Message-ID: <CAJF2gTT6bezLuxe-+8tSUyBstbnW-=FQnN8N-Gno12C07+tgYQ@mail.gmail.com>
Date: Wed, 4 Dec 2024 17:10:26 +0800
From: Guo Ren <guoren@...nel.org>
To: Alexandre Ghiti <alex@...ti.fr>
Cc: paul.walmsley@...ive.com, palmer@...belt.com, bjorn@...osinc.com, 
	conor@...nel.org, leobras@...hat.com, peterz@...radead.org, 
	parri.andrea@...il.com, will@...nel.org, longman@...hat.com, 
	boqun.feng@...il.com, arnd@...db.de, alexghiti@...osinc.com, 
	linux-riscv@...ts.infradead.org, linux-kernel@...r.kernel.org, 
	Guo Ren <guoren@...ux.alibaba.com>
Subject: Re: [RESEND PATCH] riscv: Fixup boot failure when CONFIG_DEBUG_RT_MUTEXES=y

On Wed, Dec 4, 2024 at 4:27 PM Alexandre Ghiti <alex@...ti.fr> wrote:
>
> Hi Guo,
>
> On 30/11/2024 16:33, guoren@...nel.org wrote:
> > From: Guo Ren <guoren@...ux.alibaba.com>
> >
> > When CONFIG_DEBUG_RT_MUTEXES=y, mutex_lock->rt_mutex_try_acquire
> > would change from rt_mutex_cmpxchg_acquire to
> > rt_mutex_slowtrylock():
> >       raw_spin_lock_irqsave(&lock->wait_lock, flags);
> >       ret = __rt_mutex_slowtrylock(lock);
> >       raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
> >
> > Because queued_spin_#ops to ticket_#ops is changed one by one by
> > jump_label, raw_spin_lock/unlock would cause a deadlock during the
> > changing.
> >
> > That means in arch/riscv/kernel/jump_label.c:
> > 1.
> > arch_jump_label_transform_queue() ->
> > mutex_lock(&text_mutex); +-> raw_spin_lock  -> queued_spin_lock
> >                        |-> raw_spin_unlock -> queued_spin_unlock
> > patch_insn_write -> change the raw_spin_lock to ticket_lock
> > mutex_unlock(&text_mutex);
> > ...
> >
> > 2. /* Dirty the lock value */
> > arch_jump_label_transform_queue() ->
> > mutex_lock(&text_mutex); +-> raw_spin_lock -> *ticket_lock*
> >                           |-> raw_spin_unlock -> *queued_spin_unlock*
> >                         /* BUG: ticket_lock with queued_spin_unlock */
> > patch_insn_write  ->  change the raw_spin_unlock to ticket_unlock
> > mutex_unlock(&text_mutex);
> > ...
> >
> > 3. /* Dead lock */
> > arch_jump_label_transform_queue() ->
> > mutex_lock(&text_mutex); +-> raw_spin_lock -> ticket_lock /* deadlock! */
> >                           |-> raw_spin_unlock -> ticket_unlock
> > patch_insn_write -> change other raw_spin_#op -> ticket_#op
> > mutex_unlock(&text_mutex);
> >
> > So, the solution is to disable mutex usage of
> > arch_jump_label_transform_queue() during early_boot_irqs_disabled, just
> > like we have done for stop_machine.
> >
> > Reported-by: Conor Dooley <conor@...nel.org>
> > Signed-off-by: Guo Ren <guoren@...ux.alibaba.com>
> > Signed-off-by: Guo Ren <guoren@...nel.org>
> > Fixes: ab83647fadae ("riscv: Add qspinlock support")
> > Link: https://lore.kernel.org/linux-riscv/CAJF2gTQwYTGinBmCSgVUoPv0_q4EPt_+WiyfUA1HViAKgUzxAg@mail.gmail.com/T/#mf488e6347817fca03bb93a7d34df33d8615b3775
> > Cc: Palmer Dabbelt <palmer@...belt.com>
> > Cc: Alexandre Ghiti <alexghiti@...osinc.com>
> > ---
> >   arch/riscv/kernel/jump_label.c | 12 +++++++++---
> >   1 file changed, 9 insertions(+), 3 deletions(-)
> >
> > diff --git a/arch/riscv/kernel/jump_label.c b/arch/riscv/kernel/jump_label.c
> > index 6eee6f736f68..654ed159c830 100644
> > --- a/arch/riscv/kernel/jump_label.c
> > +++ b/arch/riscv/kernel/jump_label.c
> > @@ -36,9 +36,15 @@ bool arch_jump_label_transform_queue(struct jump_entry *entry,
> >               insn = RISCV_INSN_NOP;
> >       }
> >
> > -     mutex_lock(&text_mutex);
> > -     patch_insn_write(addr, &insn, sizeof(insn));
> > -     mutex_unlock(&text_mutex);
> > +     if (early_boot_irqs_disabled) {
> > +             riscv_patch_in_stop_machine = 1;
> > +             patch_insn_write(addr, &insn, sizeof(insn));
> > +             riscv_patch_in_stop_machine = 0;
> > +     } else {
> > +             mutex_lock(&text_mutex);
> > +             patch_insn_write(addr, &insn, sizeof(insn));
> > +             mutex_unlock(&text_mutex);
> > +     }
> >
> >       return true;
> >   }
>
>
> Sorry for the late answer, I've been sick lately!
I wish you a speedy recovery.

>
> Thank you very much for looking into this and finding this not-so-bad
> solution! I remind everyone that this is a temporary solution until we
> can use an alternative instead of a static key.
Herer, I want to point out that the problem is not jump_label
infrastructure but our arch_jump_label_transform_queue
->patch_insn_write. The alternative also has a similar situation. So
you still need the modification like the below:

diff --git a/arch/riscv/kernel/cpufeature.c b/arch/riscv/kernel/cpufeature.c
index 467c5c735bf5..475f7c520be4 100644
--- a/arch/riscv/kernel/cpufeature.c
+++ b/arch/riscv/kernel/cpufeature.c
@@ -1017,10 +1017,17 @@ void __init_or_module
riscv_cpufeature_patch_func(struct alt_entry *begin,
                oldptr = ALT_OLD_PTR(alt);
                altptr = ALT_ALT_PTR(alt);

+if (early_boot_irqs_disabled) {
+               riscv_patch_in_stop_machine = 1;
+               patch_text_nosync(oldptr, altptr, alt->alt_len);
+               riscv_alternative_fix_offsets(oldptr, alt->alt_len,
oldptr - altptr);
+               riscv_patch_in_stop_machine = 0;
+} else {
                mutex_lock(&text_mutex);
                patch_text_nosync(oldptr, altptr, alt->alt_len);
                riscv_alternative_fix_offsets(oldptr, alt->alt_len,
oldptr - altptr);
                mutex_unlock(&text_mutex);
+}
        }
 }
 #endif

>
> You can add:
>
> Reviewed-by: Alexandre Ghiti <alexghiti@...osinc.com>
> Tested-by: Alexandre Ghiti <alexghiti@...osinc.com>
>
> The revert is still on the table IMO, let's Palmer decide.
Okay

-- 
Best Regards
 Guo Ren

Powered by blists - more mailing lists