From: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Date: Wed, 4 May 2022 19:03:25 +0200
Subject: [PATCH] mm/vmalloc: Use raw_cpu_ptr() for vmap_block_queue access.

The per-CPU resource vmap_block_queue is accessed via the
get_cpu_var(). That macro disables preemption and then loads the pointer
from the current CPU.
This doesn't work on PREEMPT_RT because a spinlock_t is later accessed
within the preempt-disable section.

There is no need to disable preemption while accessing the per-CPU
struct vmap_block_queue because the list is protected with a spinlock_t.
The per-CPU struct is also accessed cross-CPU in
purge_fragmented_blocks().
It is possible that by using raw_cpu_ptr() the code migrates to another
CPU and uses struct from another CPU. This is fine because the list
locked and the locked section is very short.

Use raw_cpu_ptr() to access vmap_block_queue.

Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Link: https://lore.kernel.org/r/YnKx3duAB53P7ojN@linutronix.de
---
 mm/vmalloc.c |    6 ++----
 1 file changed, 2 insertions(+), 4 deletions(-)

--- a/mm/vmalloc.c
+++ b/mm/vmalloc.c
@@ -1938,11 +1938,10 @@ static void *new_vmap_block(unsigned int
 		return ERR_PTR(err);
 	}
 
-	vbq = &get_cpu_var(vmap_block_queue);
+	vbq = raw_cpu_ptr(&vmap_block_queue);
 	spin_lock(&vbq->lock);
 	list_add_tail_rcu(&vb->free_list, &vbq->free);
 	spin_unlock(&vbq->lock);
-	put_cpu_var(vmap_block_queue);
 
 	return vaddr;
 }
@@ -2021,7 +2020,7 @@ static void *vb_alloc(unsigned long size
 	order = get_order(size);
 
 	rcu_read_lock();
-	vbq = &get_cpu_var(vmap_block_queue);
+	vbq = raw_cpu_ptr(&vmap_block_queue);
 	list_for_each_entry_rcu(vb, &vbq->free, free_list) {
 		unsigned long pages_off;
 
@@ -2044,7 +2043,6 @@ static void *vb_alloc(unsigned long size
 		break;
 	}
 
-	put_cpu_var(vmap_block_queue);
 	rcu_read_unlock();
 
 	/* Allocate new block if nothing was found */