syzbot reported the deadlock here [1]. The issue is in hugetlb cow error handling when there are not enough huge pages for the faulting task which took the original reservation. It is possible that other (child) tasks could have consumed pages associated with the reservation. In this case, we want the task which took the original reservation to succeed. So, we unmap any associated pages in children so that they can be used by the faulting task that owns the reservation. The unmapping code needs to hold i_mmap_rwsem in write mode. However, due to commit c0d0381ade79 ("hugetlbfs: use i_mmap_rwsem for more pmd sharing synchronization") we are already holding i_mmap_rwsem in read mode when hugetlb_cow is called. Technically, i_mmap_rwsem does not need to be held in read mode for COW mappings as they can not share pmd's. Modifying the fault code to not take i_mmap_rwsem in read mode for COW (and other non-sharable) mappings is too involved for a stable fix. Instead, we simply drop the hugetlb_fault_mutex and i_mmap_rwsem before unmapping. This is OK as it is technically not needed. They are reacquired after unmapping as expected by calling code. Since this is done in an uncommon error path, the overhead of dropping and reacquiring mutexes is acceptable. While making changes, remove redundant BUG_ON after unmap_ref_private. [1] https://lkml.kernel.org/r/000000000000b73ccc05b5cf8558@google.com
Unified · Split
| @@ -4105,10 +4105,30 @@ retry_avoidcopy: · 4104 unmodified lines | ||
| 4105 | 4105 | * may get SIGKILLed if it later faults. |
| 4106 | 4106 | */ |
| 4107 | 4107 | if (outside_reserve) { |
| 4108 | + struct address_space *mapping = vma->vm_file->f_mapping; | |
| 4109 | + pgoff_t idx; | |
| 4110 | + u32 hash; | |
| 4111 | + | |
| 4108 | 4112 | put_page(old_page); |
| 4109 | 4113 | BUG_ON(huge_pte_none(pte)); |
| 4114 | + /* | |
| 4115 | + * Drop hugetlb_fault_mutex and i_mmap_rwsem before | |
| 4116 | + * unmapping. unmapping needs to hold i_mmap_rwsem | |
| 4117 | + * in write mode. Dropping i_mmap_rwsem in read mode | |
| 4118 | + * here is OK as COW mappings do not interact with | |
| 4119 | + * PMD sharing. | |
| 4120 | + * | |
| 4121 | + * Reacquire both after unmap operation. | |
| 4122 | + */ | |
| 4123 | + idx = vma_hugecache_offset(h, vma, haddr); | |
| 4124 | + hash = hugetlb_fault_mutex_hash(mapping, idx); | |
| 4125 | + mutex_unlock(&hugetlb_fault_mutex_table[hash]); | |
| 4126 | + i_mmap_unlock_read(mapping); | |
| 4127 | + | |
| 4110 | 4128 | unmap_ref_private(mm, vma, old_page, haddr); |
| 4111 | − BUG_ON(huge_pte_none(pte)); | |
| 4129 | + | |
| 4130 | + i_mmap_lock_read(mapping); | |
| 4131 | + mutex_lock(&hugetlb_fault_mutex_table[hash]); | |
| 4112 | 4132 | spin_lock(ptl); |
| 4113 | 4133 | ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); |
| 4114 | 4134 | if (likely(ptep && |