support madvise(MADV_REMOVE) partially

This MADV_REMOVE works with a mapping which is
- created with shmat() and
- not sharing memobj with other mappings.

kernel/include/memobj.h

@@ -31,6 +31,7 @@ enum {
 	/* for memobj.flags */
 	MF_HAS_PAGER	= 0x0001,
 	MF_SHMDT_OK	= 0x0002,
+	MF_IS_REMOVABLE	= 0x0004,
 };
 
 struct memobj {
@@ -120,6 +121,11 @@ static inline int memobj_has_pager(struct memobj *obj)
 	return !!(obj->flags & MF_HAS_PAGER);
 }
 
+static inline int memobj_is_removable(struct memobj *obj)
+{
+	return !!(obj->flags & MF_IS_REMOVABLE);
+}
+
 int fileobj_create(int fd, struct memobj **objp, int *maxprotp);
 struct shmid_ds;
 int shmobj_create(struct shmid_ds *ds, struct memobj **objp);

kernel/shmobj.c

@@ -33,11 +33,13 @@ static ihk_spinlock_t shmobj_list_lock_body = SPIN_LOCK_UNLOCKED;
 static memobj_release_func_t shmobj_release;
 static memobj_ref_func_t shmobj_ref;
 static memobj_get_page_func_t shmobj_get_page;
+static memobj_invalidate_page_func_t shmobj_invalidate_page;
 
 static struct memobj_ops shmobj_ops = {
 	.release =	&shmobj_release,
 	.ref =		&shmobj_ref,
 	.get_page =	&shmobj_get_page,
+	.invalidate_page =	&shmobj_invalidate_page,
 };
 
 static struct shmobj *to_shmobj(struct memobj *memobj)
@@ -158,7 +160,7 @@ int shmobj_create_indexed(struct shmid_ds *ds, struct shmobj **objp)
 
 	error = shmobj_create(ds, &obj);
 	if (!error) {
-		obj->flags |= MF_SHMDT_OK;
+		obj->flags |= MF_SHMDT_OK | MF_IS_REMOVABLE;
 		*objp = to_shmobj(obj);
 	}
 	return error;
@@ -369,3 +371,30 @@ out:
 			memobj, off, p2align, physp, error);
 	return error;
 }
+
+static int shmobj_invalidate_page(struct memobj *memobj, uintptr_t phys,
+		size_t pgsize)
+{
+	struct shmobj *obj = to_shmobj(memobj);
+	int error;
+	struct page *page;
+
+	dkprintf("shmobj_invalidate_page(%p,%#lx,%#lx)\n", memobj, phys, pgsize);
+
+	if (!(page = phys_to_page(phys))
+			|| !(page = page_list_lookup(obj, page->offset))) {
+		error = 0;
+		goto out;
+	}
+
+	if (ihk_atomic_read(&page->count) == 1) {
+		if (page_unmap(page)) {
+			ihk_mc_free_pages(phys_to_virt(phys), pgsize/PAGE_SIZE);
+		}
+	}
+
+	error = 0;
+out:
+	dkprintf("shmobj_invalidate_page(%p,%#lx,%#lx):%d\n", memobj, phys, pgsize, error);
+	return error;
+}

kernel/syscall.c

@@ -2691,6 +2691,8 @@ SYSCALL_DECLARE(madvise)
 	uintptr_t addr;
 	struct vm_range *range;
 	int error;
+	uintptr_t s;
+	uintptr_t e;
 
 	dkprintf("[%d]sys_madvise(%lx,%lx,%x)\n",
 			ihk_mc_get_processor_id(), start, len0, advice);
@@ -2732,10 +2734,7 @@ SYSCALL_DECLARE(madvise)
 	case MADV_DONTNEED:
 	case MADV_DONTFORK:
 	case MADV_DOFORK:
-		break;
-
 	case MADV_REMOVE:
-		error = -EACCES;
 		break;
 
 	case MADV_HWPOISON:
@@ -2777,7 +2776,17 @@ SYSCALL_DECLARE(madvise)
 			goto out;
 		}
 
-		if (!range->memobj || !memobj_has_pager(range->memobj)) {
+		if (advice == MADV_REMOVE) {
+			if (!range->memobj || !memobj_is_removable(range->memobj)) {
+				dkprintf("sys_madvise(%lx,%lx,%x):"
+						"not removable [%lx-%lx)\n",
+						start, len0, advice,
+						range->start, range->end);
+				error = -EACCES;
+				goto out;
+			}
+		}
+		else if (!range->memobj || !memobj_has_pager(range->memobj)) {
 			dkprintf("[%d]sys_madvise(%lx,%lx,%x):has not pager"
 					"[%lx-%lx) %lx\n",
 					ihk_mc_get_processor_id(), start,
@@ -2797,6 +2806,28 @@ SYSCALL_DECLARE(madvise)
 			error = -EINVAL;
 			goto out;
 		}
+
+		s = start;
+		if (s < range->start) {
+			s = range->start;
+		}
+		e = end;
+		if (range->end < e) {
+			e = range->end;
+		}
+
+		if (advice == MADV_REMOVE) {
+			error = invalidate_process_memory_range(
+					thread->vm, range, s, e);
+			if (error) {
+				kprintf("sys_madvise(%lx,%lx,%x):[%lx-%lx):"
+						"invalidate failed. %d\n",
+						start, len0, advice,
+						range->start, range->end,
+						error);
+				goto out;
+			}
+		}
 	}
 
 	error = 0;