[PATCH] media:fimc-capture: Fix a possible data inconsistency due to a data race in fimc_subdev_set_fmt()
Accesses to ctx->s_frame.width and ctx->s_frame.height should be protected by the lock fimc->lock to guarantee that width and height are consistent. Here is an example in fimc_subdev_get_fmt():
struct fimc_frame *ff = &ctx->s_frame; // Alias mutex_lock(&fimc->lock); mf->width = ff->width; mf->height = ff->height;
However, ctx->s_frame.width and ctx->s_frame.height are accessed without holding the lock fimc->lock in fimc_subdev_set_fmt():
mf->width = ctx->s_frame.width; mf->height = ctx->s_frame.height;
And thus a harmful data race can occur, which can make ctx->s_frame.width inconsistent with ctx->s_frame.height, if ctx->s_frame.height is updated right after ctx->s_frame.width is accessed by another thread.
This possible bug is found by an experimental static analysis tool developed by our team, BassCheck[1]. This tool analyzes the locking APIs to extract function pairs that can be concurrently executed, and then analyzes the instructions in the paired functions to identify possible concurrency bugs including data races and atomicity violations. The above possible bug is reported when our tool analyzes the source code of Linux 6.2.
To fix this possible data race, the lock operation mutex_lock(&fimc->lock) is moved to the front of the accesses to these two variables. With this patch applied, our tool no longer reports the bug, with the kernel configuration allyesconfig for x86_64. Due to the lack of associated hardware, we cannot test the patch in runtime testing, and just verify it according to the code logic.
[1] https://sites.google.com/view/basscheck/
Fixes: 88fa8311ee36 ("[media] s5p-fimc: Add support for ISP Writeback ...") Signed-off-by: Tuo Li islituo@gmail.com Cc: stable@vger.kernel.org Reported-by: BassCheck bass@buaa.edu.cn --- drivers/media/platform/samsung/exynos4-is/fimc-capture.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/drivers/media/platform/samsung/exynos4-is/fimc-capture.c b/drivers/media/platform/samsung/exynos4-is/fimc-capture.c index a0d43bf892e6..5c8b67f92c65 100644 --- a/drivers/media/platform/samsung/exynos4-is/fimc-capture.c +++ b/drivers/media/platform/samsung/exynos4-is/fimc-capture.c @@ -1546,6 +1546,7 @@ static int fimc_subdev_set_fmt(struct v4l2_subdev *sd, fimc_alpha_ctrl_update(ctx);
fimc_capture_mark_jpeg_xfer(ctx, ffmt->color); + mutex_lock(&fimc->lock); if (fmt->pad == FIMC_SD_PAD_SOURCE) { ff = &ctx->d_frame; /* Sink pads crop rectangle size */ @@ -1555,7 +1556,6 @@ static int fimc_subdev_set_fmt(struct v4l2_subdev *sd, ff = &ctx->s_frame; }
- mutex_lock(&fimc->lock); set_frame_bounds(ff, mf->width, mf->height);
if (fmt->pad == FIMC_SD_PAD_SINK_FIFO)
On 23/12/2023 17:43, Tuo Li wrote:
Accesses to ctx->s_frame.width and ctx->s_frame.height should be protected by the lock fimc->lock to guarantee that width and height are consistent. Here is an example in fimc_subdev_get_fmt():
struct fimc_frame *ff = &ctx->s_frame; // Alias mutex_lock(&fimc->lock); mf->width = ff->width; mf->height = ff->height;
However, ctx->s_frame.width and ctx->s_frame.height are accessed without holding the lock fimc->lock in fimc_subdev_set_fmt():
mf->width = ctx->s_frame.width; mf->height = ctx->s_frame.height;
Other places setting parts of s_frame, like fimc_capture_try_format() or fimc_capture_try_selection(), do not have mutex.
And thus a harmful data race can occur, which can make ctx->s_frame.width
Harmful how?
inconsistent with ctx->s_frame.height, if ctx->s_frame.height is updated right after ctx->s_frame.width is accessed by another thread.
This possible bug is found by an experimental static analysis tool developed by our team, BassCheck[1]. This tool analyzes the locking APIs to extract function pairs that can be concurrently executed, and then analyzes the instructions in the paired functions to identify possible concurrency bugs including data races and atomicity violations. The above possible bug is reported when our tool analyzes the source code of Linux 6.2.
To fix this possible data race, the lock operation mutex_lock(&fimc->lock) is moved to the front of the accesses to these two variables. With this patch applied, our tool no longer reports the bug, with the kernel configuration allyesconfig for x86_64. Due to the lack of associated hardware, we cannot test the patch in runtime testing, and just verify it according to the code logic.
You wrote long four paragraphs which have basically almost zero relevant information, whether this locking is needed or not. Your bass description is not relevant... or actually making things worse because I am certain you are fixing it just to fix your report, not to fix real issue.
