Merge branch 'fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/djbw/async_tx
authorLinus Torvalds <torvalds@linux-foundation.org>
Wed, 19 Mar 2008 04:34:48 +0000 (21:34 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Wed, 19 Mar 2008 04:34:48 +0000 (21:34 -0700)
* 'fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/djbw/async_tx:
  async_tx: avoid the async xor_zero_sum path when src_cnt > device->max_xor
  fsldma: Fix the DMA halt when using DMA_INTERRUPT async_tx transfer.

drivers/acpi/ec.c
drivers/ide/ide-taskfile.c
include/linux/sched.h
include/linux/topology.h
kernel/sched.c
kernel/sched_debug.c
kernel/sched_fair.c
mm/slub.c
net/core/sock.c

index e7e197e3a4ffce2151d61f5e6b48f89c98087dfa..7222a18a03198d0bc70121d1246dc27af1d0629e 100644 (file)
@@ -129,7 +129,6 @@ static struct acpi_ec {
        struct mutex lock;
        wait_queue_head_t wait;
        struct list_head list;
-       atomic_t irq_count;
        u8 handlers_installed;
 } *boot_ec, *first_ec;
 
@@ -182,8 +181,6 @@ static int acpi_ec_wait(struct acpi_ec *ec, enum ec_event event, int force_poll)
 {
        int ret = 0;
 
-       atomic_set(&ec->irq_count, 0);
-
        if (unlikely(event == ACPI_EC_EVENT_OBF_1 &&
                     test_bit(EC_FLAGS_NO_OBF1_GPE, &ec->flags)))
                force_poll = 1;
@@ -230,7 +227,6 @@ static int acpi_ec_wait(struct acpi_ec *ec, enum ec_event event, int force_poll)
                while (time_before(jiffies, delay)) {
                        if (acpi_ec_check_status(ec, event))
                                goto end;
-                       msleep(5);
                }
        }
        pr_err(PREFIX "acpi_ec_wait timeout,"
@@ -533,13 +529,6 @@ static u32 acpi_ec_gpe_handler(void *data)
        struct acpi_ec *ec = data;
 
        pr_debug(PREFIX "~~~> interrupt\n");
-       atomic_inc(&ec->irq_count);
-       if (atomic_read(&ec->irq_count) > 5) {
-               pr_err(PREFIX "GPE storm detected, disabling EC GPE\n");
-               acpi_disable_gpe(NULL, ec->gpe, ACPI_ISR);
-               clear_bit(EC_FLAGS_GPE_MODE, &ec->flags);
-               return ACPI_INTERRUPT_HANDLED;
-       }
        clear_bit(EC_FLAGS_WAIT_GPE, &ec->flags);
        if (test_bit(EC_FLAGS_GPE_MODE, &ec->flags))
                wake_up(&ec->wait);
index 0518a2e948cf46f9317f8ef27e4a4121c6df08ae..4c86a8d84b4ccd02d8567b6b00f51429e5f9c82a 100644 (file)
@@ -422,6 +422,25 @@ void task_end_request(ide_drive_t *drive, struct request *rq, u8 stat)
                ide_end_request(drive, 1, rq->nr_sectors);
 }
 
+/*
+ * We got an interrupt on a task_in case, but no errors and no DRQ.
+ *
+ * It might be a spurious irq (shared irq), but it might be a
+ * command that had no output.
+ */
+static ide_startstop_t task_in_unexpected(ide_drive_t *drive, struct request *rq, u8 stat)
+{
+       /* Command all done? */
+       if (OK_STAT(stat, READY_STAT, BUSY_STAT)) {
+               task_end_request(drive, rq, stat);
+               return ide_stopped;
+       }
+
+       /* Assume it was a spurious irq */
+       ide_set_handler(drive, &task_in_intr, WAIT_WORSTCASE, NULL);
+       return ide_started;
+}
+
 /*
  * Handler for command with PIO data-in phase (Read/Read Multiple).
  */
@@ -431,18 +450,17 @@ static ide_startstop_t task_in_intr(ide_drive_t *drive)
        struct request *rq = HWGROUP(drive)->rq;
        u8 stat = ide_read_status(drive);
 
-       /* new way for dealing with premature shared PCI interrupts */
-       if (!OK_STAT(stat, DRQ_STAT, BAD_R_STAT)) {
-               if (stat & (ERR_STAT | DRQ_STAT))
-                       return task_error(drive, rq, __FUNCTION__, stat);
-               /* No data yet, so wait for another IRQ. */
-               ide_set_handler(drive, &task_in_intr, WAIT_WORSTCASE, NULL);
-               return ide_started;
-       }
+       /* Error? */
+       if (stat & ERR_STAT)
+               return task_error(drive, rq, __FUNCTION__, stat);
+
+       /* Didn't want any data? Odd. */
+       if (!(stat & DRQ_STAT))
+               return task_in_unexpected(drive, rq, stat);
 
        ide_pio_datablock(drive, rq, 0);
 
