s390/mm: fix asce_bits handling with dynamic pagetable levels
authorGerald Schaefer <gerald.schaefer@de.ibm.com>
Fri, 15 Apr 2016 14:38:40 +0000 (16:38 +0200)
committerMartin Schwidefsky <schwidefsky@de.ibm.com>
Thu, 21 Apr 2016 07:50:09 +0000 (09:50 +0200)
There is a race with multi-threaded applications between context switch and
pagetable upgrade. In switch_mm() a new user_asce is built from mm->pgd and
mm->context.asce_bits, w/o holding any locks. A concurrent mmap with a
pagetable upgrade on another thread in crst_table_upgrade() could already
have set new asce_bits, but not yet the new mm->pgd. This would result in a
corrupt user_asce in switch_mm(), and eventually in a kernel panic from a
translation exception.

Fix this by storing the complete asce instead of just the asce_bits, which
can then be read atomically from switch_mm(), so that it either sees the
old value or the new value, but no mixture. Both cases are OK. Having the
old value would result in a page fault on access to the higher level memory,
but the fault handler would see the new mm->pgd, if it was a valid access
after the mmap on the other thread has completed. So as worst-case scenario
we would have a page fault loop for the racing thread until the next time
slice.

Also remove dead code and simplify the upgrade/downgrade path, there are no
upgrades from 2 levels, and only downgrades from 3 levels for compat tasks.
There are also no concurrent upgrades, because the mmap_sem is held with
down_write() in do_mmap, so the flush and table checks during upgrade can
be removed.

Reported-by: Michael Munday <munday@ca.ibm.com>
Reviewed-by: Martin Schwidefsky <schwidefsky@de.ibm.com>
Signed-off-by: Gerald Schaefer <gerald.schaefer@de.ibm.com>
Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com>
arch/s390/include/asm/mmu.h
arch/s390/include/asm/mmu_context.h
arch/s390/include/asm/pgalloc.h
arch/s390/include/asm/processor.h
arch/s390/include/asm/tlbflush.h
arch/s390/mm/init.c
arch/s390/mm/mmap.c
arch/s390/mm/pgalloc.c

index d29ad95..081b2ad 100644 (file)
@@ -11,7 +11,7 @@ typedef struct {
        spinlock_t list_lock;
        struct list_head pgtable_list;
        struct list_head gmap_list;
-       unsigned long asce_bits;
+       unsigned long asce;
        unsigned long asce_limit;
        unsigned long vdso_base;
        /* The mmu context allocates 4K page tables. */
index d321469..c837b79 100644 (file)
@@ -26,12 +26,28 @@ static inline int init_new_context(struct task_struct *tsk,
        mm->context.has_pgste = 0;
        mm->context.use_skey = 0;
 #endif
-       if (mm->context.asce_limit == 0) {
+       switch (mm->context.asce_limit) {
+       case 1UL << 42:
+               /*
+                * forked 3-level task, fall through to set new asce with new
+                * mm->pgd
+                */
+       case 0:
                /* context created by exec, set asce limit to 4TB */
-               mm->context.asce_bits = _ASCE_TABLE_LENGTH |
-                       _ASCE_USER_BITS | _ASCE_TYPE_REGION3;
                mm->context.asce_limit = STACK_TOP_MAX;
-       } else if (mm->context.asce_limit == (1UL << 31)) {
+               mm->context.asce = __pa(mm->pgd) | _ASCE_TABLE_LENGTH |
+                                  _ASCE_USER_BITS | _ASCE_TYPE_REGION3;
+               break;
+       case 1UL << 53:
+               /* forked 4-level task, set new asce with new mm->pgd */
+               mm->context.asce = __pa(mm->pgd) | _ASCE_TABLE_LENGTH |
+                                  _ASCE_USER_BITS | _ASCE_TYPE_REGION2;
+               break;
+       case 1UL << 31:
+               /* forked 2-level compat task, set new asce with new mm->pgd */
+               mm->context.asce = __pa(mm->pgd) | _ASCE_TABLE_LENGTH |
+                                  _ASCE_USER_BITS | _ASCE_TYPE_SEGMENT;
+               /* pgd_alloc() did not increase mm->nr_pmds */
                mm_inc_nr_pmds(mm);
        }
        crst_table_init((unsigned long *) mm->pgd, pgd_entry_type(mm));
@@ -42,7 +58,7 @@ static inline int init_new_context(struct task_struct *tsk,
 
 static inline void set_user_asce(struct mm_struct *mm)
 {
-       S390_lowcore.user_asce = mm->context.asce_bits | __pa(mm->pgd);
+       S390_lowcore.user_asce = mm->context.asce;
        if (current->thread.mm_segment.ar4)
                __ctl_load(S390_lowcore.user_asce, 7, 7);
        set_cpu_flag(CIF_ASCE);
@@ -71,7 +87,7 @@ static inline void switch_mm(struct mm_struct *prev, struct mm_struct *next,
 {
        int cpu = smp_processor_id();
 
