Re: crypto: fscrypt: crypto_create_tfm_node memory leak

[Eric Biggers <ebiggers@kernel.org>]

On Mon, Feb 17, 2025 at 06:43:15PM +0000, Lucas Tanure wrote:
> Hi,
> 
> I am working with Android 13 and V5.15 kernel. During our development,
> I found a memory leak using kmemleak.
> 
> Steps I did to find the memleak:
> mount -t debugfs debugfs /sys/kernel/debug
> echo scan=5 > /sys/kernel/debug/kmemleak
> cat /sys/kernel/debug/kmemleak
> 
> Stack I got (hundreds of them):
> unreferenced object 0xffffff8101d31000 (size 1024):
>   comm "binder:1357_2", pid 1357, jiffies 4294899464 (age 394.468s)
>   hex dump (first 32 bytes):
>     00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00  ................
>     00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00  ................
>   backtrace:
>     [<ffffffd327cac060>] crypto_create_tfm_node+0x64/0x228
>     [<ffffffd3279f8c4c>] fscrypt_prepare_key+0xbc/0x230
>     [<ffffffd3279f9758>] fscrypt_setup_v1_file_key+0x48c/0x510
>     [<ffffffd3279f8394>] fscrypt_setup_encryption_info+0x210/0x43c
>     [<ffffffd3279f8108>] fscrypt_prepare_new_inode+0x128/0x1a4
>     [<ffffffd327bcc878>] f2fs_new_inode+0x27c/0x89c
>     [<ffffffd327bce7c4>] f2fs_mkdir+0x78/0x278
>     [<ffffffd32796a3bc>] vfs_mkdir+0x138/0x204
>     [<ffffffd32796a108>] do_mkdirat+0x88/0x204
>     [<ffffffd32796a068>] __arm64_sys_mkdirat+0x40/0x58
>     [<ffffffd3274be5d4>] invoke_syscall+0x60/0x150
>     [<ffffffd3274be528>] el0_svc_common+0xc8/0x114
>     [<ffffffd3274be3f0>] do_el0_svc+0x28/0x98
>     [<ffffffd328abcf88>] el0_svc+0x28/0x90
>     [<ffffffd328abcefc>] el0t_64_sync_handler+0x88/0xec
>     [<ffffffd32741164c>] el0t_64_sync+0x1b8/0x1bc
> 
> After checking upstream, I came up with the following:
> cff805b1518f  fscrypt: fix keyring memory leak on mount failure
> 
> But my kernel has this patch. So I continued to dig around this and
> saw the function fscrypt_prepare_key in fs/crypto/keysetup.c for
> V5.15.
> I can't see the pointer tfm being used anywhere or saved, and
> smp_store_release doesn't kfree it.
> Is smp_store_release doing something with that pointer that makes this
> memory leak a false positive?
> 
> Any help with this issue would be much appreciated.
> Thanks

The pointer to the crypto_skcipher 'tfm' is stored in the fscrypt_inode_info
(previously fscrypt_info) which is stored in inode::i_crypt_info.  It gets freed
when the inode is evicted.  I don't know why you're getting a kmemleak warning.
Perhaps f2fs in that version of the kernel has a bug that is leaking inodes.

smp_store_release is just a fancy way of doing a store that includes a memory
barrier.

- Eric