[RFC PATCH 1/2] sparc64: remove mm_cpumask clearing to fix kthread_use_mm race

[RFC PATCH 1/2] sparc64: remove mm_cpumask clearing to fix kthread_use_mm race

[Nicholas Piggin]

The de facto (and apparently uncommented) standard for using an mm had,
thanks to this code in sparc if nothing else, been that you must have a
reference on mm_users *and that reference must have been obtained with
mmget()*, i.e., from a thread with a reference to mm_users that had used
the mm.

The introduction of mmget_not_zero() in commit d2005e3f41d4
("userfaultfd: don't pin the user memory in userfaultfd_file_create()")
allowed mm_count holders to aoperate on user mappings asynchronously
from the actual threads using the mm, but they were not to load those
mappings into their TLB (i.e., walking vmas and page tables is okay,
kthread_use_mm() is not).

io_uring 2b188cc1bb857 ("Add io_uring IO interface") added code which
does a kthread_use_mm() from a mmget_not_zero() refcount.

The problem with this is code which previously assumed mm == current->mm
and mm->mm_users == 1 implies the mm will remain single-threaded at
least until this thread creates another mm_users reference, has now
broken.

arch/sparc/kernel/smp_64.c:

    if (atomic_read(&mm->mm_users) == 1) {
        cpumask_copy(mm_cpumask(mm), cpumask_of(cpu));
        goto local_flush_and_out;
    }

vs fs/io_uring.c

    if (unlikely(!(ctx->flags & IORING_SETUP_SQPOLL) ||
                 !mmget_not_zero(ctx->sqo_mm)))
        return -EFAULT;
    kthread_use_mm(ctx->sqo_mm);

mmget_not_zero() could come in right after the mm_users == 1 test, then
kthread_use_mm() which sets its CPU in the mm_cpumask. That update could
be lost if cpumask_copy() occurs afterward.

I propose we fix this by allowing mmget_not_zero() to be a first-class
reference, and not have this obscure undocumented and unchecked
restriction.

The basic fix for sparc64 is to remove its mm_cpumask clearing code. The
optimisation could be effectively restored by sending IPIs to mm_cpumask
members and having them remove themselves from mm_cpumask. This is more
tricky so I leave it as an exercise for someone with a sparc64 SMP.
powerpc has a (currently similarly broken) example.

not-yet-Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
---
 arch/sparc/kernel/smp_64.c | 58 ++++++--------------------------------
 1 file changed, 9 insertions(+), 49 deletions(-)

diff --git a/arch/sparc/kernel/smp_64.c b/arch/sparc/kernel/smp_64.c
index e286e2badc8a..41423c6077d0 100644
--- a/arch/sparc/kernel/smp_64.c
+++ b/arch/sparc/kernel/smp_64.c
@@ -1039,38 +1039,9 @@ void smp_fetch_global_pmu(void)
  * are flush_tlb_*() routines, and these run after flush_cache_*()
  * which performs the flushw.
  *
- * The SMP TLB coherency scheme we use works as follows:
- *
- * 1) mm->cpu_vm_mask is a bit mask of which cpus an address
- *    space has (potentially) executed on, this is the heuristic
- *    we use to avoid doing cross calls.
- *
- *    Also, for flushing from kswapd and also for clones, we
- *    use cpu_vm_mask as the list of cpus to make run the TLB.
- *
- * 2) TLB context numbers are shared globally across all processors
- *    in the system, this allows us to play several games to avoid
- *    cross calls.
- *
- *    One invariant is that when a cpu switches to a process, and
- *    that processes tsk->active_mm->cpu_vm_mask does not have the
- *    current cpu's bit set, that tlb context is flushed locally.
- *
- *    If the address space is non-shared (ie. mm->count == 1) we avoid
- *    cross calls when we want to flush the currently running process's
- *    tlb state.  This is done by clearing all cpu bits except the current
- *    processor's in current->mm->cpu_vm_mask and performing the
- *    flush locally only.  This will force any subsequent cpus which run
- *    this task to flush the context from the local tlb if the process
- *    migrates to another cpu (again).
- *
- * 3) For shared address spaces (threads) and swapping we bite the
- *    bullet for most cases and perform the cross call (but only to
- *    the cpus listed in cpu_vm_mask).
- *
- *    The performance gain from "optimizing" away the cross call for threads is
- *    questionable (in theory the big win for threads is the massive sharing of
- *    address space state across processors).
+ * mm->cpu_vm_mask is a bit mask of which cpus an address
+ * space has (potentially) executed on, this is the heuristic
+ * we use to limit cross calls.
  */
 
