Re: Multigrain timestamps do not work on RISC-V

[Jeff Layton <jlayton@kernel.org>]

On Wed, 2025-01-29 at 18:06 -0500, Jeff Layton wrote:
> On Wed, 2025-01-29 at 18:20 +0100, Andreas Schwab wrote:
> > On Jan 29 2025, Andreas Schwab wrote:
> > 
> > > My guess would be that something in inode_set_ctime_current is going
> > > wrong.
> > 
> > The bug is in the arch_cmpxchg macros in <asm/cmpxchg.h>, they mishandle
> > atomic exchange of u32 values:
> > 
> >     # fs/inode.c:2802: 	if (cns == now.tv_nsec && inode->i_ctime_sec == now.tv_sec) {
> >     #NO_APP
> >             ld	a5,-96(s0)		# _33, now.tv_nsec
> >     # fs/inode.c:2802: 	if (cns == now.tv_nsec && inode->i_ctime_sec == now.tv_sec) {
> >             slli	a4,s2,32	#, _49, cns
> >             srli	a4,a4,32	#, _49, _49
> >     # fs/inode.c:2802: 	if (cns == now.tv_nsec && inode->i_ctime_sec == now.tv_sec) {
> >             beq	a4,a5,.L1248	#, _49, _33,
> >     .L1205:
> >             addi	a3,s1,120	#, __ai_ptr, inode
> >     .L1221:
> >     # fs/inode.c:2809: 	if (try_cmpxchg(&inode->i_ctime_nsec, &cur, now.tv_nsec)) {
> >     #APP
> >     # 2809 "fs/inode.c" 1
> >             0:	lr.w a2, 0(a3)	# __ret, *__ai_ptr_20
> >             bne  a2, a4, 1f	# __ret, _49
> > 
> > Here the unsigned extended value of cur (a4) is compared with the sign
> > extended value of inode->i_ctime_nsec (a2).  They cannot match if cur
> > has I_CTIME_QUERIED set (the sign bit).  The lr/sc loop is exited before
> > the store conditional is executed.
> > 
> >             sc.w.rl a1, a5, 0(a3)	# __rc, _33, *__ai_ptr_20
> >             bnez a1, 0b	# __rc
> >             fence rw,rw
> >     1:
> > 
> >     # 0 "" 2
> >     #NO_APP
> >             sext.w	a5,a2	# __ret, __ret
> > 
> > A redundant sign extension of the current contents of inode->i_ctime_nsec.
> > 
> >     # fs/inode.c:2809: 	if (try_cmpxchg(&inode->i_ctime_nsec, &cur, now.tv_nsec)) {
> >             bne	a5,s2,.L1211	#, __ret, cns,
> > 
> > Here the sign extended value of inode->i_ctime_nsec (a5) is compared
> > with the sign extended expected value (s2).  They match, and try_cmpxchg
> > returns true, but the store never happend.
> > 
> 
> Wow, nice catch! Had me worried there for a bit!
> 

For the record: 

Bit 30 is also unused. When I wrote these patches, I considered using
that instead of the sign bit since I wasn't sure whether I might run
into problems using the sign bit.

Obviously, fixing the macro seems like the better solution, but if
fixing this efficiently is a problem, then moving I_CTIME_QUERIED to
bit 30 is also an option.

Cheers,
-- 
Jeff Layton <jlayton@kernel.org>

_______________________________________________
linux-riscv mailing list
linux-riscv@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-riscv