Index: head/sys/conf/files =================================================================== --- head/sys/conf/files +++ head/sys/conf/files @@ -3349,6 +3349,7 @@ kern/subr_eventhandler.c standard kern/subr_fattime.c standard kern/subr_firmware.c optional firmware +kern/subr_gtaskqueue.c standard kern/subr_hash.c standard kern/subr_hints.c standard kern/subr_kdb.c standard Index: head/sys/kern/subr_gtaskqueue.c =================================================================== --- head/sys/kern/subr_gtaskqueue.c +++ head/sys/kern/subr_gtaskqueue.c @@ -0,0 +1,864 @@ +/*- + * Copyright (c) 2000 Doug Rabson + * Copyright (c) 2014 Jeff Roberson + * Copyright (c) 2016 Matthew Macy + * All rights reserved. + * + * 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 +__FBSDID("$FreeBSD$"); + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + +static MALLOC_DEFINE(M_GTASKQUEUE, "taskqueue", "Task Queues"); +static void gtaskqueue_thread_enqueue(void *); +static void gtaskqueue_thread_loop(void *arg); + + +struct gtaskqueue_busy { + struct gtask *tb_running; + TAILQ_ENTRY(gtaskqueue_busy) tb_link; +}; + +static struct gtask * const TB_DRAIN_WAITER = (struct gtask *)0x1; + +struct gtaskqueue { + STAILQ_HEAD(, gtask) tq_queue; + gtaskqueue_enqueue_fn tq_enqueue; + void *tq_context; + char *tq_name; + TAILQ_HEAD(, gtaskqueue_busy) tq_active; + struct mtx tq_mutex; + struct thread **tq_threads; + int tq_tcount; + int tq_spin; + int tq_flags; + int tq_callouts; + taskqueue_callback_fn tq_callbacks[TASKQUEUE_NUM_CALLBACKS]; + void *tq_cb_contexts[TASKQUEUE_NUM_CALLBACKS]; +}; + +#define TQ_FLAGS_ACTIVE (1 << 0) +#define TQ_FLAGS_BLOCKED (1 << 1) +#define TQ_FLAGS_UNLOCKED_ENQUEUE (1 << 2) + +#define DT_CALLOUT_ARMED (1 << 0) + +#define TQ_LOCK(tq) \ + do { \ + if ((tq)->tq_spin) \ + mtx_lock_spin(&(tq)->tq_mutex); \ + else \ + mtx_lock(&(tq)->tq_mutex); \ + } while (0) +#define TQ_ASSERT_LOCKED(tq) mtx_assert(&(tq)->tq_mutex, MA_OWNED) + +#define TQ_UNLOCK(tq) \ + do { \ + if ((tq)->tq_spin) \ + mtx_unlock_spin(&(tq)->tq_mutex); \ + else \ + mtx_unlock(&(tq)->tq_mutex); \ + } while (0) +#define TQ_ASSERT_UNLOCKED(tq) mtx_assert(&(tq)->tq_mutex, MA_NOTOWNED) + +static __inline int +TQ_SLEEP(struct gtaskqueue *tq, void *p, struct mtx *m, int pri, const char *wm, + int t) +{ + if (tq->tq_spin) + return (msleep_spin(p, m, wm, t)); + return (msleep(p, m, pri, wm, t)); +} + +static struct gtaskqueue * +_gtaskqueue_create(const char *name, int mflags, + taskqueue_enqueue_fn enqueue, void *context, + int mtxflags, const char *mtxname __unused) +{ + struct gtaskqueue *queue; + char *tq_name; + + tq_name = malloc(TASKQUEUE_NAMELEN, M_GTASKQUEUE, mflags | M_ZERO); + if (!tq_name) + return (NULL); + + snprintf(tq_name, TASKQUEUE_NAMELEN, "%s", (name) ? name : "taskqueue"); + + queue = malloc(sizeof(struct gtaskqueue), M_GTASKQUEUE, mflags | M_ZERO); + if (!queue) + return (NULL); + + STAILQ_INIT(&queue->tq_queue); + TAILQ_INIT(&queue->tq_active); + queue->tq_enqueue = enqueue; + queue->tq_context = context; + queue->tq_name = tq_name; + queue->tq_spin = (mtxflags & MTX_SPIN) != 0; + queue->tq_flags |= TQ_FLAGS_ACTIVE; + if (enqueue == gtaskqueue_thread_enqueue) + queue->tq_flags |= TQ_FLAGS_UNLOCKED_ENQUEUE; + mtx_init(&queue->tq_mutex, tq_name, NULL, mtxflags); + + return (queue); +} + + +/* + * Signal a taskqueue thread to terminate. + */ +static void +gtaskqueue_terminate(struct thread **pp, struct gtaskqueue *tq) +{ + + while (tq->tq_tcount > 0 || tq->tq_callouts > 0) { + wakeup(tq); + TQ_SLEEP(tq, pp, &tq->tq_mutex, PWAIT, "taskqueue_destroy", 0); + } +} + +static void +gtaskqueue_free(struct gtaskqueue *queue) +{ + + TQ_LOCK(queue); + queue->tq_flags &= ~TQ_FLAGS_ACTIVE; + gtaskqueue_terminate(queue->tq_threads, queue); + KASSERT(TAILQ_EMPTY(&queue->tq_active), ("Tasks still running?")); + KASSERT(queue->tq_callouts == 0, ("Armed timeout tasks")); + mtx_destroy(&queue->tq_mutex); + free(queue->tq_threads, M_GTASKQUEUE); + free(queue->tq_name, M_GTASKQUEUE); + free(queue, M_GTASKQUEUE); +} + +int +grouptaskqueue_enqueue(struct gtaskqueue *queue, struct gtask *gtask) +{ + TQ_LOCK(queue); + if (gtask->ta_flags & TASK_ENQUEUED) { + TQ_UNLOCK(queue); + return (0); + } + STAILQ_INSERT_TAIL(&queue->tq_queue, gtask, ta_link); + gtask->ta_flags |= TASK_ENQUEUED; + TQ_UNLOCK(queue); + if ((queue->tq_flags & TQ_FLAGS_BLOCKED) == 0) + queue->tq_enqueue(queue->tq_context); + return (0); +} + +static void +gtaskqueue_task_nop_fn(void *context) +{ +} + +/* + * Block until all currently queued tasks in this taskqueue + * have begun execution. Tasks queued during execution of + * this function are ignored. + */ +static void +gtaskqueue_drain_tq_queue(struct gtaskqueue *queue) +{ + struct gtask t_barrier; + + if (STAILQ_EMPTY(&queue->tq_queue)) + return; + + /* + * Enqueue our barrier after all current tasks, but with + * the highest priority so that newly queued tasks cannot + * pass it. Because of the high priority, we can not use + * taskqueue_enqueue_locked directly (which drops the lock + * anyway) so just insert it at tail while we have the + * queue lock. + */ + GTASK_INIT(&t_barrier, 0, USHRT_MAX, gtaskqueue_task_nop_fn, &t_barrier); + STAILQ_INSERT_TAIL(&queue->tq_queue, &t_barrier, ta_link); + t_barrier.ta_flags |= TASK_ENQUEUED; + + /* + * Once the barrier has executed, all previously queued tasks + * have completed or are currently executing. + */ + while (t_barrier.ta_flags & TASK_ENQUEUED) + TQ_SLEEP(queue, &t_barrier, &queue->tq_mutex, PWAIT, "-", 0); +} + +/* + * Block until all currently executing tasks for this taskqueue + * complete. Tasks that begin execution during the execution + * of this function are ignored. + */ +static void +gtaskqueue_drain_tq_active(struct gtaskqueue *queue) +{ + struct gtaskqueue_busy tb_marker, *tb_first; + + if (TAILQ_EMPTY(&queue->tq_active)) + return; + + /* Block taskq_terminate().*/ + queue->tq_callouts++; + + /* + * Wait for all currently executing taskqueue threads + * to go idle. + */ + tb_marker.tb_running = TB_DRAIN_WAITER; + TAILQ_INSERT_TAIL(&queue->tq_active, &tb_marker, tb_link); + while (TAILQ_FIRST(&queue->tq_active) != &tb_marker) + TQ_SLEEP(queue, &tb_marker, &queue->tq_mutex, PWAIT, "-", 0); + TAILQ_REMOVE(&queue->tq_active, &tb_marker, tb_link); + + /* + * Wakeup any other drain waiter that happened to queue up + * without any intervening active thread. + */ + tb_first = TAILQ_FIRST(&queue->tq_active); + if (tb_first != NULL && tb_first->tb_running == TB_DRAIN_WAITER) + wakeup(tb_first); + + /* Release taskqueue_terminate(). */ + queue->tq_callouts--; + if ((queue->tq_flags & TQ_FLAGS_ACTIVE) == 0) + wakeup_one(queue->tq_threads); +} + +void +gtaskqueue_block(struct gtaskqueue *queue) +{ + + TQ_LOCK(queue); + queue->tq_flags |= TQ_FLAGS_BLOCKED; + TQ_UNLOCK(queue); +} + +void +gtaskqueue_unblock(struct gtaskqueue *queue) +{ + + TQ_LOCK(queue); + queue->tq_flags &= ~TQ_FLAGS_BLOCKED; + if (!STAILQ_EMPTY(&queue->tq_queue)) + queue->tq_enqueue(queue->tq_context); + TQ_UNLOCK(queue); +} + +static void +gtaskqueue_run_locked(struct gtaskqueue *queue) +{ + struct gtaskqueue_busy tb; + struct gtaskqueue_busy *tb_first; + struct gtask *gtask; + + KASSERT(queue != NULL, ("tq is NULL")); + TQ_ASSERT_LOCKED(queue); + tb.tb_running = NULL; + + while (STAILQ_FIRST(&queue->tq_queue)) { + TAILQ_INSERT_TAIL(&queue->tq_active, &tb, tb_link); + + /* + * Carefully remove the first task from the queue and + * clear its TASK_ENQUEUED flag + */ + gtask = STAILQ_FIRST(&queue->tq_queue); + KASSERT(gtask != NULL, ("task is NULL")); + STAILQ_REMOVE_HEAD(&queue->tq_queue, ta_link); + gtask->ta_flags &= ~TASK_ENQUEUED; + tb.tb_running = gtask; + TQ_UNLOCK(queue); + + KASSERT(gtask->ta_func != NULL, ("task->ta_func is NULL")); + gtask->ta_func(gtask->ta_context); + + TQ_LOCK(queue); + tb.tb_running = NULL; + wakeup(gtask); + + TAILQ_REMOVE(&queue->tq_active, &tb, tb_link); + tb_first = TAILQ_FIRST(&queue->tq_active); + if (tb_first != NULL && + tb_first->tb_running == TB_DRAIN_WAITER) + wakeup(tb_first); + } +} + +static int +task_is_running(struct gtaskqueue *queue, struct gtask *gtask) +{ + struct gtaskqueue_busy *tb; + + TQ_ASSERT_LOCKED(queue); + TAILQ_FOREACH(tb, &queue->tq_active, tb_link) { + if (tb->tb_running == gtask) + return (1); + } + return (0); +} + +static int +gtaskqueue_cancel_locked(struct gtaskqueue *queue, struct gtask *gtask) +{ + + if (gtask->ta_flags & TASK_ENQUEUED) + STAILQ_REMOVE(&queue->tq_queue, gtask, gtask, ta_link); + gtask->ta_flags &= ~TASK_ENQUEUED; + return (task_is_running(queue, gtask) ? EBUSY : 0); +} + +int +gtaskqueue_cancel(struct gtaskqueue *queue, struct gtask *gtask) +{ + int error; + + TQ_LOCK(queue); + error = gtaskqueue_cancel_locked(queue, gtask); + TQ_UNLOCK(queue); + + return (error); +} + +void +gtaskqueue_drain(struct gtaskqueue *queue, struct gtask *gtask) +{ + + if (!queue->tq_spin) + WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, __func__); + + TQ_LOCK(queue); + while ((gtask->ta_flags & TASK_ENQUEUED) || task_is_running(queue, gtask)) + TQ_SLEEP(queue, gtask, &queue->tq_mutex, PWAIT, "-", 0); + TQ_UNLOCK(queue); +} + +void +gtaskqueue_drain_all(struct gtaskqueue *queue) +{ + + if (!queue->tq_spin) + WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, __func__); + + TQ_LOCK(queue); + gtaskqueue_drain_tq_queue(queue); + gtaskqueue_drain_tq_active(queue); + TQ_UNLOCK(queue); +} + +static int +_gtaskqueue_start_threads(struct gtaskqueue **tqp, int count, int pri, + cpuset_t *mask, const char *name, va_list ap) +{ + char ktname[MAXCOMLEN + 1]; + struct thread *td; + struct gtaskqueue *tq; + int i, error; + + if (count <= 0) + return (EINVAL); + + vsnprintf(ktname, sizeof(ktname), name, ap); + tq = *tqp; + + tq->tq_threads = malloc(sizeof(struct thread *) * count, M_GTASKQUEUE, + M_NOWAIT | M_ZERO); + if (tq->tq_threads == NULL) { + printf("%s: no memory for %s threads\n", __func__, ktname); + return (ENOMEM); + } + + for (i = 0; i < count; i++) { + if (count == 1) + error = kthread_add(gtaskqueue_thread_loop, tqp, NULL, + &tq->tq_threads[i], RFSTOPPED, 0, "%s", ktname); + else + error = kthread_add(gtaskqueue_thread_loop, tqp, NULL, + &tq->tq_threads[i], RFSTOPPED, 0, + "%s_%d", ktname, i); + if (error) { + /* should be ok to continue, taskqueue_free will dtrt */ + printf("%s: kthread_add(%s): error %d", __func__, + ktname, error); + tq->tq_threads[i] = NULL; /* paranoid */ + } else + tq->tq_tcount++; + } + for (i = 0; i < count; i++) { + if (tq->tq_threads[i] == NULL) + continue; + td = tq->tq_threads[i]; + if (mask) { + error = cpuset_setthread(td->td_tid, mask); + /* + * Failing to pin is rarely an actual fatal error; + * it'll just affect performance. + */ + if (error) + printf("%s: curthread=%llu: can't pin; " + "error=%d\n", + __func__, + (unsigned long long) td->td_tid, + error); + } + thread_lock(td); + sched_prio(td, pri); + sched_add(td, SRQ_BORING); + thread_unlock(td); + } + + return (0); +} + +static int +gtaskqueue_start_threads(struct gtaskqueue **tqp, int count, int pri, + const char *name, ...) +{ + va_list ap; + int error; + + va_start(ap, name); + error = _gtaskqueue_start_threads(tqp, count, pri, NULL, name, ap); + va_end(ap); + return (error); +} + +static inline void +gtaskqueue_run_callback(struct gtaskqueue *tq, + enum taskqueue_callback_type cb_type) +{ + taskqueue_callback_fn tq_callback; + + TQ_ASSERT_UNLOCKED(tq); + tq_callback = tq->tq_callbacks[cb_type]; + if (tq_callback != NULL) + tq_callback(tq->tq_cb_contexts[cb_type]); +} + +static void +gtaskqueue_thread_loop(void *arg) +{ + struct gtaskqueue **tqp, *tq; + + tqp = arg; + tq = *tqp; + gtaskqueue_run_callback(tq, TASKQUEUE_CALLBACK_TYPE_INIT); + TQ_LOCK(tq); + while ((tq->tq_flags & TQ_FLAGS_ACTIVE) != 0) { + /* XXX ? */ + gtaskqueue_run_locked(tq); + /* + * Because taskqueue_run() can drop tq_mutex, we need to + * check if the TQ_FLAGS_ACTIVE flag wasn't removed in the + * meantime, which means we missed a wakeup. + */ + if ((tq->tq_flags & TQ_FLAGS_ACTIVE) == 0) + break; + TQ_SLEEP(tq, tq, &tq->tq_mutex, 0, "-", 0); + } + gtaskqueue_run_locked(tq); + /* + * This thread is on its way out, so just drop the lock temporarily + * in order to call the shutdown callback. This allows the callback + * to look at the taskqueue, even just before it dies. + */ + TQ_UNLOCK(tq); + gtaskqueue_run_callback(tq, TASKQUEUE_CALLBACK_TYPE_SHUTDOWN); + TQ_LOCK(tq); + + /* rendezvous with thread that asked us to terminate */ + tq->tq_tcount--; + wakeup_one(tq->tq_threads); + TQ_UNLOCK(tq); + kthread_exit(); +} + +static void +gtaskqueue_thread_enqueue(void *context) +{ + struct gtaskqueue **tqp, *tq; + + tqp = context; + tq = *tqp; + wakeup_one(tq); +} + + +static struct gtaskqueue * +gtaskqueue_create_fast(const char *name, int mflags, + taskqueue_enqueue_fn enqueue, void *context) +{ + return _gtaskqueue_create(name, mflags, enqueue, context, + MTX_SPIN, "fast_taskqueue"); +} + + +struct taskqgroup_cpu { + LIST_HEAD(, grouptask) tgc_tasks; + struct gtaskqueue *tgc_taskq; + int tgc_cnt; + int tgc_cpu; +}; + +struct taskqgroup { + struct taskqgroup_cpu tqg_queue[MAXCPU]; + struct mtx tqg_lock; + char * tqg_name; + int tqg_adjusting; + int tqg_stride; + int tqg_cnt; +}; + +struct taskq_bind_task { + struct gtask bt_task; + int bt_cpuid; +}; + +static void +taskqgroup_cpu_create(struct taskqgroup *qgroup, int idx) +{ + struct taskqgroup_cpu *qcpu; + + qcpu = &qgroup->tqg_queue[idx]; + LIST_INIT(&qcpu->tgc_tasks); + qcpu->tgc_taskq = gtaskqueue_create_fast(NULL, M_WAITOK, + taskqueue_thread_enqueue, &qcpu->tgc_taskq); + gtaskqueue_start_threads(&qcpu->tgc_taskq, 1, PI_SOFT, + "%s_%d", qgroup->tqg_name, idx); + qcpu->tgc_cpu = idx * qgroup->tqg_stride; +} + +static void +taskqgroup_cpu_remove(struct taskqgroup *qgroup, int idx) +{ + + gtaskqueue_free(qgroup->tqg_queue[idx].tgc_taskq); +} + +/* + * Find the taskq with least # of tasks that doesn't currently have any + * other queues from the uniq identifier. + */ +static int +taskqgroup_find(struct taskqgroup *qgroup, void *uniq) +{ + struct grouptask *n; + int i, idx, mincnt; + int strict; + + mtx_assert(&qgroup->tqg_lock, MA_OWNED); + if (qgroup->tqg_cnt == 0) + return (0); + idx = -1; + mincnt = INT_MAX; + /* + * Two passes; First scan for a queue with the least tasks that + * does not already service this uniq id. If that fails simply find + * the queue with the least total tasks; + */ + for (strict = 1; mincnt == INT_MAX; strict = 0) { + for (i = 0; i < qgroup->tqg_cnt; i++) { + if (qgroup->tqg_queue[i].tgc_cnt > mincnt) + continue; + if (strict) { + LIST_FOREACH(n, + &qgroup->tqg_queue[i].tgc_tasks, gt_list) + if (n->gt_uniq == uniq) + break; + if (n != NULL) + continue; + } + mincnt = qgroup->tqg_queue[i].tgc_cnt; + idx = i; + } + } + if (idx == -1) + panic("taskqgroup_find: Failed to pick a qid."); + + return (idx); +} + +void +taskqgroup_attach(struct taskqgroup *qgroup, struct grouptask *gtask, + void *uniq, int irq, char *name) +{ + cpuset_t mask; + int qid; + + gtask->gt_uniq = uniq; + gtask->gt_name = name; + gtask->gt_irq = irq; + gtask->gt_cpu = -1; + mtx_lock(&qgroup->tqg_lock); + qid = taskqgroup_find(qgroup, uniq); + qgroup->tqg_queue[qid].tgc_cnt++; + LIST_INSERT_HEAD(&qgroup->tqg_queue[qid].tgc_tasks, gtask, gt_list); + gtask->gt_taskqueue = qgroup->tqg_queue[qid].tgc_taskq; + if (irq != -1 && smp_started) { + CPU_ZERO(&mask); + CPU_SET(qgroup->tqg_queue[qid].tgc_cpu, &mask); + mtx_unlock(&qgroup->tqg_lock); + intr_setaffinity(irq, &mask); + } else + mtx_unlock(&qgroup->tqg_lock); +} + +int +taskqgroup_attach_cpu(struct taskqgroup *qgroup, struct grouptask *gtask, + void *uniq, int cpu, int irq, char *name) +{ + cpuset_t mask; + int i, qid; + + qid = -1; + gtask->gt_uniq = uniq; + gtask->gt_name = name; + gtask->gt_irq = irq; + gtask->gt_cpu = cpu; + mtx_lock(&qgroup->tqg_lock); + if (smp_started) { + for (i = 0; i < qgroup->tqg_cnt; i++) + if (qgroup->tqg_queue[i].tgc_cpu == cpu) { + qid = i; + break; + } + if (qid == -1) { + mtx_unlock(&qgroup->tqg_lock); + return (EINVAL); + } + } else + qid = 0; + qgroup->tqg_queue[qid].tgc_cnt++; + LIST_INSERT_HEAD(&qgroup->tqg_queue[qid].tgc_tasks, gtask, gt_list); + gtask->gt_taskqueue = qgroup->tqg_queue[qid].tgc_taskq; + if (irq != -1 && smp_started) { + CPU_ZERO(&mask); + CPU_SET(qgroup->tqg_queue[qid].tgc_cpu, &mask); + mtx_unlock(&qgroup->tqg_lock); + intr_setaffinity(irq, &mask); + } else + mtx_unlock(&qgroup->tqg_lock); + return (0); +} + +void +taskqgroup_detach(struct taskqgroup *qgroup, struct grouptask *gtask) +{ + int i; + + mtx_lock(&qgroup->tqg_lock); + for (i = 0; i < qgroup->tqg_cnt; i++) + if (qgroup->tqg_queue[i].tgc_taskq == gtask->gt_taskqueue) + break; + if (i == qgroup->tqg_cnt) + panic("taskqgroup_detach: task not in group\n"); + qgroup->tqg_queue[i].tgc_cnt--; + LIST_REMOVE(gtask, gt_list); + mtx_unlock(&qgroup->tqg_lock); + gtask->gt_taskqueue = NULL; +} + +static void +taskqgroup_binder(void *ctx) +{ + struct taskq_bind_task *gtask = (struct taskq_bind_task *)ctx; + cpuset_t mask; + int error; + + CPU_ZERO(&mask); + CPU_SET(gtask->bt_cpuid, &mask); + error = cpuset_setthread(curthread->td_tid, &mask); + thread_lock(curthread); + sched_bind(curthread, gtask->bt_cpuid); + thread_unlock(curthread); + + if (error) + printf("taskqgroup_binder: setaffinity failed: %d\n", + error); + free(gtask, M_DEVBUF); +} + +static void +taskqgroup_bind(struct taskqgroup *qgroup) +{ + struct taskq_bind_task *gtask; + int i; + + /* + * Bind taskqueue threads to specific CPUs, if they have been assigned + * one. + */ + for (i = 0; i < qgroup->tqg_cnt; i++) { + gtask = malloc(sizeof (*gtask), M_DEVBUF, M_NOWAIT); + GTASK_INIT(>ask->bt_task, 0, 0, taskqgroup_binder, gtask); + gtask->bt_cpuid = qgroup->tqg_queue[i].tgc_cpu; + grouptaskqueue_enqueue(qgroup->tqg_queue[i].tgc_taskq, + >ask->bt_task); + } +} + +static int +_taskqgroup_adjust(struct taskqgroup *qgroup, int cnt, int stride) +{ + LIST_HEAD(, grouptask) gtask_head = LIST_HEAD_INITIALIZER(NULL); + cpuset_t mask; + struct grouptask *gtask; + int i, old_cnt, qid; + + mtx_assert(&qgroup->tqg_lock, MA_OWNED); + + if (cnt < 1 || cnt * stride > mp_ncpus || !smp_started) { + printf("taskqgroup_adjust failed cnt: %d stride: %d mp_ncpus: %d smp_started: %d\n", + cnt, stride, mp_ncpus, smp_started); + return (EINVAL); + } + if (qgroup->tqg_adjusting) { + printf("taskqgroup_adjust failed: adjusting\n"); + return (EBUSY); + } + qgroup->tqg_adjusting = 1; + old_cnt = qgroup->tqg_cnt; + mtx_unlock(&qgroup->tqg_lock); + /* + * Set up queue for tasks added before boot. + */ + if (old_cnt == 0) { + LIST_SWAP(>ask_head, &qgroup->tqg_queue[0].tgc_tasks, + grouptask, gt_list); + qgroup->tqg_queue[0].tgc_cnt = 0; + } + + /* + * If new taskq threads have been added. + */ + for (i = old_cnt; i < cnt; i++) + taskqgroup_cpu_create(qgroup, i); + mtx_lock(&qgroup->tqg_lock); + qgroup->tqg_cnt = cnt; + qgroup->tqg_stride = stride; + + /* + * Adjust drivers to use new taskqs. + */ + for (i = 0; i < old_cnt; i++) { + while ((gtask = LIST_FIRST(&qgroup->tqg_queue[i].tgc_tasks))) { + LIST_REMOVE(gtask, gt_list); + qgroup->tqg_queue[i].tgc_cnt--; + LIST_INSERT_HEAD(>ask_head, gtask, gt_list); + } + } + + while ((gtask = LIST_FIRST(>ask_head))) { + LIST_REMOVE(gtask, gt_list); + if (gtask->gt_cpu == -1) + qid = taskqgroup_find(qgroup, gtask->gt_uniq); + else { + for (i = 0; i < qgroup->tqg_cnt; i++) + if (qgroup->tqg_queue[i].tgc_cpu == gtask->gt_cpu) { + qid = i; + break; + } + } + qgroup->tqg_queue[qid].tgc_cnt++; + LIST_INSERT_HEAD(&qgroup->tqg_queue[qid].tgc_tasks, gtask, + gt_list); + gtask->gt_taskqueue = qgroup->tqg_queue[qid].tgc_taskq; + } + /* + * Set new CPU and IRQ affinity + */ + for (i = 0; i < cnt; i++) { + qgroup->tqg_queue[i].tgc_cpu = i * qgroup->tqg_stride; + CPU_ZERO(&mask); + CPU_SET(qgroup->tqg_queue[i].tgc_cpu, &mask); + LIST_FOREACH(gtask, &qgroup->tqg_queue[i].tgc_tasks, gt_list) { + if (gtask->gt_irq == -1) + continue; + intr_setaffinity(gtask->gt_irq, &mask); + } + } + mtx_unlock(&qgroup->tqg_lock); + + /* + * If taskq thread count has been reduced. + */ + for (i = cnt; i < old_cnt; i++) + taskqgroup_cpu_remove(qgroup, i); + + mtx_lock(&qgroup->tqg_lock); + qgroup->tqg_adjusting = 0; + + taskqgroup_bind(qgroup); + + return (0); +} + +int +taskqgroup_adjust(struct taskqgroup *qgroup, int cpu, int stride) +{ + int error; + + mtx_lock(&qgroup->tqg_lock); + error = _taskqgroup_adjust(qgroup, cpu, stride); + mtx_unlock(&qgroup->tqg_lock); + + return (error); +} + +struct taskqgroup * +taskqgroup_create(char *name) +{ + struct taskqgroup *qgroup; + + qgroup = malloc(sizeof(*qgroup), M_GTASKQUEUE, M_WAITOK | M_ZERO); + mtx_init(&qgroup->tqg_lock, "taskqgroup", NULL, MTX_DEF); + qgroup->tqg_name = name; + LIST_INIT(&qgroup->tqg_queue[0].tgc_tasks); + + return (qgroup); +} + +void +taskqgroup_destroy(struct taskqgroup *qgroup) +{ + +} Index: head/sys/kern/subr_taskqueue.c =================================================================== --- head/sys/kern/subr_taskqueue.c +++ head/sys/kern/subr_taskqueue.c @@ -261,22 +261,6 @@ } int -grouptaskqueue_enqueue(struct taskqueue *queue, struct task *task) -{ - TQ_LOCK(queue); - if (task->ta_pending) { - TQ_UNLOCK(queue); - return (0); - } - STAILQ_INSERT_TAIL(&queue->tq_queue, task, ta_link); - task->ta_pending = 1; - TQ_UNLOCK(queue); - if ((queue->tq_flags & TQ_FLAGS_BLOCKED) == 0) - queue->tq_enqueue(queue->tq_context); - return (0); -} - -int taskqueue_enqueue(struct taskqueue *queue, struct task *task) { int res; @@ -806,347 +790,3 @@ } return (ret); } - -struct taskqgroup_cpu { - LIST_HEAD(, grouptask) tgc_tasks; - struct taskqueue *tgc_taskq; - int tgc_cnt; - int tgc_cpu; -}; - -struct taskqgroup { - struct taskqgroup_cpu tqg_queue[MAXCPU]; - struct mtx tqg_lock; - char * tqg_name; - int tqg_adjusting; - int tqg_stride; - int tqg_cnt; -}; - -struct taskq_bind_task { - struct task bt_task; - int bt_cpuid; -}; - -static void -taskqgroup_cpu_create(struct taskqgroup *qgroup, int idx) -{ - struct taskqgroup_cpu *qcpu; - int i, j; - - qcpu = &qgroup->tqg_queue[idx]; - LIST_INIT(&qcpu->tgc_tasks); - qcpu->tgc_taskq = taskqueue_create_fast(NULL, M_WAITOK, - taskqueue_thread_enqueue, &qcpu->tgc_taskq); - taskqueue_start_threads(&qcpu->tgc_taskq, 1, PI_SOFT, - "%s_%d", qgroup->tqg_name, idx); - - for (i = CPU_FIRST(), j = 0; j < idx * qgroup->tqg_stride; - j++, i = CPU_NEXT(i)) { - /* - * Wait: evaluate the idx * qgroup->tqg_stride'th CPU, - * potentially wrapping the actual count - */ - } - qcpu->tgc_cpu = i; -} - -static void -taskqgroup_cpu_remove(struct taskqgroup *qgroup, int idx) -{ - - taskqueue_free(qgroup->tqg_queue[idx].tgc_taskq); -} - -/* - * Find the taskq with least # of tasks that doesn't currently have any - * other queues from the uniq identifier. - */ -static int -taskqgroup_find(struct taskqgroup *qgroup, void *uniq) -{ - struct grouptask *n; - int i, idx, mincnt; - int strict; - - mtx_assert(&qgroup->tqg_lock, MA_OWNED); - if (qgroup->tqg_cnt == 0) - return (0); - idx = -1; - mincnt = INT_MAX; - /* - * Two passes; First scan for a queue with the least tasks that - * does not already service this uniq id. If that fails simply find - * the queue with the least total tasks; - */ - for (strict = 1; mincnt == INT_MAX; strict = 0) { - for (i = 0; i < qgroup->tqg_cnt; i++) { - if (qgroup->tqg_queue[i].tgc_cnt > mincnt) - continue; - if (strict) { - LIST_FOREACH(n, - &qgroup->tqg_queue[i].tgc_tasks, gt_list) - if (n->gt_uniq == uniq) - break; - if (n != NULL) - continue; - } - mincnt = qgroup->tqg_queue[i].tgc_cnt; - idx = i; - } - } - if (idx == -1) - panic("taskqgroup_find: Failed to pick a qid."); - - return (idx); -} - -void -taskqgroup_attach(struct taskqgroup *qgroup, struct grouptask *gtask, - void *uniq, int irq, char *name) -{ - cpuset_t mask; - int qid; - - gtask->gt_uniq = uniq; - gtask->gt_name = name; - gtask->gt_irq = irq; - gtask->gt_cpu = -1; - mtx_lock(&qgroup->tqg_lock); - qid = taskqgroup_find(qgroup, uniq); - qgroup->tqg_queue[qid].tgc_cnt++; - LIST_INSERT_HEAD(&qgroup->tqg_queue[qid].tgc_tasks, gtask, gt_list); - gtask->gt_taskqueue = qgroup->tqg_queue[qid].tgc_taskq; - if (irq != -1 && smp_started) { - CPU_ZERO(&mask); - CPU_SET(qgroup->tqg_queue[qid].tgc_cpu, &mask); - mtx_unlock(&qgroup->tqg_lock); - intr_setaffinity(irq, &mask); - } else - mtx_unlock(&qgroup->tqg_lock); -} - -int -taskqgroup_attach_cpu(struct taskqgroup *qgroup, struct grouptask *gtask, - void *uniq, int cpu, int irq, char *name) -{ - cpuset_t mask; - int i, qid; - - qid = -1; - gtask->gt_uniq = uniq; - gtask->gt_name = name; - gtask->gt_irq = irq; - gtask->gt_cpu = cpu; - mtx_lock(&qgroup->tqg_lock); - if (smp_started) { - for (i = 0; i < qgroup->tqg_cnt; i++) - if (qgroup->tqg_queue[i].tgc_cpu == cpu) { - qid = i; - break; - } - if (qid == -1) { - mtx_unlock(&qgroup->tqg_lock); - return (EINVAL); - } - } else - qid = 0; - qgroup->tqg_queue[qid].tgc_cnt++; - LIST_INSERT_HEAD(&qgroup->tqg_queue[qid].tgc_tasks, gtask, gt_list); - gtask->gt_taskqueue = qgroup->tqg_queue[qid].tgc_taskq; - if (irq != -1 && smp_started) { - CPU_ZERO(&mask); - CPU_SET(qgroup->tqg_queue[qid].tgc_cpu, &mask); - mtx_unlock(&qgroup->tqg_lock); - intr_setaffinity(irq, &mask); - } else - mtx_unlock(&qgroup->tqg_lock); - return (0); -} - -void -taskqgroup_detach(struct taskqgroup *qgroup, struct grouptask *gtask) -{ - int i; - - mtx_lock(&qgroup->tqg_lock); - for (i = 0; i < qgroup->tqg_cnt; i++) - if (qgroup->tqg_queue[i].tgc_taskq == gtask->gt_taskqueue) - break; - if (i == qgroup->tqg_cnt) - panic("taskqgroup_detach: task not in group\n"); - qgroup->tqg_queue[i].tgc_cnt--; - LIST_REMOVE(gtask, gt_list); - mtx_unlock(&qgroup->tqg_lock); - gtask->gt_taskqueue = NULL; -} - -static void -taskqgroup_binder(void *ctx, int pending) -{ - struct taskq_bind_task *task = (struct taskq_bind_task *)ctx; - cpuset_t mask; - int error; - - CPU_ZERO(&mask); - CPU_SET(task->bt_cpuid, &mask); - error = cpuset_setthread(curthread->td_tid, &mask); - thread_lock(curthread); - sched_bind(curthread, task->bt_cpuid); - thread_unlock(curthread); - - if (error) - printf("taskqgroup_binder: setaffinity failed: %d\n", - error); - free(task, M_DEVBUF); -} - -static void -taskqgroup_bind(struct taskqgroup *qgroup) -{ - struct taskq_bind_task *task; - int i; - - /* - * Bind taskqueue threads to specific CPUs, if they have been assigned - * one. - */ - for (i = 0; i < qgroup->tqg_cnt; i++) { - task = malloc(sizeof (*task), M_DEVBUF, M_NOWAIT); - TASK_INIT(&task->bt_task, 0, taskqgroup_binder, task); - task->bt_cpuid = qgroup->tqg_queue[i].tgc_cpu; - taskqueue_enqueue(qgroup->tqg_queue[i].tgc_taskq, - &task->bt_task); - } -} - -static int -_taskqgroup_adjust(struct taskqgroup *qgroup, int cnt, int stride) -{ - LIST_HEAD(, grouptask) gtask_head = LIST_HEAD_INITIALIZER(NULL); - cpuset_t mask; - struct grouptask *gtask; - int i, k, old_cnt, qid, cpu; - - mtx_assert(&qgroup->tqg_lock, MA_OWNED); - - if (cnt < 1 || cnt * stride > mp_ncpus || !smp_started) { - printf("taskqgroup_adjust failed cnt: %d stride: %d " - "mp_ncpus: %d smp_started: %d\n", cnt, stride, mp_ncpus, - smp_started); - return (EINVAL); - } - if (qgroup->tqg_adjusting) { - printf("taskqgroup_adjust failed: adjusting\n"); - return (EBUSY); - } - qgroup->tqg_adjusting = 1; - old_cnt = qgroup->tqg_cnt; - mtx_unlock(&qgroup->tqg_lock); - /* - * Set up queue for tasks added before boot. - */ - if (old_cnt == 0) { - LIST_SWAP(>ask_head, &qgroup->tqg_queue[0].tgc_tasks, - grouptask, gt_list); - qgroup->tqg_queue[0].tgc_cnt = 0; - } - - /* - * If new taskq threads have been added. - */ - for (i = old_cnt; i < cnt; i++) - taskqgroup_cpu_create(qgroup, i); - mtx_lock(&qgroup->tqg_lock); - qgroup->tqg_cnt = cnt; - qgroup->tqg_stride = stride; - - /* - * Adjust drivers to use new taskqs. - */ - for (i = 0; i < old_cnt; i++) { - while ((gtask = LIST_FIRST(&qgroup->tqg_queue[i].tgc_tasks))) { - LIST_REMOVE(gtask, gt_list); - qgroup->tqg_queue[i].tgc_cnt--; - LIST_INSERT_HEAD(>ask_head, gtask, gt_list); - } - } - - while ((gtask = LIST_FIRST(>ask_head))) { - LIST_REMOVE(gtask, gt_list); - if (gtask->gt_cpu == -1) - qid = taskqgroup_find(qgroup, gtask->gt_uniq); - else { - for (i = 0; i < qgroup->tqg_cnt; i++) - if (qgroup->tqg_queue[i].tgc_cpu == gtask->gt_cpu) { - qid = i; - break; - } - } - qgroup->tqg_queue[qid].tgc_cnt++; - LIST_INSERT_HEAD(&qgroup->tqg_queue[qid].tgc_tasks, gtask, - gt_list); - gtask->gt_taskqueue = qgroup->tqg_queue[qid].tgc_taskq; - } - /* - * Set new CPU and IRQ affinity - */ - cpu = CPU_FIRST(); - for (i = 0; i < cnt; i++) { - qgroup->tqg_queue[i].tgc_cpu = cpu; - for (k = 0; k < qgroup->tqg_stride; k++) - cpu = CPU_NEXT(cpu); - CPU_ZERO(&mask); - CPU_SET(qgroup->tqg_queue[i].tgc_cpu, &mask); - LIST_FOREACH(gtask, &qgroup->tqg_queue[i].tgc_tasks, gt_list) { - if (gtask->gt_irq == -1) - continue; - intr_setaffinity(gtask->gt_irq, &mask); - } - } - mtx_unlock(&qgroup->tqg_lock); - - /* - * If taskq thread count has been reduced. - */ - for (i = cnt; i < old_cnt; i++) - taskqgroup_cpu_remove(qgroup, i); - - mtx_lock(&qgroup->tqg_lock); - qgroup->tqg_adjusting = 0; - - taskqgroup_bind(qgroup); - - return (0); -} - -int -taskqgroup_adjust(struct taskqgroup *qgroup, int cpu, int stride) -{ - int error; - - mtx_lock(&qgroup->tqg_lock); - error = _taskqgroup_adjust(qgroup, cpu, stride); - mtx_unlock(&qgroup->tqg_lock); - - return (error); -} - -struct taskqgroup * -taskqgroup_create(char *name) -{ - struct taskqgroup *qgroup; - - qgroup = malloc(sizeof(*qgroup), M_TASKQUEUE, M_WAITOK | M_ZERO); - mtx_init(&qgroup->tqg_lock, "taskqgroup", NULL, MTX_DEF); - qgroup->tqg_name = name; - LIST_INIT(&qgroup->tqg_queue[0].tgc_tasks); - - return (qgroup); -} - -void -taskqgroup_destroy(struct taskqgroup *qgroup) -{ - -} Index: head/sys/net/ifdi_if.m =================================================================== --- head/sys/net/ifdi_if.m +++ head/sys/net/ifdi_if.m @@ -60,9 +60,10 @@ return (0); } - static void + static int null_queue_intr_enable(if_ctx_t _ctx __unused, uint16_t _qid __unused) { + return (ENOTSUP); } static void @@ -194,7 +195,7 @@ if_ctx_t _ctx; }; -METHOD void queue_intr_enable { +METHOD int queue_intr_enable { if_ctx_t _ctx; uint16_t _qid; } DEFAULT null_queue_intr_enable; Index: head/sys/net/iflib.h =================================================================== --- head/sys/net/iflib.h +++ head/sys/net/iflib.h @@ -35,6 +35,7 @@ #include #include #include +#include /* @@ -63,12 +64,14 @@ typedef struct if_rxd_frag { uint8_t irf_flid; uint16_t irf_idx; + uint16_t irf_len; } *if_rxd_frag_t; typedef struct if_rxd_info { /* set by iflib */ uint16_t iri_qsidx; /* qset index */ uint16_t iri_vtag; /* vlan tag - if flag set */ + /* XXX redundant with the new irf_len field */ uint16_t iri_len; /* packet length */ uint16_t iri_cidx; /* consumer index of cq */ struct ifnet *iri_ifp; /* some drivers >1 interface per softc */ @@ -156,10 +159,11 @@ void (*ift_txd_flush) (void *, uint16_t, uint32_t); int (*ift_txd_credits_update) (void *, uint16_t, uint32_t, bool); - int (*ift_rxd_available) (void *, uint16_t qsidx, uint32_t pidx); + int (*ift_rxd_available) (void *, uint16_t qsidx, uint32_t pidx, + int budget); int (*ift_rxd_pkt_get) (void *, if_rxd_info_t ri); void (*ift_rxd_refill) (void * , uint16_t qsidx, uint8_t flidx, uint32_t pidx, - uint64_t *paddrs, caddr_t *vaddrs, uint16_t count); + uint64_t *paddrs, caddr_t *vaddrs, uint16_t count, uint16_t buf_size); void (*ift_rxd_flush) (void *, uint16_t qsidx, uint8_t flidx, uint32_t pidx); int (*ift_legacy_intr) (void *); } *if_txrx_t; @@ -170,11 +174,20 @@ int isc_ntxqsets; int isc_msix_bar; /* can be model specific - initialize in attach_pre */ int isc_tx_nsegments; /* can be model specific - initialize in attach_pre */ + int isc_ntxd[8]; + int isc_nrxd[8]; + + uint32_t isc_txqsizes[8]; + uint32_t isc_rxqsizes[8]; + int isc_max_txqsets; + int isc_max_rxqsets; int isc_tx_tso_segments_max; int isc_tx_tso_size_max; int isc_tx_tso_segsize_max; int isc_rss_table_size; int isc_rss_table_mask; + int isc_nrxqsets_max; + int isc_ntxqsets_max; iflib_intr_mode_t isc_intr; uint16_t isc_max_frame_size; /* set at init time by driver */ @@ -188,8 +201,6 @@ int isc_magic; if_txrx_t isc_txrx; driver_t *isc_driver; - int isc_ntxd; - int isc_nrxd; int isc_nfl; int isc_flags; bus_size_t isc_q_align; @@ -199,14 +210,11 @@ bus_size_t isc_rx_maxsegsize; int isc_rx_nsegments; int isc_rx_process_limit; - - - uint32_t isc_txqsizes[8]; int isc_ntxqs; /* # of tx queues per tx qset - usually 1 */ - uint32_t isc_rxqsizes[8]; int isc_nrxqs; /* # of rx queues per rx qset - intel 1, chelsio 2, broadcom 3 */ int isc_admin_intrcnt; /* # of admin/link interrupts */ + int isc_tx_reclaim_thresh; /* fields necessary for probe */ @@ -215,6 +223,12 @@ /* optional function to transform the read values to match the table*/ void (*isc_parse_devinfo) (uint16_t *device_id, uint16_t *subvendor_id, uint16_t *subdevice_id, uint16_t *rev_id); + int isc_nrxd_min[8]; + int isc_nrxd_default[8]; + int isc_nrxd_max[8]; + int isc_ntxd_min[8]; + int isc_ntxd_default[8]; + int isc_ntxd_max[8]; }; typedef struct iflib_dma_info { @@ -240,9 +254,9 @@ /* - * Interface has a separate command queue + * Interface has a separate command queue for RX */ -#define IFLIB_HAS_CQ 0x1 +#define IFLIB_HAS_RXCQ 0x1 /* * Driver has already allocated vectors */ @@ -252,6 +266,10 @@ * Interface is a virtual function */ #define IFLIB_IS_VF 0x4 +/* + * Interface has a separate command queue for TX + */ +#define IFLIB_HAS_TXCQ 0x8 /* @@ -308,7 +326,10 @@ void iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, char *name); void iflib_config_gtask_init(if_ctx_t ctx, struct grouptask *gtask, - task_fn_t *fn, char *name); + gtask_fn_t *fn, char *name); + +void iflib_config_gtask_deinit(struct grouptask *gtask); + void iflib_tx_intr_deferred(if_ctx_t ctx, int txqid); @@ -317,7 +338,7 @@ void iflib_iov_intr_deferred(if_ctx_t ctx); -void iflib_link_state_change(if_ctx_t ctx, int linkstate); +void iflib_link_state_change(if_ctx_t ctx, int linkstate, uint64_t baudrate); int iflib_dma_alloc(if_ctx_t ctx, int size, iflib_dma_info_t dma, int mapflags); void iflib_dma_free(iflib_dma_info_t dma); Index: head/sys/net/iflib.c =================================================================== --- head/sys/net/iflib.c +++ head/sys/net/iflib.c @@ -49,6 +49,7 @@ #include #include #include +#include #include @@ -101,7 +102,6 @@ * Enable mbuf vectors for compressing long mbuf chains */ - /* * NB: * - Prefetching in tx cleaning should perhaps be a tunable. The distance ahead @@ -181,8 +181,10 @@ struct sysctl_oid *ifc_sysctl_node; uint16_t ifc_sysctl_ntxqs; uint16_t ifc_sysctl_nrxqs; - uint16_t ifc_sysctl_ntxds; - uint16_t ifc_sysctl_nrxds; + uint16_t ifc_sysctl_qs_eq_override; + + uint16_t ifc_sysctl_ntxds[8]; + uint16_t ifc_sysctl_nrxds[8]; struct if_txrx ifc_txrx; #define isc_txd_encap ifc_txrx.ift_txd_encap #define isc_txd_flush ifc_txrx.ift_txd_flush @@ -294,10 +296,11 @@ #define IFLIB_RESTART_BUDGET 8 -#define IFC_LEGACY 0x1 -#define IFC_QFLUSH 0x2 -#define IFC_MULTISEG 0x4 -#define IFC_DMAR 0x8 +#define IFC_LEGACY 0x01 +#define IFC_QFLUSH 0x02 +#define IFC_MULTISEG 0x04 +#define IFC_DMAR 0x08 +#define IFC_SC_ALLOCATED 0x10 #define CSUM_OFFLOAD (CSUM_IP_TSO|CSUM_IP6_TSO|CSUM_IP| \ CSUM_IP_UDP|CSUM_IP_TCP|CSUM_IP_SCTP| \ @@ -311,6 +314,7 @@ uint8_t ift_db_pending; uint8_t ift_db_pending_queued; uint8_t ift_npending; + uint8_t ift_br_offset; /* implicit pad */ uint64_t ift_processed; uint64_t ift_cleaned; @@ -414,6 +418,7 @@ uint16_t ifr_cq_cidx; uint16_t ifr_cq_pidx; uint8_t ifr_cq_gen; + uint8_t ifr_fl_offset; if_ctx_t ifr_ctx; iflib_fl_t ifr_fl; @@ -604,7 +609,7 @@ static void iflib_rx_structures_free(if_ctx_t ctx); static int iflib_queues_alloc(if_ctx_t ctx); static int iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq); -static int iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, int cidx); +static int iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, int cidx, int budget); static int iflib_qset_structures_setup(if_ctx_t ctx); static int iflib_msix_init(if_ctx_t ctx); static int iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filterarg, int *rid, char *str); @@ -875,7 +880,7 @@ for (fl = rxq->ifr_fl, i = 0; i < rxq->ifr_nfl; i++, fl++) { nic_i = fl->ifl_cidx; nm_i = netmap_idx_n2k(kring, nic_i); - avail = ctx->isc_rxd_available(ctx->ifc_softc, kring->ring_id, nic_i); + avail = ctx->isc_rxd_available(ctx->ifc_softc, kring->ring_id, nic_i, INT_MAX); for (n = 0; avail > 0; n++, avail--) { error = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri); if (error) @@ -930,7 +935,7 @@ /* * XXX we should be batching this operation - TODO */ - ctx->isc_rxd_refill(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id, nic_i, &paddr, &vaddr, 1); + ctx->isc_rxd_refill(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id, nic_i, &paddr, &vaddr, 1, fl->ifl_buf_size); bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_sds[nic_i].ifsd_map, BUS_DMASYNC_PREREAD); nm_i = nm_next(nm_i, lim); @@ -958,6 +963,7 @@ iflib_netmap_attach(if_ctx_t ctx) { struct netmap_adapter na; + if_softc_ctx_t scctx = &ctx->ifc_softc_ctx; bzero(&na, sizeof(na)); @@ -966,8 +972,8 @@ MPASS(ctx->ifc_softc_ctx.isc_ntxqsets); MPASS(ctx->ifc_softc_ctx.isc_nrxqsets); - na.num_tx_desc = ctx->ifc_sctx->isc_ntxd; - na.num_rx_desc = ctx->ifc_sctx->isc_ntxd; + na.num_tx_desc = scctx->isc_ntxd[0]; + na.num_rx_desc = scctx->isc_nrxd[0]; na.nm_txsync = iflib_netmap_txsync; na.nm_rxsync = iflib_netmap_rxsync; na.nm_register = iflib_netmap_register; @@ -986,7 +992,7 @@ if (slot == 0) return; - for (int i = 0; i < ctx->ifc_sctx->isc_ntxd; i++) { + for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxd[0]; i++) { /* * In netmap mode, set the map for the packet buffer. @@ -1011,7 +1017,7 @@ if (slot == 0) return; sd = rxq->ifr_fl[0].ifl_sds; - nrxd = ctx->ifc_sctx->isc_nrxd; + nrxd = ctx->ifc_softc_ctx.isc_nrxd[0]; for (int i = 0; i < nrxd; i++, sd++) { int sj = netmap_idx_n2k(&na->rx_rings[rxq->ifr_id], i); uint64_t paddr; @@ -1021,7 +1027,7 @@ vaddr = addr = PNMB(na, slot + sj, &paddr); netmap_load_map(na, rxq->ifr_fl[0].ifl_ifdi->idi_tag, sd->ifsd_map, addr); /* Update descriptor and the cached value */ - ctx->isc_rxd_refill(ctx->ifc_softc, rxq->ifr_id, 0 /* fl_id */, i, &paddr, &vaddr, 1); + ctx->isc_rxd_refill(ctx->ifc_softc, rxq->ifr_id, 0 /* fl_id */, i, &paddr, &vaddr, 1, rxq->ifr_fl[0].ifl_buf_size); } /* preserve queue */ if (ctx->ifc_ifp->if_capenable & IFCAP_NETMAP) { @@ -1236,7 +1242,8 @@ nsegments = scctx->isc_tx_nsegments; ntsosegments = scctx->isc_tx_tso_segments_max; - MPASS(sctx->isc_ntxd > 0); + MPASS(scctx->isc_ntxd[0] > 0); + MPASS(scctx->isc_ntxd[txq->ift_br_offset] > 0); MPASS(nsegments > 0); MPASS(ntsosegments > 0); /* @@ -1259,13 +1266,11 @@ sctx->isc_tx_maxsize, nsegments, sctx->isc_tx_maxsegsize); goto fail; } -#ifdef INVARIANTS +#ifdef IFLIB_DIAGNOSTICS device_printf(dev,"maxsize: %zd nsegments: %d maxsegsize: %zd\n", sctx->isc_tx_maxsize, nsegments, sctx->isc_tx_maxsegsize); + #endif - device_printf(dev,"TSO maxsize: %d ntsosegments: %d maxsegsize: %d\n", - scctx->isc_tx_tso_size_max, ntsosegments, - scctx->isc_tx_tso_segsize_max); if ((err = bus_dma_tag_create(bus_get_dma_tag(dev), 1, 0, /* alignment, bounds */ BUS_SPACE_MAXADDR, /* lowaddr */ @@ -1282,21 +1287,21 @@ goto fail; } -#ifdef INVARIANTS +#ifdef IFLIB_DIAGNOSTICS device_printf(dev,"TSO maxsize: %d ntsosegments: %d maxsegsize: %d\n", scctx->isc_tx_tso_size_max, ntsosegments, scctx->isc_tx_tso_segsize_max); #endif if (!(txq->ift_sds.ifsd_flags = (uint8_t *) malloc(sizeof(uint8_t) * - sctx->isc_ntxd, M_IFLIB, M_NOWAIT | M_ZERO))) { + scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) { device_printf(dev, "Unable to allocate tx_buffer memory\n"); err = ENOMEM; goto fail; } if (!(txq->ift_sds.ifsd_m = (struct mbuf **) malloc(sizeof(struct mbuf *) * - sctx->isc_ntxd, M_IFLIB, M_NOWAIT | M_ZERO))) { + scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) { device_printf(dev, "Unable to allocate tx_buffer memory\n"); err = ENOMEM; goto fail; @@ -1308,13 +1313,13 @@ return (0); if (!