 /* This currently is only used by the hugetlb arch pre-fault
@@ -1082,16 +1053,10 @@ void smp_flush_tlb_mm(struct mm_struct *mm)
 	u32 ctx = CTX_HWBITS(mm->context);
 	int cpu = get_cpu();
 
-	if (atomic_read(&mm->mm_users) == 1) {
-		cpumask_copy(mm_cpumask(mm), cpumask_of(cpu));
-		goto local_flush_and_out;
-	}
-
 	smp_cross_call_masked(&xcall_flush_tlb_mm,
 			      ctx, 0, 0,
 			      mm_cpumask(mm));
 
-local_flush_and_out:
 	__flush_tlb_mm(ctx, SECONDARY_CONTEXT);
 
 	put_cpu();
@@ -1120,11 +1085,8 @@ void smp_flush_tlb_pending(struct mm_struct *mm, unsigned long nr, unsigned long
 	info.nr = nr;
 	info.vaddrs = vaddrs;
 
-	if (mm == current->mm && atomic_read(&mm->mm_users) == 1)
-		cpumask_copy(mm_cpumask(mm), cpumask_of(cpu));
-	else
-		smp_call_function_many(mm_cpumask(mm), tlb_pending_func,
-				       &info, 1);
+	smp_call_function_many(mm_cpumask(mm), tlb_pending_func,
+			       &info, 1);
 
 	__flush_tlb_pending(ctx, nr, vaddrs);
 
@@ -1136,12 +1098,10 @@ void smp_flush_tlb_page(struct mm_struct *mm, unsigned long vaddr)
 	unsigned long context = CTX_HWBITS(mm->context);
 	int cpu = get_cpu();
 
-	if (mm == current->mm && atomic_read(&mm->mm_users) == 1)
-		cpumask_copy(mm_cpumask(mm), cpumask_of(cpu));
-	else
-		smp_cross_call_masked(&xcall_flush_tlb_page,
-				      context, vaddr, 0,
-				      mm_cpumask(mm));
+	smp_cross_call_masked(&xcall_flush_tlb_page,
+			      context, vaddr, 0,
+			      mm_cpumask(mm));
+
 	__flush_tlb_page(context, vaddr);
 
 	put_cpu();
-- 
2.23.0

[RFC PATCH 2/2] powerpc/64s/radix: Fix mm_cpumask trimming race vs kthread_use_mm

[Nicholas Piggin]

Commit 0cef77c7798a7 ("powerpc/64s/radix: flush remote CPUs out of
single-threaded mm_cpumask") added a mechanism to trim the mm_cpumask of
a process under certain conditions. One of the assumptions is that
mm_users would not be incremented via a reference outside the process
context with mmget_not_zero() then go on to kthread_use_mm() via that
reference.

That invariant was broken by io_uring code (see previous sparc64 fix),
but I'll point Fixes: to the original powerpc commit because we are
changing that assumption going forward, so this will make backports
match up.

Fix this by no longer relying on that assumption, but by having each CPU
check the mm is not being used, and clearing their own bit from the mask
if it's okay. This relies on commit 38cf307c1f20 ("mm: fix
kthread_use_mm() vs TLB invalidate") to disable irqs over the mm switch.

