]> git.ipfire.org Git - thirdparty/kernel/linux.git/commitdiff
sched_ext: Implement scx_bpf_dispatch[_vtime]_from_dsq()
authorTejun Heo <tj@kernel.org>
Mon, 9 Sep 2024 23:42:47 +0000 (13:42 -1000)
committerTejun Heo <tj@kernel.org>
Mon, 9 Sep 2024 23:42:47 +0000 (13:42 -1000)
Once a task is put into a DSQ, the allowed operations are fairly limited.
Tasks in the built-in local and global DSQs are executed automatically and,
ignoring dequeue, there is only one way a task in a user DSQ can be
manipulated - scx_bpf_consume() moves the first task to the dispatching
local DSQ. This inflexibility sometimes gets in the way and is an area where
multiple feature requests have been made.

Implement scx_bpf_dispatch[_vtime]_from_dsq(), which can be called during
DSQ iteration and can move the task to any DSQ - local DSQs, global DSQ and
user DSQs. The kfuncs can be called from ops.dispatch() and any BPF context
which dosen't hold a rq lock including BPF timers and SYSCALL programs.

This is an expansion of an earlier patch which only allowed moving into the
dispatching local DSQ:

  http://lkml.kernel.org/r/Zn4Cw4FDTmvXnhaf@slm.duckdns.org

v2: Remove @slice and @vtime from scx_bpf_dispatch_from_dsq[_vtime]() as
    they push scx_bpf_dispatch_from_dsq_vtime() over the kfunc argument
    count limit and often won't be needed anyway. Instead provide
    scx_bpf_dispatch_from_dsq_set_{slice|vtime}() kfuncs which can be called
    only when needed and override the specified parameter for the subsequent
    dispatch.

Signed-off-by: Tejun Heo <tj@kernel.org>
Cc: Daniel Hodges <hodges.daniel.scott@gmail.com>
Cc: David Vernet <void@manifault.com>
Cc: Changwoo Min <multics69@gmail.com>
Cc: Andrea Righi <andrea.righi@linux.dev>
Cc: Dan Schatzberg <schatzberg.dan@gmail.com>
kernel/sched/ext.c
tools/sched_ext/include/scx/common.bpf.h

index 44a2d1b9e8377c1db0db55d3d162abe62ed3fbd2..b6e8c224f894558fb2ae9a3f5a4e6af6c74ebfe7 100644 (file)
@@ -1158,6 +1158,11 @@ static __always_inline bool scx_kf_allowed_on_arg_tasks(u32 mask,
        return true;
 }
 
+static bool scx_kf_allowed_if_unlocked(void)
+{
+       return !current->scx.kf_mask;
+}
+
 /**
  * nldsq_next_task - Iterate to the next task in a non-local DSQ
  * @dsq: user dsq being interated
@@ -1211,13 +1216,20 @@ enum scx_dsq_iter_flags {
        /* iterate in the reverse dispatch order */
        SCX_DSQ_ITER_REV                = 1U << 16,
 
+       __SCX_DSQ_ITER_HAS_SLICE        = 1U << 30,
+       __SCX_DSQ_ITER_HAS_VTIME        = 1U << 31,
+
        __SCX_DSQ_ITER_USER_FLAGS       = SCX_DSQ_ITER_REV,
-       __SCX_DSQ_ITER_ALL_FLAGS        = __SCX_DSQ_ITER_USER_FLAGS,
+       __SCX_DSQ_ITER_ALL_FLAGS        = __SCX_DSQ_ITER_USER_FLAGS |
+                                         __SCX_DSQ_ITER_HAS_SLICE |
+                                         __SCX_DSQ_ITER_HAS_VTIME,
 };
 
 struct bpf_iter_scx_dsq_kern {
        struct scx_dsq_list_node        cursor;
        struct scx_dispatch_q           *dsq;
+       u64                             slice;
+       u64                             vtime;
 } __attribute__((aligned(8)));
 
 struct bpf_iter_scx_dsq {
@@ -5872,7 +5884,7 @@ __bpf_kfunc_start_defs();
  * scx_bpf_dispatch - Dispatch a task into the FIFO queue of a DSQ
  * @p: task_struct to dispatch
  * @dsq_id: DSQ to dispatch to
- * @slice: duration @p can run for in nsecs
+ * @slice: duration @p can run for in nsecs, 0 to keep the current value
  * @enq_flags: SCX_ENQ_*
  *
  * Dispatch @p into the FIFO queue of the DSQ identified by @dsq_id. It is safe
@@ -5922,7 +5934,7 @@ __bpf_kfunc void scx_bpf_dispatch(struct task_struct *p, u64 dsq_id, u64 slice,
  * scx_bpf_dispatch_vtime - Dispatch a task into the vtime priority queue of a DSQ
  * @p: task_struct to dispatch
  * @dsq_id: DSQ to dispatch to
- * @slice: duration @p can run for in nsecs
+ * @slice: duration @p can run for in nsecs, 0 to keep the current value
  * @vtime: @p's ordering inside the vtime-sorted queue of the target DSQ
  * @enq_flags: SCX_ENQ_*
  *
@@ -5963,6 +5975,118 @@ static const struct btf_kfunc_id_set scx_kfunc_set_enqueue_dispatch = {
        .set                    = &scx_kfunc_ids_enqueue_dispatch,
 };
 
