Skip to contentSkip to conversation

RISC-V: Fix unannoted hardirqs-on in return to userspace slow-path

PATCH Applied Andrew Bresticker 7 participants 11 messages Nov 11, 2022 – Dec 8, 2022 RISC-V

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 RISC-V: Fix unannoted hardirqs-on in return to userspace slow-path

The return to userspace path in entry.S may enable interrupts without the corresponding lockdep annotation, producing a splat[0] when DEBUG_LOCKDEP is enabled. Simply calling __trace_hardirqs_on() here gets a bit messy due to the use of RA to point back to ret_from_exception, so just move the whole slow-path loop into C. It's more readable and it lets us use local_irq_{enable,disable}(), avoiding the need for manual annotations altogether.

[0]:

  ------------[ cut here ]------------
  DEBUG_LOCKS_WARN_ON(!lockdep_hardirqs_enabled())
  WARNING: CPU: 2 PID: 1 at kernel/locking/lockdep.c:5512 check_flags+0x10a/0x1e0
  Modules linked in:
  CPU: 2 PID: 1 Comm: init Not tainted 6.1.0-rc4-00160-gb56b6e2b4f31 #53
  Hardware name: riscv-virtio,qemu (DT)
  epc : check_flags+0x10a/0x1e0
  ra : check_flags+0x10a/0x1e0
  <snip>
   status: 0000000200000100 badaddr: 0000000000000000 cause: 0000000000000003
  [<ffffffff808edb90>] lock_is_held_type+0x78/0x14e
  [<ffffffff8003dae2>] __might_resched+0x26/0x22c
  [<ffffffff8003dd24>] __might_sleep+0x3c/0x66
  [<ffffffff80022c60>] get_signal+0x9e/0xa70
  [<ffffffff800054a2>] do_notify_resume+0x6e/0x422
  [<ffffffff80003c68>] ret_from_exception+0x0/0x10
  irq event stamp: 44512
  hardirqs last  enabled at (44511): [<ffffffff808f901c>] _raw_spin_unlock_irqrestore+0x54/0x62
  hardirqs last disabled at (44512): [<ffffffff80008200>] __trace_hardirqs_off+0xc/0x14
  softirqs last  enabled at (44472): [<ffffffff808f9fbe>] __do_softirq+0x3de/0x51e
  softirqs last disabled at (44467): [<ffffffff80017760>] irq_exit+0xd6/0x104
  ---[ end trace 0000000000000000 ]---
  possible reason: unannotated irqs-on.

---
This should also theoretically be fixed by the conversion to generic entry, but it's not clear how far away that series is from landing.

---
 arch/riscv/kernel/entry.S  | 18 +++++-------------
 arch/riscv/kernel/signal.c | 34 +++++++++++++++++++++-------------
 2 files changed, 26 insertions(+), 26 deletions(-)
arch/riscv/kernel/entry.S +5 -13
index b9eda3fcbd6d..58dfa8595e19 100644
@@ -263,12 +263,11 @@ ret_from_exception:
 #endif
 	bnez s0, resume_kernel
 
-resume_userspace:
 	/* Interrupts must be disabled here so flags are checked atomically */
 	REG_L s0, TASK_TI_FLAGS(tp) /* current_thread_info->flags */
 	andi s1, s0, _TIF_WORK_MASK
-	bnez s1, work_pending
-
+	bnez s1, resume_userspace_slow
+resume_userspace:
 #ifdef CONFIG_CONTEXT_TRACKING_USER
 	call user_enter_callable
 #endif
@@ -368,19 +367,12 @@ resume_kernel:
 	j restore_all
 #endif
 
-work_pending:
+resume_userspace_slow:
 	/* Enter slow path for supplementary processing */
-	la ra, ret_from_exception
-	andi s1, s0, _TIF_NEED_RESCHED
-	bnez s1, work_resched
-work_notifysig:
-	/* Handle pending signals and notify-resume requests */
-	csrs CSR_STATUS, SR_IE /* Enable interrupts for do_notify_resume() */
 	move a0, sp /* pt_regs */
 	move a1, s0 /* current_thread_info->flags */
-	tail do_notify_resume
-work_resched:
-	tail schedule
+	call do_work_pending
+	j resume_userspace
 
 /* Slow paths for ptrace. */
 handle_syscall_trace_enter:
arch/riscv/kernel/signal.c +21 -13
index 5c591123c440..bfb2afa4135f 100644
@@ -313,19 +313,27 @@ static void do_signal(struct pt_regs *regs)
 }
 
