Re: [Linux Kernel Bug] memory leak in ubi_attach

[Zhihao Cheng <chengzhihao1@huawei.com>]

[-- Attachment #1: Type: text/plain, Size: 3282 bytes --]

在 2024/1/24 22:41, Chenyuan Yang 写道:
> Hi Zhihao,
> 
> Thanks very much for your reply and further exploration of this crash!
> 
> I have attached the config to reproduce this memory leak. For the
> command line, you can use qemu to boot the kernel (You need to build
> the bullseye-image first).
> ```
> mkdir -p logs
> qemu-system-x86_64 \
> -m 2G \
> -smp 2 \
> -kernel linux/arch/x86/boot/bzImage \
> -append "console=ttyS0 root=/dev/sda earlyprintk=serial net.ifnames=0" \
> -drive file=image/bullseye-qemu.img,format=raw \
> -net user,host=10.0.2.10,hostfwd=tcp:127.0.0.1:10021-:22 \
> -net nic,model=e1000 \
> -enable-kvm \
> -nographic \
> -pidfile vm.pid \
> ```
> 
> In qemu, you can run
> ```
> gcc -pthread repro.c -o exe
> ./exe
> ```
> to reproduce the memory leak.
> 

The bad news is that I can't reproduce the problem with your kernel 
config. The good news is that I seem to find the cause of the problem by 
code reviewing:
alloc_ai -> kmem_cache_create -> kmem_cache_create_usercopy:
  s = __kmem_cache_alias
   s = find_mergeable // cannot find mergeable kmem_cache, s = NULL
  create_cache
   s = kmem_cache_zalloc
   __kmem_cache_create(s)
    sysfs_slab_add
     kobject_init_and_add(&s->kobj)
      kobject_init(kobj) // kobj->kref = 1
      kobject_add_varg
       kobject_set_name_vargs
        kvasprintf_const
         kstrdup_const
          kstrdup
           kmalloc_track_caller // allocated memory
        kobj->name = s  // assign allocated memory to kobj->name
       kobject_add_internal
        create_dir // failed caused by ENOMEM
        // kobj->name won't be released even 's' is released

You can reproduce it by executing ubiattach -m0 with reproduce.diff applied:
[root@localhost ~]# cat /sys/kernel/debug/kmemleak
unreferenced object 0xffff88813b488080 (size 16):
   comm "ubiattach", pid 9535, jiffies 4294943095
   hex dump (first 16 bytes):
     3a 30 30 30 30 30 35 36 00 00 00 00 00 00 00 00  :0000056........
   backtrace (crc 8a117ab9):
     [<ffffffff815f91c2>] __kmalloc_node_track_caller+0x322/0x420
     [<ffffffff81575efc>] kstrdup+0x3c/0x70
     [<ffffffff81575f8f>] kstrdup_const+0x5f/0x70
     [<ffffffff82535426>] kvasprintf_const+0xc6/0x110
     [<ffffffff84b88cb0>] kobject_set_name_vargs+0x40/0xd0
     [<ffffffff84b8911b>] kobject_init_and_add+0x8b/0xf0
     [<ffffffff815faba9>] sysfs_slab_add+0x119/0x1e0
     [<ffffffff815fc2dc>] __kmem_cache_create+0x41c/0x590
     [<ffffffff81586019>] kmem_cache_create_usercopy+0x169/0x300
     [<ffffffff815861c1>] kmem_cache_create+0x11/0x20
     [<ffffffff82ec8896>] ubi_attach+0xc6/0x1de0
     [<ffffffff82eb77a3>] ubi_attach_mtd_dev+0x8a3/0x1120
     [<ffffffff82eb8bfc>] ctrl_cdev_ioctl+0x1fc/0x270
     [<ffffffff816ca312>] __x64_sys_ioctl+0xf2/0x140
     [<ffffffff84bc3970>] do_syscall_64+0x50/0x140
     [<ffffffff84c0008b>] entry_SYSCALL_64_after_hwframe+0x63/0x6b

There are two ways to fix it:
1. Release kobj->name directly in the error baranch of 
kobject_add_internal() after calling create_dir().
2. Put kobj's refcnt in the error baranch of kobject_add_internal() 
after calling create_dir().
I'm not sure which one is better or whether there are other fix methods, 
maybe you can send an email in kobj related mailiing lists to confirm 
the problem.