+static bool scx_dispatch_from_dsq(struct bpf_iter_scx_dsq_kern *kit,
+                                 struct task_struct *p, u64 dsq_id,
+                                 u64 enq_flags)
+{
+       struct scx_dispatch_q *src_dsq = kit->dsq, *dst_dsq;
+       struct rq *this_rq, *src_rq, *dst_rq, *locked_rq;
+       bool dispatched = false;
+       bool in_balance;
+       unsigned long flags;
+
+       if (!scx_kf_allowed_if_unlocked() && !scx_kf_allowed(SCX_KF_DISPATCH))
+               return false;
+
+       /*
+        * Can be called from either ops.dispatch() locking this_rq() or any
+        * context where no rq lock is held. If latter, lock @p's task_rq which
+        * we'll likely need anyway.
+        */
+       src_rq = task_rq(p);
+
+       local_irq_save(flags);
+       this_rq = this_rq();
+       in_balance = this_rq->scx.flags & SCX_RQ_IN_BALANCE;
+
+       if (in_balance) {
+               if (this_rq != src_rq) {
+                       raw_spin_rq_unlock(this_rq);
+                       raw_spin_rq_lock(src_rq);
+               }
+       } else {
+               raw_spin_rq_lock(src_rq);
+       }
+
+       locked_rq = src_rq;
+       raw_spin_lock(&src_dsq->lock);
+
+       /*
+        * Did someone else get to it? @p could have already left $src_dsq, got
+        * re-enqueud, or be in the process of being consumed by someone else.
+        */
+       if (unlikely(p->scx.dsq != src_dsq ||
+                    u32_before(kit->cursor.priv, p->scx.dsq_seq) ||
+                    p->scx.holding_cpu >= 0) ||
+           WARN_ON_ONCE(src_rq != task_rq(p))) {
+               raw_spin_unlock(&src_dsq->lock);
+               goto out;
+       }
+
+       /* @p is still on $src_dsq and stable, determine the destination */
+       dst_dsq = find_dsq_for_dispatch(this_rq, dsq_id, p);
+
+       if (dst_dsq->id == SCX_DSQ_LOCAL) {
+               dst_rq = container_of(dst_dsq, struct rq, scx.local_dsq);
+               if (!task_can_run_on_remote_rq(p, dst_rq, true)) {
+                       dst_dsq = &scx_dsq_global;
+                       dst_rq = src_rq;
+               }
+       } else {
+               /* no need to migrate if destination is a non-local DSQ */
+               dst_rq = src_rq;
+       }
+
+       /*
+        * Move @p into $dst_dsq. If $dst_dsq is the local DSQ of a different
+        * CPU, @p will be migrated.
+        */
+       if (dst_dsq->id == SCX_DSQ_LOCAL) {
+               /* @p is going from a non-local DSQ to a local DSQ */
+               if (src_rq == dst_rq) {
+                       task_unlink_from_dsq(p, src_dsq);
+                       move_local_task_to_local_dsq(p, enq_flags,
+                                                    src_dsq, dst_rq);
+                       raw_spin_unlock(&src_dsq->lock);
+               } else {
+                       raw_spin_unlock(&src_dsq->lock);
+                       move_remote_task_to_local_dsq(p, enq_flags,
+                                                     src_rq, dst_rq);
+                       locked_rq = dst_rq;
+               }
+       } else {
+               /*
+                * @p is going from a non-local DSQ to a non-local DSQ. As
+                * $src_dsq is already locked, do an abbreviated dequeue.
+                */
+               task_unlink_from_dsq(p, src_dsq);
+               p->scx.dsq = NULL;
+               raw_spin_unlock(&src_dsq->lock);
+
+               if (kit->cursor.flags & __SCX_DSQ_ITER_HAS_VTIME)
+                       p->scx.dsq_vtime = kit->vtime;
+               dispatch_enqueue(dst_dsq, p, enq_flags);
+       }
+
+       if (kit->cursor.flags & __SCX_DSQ_ITER_HAS_SLICE)
+               p->scx.slice = kit->slice;
+
+       dispatched = true;
+out:
+       if (in_balance) {
+               if (this_rq != locked_rq) {
+                       raw_spin_rq_unlock(locked_rq);
+                       raw_spin_rq_lock(this_rq);
+               }
+       } else {
+               raw_spin_rq_unlock_irqrestore(locked_rq, flags);
+       }
+
+       kit->cursor.flags &= ~(__SCX_DSQ_ITER_HAS_SLICE |
+                              __SCX_DSQ_ITER_HAS_VTIME);
+       return dispatched;
+}
+
 __bpf_kfunc_start_defs();
 