 /*
- * notification of userspace execution resumption
- * - triggered by the _TIF_WORK_MASK flags
+ * Handle any pending work on the resume-to-userspace path, as indicated by
+ * _TIF_WORK_MASK. Entered from assembly with IRQs off.
  */
-asmlinkage __visible void do_notify_resume(struct pt_regs *regs,
-					   unsigned long thread_info_flags)
+asmlinkage __visible void do_work_pending(struct pt_regs *regs,
+					  unsigned long thread_info_flags)
 {
-	if (thread_info_flags & _TIF_UPROBE)
-		uprobe_notify_resume(regs);
-
-	/* Handle pending signal delivery */
-	if (thread_info_flags & (_TIF_SIGPENDING | _TIF_NOTIFY_SIGNAL))
-		do_signal(regs);
-
-	if (thread_info_flags & _TIF_NOTIFY_RESUME)
-		resume_user_mode_work(regs);
+	do {
+		if (thread_info_flags & _TIF_NEED_RESCHED) {
+			schedule();
+		} else {
+			local_irq_enable();
+			if (thread_info_flags & _TIF_UPROBE)
+				uprobe_notify_resume(regs);
+			/* Handle pending signal delivery */
+			if (thread_info_flags & (_TIF_SIGPENDING |
+						 _TIF_NOTIFY_SIGNAL))
+				do_signal(regs);
+			if (thread_info_flags & _TIF_NOTIFY_RESUME)
+				resume_user_mode_work(regs);
+		}
+		local_irq_disable();
+		thread_info_flags = read_thread_flags();
+	} while (thread_info_flags & _TIF_WORK_MASK);
 }
Signature · 1 line
2.25.1
#1 lore raw
Quoted 33 lines · Andrew Bresticker

The return to userspace path in entry.S may enable interrupts without the corresponding lockdep annotation, producing a splat[0] when DEBUG_LOCKDEP is enabled. Simply calling __trace_hardirqs_on() here gets a bit messy due to the use of RA to point back to ret_from_exception, so just move the whole slow-path loop into C. It's more readable and it lets us use local_irq_{enable,disable}(), avoiding the need for manual annotations altogether.

[0]:

  ------------[ cut here ]------------
  DEBUG_LOCKS_WARN_ON(!lockdep_hardirqs_enabled())
  WARNING: CPU: 2 PID: 1 at kernel/locking/lockdep.c:5512 check_flags+0x10a/0x1e0
  Modules linked in:
  CPU: 2 PID: 1 Comm: init Not tainted 6.1.0-rc4-00160-gb56b6e2b4f31 #53
  Hardware name: riscv-virtio,qemu (DT)
  epc : check_flags+0x10a/0x1e0
  ra : check_flags+0x10a/0x1e0
  <snip>
   status: 0000000200000100 badaddr: 0000000000000000 cause: 0000000000000003
  [<ffffffff808edb90>] lock_is_held_type+0x78/0x14e
  [<ffffffff8003dae2>] __might_resched+0x26/0x22c
  [<ffffffff8003dd24>] __might_sleep+0x3c/0x66
  [<ffffffff80022c60>] get_signal+0x9e/0xa70
  [<ffffffff800054a2>] do_notify_resume+0x6e/0x422
  [<ffffffff80003c68>] ret_from_exception+0x0/0x10
  irq event stamp: 44512
  hardirqs last  enabled at (44511): [<ffffffff808f901c>] _raw_spin_unlock_irqrestore+0x54/0x62
  hardirqs last disabled at (44512): [<ffffffff80008200>] __trace_hardirqs_off+0xc/0x14
  softirqs last  enabled at (44472): [<ffffffff808f9fbe>] __do_softirq+0x3de/0x51e
  softirqs last disabled at (44467): [<ffffffff80017760>] irq_exit+0xd6/0x104
  ---[ end trace 0000000000000000 ]---
  possible reason: unannotated irqs-on.
Quoted 4 lines

---
This should also theoretically be fixed by the conversion to generic entry, but it's not clear how far away that series is from landing.

Heh.. In case this makes the cut instead - what commit does this fix?

Quoted 95 lines
---
 arch/riscv/kernel/entry.S  | 18 +++++-------------
 arch/riscv/kernel/signal.c | 34 +++++++++++++++++++++-------------
 2 files changed, 26 insertions(+), 26 deletions(-)
arch/riscv/kernel/entry.S +5 -13
index b9eda3fcbd6d..58dfa8595e19 100644
@@ -263,12 +263,11 @@ ret_from_exception:
 #endif
 	bnez s0, resume_kernel
 
-resume_userspace:
 	/* Interrupts must be disabled here so flags are checked atomically */
 	REG_L s0, TASK_TI_FLAGS(tp) /* current_thread_info->flags */
 	andi s1, s0, _TIF_WORK_MASK
-	bnez s1, work_pending
-
+	bnez s1, resume_userspace_slow
+resume_userspace:
 #ifdef CONFIG_CONTEXT_TRACKING_USER
 	call user_enter_callable
 #endif
@@ -368,19 +367,12 @@ resume_kernel:
 	j restore_all
 #endif
 