[-- Attachment #2: reproduce.diff --]
[-- Type: text/plain, Size: 2746 bytes --]

diff --git a/drivers/mtd/ubi/attach.c b/drivers/mtd/ubi/attach.c
index ae5abe492b52..3694fd52c530 100644
--- a/drivers/mtd/ubi/attach.c
+++ b/drivers/mtd/ubi/attach.c
@@ -1447,6 +1447,7 @@ static int scan_all(struct ubi_device *ubi, struct ubi_attach_info *ai,
 	return err;
 }
 
+extern int g_inject;
 static struct ubi_attach_info *alloc_ai(void)
 {
 	struct ubi_attach_info *ai;
@@ -1461,9 +1462,11 @@ static struct ubi_attach_info *alloc_ai(void)
 	INIT_LIST_HEAD(&ai->alien);
 	INIT_LIST_HEAD(&ai->fastmap);
 	ai->volumes = RB_ROOT;
+	g_inject = 1;
 	ai->aeb_slab_cache = kmem_cache_create("ubi_aeb_slab_cache",
 					       sizeof(struct ubi_ainf_peb),
 					       0, 0, NULL);
+	g_inject = 0;
 	if (!ai->aeb_slab_cache) {
 		kfree(ai);
 		ai = NULL;
diff --git a/lib/kobject.c b/lib/kobject.c
index 59dbcbdb1c91..37da8b767557 100644
--- a/lib/kobject.c
+++ b/lib/kobject.c
@@ -204,7 +204,8 @@ static void kobject_init_internal(struct kobject *kobj)
 	kobj->state_initialized = 1;
 }
 
-
+struct kobject *g_k;
+int g_inject;
 static int kobject_add_internal(struct kobject *kobj)
 {
 	int error = 0;
@@ -235,6 +236,13 @@ static int kobject_add_internal(struct kobject *kobj)
 		 parent ? kobject_name(parent) : "<NULL>",
 		 kobj->kset ? kobject_name(&kobj->kset->kobj) : "<NULL>");
 
+	if (g_inject) {
+		error = -ENOMEM; // Simulate failure in create_dir->sysfs_create_dir_ns->kernfs_create_dir_ns->kernfs_new_node->__kernfs_new_node-->kmem_cache_zalloc->slab_alloc_node->slab_pre_alloc_hook->should_failslab
+		g_k = kobj;
+		pr_err("inject failure, kobj ref %d\n", refcount_read(&kobj->kref.refcount));
+		smp_wmb();
+	} else
+
 	error = create_dir(kobj);
 	if (error) {
 		kobj_kset_leave(kobj);
@@ -685,6 +693,12 @@ static void kobject_cleanup(struct kobject *kobj)
 			 kobject_name(kobj), kobj);
 	}
 
+	smp_rmb();
+	if (g_k == kobj) {
+		pr_err("free name %s\n", name);
+		dump_stack();
+	}
+
 	/* free name if we allocated it */
 	if (name) {
 		pr_debug("'%s': free name\n", name);
diff --git a/mm/slab_common.c b/mm/slab_common.c
index 238293b1dbe1..c76277004bd0 100644
--- a/mm/slab_common.c
+++ b/mm/slab_common.c
@@ -158,10 +158,12 @@ int slab_unmergeable(struct kmem_cache *s)
 	return 0;
 }
 
+extern int g_inject;
 struct kmem_cache *find_mergeable(unsigned int size, unsigned int align,
 		slab_flags_t flags, const char *name, void (*ctor)(void *))
 {
 	struct kmem_cache *s;
+	static int enter = 0;
 
 	if (slab_nomerge)
 		return NULL;
@@ -184,6 +186,11 @@ struct kmem_cache *find_mergeable(unsigned int size, unsigned int align,
 		if (size > s->size)
 			continue;
 
+		if (g_inject && !enter) {
+			enter = 1;
+			break;
+		}
+
 		if ((flags & SLAB_MERGE_SAME) != (s->flags & SLAB_MERGE_SAME))
 			continue;
 		/*