 /**
@@ -6042,12 +6166,112 @@ __bpf_kfunc bool scx_bpf_consume(u64 dsq_id)
        }
 }
 
+/**
+ * scx_bpf_dispatch_from_dsq_set_slice - Override slice when dispatching from DSQ
+ * @it__iter: DSQ iterator in progress
+ * @slice: duration the dispatched task can run for in nsecs
+ *
+ * Override the slice of the next task that will be dispatched from @it__iter
+ * using scx_bpf_dispatch_from_dsq[_vtime](). If this function is not called,
+ * the previous slice duration is kept.
+ */
+__bpf_kfunc void scx_bpf_dispatch_from_dsq_set_slice(
+                               struct bpf_iter_scx_dsq *it__iter, u64 slice)
+{
+       struct bpf_iter_scx_dsq_kern *kit = (void *)it__iter;
+
+       kit->slice = slice;
+       kit->cursor.flags |= __SCX_DSQ_ITER_HAS_SLICE;
+}
+
+/**
+ * scx_bpf_dispatch_from_dsq_set_vtime - Override vtime when dispatching from DSQ
+ * @it__iter: DSQ iterator in progress
+ * @vtime: task's ordering inside the vtime-sorted queue of the target DSQ
+ *
+ * Override the vtime of the next task that will be dispatched from @it__iter
+ * using scx_bpf_dispatch_from_dsq_vtime(). If this function is not called, the
+ * previous slice vtime is kept. If scx_bpf_dispatch_from_dsq() is used to
+ * dispatch the next task, the override is ignored and cleared.
+ */
+__bpf_kfunc void scx_bpf_dispatch_from_dsq_set_vtime(
+                               struct bpf_iter_scx_dsq *it__iter, u64 vtime)
+{
+       struct bpf_iter_scx_dsq_kern *kit = (void *)it__iter;
+
+       kit->vtime = vtime;
+       kit->cursor.flags |= __SCX_DSQ_ITER_HAS_VTIME;
+}
+
+/**
+ * scx_bpf_dispatch_from_dsq - Move a task from DSQ iteration to a DSQ
+ * @it__iter: DSQ iterator in progress
+ * @p: task to transfer
+ * @dsq_id: DSQ to move @p to
+ * @enq_flags: SCX_ENQ_*
+ *
+ * Transfer @p which is on the DSQ currently iterated by @it__iter to the DSQ
+ * specified by @dsq_id. All DSQs - local DSQs, global DSQ and user DSQs - can
+ * be the destination.
+ *
+ * For the transfer to be successful, @p must still be on the DSQ and have been
+ * queued before the DSQ iteration started. This function doesn't care whether
+ * @p was obtained from the DSQ iteration. @p just has to be on the DSQ and have
+ * been queued before the iteration started.
+ *
+ * @p's slice is kept by default. Use scx_bpf_dispatch_from_dsq_set_slice() to
+ * update.
+ *
+ * Can be called from ops.dispatch() or any BPF context which doesn't hold a rq
+ * lock (e.g. BPF timers or SYSCALL programs).
+ *
+ * Returns %true if @p has been consumed, %false if @p had already been consumed
+ * or dequeued.
+ */
+__bpf_kfunc bool scx_bpf_dispatch_from_dsq(struct bpf_iter_scx_dsq *it__iter,
+                                          struct task_struct *p, u64 dsq_id,
+                                          u64 enq_flags)
+{
+       return scx_dispatch_from_dsq((struct bpf_iter_scx_dsq_kern *)it__iter,
+                                    p, dsq_id, enq_flags);
+}
+
+/**
+ * scx_bpf_dispatch_vtime_from_dsq - Move a task from DSQ iteration to a PRIQ DSQ
+ * @it__iter: DSQ iterator in progress
+ * @p: task to transfer
+ * @dsq_id: DSQ to move @p to
+ * @enq_flags: SCX_ENQ_*
+ *
+ * Transfer @p which is on the DSQ currently iterated by @it__iter to the
+ * priority queue of the DSQ specified by @dsq_id. The destination must be a
+ * user DSQ as only user DSQs support priority queue.
+ *
+ * @p's slice and vtime are kept by default. Use
+ * scx_bpf_dispatch_from_dsq_set_slice() and
+ * scx_bpf_dispatch_from_dsq_set_vtime() to update.
+ *
+ * All other aspects are identical to scx_bpf_dispatch_from_dsq(). See
+ * scx_bpf_dispatch_vtime() for more information on @vtime.
+ */
+__bpf_kfunc bool scx_bpf_dispatch_vtime_from_dsq(struct bpf_iter_scx_dsq *it__iter,
+                                                struct task_struct *p, u64 dsq_id,
+                                                u64 enq_flags)
+{
+       return scx_dispatch_from_dsq((struct bpf_iter_scx_dsq_kern *)it__iter,
+                                    p, dsq_id, enq_flags | SCX_ENQ_DSQ_PRIQ);
+}
+
 __bpf_kfunc_end_defs();
 
