Hi Luis, On 1/23/2023 10:02 PM, Mark Rutland wrote: > On Fri, Jan 20, 2023 at 03:25:57PM -0300, Luis Claudio R. Goncalves wrote: >> On Fri, Jan 20, 2023 at 03:37:35PM +0000, Mark Rutland wrote: >>> On Fri, Jan 20, 2023 at 09:51:07AM -0300, Luis Claudio R. Goncalves wrote: >>>> On Thu, Jan 19, 2023 at 08:49:28PM -0700, Jens Axboe wrote: >>>>> On 1/19/23 8:44?PM, Luis Claudio R. Goncalves wrote: >>>>>> On Thu, Jan 19, 2023 at 04:09:44PM -0700, Jens Axboe wrote: >>>>>>> On 1/19/23 2:03?PM, Salvatore Bonaccorso wrote: >>>>>>>> Hi Luis, all, >>>>>>>> >>>>>>>> On Thu, Jan 19, 2023 at 10:16:34AM -0300, Luis Claudio R. Goncalves wrote: >>>>>>>>> On Thu, Jan 19, 2023 at 12:38:25PM +0100, Pavel Machek wrote: >>>>>>>>>> Hi! >>>>>>>>>> >>>>>>>>>>>> I'm pleased to announce the 5.10.162-rt78 stable release. >>>>>>>>>>>> >>>>>>>>>>>> You can get this release via the git tree at: >>>>>>>>>>>> >>>>>>>>>>>> git://git.kernel.org/pub/scm/linux/kernel/git/rt/linux-stable-rt.git >>>>>>>>>>>> >>>>>>>>>>>> branch: v5.10-rt >>>>>>>>>>>> Head SHA1: 143ef105f40a65f3ddd57121d4b4bc36eb10cc06 >>>>>>>>>>>> >>>>>>>>>>>> Or to build 5.10.162-rt78 directly, the following patches should be applied: >>>>>>>>>> >>>>>>>>>>> I see that vanilla 5.10.162-rt78 fails to build with arm64 defconfig. [0] Full log [1] >>>>>>>>>>> Any pointers on what maybe wrong? >>>>>>>>>> >>>>>>>>>> We see the same failure. >>>>>>>>>> >>>>>>>>>>> AS arch/arm64/kernel/entry.o >>>>>>>>>>> arch/arm64/kernel/entry.S: Assembler messages: >>>>>>>>>>> arch/arm64/kernel/entry.S:763: Error: immediate out of range at operand 3 -- `and x2,x19,#((1<<1)|(1<<0)|(1<<2)|(1<<3)|(1<<4)|(1<<5)|(1<<6)|(1<<13)|(1<<7))' >>>>>>>>>>> make[2]: *** [scripts/Makefile.build:367: arch/arm64/kernel/entry.o] Error 1 >>>>>>>>>>> make[1]: *** [scripts/Makefile.build:503: arch/arm64/kernel] Error 2 >>>>>>>>>>> make: *** [Makefile:1837: arch/arm64] Error 2 >>>>>>>>>> >>>>>>>>>> The line is: >>>>>>>>>> >>>>>>>>>>> and x2, x19, #_TIF_WORK_MASK >>>>>>>>> >>>>>>>>> I believe this is related to the arch/arm64/include/asm/thread_info.h >>>>>>>>> changes in 5.10.162-rt78, specifically: >>>>>>>>> >>>>>>>>> 79a9991e87fe arm64: add support for TIF_NOTIFY_SIGNAL >>>>>>>>> 1ba44dcf789d Merge tag 'v5.10.162' into v5.10-rt >>>>>>>>> >>>>>>>>> The first one is the original change, coming from stable v5.10.162 and the >>>>>>>>> second one has the merge conflict I fixed in that file due to the existence >>>>>>>>> of TIF_NEED_RESCHED_LAZY in PREEMPT_RT. >>>>>>>>> >>>>>>>>> It escaped me that having TIF_NEED_RESCHED_LAZY set to 13 breaks the AND >>>>>>>>> statement reported above. Looking at >>>>>>>>> >>>>>>>>> b5a5a01d8e9a arm64: uaccess: remove addr_limit_user_check() >>>>>>>>> >>>>>>>>> specially this note >>>>>>>>> >>>>>>>>> To ensure that _TIF_WORK_MASK can be used as an immediate value in an >>>>>>>>> AND instruction (as it is in `ret_to_user`), TIF_MTE_ASYNC_FAULT is >>>>>>>>> renumbered to keep the constituent bits of _TIF_WORK_MASK contiguous. >>>>>>>>> >>>>>>>>> I understand that I need to either have to renumber TIF_NEED_RESCHED_LAZY >>>>>>>>> to 8, with the risk of breaking something else, or backport