Re: [PATCH 1/2] nvme-hwmon: Cache-line-align the NVME SMART log-buffer

[Guenter Roeck]

On 9/9/22 12:19, Serge Semin wrote:
> Recent commit 52fde2c07da6 ("nvme: set dma alignment to dword") has caused
> a regression on our platform. It turned out that the nvme_get_log() method
> invocation caused the nvme_hwmon_data structure instance corruption. In
> particular the nvme_hwmon_data.ctrl pointer was overwritten either with
> zeros or with garbage. After some researches we discovered that the
> problem happened even before the actual NVME DMA execution, but during the
> buffer mapping. Since our platform was DMA-noncoherent the mapping implied
> the cache-lines invalidations or write-backs depending on the
> DMA-direction parameter. In case of the NVME SMART log getting the DMA
> was performed from-device-to-memory, thus the cache-invalidation was
> activated during the buffer mapping. Since the log-buffer wasn't
> cache-line aligned the cache-invalidation caused the neighbour data
> discard. The neighbouring data turned to be the data surrounding the
> buffer in the framework of the nvme_hwmon_data structure.
>
> In order to fix that we need to make sure that the whole log-buffer is
> defined within the cache-line-aligned memory region so the
> cache-invalidation procedure wouldn't involve the adjacent data. By doing
> so we not only get rid from the denoted problem but also fulfill the
> requirement explicitly described in [1].
>
> After a deeper researches we found out that the denoted commit wasn't a
> root cause of the problem. It just revealed the invalidity by activating
> the DMA-based NVME SMART log getting performed in the framework of the
> NVME hwmon driver. The problem was here since the initial commit of the
> driver.
>
> [1] Documentation/core-api/dma-api.rst
>
> Fixes: 400b6a7b13a3 ("nvme: Add hardware monitoring support")
> Signed-off-by: Serge Semin <Sergey.Semin@xxxxxxxxxxxxxxxxxxxx>

Thanks for tracking this down and for the fix.

Reviewed-by: Guenter Roeck <linux@xxxxxxxxxxxx>

Guenter

> ---
>
> Folks, I've thoroughly studied the whole NVME subsystem looking for
> similar problems. Turned out there is one more place which may cause the
> same issue. It's connected with the opal_dev.{cmd,req} buffers passed to
> the nvme_sec_submit() method. The rest of the buffers involved in the NVME
> DMA are either allocated by kmalloc (must be cache-line-aligned by design)
> or bounced-buffered if allocated on the stack (see the blk_rq_map_kern()
> method implementation).
> ---
>   drivers/nvme/host/hwmon.c | 3 ++-
>   1 file changed, 2 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/nvme/host/hwmon.c b/drivers/nvme/host/hwmon.c
> index 0a586d712920..94192ab7a02d 100644
> --- a/drivers/nvme/host/hwmon.c
> +++ b/drivers/nvme/host/hwmon.c
> @@ -10,9 +10,10 @@
>   
>   #include "nvme.h"
>   
> +/* DMA-noncoherent platforms require the cache-aligned buffers */
>   struct nvme_hwmon_data {
> +	struct nvme_smart_log log ____cacheline_aligned;
>   	struct nvme_ctrl *ctrl;
> -	struct nvme_smart_log log;
>   	struct mutex read_lock;
>   };
>