-work_pending:
+resume_userspace_slow:
 	/* Enter slow path for supplementary processing */
-	la ra, ret_from_exception
-	andi s1, s0, _TIF_NEED_RESCHED
-	bnez s1, work_resched
-work_notifysig:
-	/* Handle pending signals and notify-resume requests */
-	csrs CSR_STATUS, SR_IE /* Enable interrupts for do_notify_resume() */
 	move a0, sp /* pt_regs */
 	move a1, s0 /* current_thread_info->flags */
-	tail do_notify_resume
-work_resched:
-	tail schedule
+	call do_work_pending
+	j resume_userspace
 
 /* Slow paths for ptrace. */
 handle_syscall_trace_enter:
arch/riscv/kernel/signal.c +21 -13
index 5c591123c440..bfb2afa4135f 100644
@@ -313,19 +313,27 @@ static void do_signal(struct pt_regs *regs)
 }
 
 /*
- * notification of userspace execution resumption
- * - triggered by the _TIF_WORK_MASK flags
+ * Handle any pending work on the resume-to-userspace path, as indicated by
+ * _TIF_WORK_MASK. Entered from assembly with IRQs off.
  */
-asmlinkage __visible void do_notify_resume(struct pt_regs *regs,
-					   unsigned long thread_info_flags)
+asmlinkage __visible void do_work_pending(struct pt_regs *regs,
+					  unsigned long thread_info_flags)
 {
-	if (thread_info_flags & _TIF_UPROBE)
-		uprobe_notify_resume(regs);
-
-	/* Handle pending signal delivery */
-	if (thread_info_flags & (_TIF_SIGPENDING | _TIF_NOTIFY_SIGNAL))
-		do_signal(regs);
-
-	if (thread_info_flags & _TIF_NOTIFY_RESUME)
-		resume_user_mode_work(regs);
+	do {
+		if (thread_info_flags & _TIF_NEED_RESCHED) {
+			schedule();
+		} else {
+			local_irq_enable();
+			if (thread_info_flags & _TIF_UPROBE)
+				uprobe_notify_resume(regs);
+			/* Handle pending signal delivery */
+			if (thread_info_flags & (_TIF_SIGPENDING |
+						 _TIF_NOTIFY_SIGNAL))
+				do_signal(regs);
+			if (thread_info_flags & _TIF_NOTIFY_RESUME)
+				resume_user_mode_work(regs);
+		}
+		local_irq_disable();
+		thread_info_flags = read_thread_flags();
+	} while (thread_info_flags & _TIF_WORK_MASK);
 }
Signature · 1 line
2.25.1

_______________________________________________
linux-riscv mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-riscv

#2 lore raw
Quoted 35 lines · Conor Dooley
Quoted 33 lines · Andrew Bresticker

The return to userspace path in entry.S may enable interrupts without the corresponding lockdep annotation, producing a splat[0] when DEBUG_LOCKDEP is enabled. Simply calling __trace_hardirqs_on() here gets a bit messy due to the use of RA to point back to ret_from_exception, so just move the whole slow-path loop into C. It's more readable and it lets us use local_irq_{enable,disable}(), avoiding the need for manual annotations altogether.

[0]:

  ------------[ cut here ]------------
  DEBUG_LOCKS_WARN_ON(!lockdep_hardirqs_enabled())
  WARNING: CPU: 2 PID: 1 at kernel/locking/lockdep.c:5512 check_flags+0x10a/0x1e0
  Modules linked in:
  CPU: 2 PID: 1 Comm: init Not tainted 6.1.0-rc4-00160-gb56b6e2b4f31 #53
  Hardware name: riscv-virtio,qemu (DT)
  epc : check_flags+0x10a/0x1e0
  ra : check_flags+0x10a/0x1e0
  <snip>
   status: 0000000200000100 badaddr: 0000000000000000 cause: 0000000000000003
  [<ffffffff808edb90>] lock_is_held_type+0x78/0x14e
  [<ffffffff8003dae2>] __might_resched+0x26/0x22c
  [<ffffffff8003dd24>] __might_sleep+0x3c/0x66
  [<ffffffff80022c60>] get_signal+0x9e/0xa70
  [<ffffffff800054a2>] do_notify_resume+0x6e/0x422
  [<ffffffff80003c68>] ret_from_exception+0x0/0x10
  irq event stamp: 44512
  hardirqs last  enabled at (44511): [<ffffffff808f901c>] _raw_spin_unlock_irqrestore+0x54/0x62
  hardirqs last disabled at (44512): [<ffffffff80008200>] __trace_hardirqs_off+0xc/0x14
  softirqs last  enabled at (44472): [<ffffffff808f9fbe>] __do_softirq+0x3de/0x51e
  softirqs last disabled at (44467): [<ffffffff80017760>] irq_exit+0xd6/0x104
  ---[ end trace 0000000000000000 ]---
  possible reason: unannotated irqs-on.
Quoted 6 lines
Quoted 4 lines

---
This should also theoretically be fixed by the conversion to generic entry, but it's not clear how far away that series is from landing.

