SMP - fixed usage of stale TLB entries

- when kernel copies from userspace, it must be sure that the TLB
  entries are not stale and thus the referenced memory is correct

- everytime we change a process' address space we set p_stale_tlb
  bits for all CPUs.

- Whenever a cpu finds its bit set when it wants to access the
  process' memory, it refreshes the TLB

- it is more conservative than it needs to be but it has low
  overhead than checking precisely

common/include/minix/bitmap.h

@@ -17,4 +17,18 @@
 #define SET_BIT(map,bit) ( MAP_CHUNK(map,bit) |= (1 << CHUNK_OFFSET(bit) ))
 #define UNSET_BIT(map,bit) ( MAP_CHUNK(map,bit) &= ~(1 << CHUNK_OFFSET(bit) ))
 
+#if defined(CONFIG_SMP) && defined(__GNUC__)
+#ifndef __ASSEMBLY__
+static inline bits_fill(bitchunk_t * chunks, unsigned bits)
+{
+	unsigned c, cnt;
+
+	cnt = BITMAP_CHUNKS(bits);
+	for (c = 0; c < cnt; c++)
+		bit_fill(chunks[c]);
+}
+#endif
+#endif
+
+
 #endif	/* _BITMAP_H */

kernel/arch/i386/memory.c

@@ -138,8 +138,8 @@ PRIVATE phys_bytes createpde(
 /*===========================================================================*
  *				lin_lin_copy				     *
  *===========================================================================*/
-PRIVATE int lin_lin_copy(const struct proc *srcproc, vir_bytes srclinaddr, 
-	const struct proc *dstproc, vir_bytes dstlinaddr, vir_bytes bytes)
+PRIVATE int lin_lin_copy(struct proc *srcproc, vir_bytes srclinaddr, 
+	struct proc *dstproc, vir_bytes dstlinaddr, vir_bytes bytes)
 {
 	u32_t addr;
 	proc_nr_t procslot;
@@ -167,6 +167,19 @@ PRIVATE int lin_lin_copy(const struct proc *srcproc, vir_bytes srclinaddr,
 		vir_bytes chunk = bytes;
 		int changed = 0;
 
+#ifdef CONFIG_SMP
+		unsigned cpu = cpuid;
+
+		if (GET_BIT(srcproc->p_stale_tlb, cpu)) {
+			changed = 1;
+			UNSET_BIT(srcproc->p_stale_tlb, cpu);
+		}
+		if (GET_BIT(dstproc->p_stale_tlb, cpu)) {
+			changed = 1;
+			UNSET_BIT(dstproc->p_stale_tlb, cpu);
+		}
+#endif
+
 		/* Set up 4MB ranges. */
 		srcptr = createpde(srcproc, srclinaddr, &chunk, 0, &changed);
 		dstptr = createpde(dstproc, dstlinaddr, &chunk, 1, &changed);

kernel/proc.h

@@ -38,6 +38,11 @@ struct proc {
   bitchunk_t p_cpu_mask[BITMAP_CHUNKS(CONFIG_MAX_CPUS)]; /* what CPUs is hte
 							    process allowed to
 							    run on */
+  bitchunk_t p_stale_tlb[BITMAP_CHUNKS(CONFIG_MAX_CPUS)]; /* On which cpu are
+				possibly stale entries from this process and has
+				to be fresed the next kernel touches this
+				processes memory
+				 */
 #endif
 
   /* Accounting statistics that get passed to the process' scheduler */

kernel/system/do_update.c

@@ -127,6 +127,11 @@ PUBLIC int do_update(struct proc * caller, message * m_ptr)
   printf("do_update: curr ptproc %d\n", get_cpulocal_var(ptproc)->p_endpoint);
 #endif
 
+#ifdef CONFIG_SMP
+  bits_fill(src_rp->p_stale_tlb, CONFIG_MAX_CPUS);
+  bits_fill(dst_rp->p_stale_tlb, CONFIG_MAX_CPUS);
+#endif
+
   return OK;
 }
 

kernel/system/do_vmctl.c

@@ -159,6 +159,16 @@ PUBLIC int do_vmctl(struct proc * caller, message * m_ptr)
 		 * cpu
 		 */
 		RTS_UNSET(p, RTS_VMINHIBIT);
+#ifdef CONFIG_SMP
+		/*
+		 * We don't know whether kernel has the changed mapping
+		 * installed to access userspace memory. And if so, on what CPU.
+		 * More over we don't know what mapping has changed and how and
+		 * therefore we must invalidate all mappings we have anywhere.
+		 * Next time we map memory, we map it fresh.
+		 */
+		bits_fill(p->p_stale_tlb, CONFIG_MAX_CPUS);
+#endif
 		return OK;
   }