Re: [PATCH net-next] net: socket: integrate sockfd_lookup() and sockfd_lookup_light()

[yajun.deng@linux.dev]

September 24, 2021 10:56 AM, "Al Viro" <viro@zeniv.linux.org.uk> wrote:

> On Wed, Sep 22, 2021 at 02:31:06PM +0800, Yajun Deng wrote:
> 
>> -#define sockfd_put(sock) fput(sock->file)
>> int net_ratelimit(void);
>> +#define sockfd_put(sock) \
>> +do { \
>> + struct fd *fd = (struct fd *)&sock->file; \
> 
> Have you even bothered to take a look at struct fd declaration?
> Or struct socket one, for that matter... What we have there is
> ...
> struct file *file;
> struct sock *sk;
> ...
> 
> You are taking the address of 'file' field. And treat it as
> a pointer to a structure consisting of struct file * and
> unsigned int.
> 
>> + \
>> + if (fd->flags & FDPUT_FPUT) \
> 
> ... so that would take first 4 bytes in the memory occupied
> by 'sk' field of struct socket and check if bit 0 is set.
> 
> And what significance would that bit have, pray tell? On
> little-endian architectures it's going to be the least
> significant bit in the first byte in 'sk' field. I.e.
> there you are testing if the contents of 'sk' (a pointer
> to struct sock) happens to be odd. It won't be. The
> same goes for 32bit big-endian - there you will be checking
> the least significant bit of the 4th byte of 'sk', which
> again is asking 'is the pointer stored there odd for some
> reason?'
> 
> On 64bit big-endian you are checking if the bit 32 of
> the address of object sock->sk points to is set. And the
> answer to that is "hell knows and how could that possibly
> be relevant to anything?"

Well, the forced conversion is wrong. sorry for that.