Heh.. In case this makes the cut instead - what commit does this fix?

Ah, right --

Fixes 3c4697982982 ("riscv: Enable LOCKDEP_SUPPORT & fixup

TRACE_IRQFLAGS_SUPPORT")

-Andrew

Quoted 96 lines
Quoted 96 lines
---
 arch/riscv/kernel/entry.S  | 18 +++++-------------
 arch/riscv/kernel/signal.c | 34 +++++++++++++++++++++-------------
 2 files changed, 26 insertions(+), 26 deletions(-)
arch/riscv/kernel/entry.S +5 -13
index b9eda3fcbd6d..58dfa8595e19 100644
@@ -263,12 +263,11 @@ ret_from_exception:
 #endif
      bnez s0, resume_kernel
 
-resume_userspace:
      /* Interrupts must be disabled here so flags are checked atomically */
      REG_L s0, TASK_TI_FLAGS(tp) /* current_thread_info->flags */
      andi s1, s0, _TIF_WORK_MASK
-     bnez s1, work_pending
-
+     bnez s1, resume_userspace_slow
+resume_userspace:
 #ifdef CONFIG_CONTEXT_TRACKING_USER
      call user_enter_callable
 #endif
@@ -368,19 +367,12 @@ resume_kernel:
      j restore_all
 #endif
 
-work_pending:
+resume_userspace_slow:
      /* Enter slow path for supplementary processing */
-     la ra, ret_from_exception
-     andi s1, s0, _TIF_NEED_RESCHED
-     bnez s1, work_resched
-work_notifysig:
-     /* Handle pending signals and notify-resume requests */
-     csrs CSR_STATUS, SR_IE /* Enable interrupts for do_notify_resume() */
      move a0, sp /* pt_regs */
      move a1, s0 /* current_thread_info->flags */
-     tail do_notify_resume
-work_resched:
-     tail schedule
+     call do_work_pending
+     j resume_userspace
 
 /* Slow paths for ptrace. */
 handle_syscall_trace_enter:
arch/riscv/kernel/signal.c +21 -14
index 5c591123c440..bfb2afa4135f 100644
@@ -313,19 +313,27 @@ static void do_signal(struct pt_regs *regs)
 }
 
 /*
- * notification of userspace execution resumption
- * - triggered by the _TIF_WORK_MASK flags
+ * Handle any pending work on the resume-to-userspace path, as indicated by
+ * _TIF_WORK_MASK. Entered from assembly with IRQs off.
  */
-asmlinkage __visible void do_notify_resume(struct pt_regs *regs,
-                                        unsigned long thread_info_flags)
+asmlinkage __visible void do_work_pending(struct pt_regs *regs,
+                                       unsigned long thread_info_flags)
 {
-     if (thread_info_flags & _TIF_UPROBE)
-             uprobe_notify_resume(regs);
-
-     /* Handle pending signal delivery */
-     if (thread_info_flags & (_TIF_SIGPENDING | _TIF_NOTIFY_SIGNAL))
-             do_signal(regs);
-
-     if (thread_info_flags & _TIF_NOTIFY_RESUME)
-             resume_user_mode_work(regs);
+     do {
+             if (thread_info_flags & _TIF_NEED_RESCHED) {
+                     schedule();
+             } else {
+                     local_irq_enable();
+                     if (thread_info_flags & _TIF_UPROBE)
+                             uprobe_notify_resume(regs);
+                     /* Handle pending signal delivery */
+                     if (thread_info_flags & (_TIF_SIGPENDING |
+                                              _TIF_NOTIFY_SIGNAL))
+                             do_signal(regs);
+                     if (thread_info_flags & _TIF_NOTIFY_RESUME)
+                             resume_user_mode_work(regs);
+             }
+             local_irq_disable();
+             thread_info_flags = read_thread_flags();
+     } while (thread_info_flags & _TIF_WORK_MASK);
 }
--

2.25.1

#3 lore raw
In reply to Andrew Bresticker
Quoted 8 lines · Andrew Bresticker

The return to userspace path in entry.S may enable interrupts without the corresponding lockdep annotation, producing a splat[0] when DEBUG_LOCKDEP is enabled. Simply calling __trace_hardirqs_on() here gets a bit messy due to the use of RA to point back to ret_from_exception, so just move the whole slow-path loop into C. It's more readable and it lets us use local_irq_{enable,disable}(), avoiding the need for manual annotations altogether.

