[RFC,05/24] MM locking API: allow for sleeping during unlock

Commit Message

Michel Lespinasse Feb. 24, 2020, 8:30 p.m. UTC

Following changes will implement fine-grained range locks for the
MM locking API, using a data structure to represent the existing
locks, and a mutex to protect that data structure.

As a result, we need to prepare for the possibility that unlocking
a memory range may need to sleep when acquiring that mutex.

Signed-off-by: Michel Lespinasse <[e-mail]>
---
 kernel/bpf/stackmap.c      | 6 +++++-
 kernel/exit.c              | 2 +-
 mm/mm_lock_rwsem_checked.c | 3 +++
 3 files changed, 9 insertions(+), 2 deletions(-)

Patch

diff --git kernel/bpf/stackmap.c kernel/bpf/stackmap.c
index ba2399ce00e4..0f483abeb94c 100644
--- kernel/bpf/stackmap.c
+++ kernel/bpf/stackmap.c
@@ -308,8 +308,12 @@  static void stack_map_get_build_id_offset(struct bpf_stack_build_id *id_offs,
 	 *
 	 * Same fallback is used for kernel stack (!user) on a stackmap
 	 * with build_id.
+	 *
+	 * FIXME - currently disabling the build_id lookup feature
+	 * as mm_read_range_unlock() may block, which is not always
+	 * possible to do here.
 	 */
-	if (!user || !current || !current->mm || irq_work_busy ||
+	if (true || !user || !current || !current->mm || irq_work_busy ||
 	    !mm_read_range_trylock(current->mm, mm_range_ptr)) {
 		/* cannot access current->mm, fall back to ips */
 		for (i = 0; i < trace_nr; i++) {
diff --git kernel/exit.c kernel/exit.c
index 9a0b72562adb..60ec6efb4e2c 100644
--- kernel/exit.c
+++ kernel/exit.c
@@ -478,9 +478,9 @@  static void exit_mm(void)
 	/* more a memory barrier than a real lock */
 	task_lock(current);
 	current->mm = NULL;
-	mm_read_unlock(mm);
 	enter_lazy_tlb(mm, current);
 	task_unlock(current);
+	mm_read_unlock(mm);
 	mm_update_next_owner(mm);
 	mmput(mm);
 	if (test_thread_flag(TIF_MEMDIE))
diff --git mm/mm_lock_rwsem_checked.c mm/mm_lock_rwsem_checked.c
index 3551deb85e3d..e45d1a598c87 100644
--- mm/mm_lock_rwsem_checked.c
+++ mm/mm_lock_rwsem_checked.c
@@ -61,6 +61,7 @@  EXPORT_SYMBOL(mm_write_range_trylock);
 
 void mm_write_range_unlock(struct mm_struct *mm, struct mm_lock_range *range)
 {
+	might_sleep();
 	if (range->mm != mm)
 		mm_lock_dump("mm_write_range_unlock");
 	range->mm = NULL;
@@ -113,6 +114,7 @@  EXPORT_SYMBOL(mm_read_range_trylock);
 
 void mm_read_range_unlock(struct mm_struct *mm, struct mm_lock_range *range)
 {
+	might_sleep();
 	if (range->mm != mm)
 		mm_lock_dump("mm_read_range_unlock");
 	range->mm = NULL;
@@ -123,6 +125,7 @@  EXPORT_SYMBOL(mm_read_range_unlock);
 void mm_read_range_unlock_non_owner(struct mm_struct *mm,
 				    struct mm_lock_range *range)
 {
+	might_sleep();
 	if (range->mm != mm)
 		mm_lock_dump("mm_read_range_unlock_non_owner");
 	range->mm = NULL;