-       /* If it was the last datablock check status and finish transfer. */
+       /* Are we done? Check status and finish transfer. */
        if (!hwif->nleft) {
                stat = wait_drive_not_busy(drive);
                if (!OK_STAT(stat, 0, BAD_STAT))
index 11d8e9a74effd57af24a7f8143b093da6d34f486..3625fcaf5d0f2e3f08ff928a567793aa8f890309 100644 (file)
@@ -929,6 +929,9 @@ struct sched_entity {
        u64                     vruntime;
        u64                     prev_sum_exec_runtime;
 
+       u64                     last_wakeup;
+       u64                     avg_overlap;
+
 #ifdef CONFIG_SCHEDSTATS
        u64                     wait_start;
        u64                     wait_max;
index 2352f46160d3f3558eb3964646a3c36685470618..2d8dac8799cf3191084eea9a185ce6ac66051c1c 100644 (file)
                                | SD_BALANCE_FORK       \
                                | SD_BALANCE_EXEC       \
                                | SD_WAKE_AFFINE        \
-                               | SD_WAKE_IDLE          \
                                | SD_SHARE_PKG_RESOURCES\
                                | BALANCE_FOR_MC_POWER, \
        .last_balance           = jiffies,              \
index d1ad69b270ca7abc418871ce3160cc4bfbc04135..3f7c5eb254e205c24bc95bf5ac264e28d918d9b8 100644 (file)
@@ -1396,6 +1396,12 @@ task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
 {
        s64 delta;
 
+       /*
+        * Buddy candidates are cache hot:
+        */
+       if (&p->se == cfs_rq_of(&p->se)->next)
+               return 1;
+
        if (p->sched_class != &fair_sched_class)
                return 0;
 
@@ -1855,10 +1861,11 @@ out_activate:
                schedstat_inc(p, se.nr_wakeups_remote);
        update_rq_clock(rq);
        activate_task(rq, p, 1);
-       check_preempt_curr(rq, p);
        success = 1;
 
 out_running:
+       check_preempt_curr(rq, p);
+
        p->state = TASK_RUNNING;
 #ifdef CONFIG_SMP
        if (p->sched_class->task_wake_up)
@@ -1892,6 +1899,8 @@ static void __sched_fork(struct task_struct *p)
        p->se.exec_start                = 0;
        p->se.sum_exec_runtime          = 0;
        p->se.prev_sum_exec_runtime     = 0;
+       p->se.last_wakeup               = 0;
+       p->se.avg_overlap               = 0;
 
 #ifdef CONFIG_SCHEDSTATS
        p->se.wait_start                = 0;
index 4b5e24cf2f4a98387c849d8461b85a8a9c79aa55..ef358ba0768353cfd2e8bde743fa4901ab9141a7 100644 (file)
@@ -288,6 +288,7 @@ void proc_sched_show_task(struct task_struct *p, struct seq_file *m)
        PN(se.exec_start);
        PN(se.vruntime);
        PN(se.sum_exec_runtime);
+       PN(se.avg_overlap);
 
        nr_switches = p->nvcsw + p->nivcsw;
 
index f2cc59080efa4f9ae1e12aa30afe2e3fe08a86f9..b85cac4b5e25627d057773316f534116d3b95905 100644 (file)
@@ -73,13 +73,13 @@ unsigned int sysctl_sched_batch_wakeup_granularity = 10000000UL;
 
 /*
  * SCHED_OTHER wake-up granularity.
- * (default: 10 msec * (1 + ilog(ncpus)), units: nanoseconds)
+ * (default: 5 msec * (1 + ilog(ncpus)), units: nanoseconds)
  *
  * This option delays the preemption effects of decoupled workloads
  * and reduces their over-scheduling. Synchronous workloads will still
  * have immediate wakeup/sleep latencies.
  */
-unsigned int sysctl_sched_wakeup_granularity = 10000000UL;
+unsigned int sysctl_sched_wakeup_granularity = 5000000UL;
 
 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
 
@@ -556,6 +556,21 @@ enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int wakeup)
        account_entity_enqueue(cfs_rq, se);
 }
 