Could generic_entry solve your problem? please try:
https://lore.kernel.org/linux-riscv/[email protected]/

Quoted 112 lines

[0]:

  ------------[ cut here ]------------
  DEBUG_LOCKS_WARN_ON(!lockdep_hardirqs_enabled())
  WARNING: CPU: 2 PID: 1 at kernel/locking/lockdep.c:5512 check_flags+0x10a/0x1e0
  Modules linked in:
  CPU: 2 PID: 1 Comm: init Not tainted 6.1.0-rc4-00160-gb56b6e2b4f31 #53
  Hardware name: riscv-virtio,qemu (DT)
  epc : check_flags+0x10a/0x1e0
  ra : check_flags+0x10a/0x1e0
  <snip>
   status: 0000000200000100 badaddr: 0000000000000000 cause: 0000000000000003
  [<ffffffff808edb90>] lock_is_held_type+0x78/0x14e
  [<ffffffff8003dae2>] __might_resched+0x26/0x22c
  [<ffffffff8003dd24>] __might_sleep+0x3c/0x66
  [<ffffffff80022c60>] get_signal+0x9e/0xa70
  [<ffffffff800054a2>] do_notify_resume+0x6e/0x422
  [<ffffffff80003c68>] ret_from_exception+0x0/0x10
  irq event stamp: 44512
  hardirqs last  enabled at (44511): [<ffffffff808f901c>] _raw_spin_unlock_irqrestore+0x54/0x62
  hardirqs last disabled at (44512): [<ffffffff80008200>] __trace_hardirqs_off+0xc/0x14
  softirqs last  enabled at (44472): [<ffffffff808f9fbe>] __do_softirq+0x3de/0x51e
  softirqs last disabled at (44467): [<ffffffff80017760>] irq_exit+0xd6/0x104
  ---[ end trace 0000000000000000 ]---
  possible reason: unannotated irqs-on.

---
This should also theoretically be fixed by the conversion to generic entry, but it's not clear how far away that series is from landing.

---
 arch/riscv/kernel/entry.S  | 18 +++++-------------
 arch/riscv/kernel/signal.c | 34 +++++++++++++++++++++-------------
 2 files changed, 26 insertions(+), 26 deletions(-)
arch/riscv/kernel/entry.S +5 -13
index b9eda3fcbd6d..58dfa8595e19 100644
@@ -263,12 +263,11 @@ ret_from_exception:
 #endif
        bnez s0, resume_kernel
 
-resume_userspace:
        /* Interrupts must be disabled here so flags are checked atomically */
        REG_L s0, TASK_TI_FLAGS(tp) /* current_thread_info->flags */
        andi s1, s0, _TIF_WORK_MASK
-       bnez s1, work_pending
-
+       bnez s1, resume_userspace_slow
+resume_userspace:
 #ifdef CONFIG_CONTEXT_TRACKING_USER
        call user_enter_callable
 #endif
@@ -368,19 +367,12 @@ resume_kernel:
        j restore_all
 #endif
 
-work_pending:
+resume_userspace_slow:
        /* Enter slow path for supplementary processing */
-       la ra, ret_from_exception
-       andi s1, s0, _TIF_NEED_RESCHED
-       bnez s1, work_resched
-work_notifysig:
-       /* Handle pending signals and notify-resume requests */
-       csrs CSR_STATUS, SR_IE /* Enable interrupts for do_notify_resume() */
        move a0, sp /* pt_regs */
        move a1, s0 /* current_thread_info->flags */
-       tail do_notify_resume
-work_resched:
-       tail schedule
+       call do_work_pending
+       j resume_userspace
 
 /* Slow paths for ptrace. */
 handle_syscall_trace_enter:
arch/riscv/kernel/signal.c +12 -13
index 5c591123c440..bfb2afa4135f 100644
@@ -313,19 +313,27 @@ static void do_signal(struct pt_regs *regs)
 }
 
 /*
- * notification of userspace execution resumption
- * - triggered by the _TIF_WORK_MASK flags
+ * Handle any pending work on the resume-to-userspace path, as indicated by
+ * _TIF_WORK_MASK. Entered from assembly with IRQs off.
  */
-asmlinkage __visible void do_notify_resume(struct pt_regs *regs,
-                                          unsigned long thread_info_flags)
+asmlinkage __visible void do_work_pending(struct pt_regs *regs,
+                                         unsigned long thread_info_flags)
 {
-       if (thread_info_flags & _TIF_UPROBE)
-               uprobe_notify_resume(regs);
-
-       /* Handle pending signal delivery */
-       if (thread_info_flags & (_TIF_SIGPENDING | _TIF_NOTIFY_SIGNAL))
-               do_signal(regs);
-
-       if (thread_info_flags & _TIF_NOTIFY_RESUME)
-               resume_user_mode_work(regs);
+       do {
+               if (thread_info_flags & _TIF_NEED_RESCHED) {
+                       schedule();
+               } else {
+                       local_irq_enable();
+                       if (thread_info_flags & _TIF_UPROBE)
+                               uprobe_notify_resume(regs);
+                       /* Handle pending signal delivery */

arch/riscv/kernel/signal.c static void do_signal(struct pt_regs *regs) · line 339

