@@ -418,6 +418,32 @@ config ARCH_WANT_IRQS_OFF_ACTIVATE_MM
irqs disabled over activate_mm. Architectures that do IPI based TLB
shootdowns should enable this.
+# Enable "lazy TLB", which means a user->kernel thread context switch does not
+# switch the mm to init_mm and the kernel thread takes a reference to the user
+# mm to provide its kernel mapping. This is how Linux has traditionally worked
+# (see Documentation/vm/active_mm.rst), for performance. Switching to and from
+# idle thread is a performance-critical case.
+#
+# If mm context switches are inexpensive or free (in the case of NOMMU) then
+# this could be disabled.
+#
+# It would make sense to have this depend on MMU, but need to audit and test
+# the NOMMU architectures for lazy mm refcounting first.
+config MMU_LAZY_TLB
+ def_bool y
+ depends on !NO_MMU_LAZY_TLB
+
+# This allows archs to disable MMU_LAZY_TLB. mmgrab/mmdrop in arch/ code has
+# to be audited and switched to _lazy_tlb postfix as necessary.
+config NO_MMU_LAZY_TLB
+ def_bool n
+
+# Use normal mm refcounting for MMU_LAZY_TLB kernel thread references.
+# For now, this must be enabled if MMU_LAZY_TLB is enabled.
+config MMU_LAZY_TLB_REFCOUNT
+ def_bool y
+ depends on MMU_LAZY_TLB
+
config ARCH_HAVE_NMI_SAFE_CMPXCHG
bool
@@ -52,12 +52,21 @@ static inline void mmdrop(struct mm_struct *mm)
/* Helpers for lazy TLB mm refcounting */
static inline void mmgrab_lazy_tlb(struct mm_struct *mm)
{
- mmgrab(mm);
+ if (IS_ENABLED(CONFIG_MMU_LAZY_TLB_REFCOUNT))
+ mmgrab(mm);
}
static inline void mmdrop_lazy_tlb(struct mm_struct *mm)
{
- mmdrop(mm);
+ if (IS_ENABLED(CONFIG_MMU_LAZY_TLB_REFCOUNT)) {
+ mmdrop(mm);
+ } else {
+ /*
+ * mmdrop_lazy_tlb must provide a full memory barrier, see the
+ * membarrier comment finish_task_switch which relies on this.
+ */
+ smp_mb();
+ }
}
/**
@@ -4171,7 +4171,7 @@ static struct rq *finish_task_switch(struct task_struct *prev)
__releases(rq->lock)
{
struct rq *rq = this_rq();
- struct mm_struct *mm = rq->prev_mm;
+ struct mm_struct *mm = NULL;
long prev_state;
/*
@@ -4190,7 +4190,10 @@ static struct rq *finish_task_switch(struct task_struct *prev)
current->comm, current->pid, preempt_count()))
preempt_count_set(FORK_PREEMPT_COUNT);
- rq->prev_mm = NULL;
+#ifdef CONFIG_MMU_LAZY_TLB_REFCOUNT
+ mm = rq->prev_lazy_mm;
+ rq->prev_lazy_mm = NULL;
+#endif
/*
* A task struct has one reference for the use as "current".
@@ -4282,22 +4285,10 @@ asmlinkage __visible void schedule_tail(struct task_struct *prev)
calculate_sigpending();
}
-/*
- * context_switch - switch to the new MM and the new thread's register state.
- */
-static __always_inline struct rq *
-context_switch(struct rq *rq, struct task_struct *prev,
- struct task_struct *next, struct rq_flags *rf)
+static __always_inline void
+context_switch_mm(struct rq *rq, struct task_struct *prev,
+ struct task_struct *next)
{
- prepare_task_switch(rq, prev, next);
-
- /*
- * For paravirt, this is coupled with an exit in switch_to to
- * combine the page table reload and the switch backend into
- * one hypercall.
- */
- arch_start_context_switch(prev);
-
/*
* kernel -> kernel lazy + transfer active
* user -> kernel lazy + mmgrab_lazy_tlb() active
@@ -4326,11 +4317,57 @@ context_switch(struct rq *rq, struct task_struct *prev,
switch_mm_irqs_off(prev->active_mm, next->mm, next);
if (!prev->mm) { // from kernel
- /* will mmdrop_lazy_tlb() in finish_task_switch(). */
- rq->prev_mm = prev->active_mm;
+#ifdef CONFIG_MMU_LAZY_TLB_REFCOUNT
+ /* Will mmdrop_lazy_tlb() in finish_task_switch(). */
+ rq->prev_lazy_mm = prev->active_mm;
prev->active_mm = NULL;
+#else
+ /*
+ * Without MMU_LAZY_REFCOUNT there is no lazy
+ * tracking (because no rq->prev_lazy_mm) in
+ * finish_task_switch, so no mmdrop_lazy_tlb(),
+ * so no memory barrier for membarrier (see the
+ * membarrier comment in finish_task_switch()).
+ * Do it here.
+ */
+ smp_mb();
+#endif
}
}
+}
+
+static __always_inline void
+context_switch_mm_nolazy(struct rq *rq, struct task_struct *prev,
+ struct task_struct *next)
+{
+ if (!next->mm)
+ next->active_mm = &init_mm;
+ membarrier_switch_mm(rq, prev->active_mm, next->active_mm);
+ switch_mm_irqs_off(prev->active_mm, next->active_mm, next);
+ if (!prev->mm)
+ prev->active_mm = NULL;
+}
+
+/*
+ * context_switch - switch to the new MM and the new thread's register state.
+ */
+static __always_inline struct rq *
+context_switch(struct rq *rq, struct task_struct *prev,
+ struct task_struct *next, struct rq_flags *rf)
+{
+ prepare_task_switch(rq, prev, next);
+
+ /*
+ * For paravirt, this is coupled with an exit in switch_to to
+ * combine the page table reload and the switch backend into
+ * one hypercall.
+ */
+ arch_start_context_switch(prev);
+
+ if (IS_ENABLED(CONFIG_MMU_LAZY_TLB))
+ context_switch_mm(rq, prev, next);
+ else
+ context_switch_mm_nolazy(rq, prev, next);
rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
@@ -961,7 +961,9 @@ struct rq {
struct task_struct *idle;
struct task_struct *stop;
unsigned long next_balance;
- struct mm_struct *prev_mm;
+#ifdef CONFIG_MMU_LAZY_TLB_REFCOUNT
+ struct mm_struct *prev_lazy_mm;
+#endif
unsigned int clock_update_flags;
u64 clock;
Add CONFIG_MMU_LAZY_TLB which can be configured out to disable the lazy tlb mechanism entirely, and switches to init_mm when switching to a kernel thread. NOMMU systems could easily go without this and save a bit of code and the refcount atomics, because their mm switch is a no-op. They have not been switched over by default because the arch code needs to be audited and tested for lazy tlb mm refcounting and converted to _lazy_tlb refcounting if necessary. CONFIG_MMU_LAZY_TLB_REFCOUNT is also added, but it must always be enabled if CONFIG_MMU_LAZY_TLB is enabled until the next patch which provides an alternate scheme. Signed-off-by: Nicholas Piggin <npiggin@gmail.com> --- arch/Kconfig | 26 ++++++++++++++ include/linux/sched/mm.h | 13 +++++-- kernel/sched/core.c | 75 ++++++++++++++++++++++++++++++---------- kernel/sched/sched.h | 4 ++- 4 files changed, 96 insertions(+), 22 deletions(-)