-       S390_lowcore.user_asce = next->context.asce_bits | __pa(next->pgd);
+       S390_lowcore.user_asce = next->context.asce;
        if (prev == next)
                return;
        if (MACHINE_HAS_TLB_LC)
index 9b3d9b6..da34cb6 100644 (file)
@@ -52,8 +52,8 @@ static inline unsigned long pgd_entry_type(struct mm_struct *mm)
        return _REGION2_ENTRY_EMPTY;
 }
 
-int crst_table_upgrade(struct mm_struct *, unsigned long limit);
-void crst_table_downgrade(struct mm_struct *, unsigned long limit);
+int crst_table_upgrade(struct mm_struct *);
+void crst_table_downgrade(struct mm_struct *);
 
 static inline pud_t *pud_alloc_one(struct mm_struct *mm, unsigned long address)
 {
index d6fd22e..18cdede 100644 (file)
@@ -175,7 +175,7 @@ extern __vector128 init_task_fpu_regs[__NUM_VXRS];
        regs->psw.mask  = PSW_USER_BITS | PSW_MASK_BA;                  \
        regs->psw.addr  = new_psw;                                      \
        regs->gprs[15]  = new_stackp;                                   \
-       crst_table_downgrade(current->mm, 1UL << 31);                   \
+       crst_table_downgrade(current->mm);                              \
        execve_tail();                                                  \
 } while (0)
 
index ca148f7..a2e6ef3 100644 (file)
@@ -110,8 +110,7 @@ static inline void __tlb_flush_asce(struct mm_struct *mm, unsigned long asce)
 static inline void __tlb_flush_kernel(void)
 {
        if (MACHINE_HAS_IDTE)
-               __tlb_flush_idte((unsigned long) init_mm.pgd |
-                                init_mm.context.asce_bits);
+               __tlb_flush_idte(init_mm.context.asce);
        else
                __tlb_flush_global();
 }
@@ -133,8 +132,7 @@ static inline void __tlb_flush_asce(struct mm_struct *mm, unsigned long asce)
 static inline void __tlb_flush_kernel(void)
 {
        if (MACHINE_HAS_TLB_LC)
-               __tlb_flush_idte_local((unsigned long) init_mm.pgd |
-                                      init_mm.context.asce_bits);
+               __tlb_flush_idte_local(init_mm.context.asce);
        else
                __tlb_flush_local();
 }
@@ -148,8 +146,7 @@ static inline void __tlb_flush_mm(struct mm_struct * mm)
         * only ran on the local cpu.
         */
        if (MACHINE_HAS_IDTE && list_empty(&mm->context.gmap_list))
-               __tlb_flush_asce(mm, (unsigned long) mm->pgd |
-                                mm->context.asce_bits);
+               __tlb_flush_asce(mm, mm->context.asce);
        else
                __tlb_flush_full(mm);
 }
index c7b0451..2489b2e 100644 (file)
@@ -89,7 +89,8 @@ void __init paging_init(void)
                asce_bits = _ASCE_TYPE_REGION3 | _ASCE_TABLE_LENGTH;
                pgd_type = _REGION3_ENTRY_EMPTY;
        }
-       S390_lowcore.kernel_asce = (__pa(init_mm.pgd) & PAGE_MASK) | asce_bits;
+       init_mm.context.asce = (__pa(init_mm.pgd) & PAGE_MASK) | asce_bits;
+       S390_lowcore.kernel_asce = init_mm.context.asce;
        clear_table((unsigned long *) init_mm.pgd, pgd_type,
                    sizeof(unsigned long)*2048);
        vmem_map_init();
index 45c4daa..89cf09e 100644 (file)
@@ -174,7 +174,7 @@ int s390_mmap_check(unsigned long addr, unsigned long len, unsigned long flags)
        if (!(flags & MAP_FIXED))
                addr = 0;
        if ((addr + len) >= TASK_SIZE)
-               return crst_table_upgrade(current->mm, TASK_MAX_SIZE);
+               return crst_table_upgrade(current->mm);
        return 0;
 }
 
@@ -191,7 +191,7 @@ s390_get_unmapped_area(struct file *filp, unsigned long addr,
                return area;
        if (area == -ENOMEM && !is_compat_task() && TASK_SIZE < TASK_MAX_SIZE) {
                /* Upgrade the page table to 4 levels and retry. */
-               rc = crst_table_upgrade(mm, TASK_MAX_SIZE);
+               rc = crst_table_upgrade(mm);
                if (rc)
                        return (unsigned long) rc;
                area = arch_get_unmapped_area(filp, addr, len, pgoff, flags);
@@ -213,7 +213,7 @@ s390_get_unmapped_area_topdown(struct file *filp, const unsigned long addr,
                return area;
        if (area == -ENOMEM && !is_compat_task() && TASK_SIZE < TASK_MAX_SIZE) {
                /* Upgrade the page table to 4 levels and retry. */
-               rc = crst_table_upgrade(mm, TASK_MAX_SIZE);
+               rc = crst_table_upgrade(mm);
                if (rc)
                        return (unsigned long) rc;
                area = arch_get_unmapped_area_topdown(filp, addr, len,
index f6c3de2..e8b5962 100644 (file)
@@ -76,81 +76,52 @@ static void __crst_table_upgrade(void *arg)
        __tlb_flush_local();
 }
 