Quoted 10 lines
arch/riscv/kernel/signal.c +9 -0
+                       if (thread_info_flags & (_TIF_SIGPENDING |
+                                                _TIF_NOTIFY_SIGNAL))
+                               do_signal(regs);
+                       if (thread_info_flags & _TIF_NOTIFY_RESUME)
+                               resume_user_mode_work(regs);
+               }
+               local_irq_disable();
+               thread_info_flags = read_thread_flags();
+       } while (thread_info_flags & _TIF_WORK_MASK);
 }

The more graceful code has been written in kernel/entry/common.c. Let's base it on that:

static unsigned long exit_to_user_mode_loop(struct pt_regs *regs,
unsigned long ti_work)

{
        /*
         * Before returning to user space ensure that all pending work
         * items have been completed.
         */
        while (ti_work & EXIT_TO_USER_MODE_WORK) {

                local_irq_enable_exit_to_user(ti_work);

                if (ti_work & _TIF_NEED_RESCHED)
                        schedule();

                if (ti_work & _TIF_UPROBE)
                        uprobe_notify_resume(regs);

                if (ti_work & _TIF_PATCH_PENDING)
                        klp_update_patch_state(current);

                if (ti_work & (_TIF_SIGPENDING | _TIF_NOTIFY_SIGNAL))
                        arch_do_signal_or_restart(regs);

                if (ti_work & _TIF_NOTIFY_RESUME)
                        resume_user_mode_work(regs);

                /* Architecture specific TIF work */
                arch_exit_to_user_mode_work(regs, ti_work);

                /*
                 * Disable interrupts and reevaluate the work flags as they
                 * might have changed while interrupts and preemption was
                 * enabled above.
                 */
                local_irq_disable_exit_to_user();

                /* Check if any of the above work has queued a
deferred wakeup */
                tick_nohz_user_enter_prepare();

                ti_work = read_thread_flags();
        }

        /* Return the latest work state for arch_exit_to_user_mode() */
        return ti_work;
}
Quoted 3 lines

--
2.25.1

Signature · 2 lines
Best Regards
 Guo Ren
#4 lore raw
Quoted 10 lines · Guo Ren
Quoted 8 lines · Andrew Bresticker

The return to userspace path in entry.S may enable interrupts without the corresponding lockdep annotation, producing a splat[0] when DEBUG_LOCKDEP is enabled. Simply calling __trace_hardirqs_on() here gets a bit messy due to the use of RA to point back to ret_from_exception, so just move the whole slow-path loop into C. It's more readable and it lets us use local_irq_{enable,disable}(), avoiding the need for manual annotations altogether.

Quoted 2 lines

Could generic_entry solve your problem? please try:
https://lore.kernel.org/linux-riscv/[email protected]/

Indeed it does, as I mentioned below, however it wasn't clear to me how close your series was to landing so I'v emailed this small fix for the existing bug in case your series does not make it into 6.2

-Andrew

Quoted 181 lines
Quoted 113 lines

[0]:

  ------------[ cut here ]------------
  DEBUG_LOCKS_WARN_ON(!lockdep_hardirqs_enabled())
  WARNING: CPU: 2 PID: 1 at kernel/locking/lockdep.c:5512 check_flags+0x10a/0x1e0
  Modules linked in:
  CPU: 2 PID: 1 Comm: init Not tainted 6.1.0-rc4-00160-gb56b6e2b4f31 #53
  Hardware name: riscv-virtio,qemu (DT)
  epc : check_flags+0x10a/0x1e0
  ra : check_flags+0x10a/0x1e0
  <snip>
   status: 0000000200000100 badaddr: 0000000000000000 cause: 0000000000000003
  [<ffffffff808edb90>] lock_is_held_type+0x78/0x14e
  [<ffffffff8003dae2>] __might_resched+0x26/0x22c
  [<ffffffff8003dd24>] __might_sleep+0x3c/0x66
  [<ffffffff80022c60>] get_signal+0x9e/0xa70
  [<ffffffff800054a2>] do_notify_resume+0x6e/0x422
  [<ffffffff80003c68>] ret_from_exception+0x0/0x10
  irq event stamp: 44512
  hardirqs last  enabled at (44511): [<ffffffff808f901c>] _raw_spin_unlock_irqrestore+0x54/0x62
  hardirqs last disabled at (44512): [<ffffffff80008200>] __trace_hardirqs_off+0xc/0x14
  softirqs last  enabled at (44472): [<ffffffff808f9fbe>] __do_softirq+0x3de/0x51e
  softirqs last disabled at (44467): [<ffffffff80017760>] irq_exit+0xd6/0x104
  ---[ end trace 0000000000000000 ]---
  possible reason: unannotated irqs-on.

---
This should also theoretically be fixed by the conversion to generic entry, but it's not clear how far away that series is from landing.

