parisc: Fix some PTE/TLB race conditions and optimize __flush_tlb_range based on timing results

The increased use of pdtlb/pitlb instructions seemed to increase the
frequency of random segmentation faults building packages. Further, we
had a number of cases where TLB inserts would repeatedly fail and all
forward progress would stop. The Haskell ghc package caused a lot of
trouble in this area. The final indication of a race in pte handling was
this syslog entry on sibaris (C8000):

 swap_free: Unused swap offset entry 00000004
 BUG: Bad page map in process mysqld  pte:00000100 pmd:019bbec5
 addr:00000000ec464000 vm_flags:00100073 anon_vma:0000000221023828 mapping: (null) index:ec464
 CPU: 1 PID: 9176 Comm: mysqld Not tainted 4.0.0-2-parisc64-smp #1 Debian 4.0.5-1
 Backtrace:
  [<0000000040173eb0>] show_stack+0x20/0x38
  [<0000000040444424>] dump_stack+0x9c/0x110
  [<00000000402a0d38>] print_bad_pte+0x1a8/0x278
  [<00000000402a28b8>] unmap_single_vma+0x3d8/0x770
  [<00000000402a4090>] zap_page_range+0xf0/0x198
  [<00000000402ba2a4>] SyS_madvise+0x404/0x8c0

Note that the pte value is 0 except for the accessed bit 0x100. This bit
shouldn't be set without the present bit.

It should be noted that the madvise system call is probably a trigger for many
of the random segmentation faults.

In looking at the kernel code, I found the following problems:

1) The pte_clear define didn't take TLB lock when clearing a pte.
2) We didn't test pte present bit inside lock in exception support.
3) The pte and tlb locks needed to merged in order to ensure consistency
between page table and TLB. This also has the effect of serializing TLB
broadcasts on SMP systems.

The attached change implements the above and a few other tweaks to try
to improve performance. Based on the timing code, TLB purges are very
slow (e.g., ~ 209 cycles per page on rp3440). Thus, I think it
beneficial to test the split_tlb variable to avoid duplicate purges.
Probably, all PA 2.0 machines have combined TLBs.

I dropped using __flush_tlb_range in flush_tlb_mm as I realized all
applications and most threads have a stack size that is too large to
make this useful. I added some comments to this effect.

Since implementing 1 through 3, I haven't had any random segmentation
faults on mx3210 (rp3440) in about one week of building code and running
as a Debian buildd.

Signed-off-by: John David Anglin <dave.anglin@bell.net>
Cc: stable@vger.kernel.org # v3.18+
Signed-off-by: Helge Deller <deller@gmx.de>

2 files changed, 66 insertions, 42 deletions

diff --git a/arch/parisc/include/asm/pgtable.h b/arch/parisc/include/asm/pgtable.h
index 0a183756d6ec..f93c4a4e6580 100644
--- a/arch/parisc/include/asm/pgtable.h
+++ b/arch/parisc/include/asm/pgtable.h
@@ -16,7 +16,7 @@
 #include <asm/processor.h>
 #include <asm/cache.h>
 
-extern spinlock_t pa_dbit_lock;
+extern spinlock_t pa_tlb_lock;
 
 /*
  * kern_addr_valid(ADDR) tests if ADDR is pointing to valid kernel
@@ -33,6 +33,19 @@ extern spinlock_t pa_dbit_lock;
  */
 #define kern_addr_valid(addr)   (1)
 
+/* Purge data and instruction TLB entries.  Must be called holding
+ * the pa_tlb_lock.  The TLB purge instructions are slow on SMP
+ * machines since the purge must be broadcast to all CPUs.
+ */
+
+static inline void purge_tlb_entries(struct mm_struct *mm, unsigned long addr)
+{
+        mtsp(mm->context, 1);
+        pdtlb(addr);
+        if (unlikely(split_tlb))
+                pitlb(addr);
+}
+
 /* Certain architectures need to do special things when PTEs
  * within a page table are directly modified.  Thus, the following
  * hook is made available.
@@ -42,15 +55,20 @@ extern spinlock_t pa_dbit_lock;
                 *(pteptr) = (pteval);                           \
         } while(0)
 
