *r)
{
union ring_state s;
s.state = r->state;
if (s.pidx_head == s.pidx_tail && s.pidx_tail == s.cidx &&
s.flags == IDLE)
return (1);
return (0);
}