Re: [PATCH v4 1/1] scsi: ufs: core: add device level exception support

[Bart Van Assche <bvanassche@acm.org>]

On 3/26/25 7:47 PM, Bao D. Nguyen wrote:
> On 3/26/2025 3:49 AM, Bart Van Assche wrote:
>> On 3/25/25 6:15 PM, Bao D. Nguyen wrote:
>>> The existing "struct utp_upiu_query_v4_0" probably has a bug in it. 
>>> It does not use the  __attribute__((__packed__)) attribute. The 
>>> compiler is free to add padding in this structure, resulting in the 
>>> read attribute value being incorrect. I plan to provide a separate 
>>> patch to fix this issue.
>>
>> Adding __attribute__((__packed__)) or __packed to data structures that
>> don't need it is not an improvement but is a change that makes
>> processing slower on architectures that do not support unaligned
>> accesses. Instead of adding __packed to data structures in their
>> entirety, only add it to those members that need it and check the
>> structure size as follows:
>>
>> static_assert(sizeof(...) == ...);
>>
> Thank you for the info on this, Bart.
> IMO, this response upiu data should be __packed because the data coming 
> from the hardware follows a strict format as defined by the spec. If we 
> support __pack each individual field which data may be read by the 
> driver (the attribute read commands) and check the validity of their 
> sizes, it may add some complexity?

Hi Bao,

As explained in my previous email, adding __packed to data structures in
their entirety is a bad practice. Please don't do this.

Regarding your question: I have not yet seen any data structure that
represents an on-the-wire data format where every single data member
has to be annotated with __packed. Only data members that are not
aligned to a natural boundary need this annotation. Examples are
available in this header file: include/scsi/srp.h.

Thanks,

Bart.