commit >>>>>>>>> b5a5a01d8e9a in order to remove TIF_FSCHECK and then safely renumber >>>>>>>>> TIF_NEED_RESCHED_LAZY. >>>>>>>>> >>>>>>>>> Guidance is welcome here :) >>>>>>>> >>>>>>>> Should we loop in here Jens, as having some overview of the needed >>>>>>>> changes for io_uring rebase in the 5.10.y version? (doing so in the >>>>>>>> mail). >>>>>>> >>>>>>> Huh that's funky, I built and (runtime) tested this on arm64 >>>>>>> specifically. But I do remember some details about the first 8 bits on >>>>>>> arm, but not arm64. >>>>>>> >>>>>>> I guess we need to twiddle that asm to deal with eg 16 bits, rather than >>>>>>> attempt to backport any TIF removal patches. >>>>>> >>>>>> One simple solution, tested with defconfig plus FTRACE options (including >>>>>> FTRACE_SYSCALLS) enabled, is: >>>>>> >>>>>> diff --git a/arch/arm64/include/asm/thread_info.h b/arch/arm64/include/asm/thread_info.h >>>>>> index 6eb36a2126e8..37f19bb49d38 100644 >>>>>> --- a/arch/arm64/include/asm/thread_info.h >>>>>> +++ b/arch/arm64/include/asm/thread_info.h >>>>>> @@ -70,12 +70,12 @@ void arch_release_task_struct(struct task_struct *tsk); >>>>>> #define TIF_FSCHECK 5 /* Check FS is USER_DS on return */ >>>>>> #define TIF_MTE_ASYNC_FAULT 6 /* MTE Asynchronous Tag Check Fault */ >>>>>> #define TIF_NOTIFY_SIGNAL 7 /* signal notifications exist */ >>>>>> -#define TIF_SYSCALL_TRACE 8 /* syscall trace active */ >>>>>> +#define TIF_NEED_RESCHED_LAZY 8 >>>>>> #define TIF_SYSCALL_AUDIT 9 /* syscall auditing */ >>>>>> #define TIF_SYSCALL_TRACEPOINT 10 /* syscall tracepoint for ftrace */ >>>>>> #define TIF_SECCOMP 11 /* syscall secure computing */ >>>>>> #define TIF_SYSCALL_EMU 12 /* syscall emulation active */ >>>>>> -#define TIF_NEED_RESCHED_LAZY 13 >>>>>> +#define TIF_SYSCALL_TRACE 13 /* syscall trace active */ >>>>>> #define TIF_MEMDIE 18 /* is terminating due to OOM killer */ >>>>>> #define TIF_FREEZE 19 >>>>>> #define TIF_RESTORE_SIGMASK 20 >>>>>> >>>>>> Would that be acceptable? With that we ensure the bits in >>>>>> _TIF_WORK_MASK are contiguous and within the 8 bits limit you >>>>>> mentioned. And TIF_SYSCALL_TRACE did not seem to have any (build) >>>>>> problem with the new value. >>>>> >>>>> That should work too, the _TIF_WORK_MASK bits being in the lower 8 bits >>>>> is really all we should care about. >>>> >>>> Jens, Salvatore, Mike, I ran a few tests and confirmed that the current asm >>>> code is not restricted to 8 bits. The problems is that there is a >>>> requirement for the mask bits to be contiguous in that specific context. >>> >>> Just to confirm from the arm64 side, the instruction using this just requires >>> the bits to be contiguous, there's no restriction on *which* bits those are. >> >> Thank you, that's really helpful! >> >>> If you're going to mess around with the arm64 bits, please could you Cc someone >>> form the arm64 side? e.g. I fixed a similar issue in mainline in commit: >>> >>> 870d16757ba8918c ("arm64: make _TIF_WORK_MASK bits contiguous") >>> >>> ... and either Will Deacon or Catalin Marinas may have had comments as they're >>> the arm64 maintainers... >> >> Just to avoid confusion here, this