-extern void purge_tlb_entries(struct mm_struct *, unsigned long);
+#define pte_inserted(x)                                         \
+        ((pte_val(x) & (_PAGE_PRESENT|_PAGE_ACCESSED))          \
+         == (_PAGE_PRESENT|_PAGE_ACCESSED))
 
-#define set_pte_at(mm, addr, ptep, pteval)                      \
-        do {                                                    \
+#define set_pte_at(mm, addr, ptep, pteval)                      \
+        do {                                                    \
+                pte_t old_pte;                                  \
                 unsigned long flags;                            \
-                spin_lock_irqsave(&pa_dbit_lock, flags);        \
-                set_pte(ptep, pteval);                          \
-                purge_tlb_entries(mm, addr);                    \
-                spin_unlock_irqrestore(&pa_dbit_lock, flags);   \
+                spin_lock_irqsave(&pa_tlb_lock, flags);         \
+                old_pte = *ptep;                                \
+                set_pte(ptep, pteval);                          \
+                if (pte_inserted(old_pte))                      \
+                        purge_tlb_entries(mm, addr);            \
+                spin_unlock_irqrestore(&pa_tlb_lock, flags);    \
         } while (0)
 
 #endif /* !__ASSEMBLY__ */
@@ -268,7 +286,7 @@ extern unsigned long *empty_zero_page;
 
 #define pte_none(x)     (pte_val(x) == 0)
 #define pte_present(x)  (pte_val(x) & _PAGE_PRESENT)
-#define pte_clear(mm,addr,xp)   do { pte_val(*(xp)) = 0; } while (0)
+#define pte_clear(mm, addr, xp)  set_pte_at(mm, addr, xp, __pte(0))
 
 #define pmd_flag(x)     (pmd_val(x) & PxD_FLAG_MASK)
 #define pmd_address(x)  ((unsigned long)(pmd_val(x) &~ PxD_FLAG_MASK) << PxD_VALUE_SHIFT)
@@ -435,15 +453,15 @@ static inline int ptep_test_and_clear_young(struct vm_area_struct *vma, unsigned
         if (!pte_young(*ptep))
                 return 0;
 
-        spin_lock_irqsave(&pa_dbit_lock, flags);
+        spin_lock_irqsave(&pa_tlb_lock, flags);
         pte = *ptep;
         if (!pte_young(pte)) {
-                spin_unlock_irqrestore(&pa_dbit_lock, flags);
+                spin_unlock_irqrestore(&pa_tlb_lock, flags);
                 return 0;
         }
         set_pte(ptep, pte_mkold(pte));
         purge_tlb_entries(vma->vm_mm, addr);
-        spin_unlock_irqrestore(&pa_dbit_lock, flags);
+        spin_unlock_irqrestore(&pa_tlb_lock, flags);
         return 1;
 }
 
@@ -453,11 +471,12 @@ static inline pte_t ptep_get_and_clear(struct mm_struct *mm, unsigned long addr,
         pte_t old_pte;
         unsigned long flags;
 
-        spin_lock_irqsave(&pa_dbit_lock, flags);
+        spin_lock_irqsave(&pa_tlb_lock, flags);
         old_pte = *ptep;
-        pte_clear(mm,addr,ptep);
-        purge_tlb_entries(mm, addr);
-        spin_unlock_irqrestore(&pa_dbit_lock, flags);
+        set_pte(ptep, __pte(0));
+        if (pte_inserted(old_pte))
+                purge_tlb_entries(mm, addr);
+        spin_unlock_irqrestore(&pa_tlb_lock, flags);
 
         return old_pte;
 }
@@ -465,10 +484,10 @@ static inline pte_t ptep_get_and_clear(struct mm_struct *mm, unsigned long addr,
 static inline void ptep_set_wrprotect(struct mm_struct *mm, unsigned long addr, pte_t *ptep)
 {
         unsigned long flags;
-        spin_lock_irqsave(&pa_dbit_lock, flags);
+        spin_lock_irqsave(&pa_tlb_lock, flags);
         set_pte(ptep, pte_wrprotect(*ptep));
         purge_tlb_entries(mm, addr);
-        spin_unlock_irqrestore(&pa_dbit_lock, flags);
+        spin_unlock_irqrestore(&pa_tlb_lock, flags);
 }
 