 BTF_KFUNCS_START(scx_kfunc_ids_dispatch)
 BTF_ID_FLAGS(func, scx_bpf_dispatch_nr_slots)
 BTF_ID_FLAGS(func, scx_bpf_dispatch_cancel)
 BTF_ID_FLAGS(func, scx_bpf_consume)
+BTF_ID_FLAGS(func, scx_bpf_dispatch_from_dsq_set_slice)
+BTF_ID_FLAGS(func, scx_bpf_dispatch_from_dsq_set_vtime)
+BTF_ID_FLAGS(func, scx_bpf_dispatch_from_dsq, KF_RCU)
+BTF_ID_FLAGS(func, scx_bpf_dispatch_vtime_from_dsq, KF_RCU)
 BTF_KFUNCS_END(scx_kfunc_ids_dispatch)
 
 static const struct btf_kfunc_id_set scx_kfunc_set_dispatch = {
@@ -6144,6 +6368,8 @@ __bpf_kfunc_end_defs();
 
 BTF_KFUNCS_START(scx_kfunc_ids_unlocked)
 BTF_ID_FLAGS(func, scx_bpf_create_dsq, KF_SLEEPABLE)
+BTF_ID_FLAGS(func, scx_bpf_dispatch_from_dsq, KF_RCU)
+BTF_ID_FLAGS(func, scx_bpf_dispatch_vtime_from_dsq, KF_RCU)
 BTF_KFUNCS_END(scx_kfunc_ids_unlocked)
 
 static const struct btf_kfunc_id_set scx_kfunc_set_unlocked = {
index 457462b1996690818c36b75f64d250685e692137..f538c75db1835dca7742d86650a9ff35c3843694 100644 (file)
@@ -35,6 +35,10 @@ void scx_bpf_dispatch_vtime(struct task_struct *p, u64 dsq_id, u64 slice, u64 vt
 u32 scx_bpf_dispatch_nr_slots(void) __ksym;
 void scx_bpf_dispatch_cancel(void) __ksym;
 bool scx_bpf_consume(u64 dsq_id) __ksym;
+void scx_bpf_dispatch_from_dsq_set_slice(struct bpf_iter_scx_dsq *it__iter, u64 slice) __ksym;
+void scx_bpf_dispatch_from_dsq_set_vtime(struct bpf_iter_scx_dsq *it__iter, u64 vtime) __ksym;
+bool scx_bpf_dispatch_from_dsq(struct bpf_iter_scx_dsq *it__iter, struct task_struct *p, u64 dsq_id, u64 enq_flags) __ksym __weak;
+bool scx_bpf_dispatch_vtime_from_dsq(struct bpf_iter_scx_dsq *it__iter, struct task_struct *p, u64 dsq_id, u64 enq_flags) __ksym __weak;
 u32 scx_bpf_reenqueue_local(void) __ksym;
 void scx_bpf_kick_cpu(s32 cpu, u64 flags) __ksym;
 s32 scx_bpf_dsq_nr_queued(u64 dsq_id) __ksym;
@@ -63,6 +67,12 @@ s32 scx_bpf_task_cpu(const struct task_struct *p) __ksym;
 struct rq *scx_bpf_cpu_rq(s32 cpu) __ksym;
 struct cgroup *scx_bpf_task_cgroup(struct task_struct *p) __ksym;
 
+/*
+ * Use the following as @it__iter when calling
+ * scx_bpf_dispatch[_vtime]_from_dsq() from within bpf_for_each() loops.
+ */
+#define BPF_FOR_EACH_ITER      (&___it)
+
 static inline __attribute__((format(printf, 1, 2)))
 void ___scx_bpf_bstr_format_checker(const char *fmt, ...) {}