Hi Mikulas,
On 7/22/2025 4:14 AM, Mikulas Patocka wrote:
> Hi
>
> I think that the race condition between dm_ima_measure_on_device_resume
> and dm_ima_measure_on_device_remove is still possible. The patch prevents
> it from triggering the use-after-free bug, but the race may still result
> in these two events being randomly swapped.
Not exactly. After the invocation of dm_ima_measure_on_device_remove(),
the invocation of dm_ima_measure_on_device_resume will fail because both
the inactive and active tables have been cleared, the hash cell of has
been cleared (namely dm_get_mdptr() will return NULL), and
dm_ima_measure_on_device_resume() will fail when trying to invoke
dm_ima_alloc_and_copy_name_uuid(). However I agree that the lock can not
guarantee the order of the measure as shown below.
>
> I think that a proper fix would be to move dm_ima_measure_on_device* calls
> to a position where device mapper locks are held, so that the calls will
> be serialized by the dm locks.
>
> Would it be possible to move them inside down_write(&_hash_lock)? Or,
> would there be some problems with that?
>
> Mikulas
Glad that you mention about the _hash_lock, because I have a pending
patch set (not completed yet) in which it uses _hash_lock to prevent the
potential out-of-order measure. Using _hash_lock() for
dm_ima_measure_on_table_clear() and dm_ima_measure_on_table_load() will
prevent the measure procedure measures an incorrect inactive table (or
other similar out-of-order) when there are concurrent table load and
clear processes as shown below:
1) load table X (set inactive table as V1)
2) clear table X
old_map = X
new_map = NULL
ready to invoke dm_ima_measure_on_table_clear(), but it is interrupted
by a concurrent load table Y
3) load table Y (set inactive table as V2)
4) clear table X (continue)
read inactive table V2 instead of V1
The tricky ones are dm_ima_measure_on_device_resume(). For
dm_ima_measure_on_device_resume(), using _hash_lock to protect the whole
resume procedure is a bad idea. Using _hash_lock to protect
dm_ima_measure_on_device_resume() is not enough as shown below:
1) load inactive_v1 / table X
2) resume table X
preempted before invoke dm_ima_measure_on_device_resume()
3) load inactive_v2 / table Y
4) resume table X (continue)
measure resume of inactive_v2 instead of inactive_v1
To prevent the out-of-order of dm_ima_measure_on_device_resume() and
_load()/_clear(), I think we need to track dm_ima_measurements in
dm_table, so when we load, clear and resume the table, we could know
which one to use. Will send the RFC patch set for further discussion.
>
>
> On Sun, 20 Jul 2025, Li Lingfeng wrote:
>
>> There is a window between freeing md->ima.active_table.hash and setting
>> md->ima.active_table.hash to NULL in dm_ima_measure_on_device_resume().
>> If dm_ima_measure_on_device_remove() accesses md->ima.active_table.hash
>> concurrently during this window, it could lead to a double free or UAF,
>> as shown below:
>>
>> BUG: KASAN: slab-use-after-free in dm_ima_measure_on_device_remove...
>> Read of size 71 at addr ffff88817bb9e220 by task dmsetup/2303
>>
>> CPU: 2 UID: 0 PID: 2303 Comm: dmsetup Not tainted 6.16.0-rc6-dirty...
>> Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.16.3...
>> Call Trace:
>> <TASK>
>> dump_stack_lvl+0x5b/0x80
>> print_address_description.constprop.0+0x88/0x310
>> print_report+0x12f/0x21d
>> kasan_report+0xcc/0x190
>> kasan_check_range+0x104/0x1b0
>> __asan_memcpy+0x23/0x60
>> dm_ima_measure_on_device_remove+0x3fc/0x6c0
>> dev_remove+0x123/0x1e0
>> ctl_ioctl+0x2a2/0x480
>> dm_ctl_ioctl+0xe/0x20
>> __x64_sys_ioctl+0xc7/0x110
>> do_syscall_64+0x72/0x390
>> entry_SYSCALL_64_after_hwframe+0x76/0x7e
>>
>> To reproduce this issue, add a delay between freeing
>> md->ima.active_table.hash and setting it to NULL, using the following
>> steps:
>> dmsetup create mydevice --table "0 2097152 linear /dev/sda 0"
>> dmsetup suspend mydevice
>> dmsetup reload mydevice --table "0 2097152 linear /dev/sdb 0"
>> dmsetup resume mydevice &
>> dmsetup remove mydevice
>>
>> Introduce ima_lock to prevent concurrent calls to dm_ima_measure_*
>> functions to fix it.