 #define pte_same(A,B)   (pte_val(A) == pte_val(B))
diff --git a/arch/parisc/include/asm/tlbflush.h b/arch/parisc/include/asm/tlbflush.h
index 9d086a599fa0..e84b96478193 100644
--- a/arch/parisc/include/asm/tlbflush.h
+++ b/arch/parisc/include/asm/tlbflush.h
@@ -13,6 +13,9 @@
  * active at any one time on the Merced bus.  This tlb purge
  * synchronisation is fairly lightweight and harmless so we activate
  * it on all systems not just the N class.
+
+ * It is also used to ensure PTE updates are atomic and consistent
+ * with the TLB.
  */
 extern spinlock_t pa_tlb_lock;
 
@@ -24,20 +27,24 @@ extern void flush_tlb_all_local(void *);
 
 #define smp_flush_tlb_all()     flush_tlb_all()
 
+int __flush_tlb_range(unsigned long sid,
+        unsigned long start, unsigned long end);
+
+#define flush_tlb_range(vma, start, end) \
+        __flush_tlb_range((vma)->vm_mm->context, start, end)
+
+#define flush_tlb_kernel_range(start, end) \
+        __flush_tlb_range(0, start, end)
+
 /*
  * flush_tlb_mm()
  *
- * XXX This code is NOT valid for HP-UX compatibility processes,
- * (although it will probably work 99% of the time). HP-UX
- * processes are free to play with the space id's and save them
- * over long periods of time, etc. so we have to preserve the
- * space and just flush the entire tlb. We need to check the
- * personality in order to do that, but the personality is not
- * currently being set correctly.
- *
- * Of course, Linux processes could do the same thing, but
- * we don't support that (and the compilers, dynamic linker,
- * etc. do not do that).
+ * The code to switch to a new context is NOT valid for processes
+ * which play with the space id's.  Thus, we have to preserve the
+ * space and just flush the entire tlb.  However, the compilers,
+ * dynamic linker, etc, do not manipulate space id's, so there
+ * could be a significant performance benefit in switching contexts
+ * and not flushing the whole tlb.
  */
 
 static inline void flush_tlb_mm(struct mm_struct *mm)
@@ -45,10 +52,18 @@ static inline void flush_tlb_mm(struct mm_struct *mm)
         BUG_ON(mm == &init_mm); /* Should never happen */
 
 #if 1 || defined(CONFIG_SMP)
+        /* Except for very small threads, flushing the whole TLB is
+         * faster than using __flush_tlb_range.  The pdtlb and pitlb
+         * instructions are very slow because of the TLB broadcast.
+         * It might be faster to do local range flushes on all CPUs
+         * on PA 2.0 systems.
+         */
         flush_tlb_all();
 #else
         /* FIXME: currently broken, causing space id and protection ids
-         *  to go out of sync, resulting in faults on userspace accesses.
+         * to go out of sync, resulting in faults on userspace accesses.
+         * This approach needs further investigation since running many
+         * small applications (e.g., GCC testsuite) is faster on HP-UX.
          */
         if (mm) {
                 if (mm->context != 0)
@@ -65,22 +80,12 @@ static inline void flush_tlb_page(struct vm_area_struct *vma,
 {
         unsigned long flags, sid;
 
-        /* For one page, it's not worth testing the split_tlb variable */
-
-        mb();
         sid = vma->vm_mm->context;
         purge_tlb_start(flags);
         mtsp(sid, 1);
         pdtlb(addr);
-        pitlb(addr);
+        if (unlikely(split_tlb))
+                pitlb(addr);
         purge_tlb_end(flags);
 }
-
-void __flush_tlb_range(unsigned long sid,
-        unsigned long start, unsigned long end);
-
-#define flush_tlb_range(vma,start,end) __flush_tlb_range((vma)->vm_mm->context,start,end)
-
-#define flush_tlb_kernel_range(start, end) __flush_tlb_range(0,start,end)
-
 #endif