(txq->ift_sds.ifsd_map = - (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * sctx->isc_ntxd, M_IFLIB, M_NOWAIT | M_ZERO))) { + (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) { device_printf(dev, "Unable to allocate tx_buffer map memory\n"); err = ENOMEM; goto fail; } - for (int i = 0; i < sctx->isc_ntxd; i++) { + for (int i = 0; i < scctx->isc_ntxd[txq->ift_br_offset]; i++) { err = bus_dmamap_create(txq->ift_desc_tag, 0, &txq->ift_sds.ifsd_map[i]); if (err != 0) { device_printf(dev, "Unable to create TX DMA map\n"); @@ -1348,9 +1353,8 @@ iflib_txq_destroy(iflib_txq_t txq) { if_ctx_t ctx = txq->ift_ctx; - if_shared_ctx_t sctx = ctx->ifc_sctx; - for (int i = 0; i < sctx->isc_ntxd; i++) + for (int i = 0; i < txq->ift_size; i++) iflib_txsd_destroy(ctx, txq, i); if (txq->ift_sds.ifsd_map != NULL) { free(txq->ift_sds.ifsd_map, M_IFLIB); @@ -1390,7 +1394,7 @@ bus_dmamap_unload(txq->ift_desc_tag, txq->ift_sds.ifsd_map[i]); } - m_freem(*mp); + m_free(*mp); DBG_COUNTER_INC(tx_frees); *mp = NULL; } @@ -1399,7 +1403,7 @@ iflib_txq_setup(iflib_txq_t txq) { if_ctx_t ctx = txq->ift_ctx; - if_shared_ctx_t sctx = ctx->ifc_sctx; + if_softc_ctx_t scctx = &ctx->ifc_softc_ctx; iflib_dma_info_t di; int i; @@ -1408,7 +1412,7 @@ /* Reset indices */ txq->ift_cidx_processed = txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0; - txq->ift_size = sctx->isc_ntxd; + txq->ift_size = scctx->isc_ntxd[txq->ift_br_offset]; for (i = 0, di = txq->ift_ifdi; i < ctx->ifc_nhwtxqs; i++, di++) bzero((void *)di->idi_vaddr, di->idi_size); @@ -1433,22 +1437,25 @@ { if_ctx_t ctx = rxq->ifr_ctx; if_shared_ctx_t sctx = ctx->ifc_sctx; + if_softc_ctx_t scctx = &ctx->ifc_softc_ctx; device_t dev = ctx->ifc_dev; iflib_fl_t fl; iflib_rxsd_t rxsd; int err; - MPASS(sctx->isc_nrxd > 0); + MPASS(scctx->isc_nrxd[0] > 0); + MPASS(scctx->isc_nrxd[rxq->ifr_fl_offset] > 0); fl = rxq->ifr_fl; for (int i = 0; i < rxq->ifr_nfl; i++, fl++) { fl->ifl_sds = malloc(sizeof(struct iflib_sw_rx_desc) * - sctx->isc_nrxd, M_IFLIB, M_WAITOK | M_ZERO); + scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, + M_WAITOK | M_ZERO); if (fl->ifl_sds == NULL) { device_printf(dev, "Unable to allocate rx sw desc memory\n"); return (ENOMEM); } - fl->ifl_size = sctx->isc_nrxd; /* this isn't necessarily the same */ + fl->ifl_size = scctx->isc_nrxd[rxq->ifr_fl_offset]; /* this isn't necessarily the same */ err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */ 1, 0, /* alignment, bounds */ BUS_SPACE_MAXADDR, /* lowaddr */ @@ -1468,7 +1475,7 @@ } rxsd = fl->ifl_sds; - for (int i = 0; i < sctx->isc_nrxd; i++, rxsd++) { + for (int i = 0; i < scctx->isc_nrxd[rxq->ifr_fl_offset]; i++, rxsd++) { err = bus_dmamap_create(fl->ifl_desc_tag, 0, &rxsd->ifsd_map); if (err) { device_printf(dev, "%s: bus_dmamap_create failed: %d\n", @@ -1626,7 +1633,7 @@ } if (n == 0 || i == IFLIB_MAX_RX_REFRESH) { ctx->isc_rxd_refill(ctx->ifc_softc, fl->ifl_rxq->ifr_id, fl->ifl_id, pidx, - fl->ifl_bus_addrs, fl->ifl_vm_addrs, i); + fl->ifl_bus_addrs, fl->ifl_vm_addrs, i, fl->ifl_buf_size); i = 0; pidx = fl->ifl_pidx; } @@ -1854,7 +1861,11 @@ for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) { iflib_netmap_rxq_init(ctx, rxq); } +#ifdef INVARIANTS + i = if_getdrvflags(ifp); +#endif IFDI_INIT(ctx); + MPASS(if_getdrvflags(ifp) == i); for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) { for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) { if (iflib_fl_setup(fl)) { @@ -1902,7 +1913,6 @@ iflib_txq_t txq = ctx->ifc_txqs; iflib_rxq_t rxq = ctx->ifc_rxqs; if_softc_ctx_t scctx = &ctx->ifc_softc_ctx; - if_shared_ctx_t sctx = ctx->ifc_sctx; iflib_dma_info_t di; iflib_fl_t fl; int i, j; @@ -1920,7 +1930,7 @@ /* clean any enqueued buffers */ iflib_txq_check_drain(txq, 0); /* Free any existing tx buffers. */ - for (j = 0; j < sctx->isc_ntxd; j++) { + for (j = 0; j < txq->ift_size; j++) { iflib_txsd_free(ctx, txq, j); } txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0; @@ -1990,13 +2000,24 @@ caddr_t cl; i = 0; + mh = NULL; do { sd = rxd_frag_to_sd(rxq, &ri->iri_frags[i], &cltype, TRUE); MPASS(sd->ifsd_cl != NULL); MPASS(sd->ifsd_m != NULL); + + /* Don't include zero-length frags */ + if (ri->iri_frags[i].irf_len == 0) { + /* XXX we can save the cluster here, but not the mbuf */ + m_init(sd->ifsd_m, M_NOWAIT, MT_DATA, 0); + m_free(sd->ifsd_m); + sd->ifsd_m = NULL; + continue; + } + m = sd->ifsd_m; - if (i == 0) { + if (mh == NULL) { flags = M_PKTHDR|M_EXT; mh = mt = m; padlen = ri->iri_pad; @@ -2019,14 +2040,12 @@ */ m->m_data += padlen; ri->iri_len -= padlen; - m->m_len = ri->iri_len; + m->m_len = ri->iri_frags[i].irf_len; } while (++i < ri->iri_nfrags); return (mh); } - - /* * Process one software descriptor */ @@ -2037,13 +2056,14 @@ iflib_rxsd_t sd; /* should I merge this back in now that the two paths are basically duplicated? */ - if (ri->iri_len <= IFLIB_RX_COPY_THRESH) { + if (ri->iri_nfrags == 1 && + ri->iri_frags[0].irf_len <= IFLIB_RX_COPY_THRESH) { sd = rxd_frag_to_sd(rxq, &ri->iri_frags[0], NULL, FALSE); m = sd->ifsd_m; sd->ifsd_m = NULL; m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR); memcpy(m->m_data, sd->ifsd_cl, ri->iri_len); - m->m_len = ri->iri_len; + m->m_len = ri->iri_frags[0].irf_len; } else { m = assemble_segments(rxq, ri); } @@ -2063,13 +2083,13 @@ { if_ctx_t ctx = rxq->ifr_ctx; if_shared_ctx_t sctx = ctx->ifc_sctx; + if_softc_ctx_t scctx = &ctx->ifc_softc_ctx; int avail, i; uint16_t *cidxp; struct if_rxd_info ri; int err, budget_left, rx_bytes, rx_pkts; iflib_fl_t fl; struct ifnet *ifp; - struct lro_entry *queued; int lro_enabled; /* * XXX early demux data packets so that if_input processing only handles @@ -2084,11 +2104,11 @@ mh = mt = NULL; MPASS(budget > 0); rx_pkts = rx_bytes = 0; - if (sctx->isc_flags & IFLIB_HAS_CQ) + if (sctx->isc_flags & IFLIB_HAS_RXCQ) cidxp = &rxq->ifr_cq_cidx; else cidxp = &rxq->ifr_fl[0].ifl_cidx; - if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp)) == 0) { + if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) { for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++) __iflib_fl_refill_lt(ctx, fl, budget + 8); DBG_COUNTER_INC(rx_unavail); @@ -2112,10 +2132,11 @@ /* in lieu of handling correctly - make sure it isn't being unhandled */ MPASS(err == 0); - if (sctx->isc_flags & IFLIB_HAS_CQ) { - /* we know we consumed _one_ CQ entry */ - if (++rxq->ifr_cq_cidx == sctx->isc_nrxd) { - rxq->ifr_cq_cidx = 0; + if (sctx->isc_flags & IFLIB_HAS_RXCQ) { + *cidxp = ri.iri_cidx; + /* Update our consumer index */ + while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0]) { + rxq->ifr_cq_cidx -= scctx->isc_nrxd[0]; rxq->ifr_cq_gen = 0; } /* was this only a completion queue message? */ @@ -2128,7 +2149,7 @@ /* will advance the cidx on the corresponding free lists */ m = iflib_rxd_pkt_get(rxq, &ri); if (avail == 0 && budget_left) - avail = iflib_rxd_avail(ctx, rxq, *cidxp); + avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left); if (__predict_false(m == NULL)) { DBG_COUNTER_INC(rx_mbuf_null); @@ -2148,7 +2169,6 @@ ifp = ctx->ifc_ifp; lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO); - while (mh != NULL) { m = mh; mh = mh->m_nextpkt; @@ -2162,32 +2182,33 @@ DBG_COUNTER_INC(rx_if_input); ifp->if_input(ifp, m); } + if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes); if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts); /* * Flush any outstanding LRO work */ - while ((queued = LIST_FIRST(&rxq->ifr_lc.lro_active)) != NULL) { - LIST_REMOVE(queued, next); #if defined(INET6) || defined(INET) - tcp_lro_flush(&rxq->ifr_lc, queued); + tcp_lro_flush_all(&rxq->ifr_lc); #endif - } - return (iflib_rxd_avail(ctx, rxq, *cidxp)); + if (avail) + return true; + return (iflib_rxd_avail(ctx, rxq, *cidxp, 1)); } #define M_CSUM_FLAGS(m) ((m)->m_pkthdr.csum_flags) #define M_HAS_VLANTAG(m) (m->m_flags & M_VLANTAG) -#define TXQ_MAX_DB_DEFERRED(ctx) (ctx->ifc_sctx->isc_ntxd >> 5) -#define TXQ_MAX_DB_CONSUMED(ctx) (ctx->ifc_sctx->isc_ntxd >> 4) +#define TXQ_MAX_DB_DEFERRED(size) (size >> 5) +#define TXQ_MAX_DB_CONSUMED(size) (size >> 4) static __inline void iflib_txd_db_check(if_ctx_t ctx, iflib_txq_t txq, int ring) { uint32_t dbval; - if (ring || txq->ift_db_pending >= TXQ_MAX_DB_DEFERRED(ctx)) { + if (ring || txq->ift_db_pending >= + TXQ_MAX_DB_DEFERRED(txq->ift_size)) { /* the lock will only ever be contended in the !min_latency case */ if (!TXDB_TRYLOCK(txq)) @@ -2233,9 +2254,9 @@ iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp) { struct ether_vlan_header *eh; - struct mbuf *m; + struct mbuf *m, *n; - m = *mp; + n = m = *mp; /* * Determine where frame payload starts. * Jump over vlan headers if already present, @@ -2261,7 +2282,6 @@ { struct ip *ip = NULL; struct tcphdr *th = NULL; - struct mbuf *n; int minthlen; minthlen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip) + sizeof(*th)); @@ -2403,37 +2423,31 @@ /* * If dodgy hardware rejects the scatter gather chain we've handed it - * we'll need to rebuild the mbuf chain before we can call m_defrag + * we'll need to remove the mbuf chain from ifsg_m[] before we can add the + * m_defrag'd mbufs */ static __noinline struct mbuf * -iflib_rebuild_mbuf(iflib_txq_t txq) +iflib_remove_mbuf(iflib_txq_t txq) { - - int ntxd, mhlen, len, i, pidx; + int ntxd, i, pidx; struct mbuf *m, *mh, **ifsd_m; - if_shared_ctx_t sctx; pidx = txq->ift_pidx; ifsd_m = txq->ift_sds.ifsd_m; - sctx = txq->ift_ctx->ifc_sctx; - ntxd = sctx->isc_ntxd; + ntxd = txq->ift_size; mh = m = ifsd_m[pidx]; ifsd_m[pidx] = NULL; #if MEMORY_LOGGING txq->ift_dequeued++; #endif - len = m->m_len; - mhlen = m->m_pkthdr.len; i = 1; - while (len < mhlen && (m->m_next == NULL)) { - m->m_next = ifsd_m[(pidx + i) & (ntxd-1)]; + while (m) { ifsd_m[(pidx + i) & (ntxd -1)] = NULL; #if MEMORY_LOGGING txq->ift_dequeued++; #endif m = m->m_next; - len += m->m_len; i++; } return (mh); @@ -2446,6 +2460,7 @@ { if_ctx_t ctx; if_shared_ctx_t sctx; + if_softc_ctx_t scctx; int i, next, pidx, mask, err, maxsegsz, ntxd, count; struct mbuf *m, *tmp, **ifsd_m, **mp; @@ -2459,8 +2474,9 @@ ctx = txq->ift_ctx; sctx = ctx->ifc_sctx; + scctx = &ctx->ifc_softc_ctx; ifsd_m = txq->ift_sds.ifsd_m; - ntxd = sctx->isc_ntxd; + ntxd = txq->ift_size; pidx = txq->ift_pidx; if (map != NULL) { uint8_t *ifsd_flags = txq->ift_sds.ifsd_flags; @@ -2472,13 +2488,12 @@ ifsd_flags[pidx] |= TX_SW_DESC_MAPPED; i = 0; next = pidx; - mask = (sctx->isc_ntxd-1); + mask = (txq->ift_size-1); m = *m0; do { mp = &ifsd_m[next]; *mp = m; m = m->m_next; - (*mp)->m_next = NULL; if (__predict_false((*mp)->m_len == 0)) { m_free(*mp); *mp = NULL; @@ -2529,13 +2544,12 @@ count++; tmp = m; m = m->m_next; - tmp->m_next = NULL; } while (m != NULL); *nsegs = i; } return (0); err: - *m0 = iflib_rebuild_mbuf(txq); + *m0 = iflib_remove_mbuf(txq); return (EFBIG); } @@ -2558,7 +2572,7 @@ sctx = ctx->ifc_sctx; scctx = &ctx->ifc_softc_ctx; segs = txq->ift_segs; - ntxd = sctx->isc_ntxd; + ntxd = txq->ift_size; m_head = *m_headp; map = NULL; @@ -2645,14 +2659,14 @@ if (map != NULL) bus_dmamap_unload(desc_tag, map); DBG_COUNTER_INC(encap_txq_avail_fail); - if (txq->ift_task.gt_task.ta_pending == 0) + if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0) GROUPTASK_ENQUEUE(&txq->ift_task); return (ENOBUFS); } pi.ipi_segs = segs; pi.ipi_nsegs = nsegs; - MPASS(pidx >= 0 && pidx < sctx->isc_ntxd); + MPASS(pidx >= 0 && pidx < txq->ift_size); #ifdef PKT_DEBUG print_pkt(&pi); #endif @@ -2661,11 +2675,12 @@ BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE); DBG_COUNTER_INC(tx_encap); - MPASS(pi.ipi_new_pidx >= 0 && pi.ipi_new_pidx < sctx->isc_ntxd); + MPASS(pi.ipi_new_pidx >= 0 && + pi.ipi_new_pidx < txq->ift_size); ndesc = pi.ipi_new_pidx - pi.ipi_pidx; if (pi.ipi_new_pidx < pi.ipi_pidx) { - ndesc += sctx->isc_ntxd; + ndesc += txq->ift_size; txq->ift_gen = 1; } MPASS(pi.ipi_new_pidx != pidx); @@ -2678,7 +2693,7 @@ txq->ift_pidx = pi.ipi_new_pidx; txq->ift_npending += pi.ipi_ndescs; } else if (__predict_false(err == EFBIG && remap < 2)) { - *m_headp = m_head = iflib_rebuild_mbuf(txq); + *m_headp = m_head = iflib_remove_mbuf(txq); remap = 1; txq->ift_txd_encap_efbig++; goto defrag; @@ -2700,7 +2715,7 @@ #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets) #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets) -#define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NRXQSETS(ctx)) + FIRST_QSET(ctx)) +#define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx)) #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments)) #define RECLAIM_THRESH(ctx) ((ctx)->ifc_sctx->isc_tx_reclaim_thresh) #define MAX_TX_DESC(ctx) ((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max) @@ -2712,7 +2727,7 @@ * * ORing with 2 assures that min occupancy is never less than 2 without any conditional logic */ -#define TXQ_MIN_OCCUPANCY(ctx) ((ctx->ifc_sctx->isc_ntxd >> 6)| 0x2) +#define TXQ_MIN_OCCUPANCY(size) ((size >> 6)| 0x2) static inline int iflib_txq_min_occupancy(iflib_txq_t txq) @@ -2720,7 +2735,9 @@ if_ctx_t ctx; ctx = txq->ift_ctx; - return (get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx, txq->ift_gen) < TXQ_MIN_OCCUPANCY(ctx) + MAX_TX_DESC(ctx)); + return (get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx, + txq->ift_gen) < TXQ_MIN_OCCUPANCY(txq->ift_size) + + MAX_TX_DESC(ctx)); } static void @@ -2734,7 +2751,7 @@ cidx = txq->ift_cidx; gen = txq->ift_gen; - qsize = txq->ift_ctx->ifc_sctx->isc_ntxd; + qsize = txq->ift_size; mask = qsize-1; hasmap = txq->ift_sds.ifsd_map != NULL; ifsd_flags = txq->ift_sds.ifsd_flags; @@ -2760,7 +2777,7 @@ /* XXX we don't support any drivers that batch packets yet */ MPASS(m->m_nextpkt == NULL); - m_freem(m); + m_free(m); ifsd_m[cidx] = NULL; #if MEMORY_LOGGING txq->ift_dequeued++; @@ -2856,7 +2873,7 @@ if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) { DBG_COUNTER_INC(txq_drain_flushing); for (i = 0; i < avail; i++) { - m_freem(r->items[(cidx + i) & (r->size-1)]); + m_free(r->items[(cidx + i) & (r->size-1)]); r->items[(cidx + i) & (r->size-1)] = NULL; } return (avail); @@ -2903,7 +2920,7 @@ if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))) break; - if (desc_used > TXQ_MAX_DB_CONSUMED(ctx)) + if (desc_used > TXQ_MAX_DB_CONSUMED(txq->ift_size)) break; } @@ -2924,7 +2941,7 @@ } static void -_task_fn_tx(void *context, int pending) +_task_fn_tx(void *context) { iflib_txq_t txq = context; if_ctx_t ctx = txq->ift_ctx; @@ -2935,11 +2952,12 @@ } static void -_task_fn_rx(void *context, int pending) +_task_fn_rx(void *context) { iflib_rxq_t rxq = context; if_ctx_t ctx = rxq->ifr_ctx; bool more; + int rc; DBG_COUNTER_INC(task_fn_rxs); if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))) @@ -2950,7 +2968,8 @@ IFDI_INTR_ENABLE(ctx); else { DBG_COUNTER_INC(rx_intr_enables); - IFDI_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id); + rc = IFDI_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id); + KASSERT(rc != ENOTSUP, ("MSI-X support requires queue_intr_enable, but not implemented in driver")); } } if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))) @@ -2960,7 +2979,7 @@ } static void -_task_fn_admin(void *context, int pending) +_task_fn_admin(void *context) { if_ctx_t ctx = context; if_softc_ctx_t sctx = &ctx->ifc_softc_ctx; @@ -2990,7 +3009,7 @@ static void -_task_fn_iov(void *context, int pending) +_task_fn_iov(void *context) { if_ctx_t ctx = context; @@ -3049,8 +3068,7 @@ if_ctx_t ctx = if_getsoftc(ifp); iflib_txq_t txq; - struct mbuf *marr[8], **mp, *next; - int err, i, count, qidx; + int err, qidx; if (__predict_false((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) { DBG_COUNTER_INC(tx_frees); @@ -3058,6 +3076,7 @@ return (0); } + MPASS(m->m_nextpkt == NULL); qidx = 0; if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m)) qidx = QIDX(ctx, m); @@ -3077,6 +3096,7 @@ return (ENOBUFS); } #endif +#ifdef notyet qidx = count = 0; mp = marr; next = m; @@ -3098,22 +3118,21 @@ next = next->m_nextpkt; mp[i]->m_nextpkt = NULL; } +#endif DBG_COUNTER_INC(tx_seen); - err = ifmp_ring_enqueue(txq->ift_br[0], (void **)mp, count, TX_BATCH_SIZE); + err = ifmp_ring_enqueue(txq->ift_br[0], (void **)&m, 1, TX_BATCH_SIZE); - if (iflib_txq_can_drain(txq->ift_br[0])) - GROUPTASK_ENQUEUE(&txq->ift_task); if (err) { + GROUPTASK_ENQUEUE(&txq->ift_task); /* support forthcoming later */ #ifdef DRIVER_BACKPRESSURE txq->ift_closed = TRUE; #endif - for (i = 0; i < count; i++) - m_freem(mp[i]); ifmp_ring_check_drainage(txq->ift_br[0], TX_BATCH_SIZE); + m_freem(m); + } else if (TXQ_AVAIL(txq) < (txq->ift_size >> 1)) { + GROUPTASK_ENQUEUE(&txq->ift_task); } - if (count > nitems(marr)) - free(mp, M_IFLIB); return (err); } @@ -3138,13 +3157,13 @@ if_qflush(ifp); } -#define IFCAP_REINIT (IFCAP_HWCSUM|IFCAP_TSO4|IFCAP_TSO6|IFCAP_VLAN_HWTAGGING|IFCAP_VLAN_MTU | \ - IFCAP_VLAN_HWFILTER | IFCAP_VLAN_HWTSO) -#define IFCAP_FLAGS (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \ +#define IFCAP_FLAGS (IFCAP_TXCSUM_IPV6 | IFCAP_RXCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \ IFCAP_TSO4 | IFCAP_TSO6 | IFCAP_VLAN_HWTAGGING | \ IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | IFCAP_VLAN_HWTSO) +#define IFCAP_REINIT IFCAP_FLAGS + static int iflib_if_ioctl(if_t ifp, u_long command, caddr_t data) { @@ -3428,6 +3447,9 @@ if_ctx_t ctx; if_t ifp; if_softc_ctx_t scctx; + int i; + uint16_t main_txq; + uint16_t main_rxq; ctx = malloc(sizeof(* ctx), M_IFLIB, M_WAITOK|M_ZERO); @@ -3435,6 +3457,7 @@ if (sc == NULL) { sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO); device_set_softc(dev, ctx); + ctx->ifc_flags |= IFC_SC_ALLOCATED; } ctx->ifc_sctx = sctx; @@ -3447,28 +3470,112 @@ return (err); } iflib_add_device_sysctl_pre(ctx); + + scctx = &ctx->ifc_softc_ctx; + /* + * XXX sanity check that ntxd & nrxd are a power of 2 + */ + if (ctx->ifc_sysctl_ntxqs != 0) + scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs; + if (ctx->ifc_sysctl_nrxqs != 0) + scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs; + + for (i = 0; i < sctx->isc_ntxqs; i++) { + if (ctx->ifc_sysctl_ntxds[i] != 0) + scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i]; + else + scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i]; + } + + for (i = 0; i < sctx->isc_nrxqs; i++) { + if (ctx->ifc_sysctl_nrxds[i] != 0) + scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i]; + else + scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i]; + } + + for (i = 0; i < sctx->isc_nrxqs; i++) { + if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) { + device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n", + i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]); + scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i]; + } + if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) { + device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n", + i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]); + scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i]; + } + } + + for (i = 0; i < sctx->isc_ntxqs; i++) { + if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) { + device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n", + i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]); + scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i]; + } + if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) { + device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n", + i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]); + scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i]; + } + } + if ((err = IFDI_ATTACH_PRE(ctx)) != 0) { device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err); return (err); } + if (scctx->isc_ntxqsets_max) + scctx->isc_ntxqsets = min(scctx->isc_ntxqsets, scctx->isc_ntxqsets_max); + if (scctx->isc_nrxqsets_max) + scctx->isc_nrxqsets = min(scctx->isc_nrxqsets, scctx->isc_nrxqsets_max); + #ifdef ACPI_DMAR if (dmar_get_dma_tag(device_get_parent(dev), dev) != NULL) ctx->ifc_flags |= IFC_DMAR; #endif - scctx = &ctx->ifc_softc_ctx; msix_bar = scctx->isc_msix_bar; - if (scctx->isc_tx_nsegments > sctx->isc_ntxd / MAX_SINGLE_PACKET_FRACTION) - scctx->isc_tx_nsegments = max(1, sctx->isc_ntxd / MAX_SINGLE_PACKET_FRACTION); - if (scctx->isc_tx_tso_segments_max > sctx->isc_ntxd / MAX_SINGLE_PACKET_FRACTION) - scctx->isc_tx_tso_segments_max = max(1, sctx->isc_ntxd / MAX_SINGLE_PACKET_FRACTION); - ifp = ctx->ifc_ifp; - /* - * XXX sanity check that ntxd & nrxd are a power of 2 - */ + if(sctx->isc_flags & IFLIB_HAS_TXCQ) + main_txq = 1; + else + main_txq = 0; + + if(sctx->isc_flags & IFLIB_HAS_RXCQ) + main_rxq = 1; + else + main_rxq = 0; + + /* XXX change for per-queue sizes */ + device_printf(dev, "using %d tx descriptors and %d rx descriptors\n", + scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]); + for (i = 0; i < sctx->isc_nrxqs; i++) { + if (!powerof2(scctx->isc_nrxd[i])) { + /* round down instead? */ + device_printf(dev, "# rx descriptors must be a power of 2\n"); + err = EINVAL; + goto fail; + } + } + for (i = 0; i < sctx->isc_ntxqs; i++) { + if (!powerof2(scctx->isc_ntxd[i])) { + device_printf(dev, + "# tx descriptors must be a power of 2"); + err = EINVAL; + goto fail; + } + } + + if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] / + MAX_SINGLE_PACKET_FRACTION) + scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] / + MAX_SINGLE_PACKET_FRACTION); + if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] / + MAX_SINGLE_PACKET_FRACTION) + scctx->isc_tx_tso_segments_max = max(1, + scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION); /* * Protect the stack against modern hardware @@ -3482,7 +3589,7 @@ ifp->if_hw_tsomaxsegsize = scctx->isc_tx_tso_segsize_max; if (scctx->isc_rss_table_size == 0) scctx->isc_rss_table_size = 64; - scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;; + scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1; /* ** Now setup MSI or MSI/X, should ** return us the number of supported @@ -3520,7 +3627,7 @@ MPASS(msix == 1); rid = 1; } - if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx, &rid, "irq0")) != 0) { + if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) { device_printf(dev, "iflib_legacy_setup failed %d\n", err); goto fail_intr_free; } @@ -3536,6 +3643,7 @@ } *ctxp = ctx; + if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter); iflib_add_device_sysctl_post(ctx); return (0); fail_detach: @@ -3599,7 +3707,7 @@ led_destroy(ctx->ifc_led_dev); /* XXX drain any dependent tasks */ tqg = qgroup_if_io_tqg; - for (txq = ctx->ifc_txqs, i = 0, rxq = ctx->ifc_rxqs; i < NTXQSETS(ctx); i++, txq++) { + for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) { callout_drain(&txq->ift_timer); callout_drain(&txq->ift_db_check); if (txq->ift_task.gt_uniq != NULL) @@ -3616,6 +3724,7 @@ taskqgroup_detach(tqg, &ctx->ifc_vflr_task); IFDI_DETACH(ctx); + device_set_softc(ctx->ifc_dev, NULL); if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) { pci_release_msi(dev); } @@ -3633,6 +3742,9 @@ iflib_tx_structures_free(ctx); iflib_rx_structures_free(ctx); + if (ctx->ifc_flags & IFC_SC_ALLOCATED) + free(ctx->ifc_softc, M_IFLIB); + free(ctx, M_IFLIB); return (0); } @@ -3782,7 +3894,13 @@ MPASS(sctx->isc_txrx->ift_rxd_pkt_get); MPASS(sctx->isc_txrx->ift_rxd_refill); MPASS(sctx->isc_txrx->ift_rxd_flush); - MPASS(sctx->isc_nrxd); + + MPASS(sctx->isc_nrxd_min[0]); + MPASS(sctx->isc_nrxd_max[0]); + MPASS(sctx->isc_nrxd_default[0]); + MPASS(sctx->isc_ntxd_min[0]); + MPASS(sctx->isc_ntxd_max[0]); + MPASS(sctx->isc_ntxd_default[0]); } static int @@ -3796,7 +3914,6 @@ _iflib_assert(sctx); CTX_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev)); - MPASS(ctx->ifc_flags == 0); ifp = ctx->ifc_ifp = if_gethandle(IFT_ETHER); if (ifp == NULL) { @@ -3818,7 +3935,6 @@ if_setioctlfn(ifp, iflib_if_ioctl); if_settransmitfn(ifp, iflib_if_transmit); if_setqflushfn(ifp, iflib_if_qflush); - if_setgetcounterfn(ifp, iflib_if_get_counter); if_setflags(ifp, IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST); if_setcapabilities(ifp, 0); @@ -3842,16 +3958,17 @@ iflib_queues_alloc(if_ctx_t ctx) { if_shared_ctx_t sctx = ctx->ifc_sctx; + if_softc_ctx_t scctx = &ctx->ifc_softc_ctx; device_t dev = ctx->ifc_dev; - int nrxqsets = ctx->ifc_softc_ctx.isc_nrxqsets; - int ntxqsets = ctx->ifc_softc_ctx.isc_ntxqsets; + int nrxqsets = scctx->isc_nrxqsets; + int ntxqsets = scctx->isc_ntxqsets; iflib_txq_t txq; iflib_rxq_t rxq; iflib_fl_t fl = NULL; - int i, j, cpu, err, txconf, rxconf, fl_ifdi_offset; + int i, j, cpu, err, txconf, rxconf; iflib_dma_info_t ifdip; - uint32_t *rxqsizes = sctx->isc_rxqsizes; - uint32_t *txqsizes = sctx->isc_txqsizes; + uint32_t *rxqsizes = scctx->isc_rxqsizes; + uint32_t *txqsizes = scctx->isc_txqsizes; uint8_t nrxqs = sctx->isc_nrxqs; uint8_t ntxqs = sctx->isc_ntxqs; int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1; @@ -3860,10 +3977,11 @@ struct ifmp_ring **brscp; int nbuf_rings = 1; /* XXX determine dynamically */ - KASSERT(ntxqs > 0, ("number of queues must be at least 1")); - KASSERT(nrxqs > 0, ("number of queues must be at least 1")); + KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1")); + KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1")); brscp = NULL; + txq = NULL; rxq = NULL; /* Allocate the TX ring struct memory */ @@ -3891,8 +4009,6 @@ ctx->ifc_txqs = txq; ctx->ifc_rxqs = rxq; - txq = NULL; - rxq = NULL; /* * XXX handle allocation failure @@ -3916,6 +4032,11 @@ } txq->ift_ctx = ctx; txq->ift_id = i; + if (sctx->isc_flags & IFLIB_HAS_TXCQ) { + txq->ift_br_offset = 1; + } else { + txq->ift_br_offset = 0; + } /* XXX fix this */ txq->ift_timer.c_cpu = cpu; txq->ift_db_check.c_cpu = cpu; @@ -3970,10 +4091,10 @@ } rxq->ifr_ctx = ctx; rxq->ifr_id = i; - if (sctx->isc_flags & IFLIB_HAS_CQ) { - fl_ifdi_offset = 1; + if (sctx->isc_flags & IFLIB_HAS_RXCQ) { + rxq->ifr_fl_offset = 1; } else { - fl_ifdi_offset = 0; + rxq->ifr_fl_offset = 0; } rxq->ifr_nfl = nfree_lists; if (!(fl = @@ -3986,7 +4107,8 @@ for (j = 0; j < nfree_lists; j++) { rxq->ifr_fl[j].ifl_rxq = rxq; rxq->ifr_fl[j].ifl_id = j; - rxq->ifr_fl[j].ifl_ifdi = &rxq->ifr_ifdi[j + fl_ifdi_offset]; + rxq->ifr_fl[j].ifl_ifdi = + &rxq->ifr_ifdi[j + rxq->ifr_fl_offset]; } /* Allocate receive buffers for the ring*/ if (iflib_rxsd_alloc(rxq)) { @@ -4106,12 +4228,13 @@ for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) { #if defined(INET6) || defined(INET) tcp_lro_free(&rxq->ifr_lc); - if ((err = tcp_lro_init(&rxq->ifr_lc)) != 0) { + if ((err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp, + TCP_LRO_ENTRIES, min(1024, + ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset]))) != 0) { device_printf(ctx->ifc_dev, "LRO Initialization failed!\n"); goto fail; } rxq->ifr_lro_enabled = TRUE; - rxq->ifr_lc.ifp = ctx->ifc_ifp; #endif IFDI_RXQ_SETUP(ctx, rxq->ifr_id); } @@ -4142,7 +4265,7 @@ { iflib_rxq_t rxq = ctx->ifc_rxqs; - for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, rxq++) { + for (int i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) { iflib_rx_sds_free(rxq); } } @@ -4193,7 +4316,7 @@ struct taskqgroup *tqg; iflib_filter_info_t info; cpuset_t cpus; - task_fn_t *fn; + gtask_fn_t *fn; int tqrid, err; void *q; @@ -4254,7 +4377,7 @@ { struct grouptask *gtask; struct taskqgroup *tqg; - task_fn_t *fn; + gtask_fn_t *fn; void *q; switch (type) { @@ -4310,7 +4433,7 @@ iflib_filter_info_t info; struct grouptask *gtask; struct taskqgroup *tqg; - task_fn_t *fn; + gtask_fn_t *fn; int tqrid; void *q; int err; @@ -4385,7 +4508,7 @@ } void -iflib_config_gtask_init(if_ctx_t ctx, struct grouptask *gtask, task_fn_t *fn, +iflib_config_gtask_init(if_ctx_t ctx, struct grouptask *gtask, gtask_fn_t *fn, char *name) { @@ -4394,14 +4517,21 @@ } void -iflib_link_state_change(if_ctx_t ctx, int link_state) +iflib_config_gtask_deinit(struct grouptask *gtask) +{ + + taskqgroup_detach(qgroup_if_config_tqg, gtask); +} + +void +iflib_link_state_change(if_ctx_t ctx, int link_state, uint64_t baudrate) { if_t ifp = ctx->ifc_ifp; iflib_txq_t txq = ctx->ifc_txqs; -#if 0 + if_setbaudrate(ifp, baudrate); -#endif + /* If link down, disable watchdog */ if ((ctx->ifc_link_state == LINK_STATE_UP) && (link_state == LINK_STATE_DOWN)) { for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, txq++) @@ -4431,10 +4561,11 @@ } static int -iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, int cidx) +iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, int cidx, int budget) { - return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx)); + return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx, + budget)); } void @@ -4468,8 +4599,9 @@ int iflib_num_tx_queues, iflib_num_rx_queues; int err, admincnt, bar; - iflib_num_tx_queues = ctx->ifc_sysctl_ntxqs; - iflib_num_rx_queues = ctx->ifc_sysctl_nrxqs; + iflib_num_tx_queues = scctx->isc_ntxqsets; + iflib_num_rx_queues = scctx->isc_nrxqsets; + bar = ctx->ifc_softc_ctx.isc_msix_bar; admincnt = sctx->isc_admin_intrcnt; /* Override by tuneable */ @@ -4549,18 +4681,31 @@ if (queues > rss_getnumbuckets()) queues = rss_getnumbuckets(); #endif - if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queues) - queues = rx_queues = iflib_num_rx_queues; + if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt) + rx_queues = iflib_num_rx_queues; else rx_queues = queues; + /* + * We want this to be all logical CPUs by default + */ if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues) tx_queues = iflib_num_tx_queues; else - tx_queues = queues; + tx_queues = mp_ncpus; + + if (ctx->ifc_sysctl_qs_eq_override == 0) { +#ifdef INVARIANTS + if (tx_queues != rx_queues) + device_printf(dev, "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n", + min(rx_queues, tx_queues), min(rx_queues, tx_queues)); +#endif + tx_queues = min(rx_queues, tx_queues); + rx_queues = min(rx_queues, tx_queues); + } device_printf(dev, "using %d rx queues %d tx queues \n", rx_queues, tx_queues); - vectors = queues + admincnt; + vectors = rx_queues + admincnt; if ((err = pci_alloc_msix(dev, &vectors)) == 0) { device_printf(dev, "Using MSIX interrupts with %d vectors\n", vectors); @@ -4568,6 +4713,7 @@ scctx->isc_nrxqsets = rx_queues; scctx->isc_ntxqsets = tx_queues; scctx->isc_intr = IFLIB_INTR_MSIX; + return (vectors); } else { device_printf(dev, "failed to allocate %d msix