Fixes: 0cef77c7798a7 ("powerpc/64s/radix: flush remote CPUs out of single-threaded mm_cpumask")
not-yet-Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
---
 arch/powerpc/include/asm/tlb.h       | 13 -------------
 arch/powerpc/mm/book3s64/radix_tlb.c | 23 ++++++++++++++++-------
 2 files changed, 16 insertions(+), 20 deletions(-)

diff --git a/arch/powerpc/include/asm/tlb.h b/arch/powerpc/include/asm/tlb.h
index fbc6f3002f23..d97f061fecac 100644
--- a/arch/powerpc/include/asm/tlb.h
+++ b/arch/powerpc/include/asm/tlb.h
@@ -66,19 +66,6 @@ static inline int mm_is_thread_local(struct mm_struct *mm)
 		return false;
 	return cpumask_test_cpu(smp_processor_id(), mm_cpumask(mm));
 }
-static inline void mm_reset_thread_local(struct mm_struct *mm)
-{
-	WARN_ON(atomic_read(&mm->context.copros) > 0);
-	/*
-	 * It's possible for mm_access to take a reference on mm_users to
-	 * access the remote mm from another thread, but it's not allowed
-	 * to set mm_cpumask, so mm_users may be > 1 here.
-	 */
-	WARN_ON(current->mm != mm);
-	atomic_set(&mm->context.active_cpus, 1);
-	cpumask_clear(mm_cpumask(mm));
-	cpumask_set_cpu(smp_processor_id(), mm_cpumask(mm));
-}
 #else /* CONFIG_PPC_BOOK3S_64 */
 static inline int mm_is_thread_local(struct mm_struct *mm)
 {
diff --git a/arch/powerpc/mm/book3s64/radix_tlb.c b/arch/powerpc/mm/book3s64/radix_tlb.c
index 0d233763441f..a421a0e3f930 100644
--- a/arch/powerpc/mm/book3s64/radix_tlb.c
+++ b/arch/powerpc/mm/book3s64/radix_tlb.c
@@ -645,19 +645,29 @@ static void do_exit_flush_lazy_tlb(void *arg)
 	struct mm_struct *mm = arg;
 	unsigned long pid = mm->context.id;
 
+	/*
+	 * A kthread could have done a mmget_not_zero() after the flushing CPU
+	 * checked mm_users == 1, and be in the process of kthread_use_mm when
+	 * interrupted here. In that case, current->mm will be set to mm,
+	 * because kthread_use_mm() setting ->mm and switching to the mm is
+	 * done with interrupts off.
+	 */
 	if (current->mm == mm)
-		return; /* Local CPU */
+		goto out_flush;
 
 	if (current->active_mm == mm) {
-		/*
-		 * Must be a kernel thread because sender is single-threaded.
-		 */
-		BUG_ON(current->mm);
+		WARN_ON_ONCE(current->mm != NULL);
+		/* Is a kernel thread and is using mm as the lazy tlb */
 		mmgrab(&init_mm);
-		switch_mm(mm, &init_mm, current);
 		current->active_mm = &init_mm;
+		switch_mm_irqs_off(mm, &init_mm, current);
 		mmdrop(mm);
 	}
+
+	atomic_dec(&mm->context.active_cpus);
+	cpumask_clear_cpu(smp_processor_id(), mm_cpumask(mm));
+
+out_flush:
 	_tlbiel_pid(pid, RIC_FLUSH_ALL);
 }
 
@@ -672,7 +682,6 @@ static void exit_flush_lazy_tlbs(struct mm_struct *mm)
 	 */
 	smp_call_function_many(mm_cpumask(mm), do_exit_flush_lazy_tlb,
 				(void *)mm, 1);
-	mm_reset_thread_local(mm);
 }
 
 void radix__flush_tlb_mm(struct mm_struct *mm)
-- 
2.23.0