>>
>> Fixes: 91ccbbac1747 ("dm ima: measure data on table load")
>> Signed-off-by: Li Lingfeng <lilingfe...@huawei.com>
>> ---
>> drivers/md/dm-core.h | 1 +
>> drivers/md/dm-ima.c | 38 +++++++++++++++++++++++++++++++++-----
>> drivers/md/dm-ima.h | 4 ++++
>> drivers/md/dm.c | 3 ++-
>> 4 files changed, 40 insertions(+), 6 deletions(-)
>>
>> diff --git a/drivers/md/dm-core.h b/drivers/md/dm-core.h
>> index c889332e533b..24f321b2c2c4 100644
>> --- a/drivers/md/dm-core.h
>> +++ b/drivers/md/dm-core.h
>> @@ -145,6 +145,7 @@ struct mapped_device {
>> #endif
>>
>> #ifdef CONFIG_IMA
>> + struct mutex ima_lock;
>> struct dm_ima_measurements ima;
>> #endif
>> };
>> diff --git a/drivers/md/dm-ima.c b/drivers/md/dm-ima.c
>> index b90f34259fbb..0fefad6fbb9f 100644
>> --- a/drivers/md/dm-ima.c
>> +++ b/drivers/md/dm-ima.c
>> @@ -164,7 +164,7 @@ static int dm_ima_alloc_and_copy_capacity_str(struct
>> mapped_device *md, char **c
>> }
>>
>> /*
>> - * Initialize/reset the dm ima related data structure variables.
>> + * Reset the dm ima related data structure variables.
>> */
>> void dm_ima_reset_data(struct mapped_device *md)
>> {
>> @@ -172,6 +172,23 @@ void dm_ima_reset_data(struct mapped_device *md)
>> md->ima.dm_version_str_len = strlen(DM_IMA_VERSION_STR);
>> }
>>
>> +/*
>> + * Initialize the dm ima.
>> + */
>> +void dm_ima_init(struct mapped_device *md)
>> +{
>> + dm_ima_reset_data(md);
>> + mutex_init(&md->ima_lock);
>> +}
>> +
>> +/*
>> + * Destroy the dm ima related data structure.
>> + */
>> +void dm_ima_destroy(struct mapped_device *md)
>> +{
>> + mutex_destroy(&md->ima_lock);
>> +}
>> +
>> /*
>> * Build up the IMA data for each target, and finally measure.
>> */
>> @@ -195,9 +212,10 @@ void dm_ima_measure_on_table_load(struct dm_table
>> *table, unsigned int status_fl
>> const size_t hash_alg_prefix_len = strlen(DM_IMA_TABLE_HASH_ALG) + 1;
>> char table_load_event_name[] = "dm_table_load";
>>
>> + mutex_lock(&table->md->ima_lock);
>> ima_buf = dm_ima_alloc(DM_IMA_MEASUREMENT_BUF_LEN, GFP_KERNEL, noio);
>> if (!ima_buf)
>> - return;
>> + goto error;
>>
>> target_metadata_buf = dm_ima_alloc(DM_IMA_TARGET_METADATA_BUF_LEN,
>> GFP_KERNEL, noio);
>> if (!target_metadata_buf)
>> @@ -361,6 +379,7 @@ void dm_ima_measure_on_table_load(struct dm_table
>> *table, unsigned int status_fl
>> kfree(ima_buf);
>> kfree(target_metadata_buf);
>> kfree(target_data_buf);
>> + mutex_unlock(&table->md->ima_lock);
>> }
>>
>> /*
>> @@ -376,9 +395,10 @@ void dm_ima_measure_on_device_resume(struct
>> mapped_device *md, bool swap)
>> bool nodata = true;
>> int r;
>>
>> + mutex_lock(&md->ima_lock);
>> device_table_data = dm_ima_alloc(DM_IMA_DEVICE_BUF_LEN, GFP_KERNEL,
>> noio);
>> if (!device_table_data)
>> - return;
>> + goto error;
>>
>> r = dm_ima_alloc_and_copy_capacity_str(md, &capacity_str, noio);
>> if (r)
>> @@ -466,6 +486,7 @@ void dm_ima_measure_on_device_resume(struct
>> mapped_device *md, bool swap)
>> error:
>> kfree(capacity_str);
>> kfree(device_table_data);
>> + mutex_unlock(&md->ima_lock);
>> }
>>
>> /*
>> @@ -490,6 +511,7 @@ void dm_ima_measure_on_device_remove(struct
>> mapped_device *md, bool remove_all)
>> bool nodata = true;
>> int r;
>>
>> + mutex_lock(&md->ima_lock);
>> device_table_data = dm_ima_alloc(DM_IMA_DEVICE_BUF_LEN*2, GFP_KERNEL,