Instead provide the report.
Fixes: 88fa8311ee36 ("[media] s5p-fimc: Add support for ISP Writeback ...") Signed-off-by: Tuo Li islituo@gmail.com Cc: stable@vger.kernel.org Reported-by: BassCheck bass@buaa.edu.cn
Run checkpatch, you will see the warning.
Best regards, Krzysztof
Hi Krzysztof,
Thanks for your reply very much. It is very helpful. I am really sorry to confuse you.
On 2023/12/24 18:11, Krzysztof Kozlowski wrote:
On 23/12/2023 17:43, Tuo Li wrote:
Accesses to ctx->s_frame.width and ctx->s_frame.height should be protected by the lock fimc->lock to guarantee that width and height are consistent. Here is an example in fimc_subdev_get_fmt():
struct fimc_frame *ff = &ctx->s_frame; // Alias mutex_lock(&fimc->lock); mf->width = ff->width; mf->height = ff->height;
However, ctx->s_frame.width and ctx->s_frame.height are accessed without holding the lock fimc->lock in fimc_subdev_set_fmt():
mf->width = ctx->s_frame.width; mf->height = ctx->s_frame.height;
Other places setting parts of s_frame, like fimc_capture_try_format() or fimc_capture_try_selection(), do not have mutex.
And thus a harmful data race can occur, which can make ctx->s_frame.width
Harmful how?
The function set_frame_crop() which updates s_frame.width and s_frame.height is called by fimc_cap_s_selection(). fimc_cap_s_selection() is an ioctl function and it is likely to be able to execute concurrently with other functions such as fimc_subdev_set_fmt(). However, in fimc_subdev_set_fmt(), the accesses to s_frame.width and s_frame.height are not protected by the mutex lock fimc->lock.
If fimc_cap_s_selection() and fimc_subdev_set_fmt() can execute concurrently and the execution orders is like this:
1. ctx->s_frame.width is accessed and assigned to mf->width in fimc_subdev_set_fmt() Line 1552 in fimc-capture.c 2. ctx->s_frame.width and ctx->s_frame.height is updated by fimc_cap_s_selection() Line 1329 in fimc-capture.c 3. ctx->s_frame.height is accessed and assigned to mf->height in fimc_subdev_set_fmt() Line 1553 in fimc-capture.c
The width and height assigned to mf are not coming from the same ctx->s_frame configuration and can cause data inconsistency.
Besides, the functions fimc_subdev_set_selection() and fimc_subdev_set_fmt() exist in the same driver interface named fimc_subdev_pad_ops. Therefore, it seems that fimc_subdev_set_fmt() and fimc_subdev_set_selection() can also execute concurrently. However, if the execution order of fimc_subdev_set_selection() and fimc_subdev_set_fmt() is like mentioned above, a data inconsistency can also occur.
I analyze this possible data race manually according to the code logic, and I am not sure whether all accesses to configurations such as width, height, f_width, and f_height should be protected by a mutex lock to make them consistent. I am really sorry to trouble you, and any feedback will be appreciated!
inconsistent with ctx->s_frame.height, if ctx->s_frame.height is updated right after ctx->s_frame.width is accessed by another thread.
This possible bug is found by an experimental static analysis tool developed by our team, BassCheck[1]. This tool analyzes the locking APIs to extract function pairs that can be concurrently executed, and then analyzes the instructions in the paired functions to identify possible concurrency bugs including data races and atomicity violations. The above possible bug is reported when our tool analyzes the source code of Linux 6.2.
To fix this possible data race, the lock operation mutex_lock(&fimc->lock) is moved to the front of the accesses to these two variables. With this patch applied, our tool no longer reports the bug, with the kernel configuration allyesconfig for x86_64. Due to the lack of associated hardware, we cannot test the patch in runtime testing, and just verify it according to the code logic.
You wrote long four paragraphs which have basically almost zero relevant information, whether this locking is needed or not. Your bass description is not relevant... or actually making things worse because I am certain you are fixing it just to fix your report, not to fix real issue.
Instead provide the report.
I am sorry to confuse you, and I wrote these descriptions according to researcher guidelines in the kernel documentation Documentation/process/researcher-guidelines.rst I will provide a more concise patch in the future.
Fixes: 88fa8311ee36 ("[media] s5p-fimc: Add support for ISP Writeback ...") Signed-off-by: Tuo Li islituo@gmail.com Cc: stable@vger.kernel.org Reported-by: BassCheck bass@buaa.edu.cn
Run checkpatch, you will see the warning.
Thanks for your advice. I am sorry to bother you, and I will be careful in the subsequent patches.
Best regards, Krzysztof
Sincerely, Tuo Li
linux-stable-mirror@lists.linaro.org
-
Krzysztof Kozlowski -
Li Tuo -
Tuo Li