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 <aik@ozlabs.ru>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>

diff --git a/arch/powerpc/mm/mmu_context_iommu.c b/arch/powerpc/mm/mmu_context_iommu.c
index e7a9c4f6bfca49585beffcb6fc3dc755eb054e8f..9d9be850f8c20e3a81a5418868b3afdbc62adbbd 100644 (file)
--- a/arch/powerpc/mm/mmu_context_iommu.c
+++ b/arch/powerpc/mm/mmu_context_iommu.c
@@ -95,28 +95,14 @@ static long mm_iommu_do_alloc(struct mm_struct *mm, unsigned long ua,
                              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;
        }
@@ -150,15 +136,10 @@ static long mm_iommu_do_alloc(struct mm_struct *mm, unsigned long ua,
        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;
@@ -183,21 +164,43 @@ static long mm_iommu_do_alloc(struct mm_struct *mm, unsigned long ua,
        }
 
 good_exit:
-       ret = 0;
        atomic64_set(&mem->mapped, 1);
        mem->used = 1;
        mem->ua = ua;
        mem->entries = entries;
-       *pmem = mem;
 
-       list_add_rcu(&mem->next, &mm->context.iommu_group_mem_list);
+       mutex_lock(&mem_list_mutex);
 
-unlock_exit:
-       if (locked_entries && ret)
-               mm_iommu_adjust_locked_vm(mm, locked_entries, false);
+       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);
 
        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;
 }
 
@@ -266,7 +269,7 @@ static void mm_iommu_release(struct mm_iommu_table_group_mem_t *mem)
 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);
 
@@ -287,17 +290,17 @@ long mm_iommu_put(struct mm_struct *mm, struct mm_iommu_table_group_mem_t *mem)
                goto unlock_exit;
        }
 
+       if (mem->dev_hpa == MM_IOMMU_TABLE_INVALID_HPA)
+               unlock_entries = mem->entries;
+
        /* @mapped became 0 so now mappings are disabled, release the region */
-       entries = mem->entries;
-       dev_hpa = mem->dev_hpa;
        mm_iommu_release(mem);
 
-       if (dev_hpa == MM_IOMMU_TABLE_INVALID_HPA)
-               mm_iommu_adjust_locked_vm(mm, entries, false);
-
 unlock_exit:
        mutex_unlock(&mem_list_mutex);
 
+       mm_iommu_adjust_locked_vm(mm, unlock_entries, false);
+
        return ret;
 }
 EXPORT_SYMBOL_GPL(mm_iommu_put);