vectors, err: %d - using MSI\n", vectors, err); @@ -4617,7 +4763,58 @@ return(rc); } +enum iflib_ndesc_handler { + IFLIB_NTXD_HANDLER, + IFLIB_NRXD_HANDLER, +}; + +static int +mp_ndesc_handler(SYSCTL_HANDLER_ARGS) +{ + if_ctx_t ctx = (void *)arg1; + enum iflib_ndesc_handler type = arg2; + char buf[256] = {0}; + uint16_t *ndesc; + char *p, *next; + int nqs, rc, i; + + MPASS(type == IFLIB_NTXD_HANDLER || type == IFLIB_NRXD_HANDLER); + + nqs = 8; + switch(type) { + case IFLIB_NTXD_HANDLER: + ndesc = ctx->ifc_sysctl_ntxds; + if (ctx->ifc_sctx) + nqs = ctx->ifc_sctx->isc_ntxqs; + break; + case IFLIB_NRXD_HANDLER: + ndesc = ctx->ifc_sysctl_nrxds; + if (ctx->ifc_sctx) + nqs = ctx->ifc_sctx->isc_nrxqs; + break; + } + if (nqs == 0) + nqs = 8; + + for (i=0; i<8; i++) { + if (i >= nqs) + break; + if (i) + strcat(buf, ","); + sprintf(strchr(buf, 0), "%d", ndesc[i]); + } + + rc = sysctl_handle_string(oidp, buf, sizeof(buf), req); + if (rc || req->newptr == NULL) + return rc; + for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p; + i++, p = strsep(&next, " ,")) { + ndesc[i] = strtoul(p, NULL, 10); + } + + return(rc); +} #define NAME_BUFLEN 32 static void @@ -4634,19 +4831,29 @@ CTLFLAG_RD, NULL, "IFLIB fields"); oid_list = SYSCTL_CHILDREN(node); + SYSCTL_ADD_STRING(ctx_list, oid_list, OID_AUTO, "driver_version", + CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version, 0, + "driver version"); + SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxqs", CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0, "# of txqs to use, 0 => use default #"); SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxqs", - CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0, - "# of txqs to use, 0 => use default #"); - SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxds", - CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxds, 0, - "# of tx descriptors to use, 0 => use default #"); - SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxds", - CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxds, 0, - "# of rx descriptors to use, 0 => use default #"); - + CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0, + "# of rxqs to use, 0 => use default #"); + SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_qs_enable", + CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0, + "permit #txq != #rxq"); + + /* XXX change for per-queue sizes */ + SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_ntxds", + CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NTXD_HANDLER, + mp_ndesc_handler, "A", + "list of # of tx descriptors to use, 0 = use default #"); + SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_nrxds", + CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NRXD_HANDLER, + mp_ndesc_handler, "A", + "list of # of rx descriptors to use, 0 = use default #"); } static void @@ -4700,7 +4907,7 @@ &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed"); SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_desc_avail", CTLFLAG_RD, - &txq->ift_mbuf_defrag_failed, "# of times no descriptors were available"); + &txq->ift_no_desc_avail, "# of times no descriptors were available"); SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "tx_map_failed", CTLFLAG_RD, &txq->ift_map_failed, "# of times dma map failed"); @@ -4763,7 +4970,7 @@ queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf, CTLFLAG_RD, NULL, "Queue Name"); queue_list = SYSCTL_CHILDREN(queue_node); - if (sctx->isc_flags & IFLIB_HAS_CQ) { + if (sctx->isc_flags & IFLIB_HAS_RXCQ) { SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_pidx", CTLFLAG_RD, &rxq->ifr_cq_pidx, 1, "Producer Index"); Index: head/sys/sys/_task.h =================================================================== --- head/sys/sys/_task.h +++ head/sys/sys/_task.h @@ -42,6 +42,7 @@ * (q) taskqueue lock */ typedef void task_fn_t(void *context, int pending); +typedef void gtask_fn_t(void *context); struct task { STAILQ_ENTRY(task) ta_link; /* (q) link for queue */ @@ -51,8 +52,16 @@ void *ta_context; /* (c) argument for handler */ }; +struct gtask { + STAILQ_ENTRY(gtask) ta_link; /* (q) link for queue */ + uint16_t ta_flags; /* (q) state flags */ + u_short ta_priority; /* (c) Priority */ + gtask_fn_t *ta_func; /* (c) task handler */ + void *ta_context; /* (c) argument for handler */ +}; + struct grouptask { - struct task gt_task; + struct gtask gt_task; void *gt_taskqueue; LIST_ENTRY(grouptask) gt_list; void *gt_uniq; Index: head/sys/sys/gtaskqueue.h =================================================================== --- head/sys/sys/gtaskqueue.h +++ head/sys/sys/gtaskqueue.h @@ -0,0 +1,125 @@ +/*- + * Copyright (c) 2014 Jeffrey Roberson + * Copyright (c) 2016 Matthew Macy + * All rights reserved. + * + * 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. + * + * $FreeBSD$ + */ + +#ifndef _SYS_GTASKQUEUE_H_ +#define _SYS_GTASKQUEUE_H_ +#include + +#ifndef _KERNEL +#error "no user-servicable parts inside" +#endif + +struct gtaskqueue; +typedef void (*gtaskqueue_enqueue_fn)(void *context); + +/* + * Taskqueue groups. Manages dynamic thread groups and irq binding for + * device and other tasks. + */ + +void gtaskqueue_block(struct gtaskqueue *queue); +void gtaskqueue_unblock(struct gtaskqueue *queue); + +int gtaskqueue_cancel(struct gtaskqueue *queue, struct gtask *gtask); +void gtaskqueue_drain(struct gtaskqueue *queue, struct gtask *task); +void gtaskqueue_drain_all(struct gtaskqueue *queue); + +int grouptaskqueue_enqueue(struct gtaskqueue *queue, struct gtask *task); +void taskqgroup_attach(struct taskqgroup *qgroup, struct grouptask *grptask, + void *uniq, int irq, char *name); +int taskqgroup_attach_cpu(struct taskqgroup *qgroup, struct grouptask *grptask, + void *uniq, int cpu, int irq, char *name); +void taskqgroup_detach(struct taskqgroup *qgroup, struct grouptask *gtask); +struct taskqgroup *taskqgroup_create(char *name); +void taskqgroup_destroy(struct taskqgroup *qgroup); +int taskqgroup_adjust(struct taskqgroup *qgroup, int cnt, int stride); + +#define TASK_ENQUEUED 0x1 +#define TASK_SKIP_WAKEUP 0x2 + + +#define GTASK_INIT(task, flags, priority, func, context) do { \ + (task)->ta_flags = flags; \ + (task)->ta_priority = (priority); \ + (task)->ta_func = (func); \ + (task)->ta_context = (context); \ +} while (0) + +#define GROUPTASK_INIT(gtask, priority, func, context) \ + GTASK_INIT(&(gtask)->gt_task, TASK_SKIP_WAKEUP, priority, func, context) + +#define GROUPTASK_ENQUEUE(gtask) \ + grouptaskqueue_enqueue((gtask)->gt_taskqueue, &(gtask)->gt_task) + +#define TASKQGROUP_DECLARE(name) \ +extern struct taskqgroup *qgroup_##name + + +#ifdef EARLY_AP_STARTUP +#define TASKQGROUP_DEFINE(name, cnt, stride) \ + \ +struct taskqgroup *qgroup_##name; \ + \ +static void \ +taskqgroup_define_##name(void *arg) \ +{ \ + qgroup_##name = taskqgroup_create(#name); \ + taskqgroup_adjust(qgroup_##name, (cnt), (stride)); \ +} \ + \ +SYSINIT(taskqgroup_##name, SI_SUB_INIT_IF, SI_ORDER_FIRST, \ + taskqgroup_define_##name, NULL) +#else +#define TASKQGROUP_DEFINE(name, cnt, stride) \ + \ +struct taskqgroup *qgroup_##name; \ + \ +static void \ +taskqgroup_define_##name(void *arg) \ +{ \ + qgroup_##name = taskqgroup_create(#name); \ +} \ + \ +SYSINIT(taskqgroup_##name, SI_SUB_INIT_IF, SI_ORDER_FIRST, \ + taskqgroup_define_##name, NULL); \ + \ +static void \ +taskqgroup_adjust_##name(void *arg) \ +{ \ + taskqgroup_adjust(qgroup_##name, (cnt), (stride)); \ +} \ + \ +SYSINIT(taskqgroup_adj_##name, SI_SUB_SMP, SI_ORDER_ANY, \ + taskqgroup_adjust_##name, NULL); \ + \ +struct __hack +#endif +TASKQGROUP_DECLARE(net); + +#endif /* !_SYS_GTASKQUEUE_H_ */ Index: head/sys/sys/taskqueue.h =================================================================== --- head/sys/sys/taskqueue.h +++ head/sys/sys/taskqueue.h @@ -204,78 +204,4 @@ taskqueue_enqueue_fn enqueue, void *context); -/* - * Taskqueue groups. Manages dynamic thread groups and irq binding for - * device and other tasks. - */ -int grouptaskqueue_enqueue(struct taskqueue *queue, struct task *task); -void taskqgroup_attach(struct taskqgroup *qgroup, struct grouptask *gtask, - void *uniq, int irq, char *name); -int taskqgroup_attach_cpu(struct taskqgroup *qgroup, struct grouptask *gtask, - void *uniq, int cpu, int irq, char *name); -void taskqgroup_detach(struct taskqgroup *qgroup, struct grouptask *gtask); -struct taskqgroup *taskqgroup_create(char *name); -void taskqgroup_destroy(struct taskqgroup *qgroup); -int taskqgroup_adjust(struct taskqgroup *qgroup, int cnt, int stride); - -#define TASK_SKIP_WAKEUP 0x1 - -#define GTASK_INIT(task, priority, func, context) do { \ - (task)->ta_pending = 0; \ - (task)->ta_priority = (priority); \ - (task)->ta_func = (func); \ - (task)->ta_context = (context); \ -} while (0) - -#define GROUPTASK_INIT(gtask, priority, func, context) \ - GTASK_INIT(&(gtask)->gt_task, priority, func, context) - -#define GROUPTASK_ENQUEUE(gtask) \ - grouptaskqueue_enqueue((gtask)->gt_taskqueue, &(gtask)->gt_task) - -#define TASKQGROUP_DECLARE(name) \ -extern struct taskqgroup *qgroup_##name - -#ifdef EARLY_AP_STARTUP -#define TASKQGROUP_DEFINE(name, cnt, stride) \ - \ -struct taskqgroup *qgroup_##name; \ - \ -static void \ -taskqgroup_define_##name(void *arg) \ -{ \ - qgroup_##name = taskqgroup_create(#name); \ - taskqgroup_adjust(qgroup_##name, (cnt), (stride)); \ -} \ - \ -SYSINIT(taskqgroup_##name, SI_SUB_INIT_IF, SI_ORDER_FIRST, \ - taskqgroup_define_##name, NULL) -#else -#define TASKQGROUP_DEFINE(name, cnt, stride) \ - \ -struct taskqgroup *qgroup_##name; \ - \ -static void \ -taskqgroup_define_##name(void *arg) \ -{ \ - qgroup_##name = taskqgroup_create(#name); \ -} \ - \ -SYSINIT(taskqgroup_##name, SI_SUB_INIT_IF, SI_ORDER_FIRST, \ - taskqgroup_define_##name, NULL); \ - \ -static void \ -taskqgroup_adjust_##name(void *arg) \ -{ \ - taskqgroup_adjust(qgroup_##name, (cnt), (stride)); \ -} \ - \ -SYSINIT(taskqgroup_adj_##name, SI_SUB_SMP, SI_ORDER_ANY, \ - taskqgroup_adjust_##name, NULL); \ - \ -struct __hack -#endif - -TASKQGROUP_DECLARE(net); - #endif /* !_SYS_TASKQUEUE_H_ */