+static void update_avg(u64 *avg, u64 sample)
+{
+       s64 diff = sample - *avg;
+       *avg += diff >> 3;
+}
+
+static void update_avg_stats(struct cfs_rq *cfs_rq, struct sched_entity *se)
+{
+       if (!se->last_wakeup)
+               return;
+
+       update_avg(&se->avg_overlap, se->sum_exec_runtime - se->last_wakeup);
+       se->last_wakeup = 0;
+}
+
 static void
 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int sleep)
 {
@@ -566,6 +581,7 @@ dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int sleep)
 
        update_stats_dequeue(cfs_rq, se);
        if (sleep) {
+               update_avg_stats(cfs_rq, se);
 #ifdef CONFIG_SCHEDSTATS
                if (entity_is_task(se)) {
                        struct task_struct *tsk = task_of(se);
@@ -980,96 +996,121 @@ static inline int wake_idle(int cpu, struct task_struct *p)
 #endif
 
 #ifdef CONFIG_SMP
-static int select_task_rq_fair(struct task_struct *p, int sync)
+
+static const struct sched_class fair_sched_class;
+
+static int
+wake_affine(struct rq *rq, struct sched_domain *this_sd, struct rq *this_rq,
+           struct task_struct *p, int prev_cpu, int this_cpu, int sync,
+           int idx, unsigned long load, unsigned long this_load,
+           unsigned int imbalance)
 {
-       int cpu, this_cpu;
-       struct rq *rq;
-       struct sched_domain *sd, *this_sd = NULL;
-       int new_cpu;
+       struct task_struct *curr = this_rq->curr;
+       unsigned long tl = this_load;
+       unsigned long tl_per_task;
+
+       if (!(this_sd->flags & SD_WAKE_AFFINE))
+               return 0;
+
+       /*
+        * If the currently running task will sleep within
+        * a reasonable amount of time then attract this newly
+        * woken task:
+        */
+       if (sync && curr->sched_class == &fair_sched_class) {
+               if (curr->se.avg_overlap < sysctl_sched_migration_cost &&
+                               p->se.avg_overlap < sysctl_sched_migration_cost)
+                       return 1;
+       }
 
-       cpu      = task_cpu(p);
-       rq       = task_rq(p);
-       this_cpu = smp_processor_id();
-       new_cpu  = cpu;
+       schedstat_inc(p, se.nr_wakeups_affine_attempts);
+       tl_per_task = cpu_avg_load_per_task(this_cpu);
 
-       if (cpu == this_cpu)
-               goto out_set_cpu;
+       /*
+        * If sync wakeup then subtract the (maximum possible)
+        * effect of the currently running task from the load
+        * of the current CPU:
+        */
+       if (sync)
+               tl -= current->se.load.weight;
+
+       if ((tl <= load && tl + target_load(prev_cpu, idx) <= tl_per_task) ||
+                       100*(tl + p->se.load.weight) <= imbalance*load) {
+               /*
+                * This domain has SD_WAKE_AFFINE and
+                * p is cache cold in this domain, and
+                * there is no bad imbalance.
+                */
+               schedstat_inc(this_sd, ttwu_move_affine);
+               schedstat_inc(p, se.nr_wakeups_affine);
 
+               return 1;
+       }
+       return 0;
+}
+
+static int select_task_rq_fair(struct task_struct *p, int sync)
+{
+       struct sched_domain *sd, *this_sd = NULL;
+       int prev_cpu, this_cpu, new_cpu;
+       unsigned long load, this_load;
+       struct rq *rq, *this_rq;
+       unsigned int imbalance;
+       int idx;
+
+       prev_cpu        = task_cpu(p);
+       rq              = task_rq(p);
+       this_cpu        = smp_processor_id();
+       this_rq         = cpu_rq(this_cpu);
+       new_cpu         = prev_cpu;
+
+       /*
+        * 'this_sd' is the first domain that both
+        * this_cpu and prev_cpu are present in:
+        */
        for_each_domain(this_cpu, sd) {
-               if (cpu_isset(cpu, sd->span)) {
+               if (cpu_isset(prev_cpu, sd->span)) {
                        this_sd = sd;
                        break;
                }
        }
 
        if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
-               goto out_set_cpu;
+               goto out;
 