>> noio);
>> if (!device_table_data)
>> goto exit;
>> @@ -597,6 +619,7 @@ void dm_ima_measure_on_device_remove(struct
>> mapped_device *md, bool remove_all)
>>
>> kfree(dev_name);
>> kfree(dev_uuid);
>> + mutex_unlock(&md->ima_lock);
>> }
>>
>> /*
>> @@ -612,9 +635,10 @@ void dm_ima_measure_on_table_clear(struct mapped_device
>> *md, bool new_map)
>> bool nodata = true;
>> int r;
>>
>> + mutex_lock(&md->ima_lock);
>> device_table_data = dm_ima_alloc(DM_IMA_DEVICE_BUF_LEN, GFP_KERNEL,
>> noio);
>> if (!device_table_data)
>> - return;
>> + goto error;
>>
>> r = dm_ima_alloc_and_copy_capacity_str(md, &capacity_str, noio);
>> if (r)
>> @@ -696,6 +720,8 @@ void dm_ima_measure_on_table_clear(struct mapped_device
>> *md, bool new_map)
>> kfree(capacity_str);
>> error1:
>> kfree(device_table_data);
>> +error:
>> + mutex_unlock(&md->ima_lock);
>> }
>>
>> /*
>> @@ -708,9 +734,10 @@ void dm_ima_measure_on_device_rename(struct
>> mapped_device *md)
>> bool noio = true;
>> int r;
>>
>> + mutex_lock(&md->ima_lock);
>> if (dm_ima_alloc_and_copy_device_data(md, &new_device_data,
>> md->ima.active_table.num_targets,
>> noio))
>> - return;
>> + goto error;
>>
>> if (dm_ima_alloc_and_copy_name_uuid(md, &new_dev_name, &new_dev_uuid,
>> noio))
>> goto error;
>> @@ -745,4 +772,5 @@ void dm_ima_measure_on_device_rename(struct
>> mapped_device *md)
>> kfree(old_device_data);
>> kfree(new_dev_name);
>> kfree(new_dev_uuid);
>> + mutex_unlock(&md->ima_lock);
>> }
>> diff --git a/drivers/md/dm-ima.h b/drivers/md/dm-ima.h
>> index 568870a1a145..36bbcf1b25a0 100644
>> --- a/drivers/md/dm-ima.h
>> +++ b/drivers/md/dm-ima.h
>> @@ -57,6 +57,8 @@ struct dm_ima_measurements {
>> };
>>
>> void dm_ima_reset_data(struct mapped_device *md);
>> +void dm_ima_init(struct mapped_device *md);
>> +void dm_ima_destroy(struct mapped_device *md);
>> void dm_ima_measure_on_table_load(struct dm_table *table, unsigned int
>> status_flags);
>> void dm_ima_measure_on_device_resume(struct mapped_device *md, bool swap);
>> void dm_ima_measure_on_device_remove(struct mapped_device *md, bool
>> remove_all);
>> @@ -66,6 +68,8 @@ void dm_ima_measure_on_device_rename(struct mapped_device
>> *md);
>> #else
>>
>> static inline void dm_ima_reset_data(struct mapped_device *md) {}
>> +static inline void dm_ima_init(struct mapped_device *md) {}
>> +static inline void dm_ima_destroy(struct mapped_device *md) {}
>> static inline void dm_ima_measure_on_table_load(struct dm_table *table,
>> unsigned int status_flags) {}
>> static inline void dm_ima_measure_on_device_resume(struct mapped_device
>> *md, bool swap) {}
>> static inline void dm_ima_measure_on_device_remove(struct mapped_device
>> *md, bool remove_all) {}
>> diff --git a/drivers/md/dm.c b/drivers/md/dm.c
>> index 1726f0f828cc..b7eab324804c 100644
>> --- a/drivers/md/dm.c
>> +++ b/drivers/md/dm.c
>> @@ -2371,6 +2371,7 @@ static void free_dev(struct mapped_device *md)
>> unlock_fs(md);
>>
>> cleanup_mapped_device(md);
>> + dm_ima_destroy(md);
>>
>> WARN_ON_ONCE(!list_empty(&md->table_devices));
>> dm_stats_cleanup(&md->stats);
>> @@ -2506,7 +2507,7 @@ int dm_create(int minor, struct mapped_device **result)
>> if (!md)
>> return -ENXIO;
>>
>> - dm_ima_reset_data(md);
>> + dm_ima_init(md);
>>
>> *result = md;
>> return 0;
>> --
>> 2.46.1
>
> .