change is specific to the v5.10-rt, not >> applicable upstream nor to newer RT. We only saw the problem because >> TIF_NOTIFY_SIGNAL was mapped to a bit used by TIF_NEED_RESCHED_LAZY in >> v5.10-rt (the PREEMPT_RT changes on top of stable v5.10). This is why >> nobody from the arm64 side was copied initially, as we were trying to >> assert what was the problem. > > Sure, and sorry, I think my reply came across a bit stronger than I intended. I > probably should have said something like: "please feel free to rope in one of > us from the arm64 side". > > I know from experience that this area is fairly subtle, and I'd like to help to > ensure that the fix doesn't introduce a subtle breakage or interact poorly with > future backports. > >>>> The explanation from commit b5a5a01d8e9a ("arm64: uaccess: remove >>>> addr_limit_user_check()") describes quite well our case: >>>> >>>> To ensure that _TIF_WORK_MASK can be used as an immediate value in an >>>> AND instruction (as it is in `ret_to_user`), TIF_MTE_ASYNC_FAULT is >>>> renumbered to keep the constituent bits of _TIF_WORK_MASK contiguous. >>>> >>>> My question is: do you prefer renumbering the bits or the neat asm hack >>>> that Mike proposed? >>> >>> I would strongly recommend renumbering the bits over changing the asm. That's >>> going to be closer to what mainline has already done, and it avoids introducing >>> weird ifdeffery. >>> >>> That said, rather than swapping TIF_SYSCALL_TRACE and TIF_NEED_RESCHED_LAZY, >>> you could just shuffle the bits down-by-one, keeping all the existing >>> contiguity, e.g. >>> >>> #define TIF_NEED_RESCHED_LAZY 8 >>> #define TIF_SYSCALL_TRACE 9 >>> #define TIF_SYSCALL_AUDIT 10 >>> #define TIF_SYSCALL_TRACEPOINT 11 >>> >>> ... and so on. >> >> Would something like this be a good interpretation of your suggestion? >> >> diff --git a/arch/arm64/include/asm/thread_info.h b/arch/arm64/include/asm/thread_info.h >> index 6eb36a212..2afd9ceb6 100644 >> --- a/arch/arm64/include/asm/thread_info.h >> +++ b/arch/arm64/include/asm/thread_info.h >> @@ -70,12 +70,12 @@ void arch_release_task_struct(struct task_struct *tsk); >> #define TIF_FSCHECK 5 /* Check FS is USER_DS on return */ >> #define TIF_MTE_ASYNC_FAULT 6 /* MTE Asynchronous Tag Check Fault */ >> #define TIF_NOTIFY_SIGNAL 7 /* signal notifications exist */ >> -#define TIF_SYSCALL_TRACE 8 /* syscall trace active */ >> -#define TIF_SYSCALL_AUDIT 9 /* syscall auditing */ >> -#define TIF_SYSCALL_TRACEPOINT 10 /* syscall tracepoint for ftrace */ >> -#define TIF_SECCOMP 11 /* syscall secure computing */ >> -#define TIF_SYSCALL_EMU 12 /* syscall emulation active */ >> -#define TIF_NEED_RESCHED_LAZY 13 >> +#define TIF_NEED_RESCHED_LAZY 8 >> +#define TIF_SYSCALL_TRACE 9 /* syscall trace active */ >> +#define TIF_SYSCALL_AUDIT 10 /* syscall auditing */ >> +#define TIF_SYSCALL_TRACEPOINT 11 /* syscall tracepoint for ftrace */ >> +#define TIF_SECCOMP 12 /* syscall secure computing */ >> +#define TIF_SYSCALL_EMU 13 /* syscall emulation active */ > I can confirm v5.10.162-rt179 which has above fix builds fine for arm64 defconfig. Thanks everyone addressing this issue! Regards Vignesh