        /*
         * Check for affine wakeup and passive balancing possibilities.
         */
-       if (this_sd) {
-               int idx = this_sd->wake_idx;
-               unsigned int imbalance;
-               unsigned long load, this_load;
-
-               imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
-
-               load = source_load(cpu, idx);
-               this_load = target_load(this_cpu, idx);
-
-               new_cpu = this_cpu; /* Wake to this CPU if we can */
-
-               if (this_sd->flags & SD_WAKE_AFFINE) {
-                       unsigned long tl = this_load;
-                       unsigned long tl_per_task;
-
-                       /*
-                        * Attract cache-cold tasks on sync wakeups:
-                        */
-                       if (sync && !task_hot(p, rq->clock, this_sd))
-                               goto out_set_cpu;
-
-                       schedstat_inc(p, se.nr_wakeups_affine_attempts);
-                       tl_per_task = cpu_avg_load_per_task(this_cpu);
-
-                       /*
-                        * If sync wakeup then subtract the (maximum possible)
-                        * effect of the currently running task from the load
-                        * of the current CPU:
-                        */
-                       if (sync)
-                               tl -= current->se.load.weight;
-
-                       if ((tl <= load &&
-                               tl + target_load(cpu, idx) <= tl_per_task) ||
-                              100*(tl + p->se.load.weight) <= imbalance*load) {
-                               /*
-                                * This domain has SD_WAKE_AFFINE and
-                                * p is cache cold in this domain, and
-                                * there is no bad imbalance.
-                                */
-                               schedstat_inc(this_sd, ttwu_move_affine);
-                               schedstat_inc(p, se.nr_wakeups_affine);
-                               goto out_set_cpu;
-                       }
-               }
+       if (!this_sd)
+               goto out;
 
-               /*
-                * Start passive balancing when half the imbalance_pct
-                * limit is reached.
-                */
-               if (this_sd->flags & SD_WAKE_BALANCE) {
-                       if (imbalance*this_load <= 100*load) {
-                               schedstat_inc(this_sd, ttwu_move_balance);
-                               schedstat_inc(p, se.nr_wakeups_passive);
-                               goto out_set_cpu;
-                       }
+       idx = this_sd->wake_idx;
+
+       imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
+
+       load = source_load(prev_cpu, idx);
+       this_load = target_load(this_cpu, idx);
+
+       if (wake_affine(rq, this_sd, this_rq, p, prev_cpu, this_cpu, sync, idx,
+                                    load, this_load, imbalance))
+               return this_cpu;
+
+       if (prev_cpu == this_cpu)
+               goto out;
+
+       /*
+        * Start passive balancing when half the imbalance_pct
+        * limit is reached.
+        */
+       if (this_sd->flags & SD_WAKE_BALANCE) {
+               if (imbalance*this_load <= 100*load) {
+                       schedstat_inc(this_sd, ttwu_move_balance);
+                       schedstat_inc(p, se.nr_wakeups_passive);
+                       return this_cpu;
                }
        }
 
-       new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */
-out_set_cpu:
+out:
        return wake_idle(new_cpu, p);
 }
 #endif /* CONFIG_SMP */
@@ -1092,6 +1133,10 @@ static void check_preempt_wakeup(struct rq *rq, struct task_struct *p)
                return;
        }
 
+       se->last_wakeup = se->sum_exec_runtime;
+       if (unlikely(se == pse))
+               return;
+
        cfs_rq_of(pse)->next = pse;
 
        /*
index 96d63eb3ab179528efd1ea8c64cbc6358212a967..ca71d5b81e4a436b43bb183728650887eef4f0e9 100644 (file)
--- a/mm/slub.c
+++ b/mm/slub.c
@@ -1536,9 +1536,15 @@ new_slab:
         * That is only possible if certain conditions are met that are being
         * checked when a slab is created.
         */
-       if (!(gfpflags & __GFP_NORETRY) && (s->flags & __PAGE_ALLOC_FALLBACK))
-               return kmalloc_large(s->objsize, gfpflags);
-
+       if (!(gfpflags & __GFP_NORETRY) &&
+                               (s->flags & __PAGE_ALLOC_FALLBACK)) {
+               if (gfpflags & __GFP_WAIT)
+                       local_irq_enable();
+               object = kmalloc_large(s->objsize, gfpflags);
+               if (gfpflags & __GFP_WAIT)
+                       local_irq_disable();
+               return object;
+       }
        return NULL;
 debug:
        if (!alloc_debug_processing(s, c->page, object, addr))
index 09cb3a74de7f35a976f5b21fbc33a67b3a8224b9..2654c147c004a736ed0693f7b2b60e915b7505ea 100644 (file)
@@ -1621,7 +1621,7 @@ static void sock_def_readable(struct sock *sk, int len)
 {
        read_lock(&sk->sk_callback_lock);
        if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
-               wake_up_interruptible(sk->sk_sleep);
+               wake_up_interruptible_sync(sk->sk_sleep);
        sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
        read_unlock(&sk->sk_callback_lock);
 }
@@ -1635,7 +1635,7 @@ static void sock_def_write_space(struct sock *sk)
         */
        if ((atomic_read(&sk->sk_wmem_alloc) << 1) <= sk->sk_sndbuf) {
                if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
-                       wake_up_interruptible(sk->sk_sleep);
+                       wake_up_interruptible_sync(sk->sk_sleep);
 
                /* Should agree with poll, otherwise some programs break */
                if (sock_writeable(sk))