@@ -62,7 +62,7 @@ static int do_adjust_pte(struct vm_area_struct *vma, unsigned long address,
6262}
6363
6464static int adjust_pte (struct vm_area_struct * vma , unsigned long address ,
65- unsigned long pfn , struct vm_fault * vmf )
65+ unsigned long pfn , bool need_lock )
6666{
6767 spinlock_t * ptl ;
6868 pgd_t * pgd ;
@@ -99,12 +99,11 @@ static int adjust_pte(struct vm_area_struct *vma, unsigned long address,
9999 if (!pte )
100100 return 0 ;
101101
102- /*
103- * If we are using split PTE locks, then we need to take the page
104- * lock here. Otherwise we are using shared mm->page_table_lock
105- * which is already locked, thus cannot take it.
106- */
107- if (ptl != vmf -> ptl ) {
102+ if (need_lock ) {
103+ /*
104+ * Use nested version here to indicate that we are already
105+ * holding one similar spinlock.
106+ */
108107 spin_lock_nested (ptl , SINGLE_DEPTH_NESTING );
109108 if (unlikely (!pmd_same (pmdval , pmdp_get_lockless (pmd )))) {
110109 pte_unmap_unlock (pte , ptl );
@@ -114,7 +113,7 @@ static int adjust_pte(struct vm_area_struct *vma, unsigned long address,
114113
115114 ret = do_adjust_pte (vma , address , pfn , pte );
116115
117- if (ptl != vmf -> ptl )
116+ if (need_lock )
118117 spin_unlock (ptl );
119118 pte_unmap (pte );
120119
@@ -123,9 +122,10 @@ static int adjust_pte(struct vm_area_struct *vma, unsigned long address,
123122
124123static void
125124make_coherent (struct address_space * mapping , struct vm_area_struct * vma ,
126- unsigned long addr , pte_t * ptep , unsigned long pfn ,
127- struct vm_fault * vmf )
125+ unsigned long addr , pte_t * ptep , unsigned long pfn )
128126{
127+ const unsigned long pmd_start_addr = ALIGN_DOWN (addr , PMD_SIZE );
128+ const unsigned long pmd_end_addr = pmd_start_addr + PMD_SIZE ;
129129 struct mm_struct * mm = vma -> vm_mm ;
130130 struct vm_area_struct * mpnt ;
131131 unsigned long offset ;
@@ -141,6 +141,14 @@ make_coherent(struct address_space *mapping, struct vm_area_struct *vma,
141141 */
142142 flush_dcache_mmap_lock (mapping );
143143 vma_interval_tree_foreach (mpnt , & mapping -> i_mmap , pgoff , pgoff ) {
144+ /*
145+ * If we are using split PTE locks, then we need to take the pte
146+ * lock. Otherwise we are using shared mm->page_table_lock which
147+ * is already locked, thus cannot take it.
148+ */
149+ bool need_lock = IS_ENABLED (CONFIG_SPLIT_PTE_PTLOCKS );
150+ unsigned long mpnt_addr ;
151+
144152 /*
145153 * If this VMA is not in our MM, we can ignore it.
146154 * Note that we intentionally mask out the VMA
@@ -151,7 +159,12 @@ make_coherent(struct address_space *mapping, struct vm_area_struct *vma,
151159 if (!(mpnt -> vm_flags & VM_MAYSHARE ))
152160 continue ;
153161 offset = (pgoff - mpnt -> vm_pgoff ) << PAGE_SHIFT ;
154- aliases += adjust_pte (mpnt , mpnt -> vm_start + offset , pfn , vmf );
162+ mpnt_addr = mpnt -> vm_start + offset ;
163+
164+ /* Avoid deadlocks by not grabbing the same PTE lock again. */
165+ if (mpnt_addr >= pmd_start_addr && mpnt_addr < pmd_end_addr )
166+ need_lock = false;
167+ aliases += adjust_pte (mpnt , mpnt_addr , pfn , need_lock );
155168 }
156169 flush_dcache_mmap_unlock (mapping );
157170 if (aliases )
@@ -194,7 +207,7 @@ void update_mmu_cache_range(struct vm_fault *vmf, struct vm_area_struct *vma,
194207 __flush_dcache_folio (mapping , folio );
195208 if (mapping ) {
196209 if (cache_is_vivt ())
197- make_coherent (mapping , vma , addr , ptep , pfn , vmf );
210+ make_coherent (mapping , vma , addr , ptep , pfn );
198211 else if (vma -> vm_flags & VM_EXEC )
199212 __flush_icache_all ();
200213 }
0 commit comments