-int crst_table_upgrade(struct mm_struct *mm, unsigned long limit)
+int crst_table_upgrade(struct mm_struct *mm)
 {
        unsigned long *table, *pgd;
-       unsigned long entry;
-       int flush;
 
-       BUG_ON(limit > TASK_MAX_SIZE);
-       flush = 0;
-repeat:
+       /* upgrade should only happen from 3 to 4 levels */
+       BUG_ON(mm->context.asce_limit != (1UL << 42));
+
        table = crst_table_alloc(mm);
        if (!table)
                return -ENOMEM;
+
        spin_lock_bh(&mm->page_table_lock);
-       if (mm->context.asce_limit < limit) {
-               pgd = (unsigned long *) mm->pgd;
-               if (mm->context.asce_limit <= (1UL << 31)) {
-                       entry = _REGION3_ENTRY_EMPTY;
-                       mm->context.asce_limit = 1UL << 42;
-                       mm->context.asce_bits = _ASCE_TABLE_LENGTH |
-                                               _ASCE_USER_BITS |
-                                               _ASCE_TYPE_REGION3;
-               } else {
-                       entry = _REGION2_ENTRY_EMPTY;
-                       mm->context.asce_limit = 1UL << 53;
-                       mm->context.asce_bits = _ASCE_TABLE_LENGTH |
-                                               _ASCE_USER_BITS |
-                                               _ASCE_TYPE_REGION2;
-               }
-               crst_table_init(table, entry);
-               pgd_populate(mm, (pgd_t *) table, (pud_t *) pgd);
-               mm->pgd = (pgd_t *) table;
-               mm->task_size = mm->context.asce_limit;
-               table = NULL;
-               flush = 1;
-       }
+       pgd = (unsigned long *) mm->pgd;
+       crst_table_init(table, _REGION2_ENTRY_EMPTY);
+       pgd_populate(mm, (pgd_t *) table, (pud_t *) pgd);
+       mm->pgd = (pgd_t *) table;
+       mm->context.asce_limit = 1UL << 53;
+       mm->context.asce = __pa(mm->pgd) | _ASCE_TABLE_LENGTH |
+                          _ASCE_USER_BITS | _ASCE_TYPE_REGION2;
+       mm->task_size = mm->context.asce_limit;
        spin_unlock_bh(&mm->page_table_lock);
-       if (table)
-               crst_table_free(mm, table);
-       if (mm->context.asce_limit < limit)
-               goto repeat;
-       if (flush)
-               on_each_cpu(__crst_table_upgrade, mm, 0);
+
+       on_each_cpu(__crst_table_upgrade, mm, 0);
        return 0;
 }
 
-void crst_table_downgrade(struct mm_struct *mm, unsigned long limit)
+void crst_table_downgrade(struct mm_struct *mm)
 {
        pgd_t *pgd;
 
+       /* downgrade should only happen from 3 to 2 levels (compat only) */
+       BUG_ON(mm->context.asce_limit != (1UL << 42));
+
        if (current->active_mm == mm) {
                clear_user_asce();
                __tlb_flush_mm(mm);
        }
-       while (mm->context.asce_limit > limit) {
-               pgd = mm->pgd;
-               switch (pgd_val(*pgd) & _REGION_ENTRY_TYPE_MASK) {
-               case _REGION_ENTRY_TYPE_R2:
-                       mm->context.asce_limit = 1UL << 42;
-                       mm->context.asce_bits = _ASCE_TABLE_LENGTH |
-                                               _ASCE_USER_BITS |
-                                               _ASCE_TYPE_REGION3;
-                       break;
-               case _REGION_ENTRY_TYPE_R3:
-                       mm->context.asce_limit = 1UL << 31;
-                       mm->context.asce_bits = _ASCE_TABLE_LENGTH |
-                                               _ASCE_USER_BITS |
-                                               _ASCE_TYPE_SEGMENT;
-                       break;
-               default:
-                       BUG();
-               }
-               mm->pgd = (pgd_t *) (pgd_val(*pgd) & _REGION_ENTRY_ORIGIN);
-               mm->task_size = mm->context.asce_limit;
-               crst_table_free(mm, (unsigned long *) pgd);
-       }
+
+       pgd = mm->pgd;
+       mm->pgd = (pgd_t *) (pgd_val(*pgd) & _REGION_ENTRY_ORIGIN);
+       mm->context.asce_limit = 1UL << 31;
+       mm->context.asce = __pa(mm->pgd) | _ASCE_TABLE_LENGTH |
+                          _ASCE_USER_BITS | _ASCE_TYPE_SEGMENT;
+       mm->task_size = mm->context.asce_limit;
+       crst_table_free(mm, (unsigned long *) pgd);
+
        if (current->active_mm == mm)
                set_user_asce(mm);
 }