---
 arch/riscv/kernel/entry.S  | 18 +++++-------------
 arch/riscv/kernel/signal.c | 34 +++++++++++++++++++++-------------
 2 files changed, 26 insertions(+), 26 deletions(-)
arch/riscv/kernel/entry.S +5 -13
index b9eda3fcbd6d..58dfa8595e19 100644
@@ -263,12 +263,11 @@ ret_from_exception:
 #endif
        bnez s0, resume_kernel
 
-resume_userspace:
        /* Interrupts must be disabled here so flags are checked atomically */
        REG_L s0, TASK_TI_FLAGS(tp) /* current_thread_info->flags */
        andi s1, s0, _TIF_WORK_MASK
-       bnez s1, work_pending
-
+       bnez s1, resume_userspace_slow
+resume_userspace:
 #ifdef CONFIG_CONTEXT_TRACKING_USER
        call user_enter_callable
 #endif
@@ -368,19 +367,12 @@ resume_kernel:
        j restore_all
 #endif
 
-work_pending:
+resume_userspace_slow:
        /* Enter slow path for supplementary processing */
-       la ra, ret_from_exception
-       andi s1, s0, _TIF_NEED_RESCHED
-       bnez s1, work_resched
-work_notifysig:
-       /* Handle pending signals and notify-resume requests */
-       csrs CSR_STATUS, SR_IE /* Enable interrupts for do_notify_resume() */
        move a0, sp /* pt_regs */
        move a1, s0 /* current_thread_info->flags */
-       tail do_notify_resume
-work_resched:
-       tail schedule
+       call do_work_pending
+       j resume_userspace
 
 /* Slow paths for ptrace. */
 handle_syscall_trace_enter:
arch/riscv/kernel/signal.c +12 -13
index 5c591123c440..bfb2afa4135f 100644
@@ -313,19 +313,27 @@ static void do_signal(struct pt_regs *regs)
 }
 
 /*
- * notification of userspace execution resumption
- * - triggered by the _TIF_WORK_MASK flags
+ * Handle any pending work on the resume-to-userspace path, as indicated by
+ * _TIF_WORK_MASK. Entered from assembly with IRQs off.
  */
-asmlinkage __visible void do_notify_resume(struct pt_regs *regs,
-                                          unsigned long thread_info_flags)
+asmlinkage __visible void do_work_pending(struct pt_regs *regs,
+                                         unsigned long thread_info_flags)
 {
-       if (thread_info_flags & _TIF_UPROBE)
-               uprobe_notify_resume(regs);
-
-       /* Handle pending signal delivery */
-       if (thread_info_flags & (_TIF_SIGPENDING | _TIF_NOTIFY_SIGNAL))
-               do_signal(regs);
-
-       if (thread_info_flags & _TIF_NOTIFY_RESUME)
-               resume_user_mode_work(regs);
+       do {
+               if (thread_info_flags & _TIF_NEED_RESCHED) {
+                       schedule();
+               } else {
+                       local_irq_enable();
+                       if (thread_info_flags & _TIF_UPROBE)
+                               uprobe_notify_resume(regs);
+                       /* Handle pending signal delivery */

arch/riscv/kernel/signal.c static void do_signal(struct pt_regs *regs) · line 339

Quoted 10 lines
arch/riscv/kernel/signal.c +9 -0
+                       if (thread_info_flags & (_TIF_SIGPENDING |
+                                                _TIF_NOTIFY_SIGNAL))
+                               do_signal(regs);
+                       if (thread_info_flags & _TIF_NOTIFY_RESUME)
+                               resume_user_mode_work(regs);
+               }
+               local_irq_disable();
+               thread_info_flags = read_thread_flags();
+       } while (thread_info_flags & _TIF_WORK_MASK);
 }

The more graceful code has been written in kernel/entry/common.c. Let's base it on that:

static unsigned long exit_to_user_mode_loop(struct pt_regs *regs,
unsigned long ti_work)

{
        /*
         * Before returning to user space ensure that all pending work
         * items have been completed.
         */
        while (ti_work & EXIT_TO_USER_MODE_WORK) {

                local_irq_enable_exit_to_user(ti_work);

                if (ti_work & _TIF_NEED_RESCHED)
                        schedule();

                if (ti_work & _TIF_UPROBE)
                        uprobe_notify_resume(regs);

                if (ti_work & _TIF_PATCH_PENDING)
                        klp_update_patch_state(current);

                if (ti_work & (_TIF_SIGPENDING | _TIF_NOTIFY_SIGNAL))
                        arch_do_signal_or_restart(regs);

                if (ti_work & _TIF_NOTIFY_RESUME)
                        resume_user_mode_work(regs);

                /* Architecture specific TIF work */
                arch_exit_to_user_mode_work(regs, ti_work);

