powerpc/mm_iommu: Fix potential deadlock

Currently mm_iommu_do_alloc() is called in 2 cases:
- VFIO_IOMMU_SPAPR_REGISTER_MEMORY ioctl() for normal memory:
this locks &mem_list_mutex and then locks mm::mmap_sem
several times when adjusting locked_vm or pinning pages;
- vfio_pci_nvgpu_regops::mmap() for GPU memory:
this is called with mm::mmap_sem held already and it locks
&mem_list_mutex.

So one can craft a userspace program to do special ioctl and mmap in
2 threads concurrently and cause a deadlock which lockdep warns about
(below).

We did not hit this yet because QEMU constructs the machine in a single
thread.

This moves the overlap check next to where the new entry is added and
reduces the amount of time spent with &mem_list_mutex held.

This moves locked_vm adjustment from under &mem_list_mutex.

This relies on mm_iommu_adjust_locked_vm() doing nothing when entries==0.

This is one of the lockdep warnings:

======================================================
WARNING: possible circular locking dependency detected
5.1.0-rc2-le_nv2_aikATfstn1-p1 #363 Not tainted
------------------------------------------------------
qemu-system-ppc/8038 is trying to acquire lock:
000000002ec6c453 (mem_list_mutex){+.+.}, at: mm_iommu_do_alloc+0x70/0x490

but task is already holding lock:
00000000fd7da97f (&mm->mmap_sem){++++}, at: vm_mmap_pgoff+0xf0/0x160

which lock already depends on the new lock.

the existing dependency chain (in reverse order) is:

-> #1 (&mm->mmap_sem){++++}:
lock_acquire+0xf8/0x260
down_write+0x44/0xa0
mm_iommu_adjust_locked_vm.part.1+0x4c/0x190
mm_iommu_do_alloc+0x310/0x490
tce_iommu_ioctl.part.9+0xb84/0x1150 [vfio_iommu_spapr_tce]
vfio_fops_unl_ioctl+0x94/0x430 [vfio]
do_vfs_ioctl+0xe4/0x930
ksys_ioctl+0xc4/0x110
sys_ioctl+0x28/0x80
system_call+0x5c/0x70

-> #0 (mem_list_mutex){+.+.}:
__lock_acquire+0x1484/0x1900
lock_acquire+0xf8/0x260
__mutex_lock+0x88/0xa70
mm_iommu_do_alloc+0x70/0x490
vfio_pci_nvgpu_mmap+0xc0/0x130 [vfio_pci]
vfio_pci_mmap+0x198/0x2a0 [vfio_pci]
vfio_device_fops_mmap+0x44/0x70 [vfio]
mmap_region+0x5d4/0x770
do_mmap+0x42c/0x650
vm_mmap_pgoff+0x124/0x160
ksys_mmap_pgoff+0xdc/0x2f0
sys_mmap+0x40/0x80
system_call+0x5c/0x70

other info that might help us debug this:

Possible unsafe locking scenario:

CPU0 CPU1
---- ----
lock(&mm->mmap_sem);
lock(mem_list_mutex);
lock(&mm->mmap_sem);
lock(mem_list_mutex);

*** DEADLOCK ***

1 lock held by qemu-system-ppc/8038:
#0: 00000000fd7da97f (&mm->mmap_sem){++++}, at: vm_mmap_pgoff+0xf0/0x160

Fixes: c10c21efa4bc ("powerpc/vfio/iommu/kvm: Do not pin device memory", 2018-12-19)
Signed-off-by: Alexey Kardashevskiy <[e-mail]>
Signed-off-by: Michael Ellerman <[e-mail]>

