| Message ID | YFnsqPphqvItA3z2@mwanda (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | [v2] cifsd: fix error handling in ksmbd_server_init() | expand |
> > The error handling in ksmbd_server_init() uses "one function to free everything style" which is > impossible to audit and leads to several canonical bugs. When we free something that wasn't allocated > it may be uninitialized, an error pointer, freed in a different function or we try freeing "foo->bar" > when "foo" is a NULL pointer. And since the code is impossible to audit then it leads to memory leaks. > > In the ksmbd_server_init() function, every goto will lead to a crash because we have not allocated the > work queue but we call > ksmbd_workqueue_destroy() which tries to flush a NULL work queue. > Another bug is if ksmbd_init_buffer_pools() fails then it leads to a double free because we free > "work_cache" twice. A third type of bug is that we forgot to call ksmbd_release_inode_hash() so that > is a resource leak. > > A better way to write error handling is for every function to clean up after itself and never leave > things partially allocated. Then we can use "free the last successfully allocated resource" style. > That way when someone is reading the code they can just track the last resource in their head and > verify that the goto matches what they expect. > > In this patch I modified ksmbd_ipc_init() to clean up after itself and then I converted > ksmbd_server_init() to use gotos to clean up. > > Fixes: cabcebc31de4 ("cifsd: introduce SMB3 kernel server") > Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com> > --- > v2: remove __exit annotation from ksmbd_release_inode_hash() as detected by the kbuild-bot I will apply. Thanks for your patch!
diff --git a/fs/cifsd/server.c b/fs/cifsd/server.c index 85862c3ea7c0..31e454cb3ce2 100644 --- a/fs/cifsd/server.c +++ b/fs/cifsd/server.c @@ -567,39 +567,52 @@ static int __init ksmbd_server_init(void) ret = server_conf_init(); if (ret) - return ret; + goto err_unregister; ret = ksmbd_init_buffer_pools(); if (ret) - return ret; + goto err_unregister; ret = ksmbd_init_session_table(); if (ret) - goto error; + goto err_destroy_pools; ret = ksmbd_ipc_init(); if (ret) - goto error; + goto err_free_session_table; ret = ksmbd_init_global_file_table(); if (ret) - goto error; + goto err_ipc_release; ret = ksmbd_inode_hash_init(); if (ret) - goto error; + goto err_destroy_file_table; ret = ksmbd_crypto_create(); if (ret) - goto error; + goto err_release_inode_hash; ret = ksmbd_workqueue_init(); if (ret) - goto error; + goto err_crypto_destroy; return 0; -error: - ksmbd_server_shutdown(); +err_crypto_destroy: + ksmbd_crypto_destroy(); +err_release_inode_hash: + ksmbd_release_inode_hash(); +err_destroy_file_table: + ksmbd_free_global_file_table(); +err_ipc_release: + ksmbd_ipc_release(); +err_free_session_table: + ksmbd_free_session_table(); +err_destroy_pools: + ksmbd_destroy_buffer_pools(); +err_unregister: + class_unregister(&ksmbd_control_class); + return ret; } diff --git a/fs/cifsd/transport_ipc.c b/fs/cifsd/transport_ipc.c index c49e46fda9b1..e5f4d97b2924 100644 --- a/fs/cifsd/transport_ipc.c +++ b/fs/cifsd/transport_ipc.c @@ -887,11 +887,19 @@ int ksmbd_ipc_init(void) if (ret) { ksmbd_err("Failed to register KSMBD netlink interface %d\n", ret); - return ret; + goto cancel_work; } ida = ksmbd_ida_alloc(); - if (!ida) - return -ENOMEM; + if (!ida) { + ret = -ENOMEM; + goto unregister; + } return 0; + +unregister: + genl_unregister_family(&ksmbd_genl_family); +cancel_work: + cancel_delayed_work_sync(&ipc_timer_work); + return ret; } diff --git a/fs/cifsd/vfs_cache.c b/fs/cifsd/vfs_cache.c index af92fab5b7ae..34e045f27230 100644 --- a/fs/cifsd/vfs_cache.c +++ b/fs/cifsd/vfs_cache.c @@ -236,7 +236,7 @@ int __init ksmbd_inode_hash_init(void) return 0; } -void __exit ksmbd_release_inode_hash(void) +void ksmbd_release_inode_hash(void) { vfree(inode_hashtable); } diff --git a/fs/cifsd/vfs_cache.h b/fs/cifsd/vfs_cache.h index 7d23657c86c6..04ab5967a9ae 100644 --- a/fs/cifsd/vfs_cache.h +++ b/fs/cifsd/vfs_cache.h @@ -194,7 +194,7 @@ void ksmbd_set_fd_limit(unsigned long limit); */ int __init ksmbd_inode_hash_init(void); -void __exit ksmbd_release_inode_hash(void); +void ksmbd_release_inode_hash(void); enum KSMBD_INODE_STATUS { KSMBD_INODE_STATUS_OK,
The error handling in ksmbd_server_init() uses "one function to free everything style" which is impossible to audit and leads to several canonical bugs. When we free something that wasn't allocated it may be uninitialized, an error pointer, freed in a different function or we try freeing "foo->bar" when "foo" is a NULL pointer. And since the code is impossible to audit then it leads to memory leaks. In the ksmbd_server_init() function, every goto will lead to a crash because we have not allocated the work queue but we call ksmbd_workqueue_destroy() which tries to flush a NULL work queue. Another bug is if ksmbd_init_buffer_pools() fails then it leads to a double free because we free "work_cache" twice. A third type of bug is that we forgot to call ksmbd_release_inode_hash() so that is a resource leak. A better way to write error handling is for every function to clean up after itself and never leave things partially allocated. Then we can use "free the last successfully allocated resource" style. That way when someone is reading the code they can just track the last resource in their head and verify that the goto matches what they expect. In this patch I modified ksmbd_ipc_init() to clean up after itself and then I converted ksmbd_server_init() to use gotos to clean up. Fixes: cabcebc31de4 ("cifsd: introduce SMB3 kernel server") Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com> --- v2: remove __exit annotation from ksmbd_release_inode_hash() as detected by the kbuild-bot fs/cifsd/server.c | 33 +++++++++++++++++++++++---------- fs/cifsd/transport_ipc.c | 14 +++++++++++--- fs/cifsd/vfs_cache.c | 2 +- fs/cifsd/vfs_cache.h | 2 +- 4 files changed, 36 insertions(+), 15 deletions(-)