arch/powerpc/mm/mmu_context_iommu.c

···
 			      unsigned long entries, unsigned long dev_hpa,
 			      struct mm_iommu_table_group_mem_t **pmem)
 {
-	struct mm_iommu_table_group_mem_t *mem;
-	long i, ret, locked_entries = 0;
+	struct mm_iommu_table_group_mem_t *mem, *mem2;
+	long i, ret, locked_entries = 0, pinned = 0;
 	unsigned int pageshift;
-
-	mutex_lock(&mem_list_mutex);
-
-	list_for_each_entry_rcu(mem, &mm->context.iommu_group_mem_list,
-			next) {
-		/* Overlap? */
-		if ((mem->ua < (ua + (entries << PAGE_SHIFT))) &&
-				(ua < (mem->ua +
-				       (mem->entries << PAGE_SHIFT)))) {
-			ret = -EINVAL;
-			goto unlock_exit;
-		}
-
-	}
 
 	if (dev_hpa == MM_IOMMU_TABLE_INVALID_HPA) {
 		ret = mm_iommu_adjust_locked_vm(mm, entries, true);
 		if (ret)
-			goto unlock_exit;
+			return ret;
 
 		locked_entries = entries;
 	}
···
 	down_read(&mm->mmap_sem);
 	ret = get_user_pages_longterm(ua, entries, FOLL_WRITE, mem->hpages, NULL);
 	up_read(&mm->mmap_sem);
+	pinned = ret > 0 ? ret : 0;
 	if (ret != entries) {
-		/* free the reference taken */
-		for (i = 0; i < ret; i++)
-			put_page(mem->hpages[i]);
-
-		vfree(mem->hpas);
-		kfree(mem);
 		ret = -EFAULT;
-		goto unlock_exit;
+		goto free_exit;
 	}
 
 	pageshift = PAGE_SHIFT;
···
 	}
 
 good_exit:
-	ret = 0;
 	atomic64_set(&mem->mapped, 1);
 	mem->used = 1;
 	mem->ua = ua;
 	mem->entries = entries;
-	*pmem = mem;
+
+	mutex_lock(&mem_list_mutex);
+
+	list_for_each_entry_rcu(mem2, &mm->context.iommu_group_mem_list, next) {
+		/* Overlap? */
+		if ((mem2->ua < (ua + (entries << PAGE_SHIFT))) &&
+				(ua < (mem2->ua +
+				       (mem2->entries << PAGE_SHIFT)))) {
+			ret = -EINVAL;
+			mutex_unlock(&mem_list_mutex);
+			goto free_exit;
+		}
+	}
 
 	list_add_rcu(&mem->next, &mm->context.iommu_group_mem_list);
 
-unlock_exit:
-	if (locked_entries && ret)
-		mm_iommu_adjust_locked_vm(mm, locked_entries, false);
-
 	mutex_unlock(&mem_list_mutex);
+
+	*pmem = mem;
+
+	return 0;
+
+free_exit:
+	/* free the reference taken */
+	for (i = 0; i < pinned; i++)
+		put_page(mem->hpages[i]);
+
+	vfree(mem->hpas);
+	kfree(mem);
+
+unlock_exit:
+	mm_iommu_adjust_locked_vm(mm, locked_entries, false);
 
 	return ret;
 }
···
 long mm_iommu_put(struct mm_struct *mm, struct mm_iommu_table_group_mem_t *mem)
 {
 	long ret = 0;
-	unsigned long entries, dev_hpa;
+	unsigned long unlock_entries = 0;
 
 	mutex_lock(&mem_list_mutex);
 
···
 		goto unlock_exit;
 	}
 
-	/* @mapped became 0 so now mappings are disabled, release the region */
-	entries = mem->entries;
-	dev_hpa = mem->dev_hpa;
-	mm_iommu_release(mem);
+	if (mem->dev_hpa == MM_IOMMU_TABLE_INVALID_HPA)
+		unlock_entries = mem->entries;
 
-	if (dev_hpa == MM_IOMMU_TABLE_INVALID_HPA)
-		mm_iommu_adjust_locked_vm(mm, entries, false);
+	/* @mapped became 0 so now mappings are disabled, release the region */
+	mm_iommu_release(mem);
 
 unlock_exit:
 	mutex_unlock(&mem_list_mutex);
+
+	mm_iommu_adjust_locked_vm(mm, unlock_entries, false);
 
 	return ret;
 }