                /*
                 * Disable interrupts and reevaluate the work flags as they
                 * might have changed while interrupts and preemption was
                 * enabled above.
                 */
                local_irq_disable_exit_to_user();

                /* Check if any of the above work has queued a
deferred wakeup */
                tick_nohz_user_enter_prepare();

                ti_work = read_thread_flags();
        }

        /* Return the latest work state for arch_exit_to_user_mode() */
        return ti_work;
}
Quoted 3 lines

--
2.25.1

--
Best Regards
Guo Ren

#5 lore raw
Quoted 11 lines · Andrew Bresticker
Quoted 10 lines · Guo Ren
Quoted 8 lines · Andrew Bresticker

The return to userspace path in entry.S may enable interrupts without the corresponding lockdep annotation, producing a splat[0] when DEBUG_LOCKDEP is enabled. Simply calling __trace_hardirqs_on() here gets a bit messy due to the use of RA to point back to ret_from_exception, so just move the whole slow-path loop into C. It's more readable and it lets us use local_irq_{enable,disable}(), avoiding the need for manual annotations altogether.

Quoted 6 lines
Quoted 2 lines

Could generic_entry solve your problem? please try:
https://lore.kernel.org/linux-riscv/[email protected]/

Indeed it does, as I mentioned below, however it wasn't clear to me how close your series was to landing so I'v emailed this small fix for the existing bug in case your series does not make it into 6.2

The backportablilty of this 26+, 26- change has to be considered too. The generic entry patchset has 20x the changes and would not ordinarily be stable material. How much of that series would be required to solve the problem?

Quoted 186 lines
Quoted 173 lines
Quoted 113 lines

[0]:

  ------------[ cut here ]------------
  DEBUG_LOCKS_WARN_ON(!lockdep_hardirqs_enabled())
  WARNING: CPU: 2 PID: 1 at kernel/locking/lockdep.c:5512 check_flags+0x10a/0x1e0
  Modules linked in:
  CPU: 2 PID: 1 Comm: init Not tainted 6.1.0-rc4-00160-gb56b6e2b4f31 #53
  Hardware name: riscv-virtio,qemu (DT)
  epc : check_flags+0x10a/0x1e0
  ra : check_flags+0x10a/0x1e0
  <snip>
   status: 0000000200000100 badaddr: 0000000000000000 cause: 0000000000000003
  [<ffffffff808edb90>] lock_is_held_type+0x78/0x14e
  [<ffffffff8003dae2>] __might_resched+0x26/0x22c
  [<ffffffff8003dd24>] __might_sleep+0x3c/0x66
  [<ffffffff80022c60>] get_signal+0x9e/0xa70
  [<ffffffff800054a2>] do_notify_resume+0x6e/0x422
  [<ffffffff80003c68>] ret_from_exception+0x0/0x10
  irq event stamp: 44512
  hardirqs last  enabled at (44511): [<ffffffff808f901c>] _raw_spin_unlock_irqrestore+0x54/0x62
  hardirqs last disabled at (44512): [<ffffffff80008200>] __trace_hardirqs_off+0xc/0x14
  softirqs last  enabled at (44472): [<ffffffff808f9fbe>] __do_softirq+0x3de/0x51e
  softirqs last disabled at (44467): [<ffffffff80017760>] irq_exit+0xd6/0x104
  ---[ end trace 0000000000000000 ]---
  possible reason: unannotated irqs-on.

---
This should also theoretically be fixed by the conversion to generic entry, but it's not clear how far away that series is from landing.

---
 arch/riscv/kernel/entry.S  | 18 +++++-------------
 arch/riscv/kernel/signal.c | 34 +++++++++++++++++++++-------------
 2 files changed, 26 insertions(+), 26 deletions(-)
arch/riscv/kernel/entry.S +5 -13
index b9eda3fcbd6d..58dfa8595e19 100644
@@ -263,12 +263,11 @@ ret_from_exception:
 #endif
        bnez s0, resume_kernel
 
-resume_userspace:
        /* Interrupts must be disabled here so flags are checked atomically */
        REG_L s0, TASK_TI_FLAGS(tp) /* current_thread_info->flags */
        andi s1, s0, _TIF_WORK_MASK
-       bnez s1, work_pending
-
+       bnez s1, resume_userspace_slow
+resume_userspace:
 #ifdef CONFIG_CONTEXT_TRACKING_USER
        call user_enter_callable
 #endif
@@ -368,19 +367,12 @@ resume_kernel:
        j restore_all
 #endif
 
-work_pending:
+resume_userspace_slow:
        /* Enter slow path for supplementary processing */
-       la ra, ret_from_exception
-       andi s1, s0, _TIF_NEED_RESCHED
-       bnez s1, work_resched
-work_notifysig:
-       /* Handle pending signals and notify-resume requests */
-       csrs CSR_STATUS, SR_IE /* Enable interrupts for do_notify_resume() */
        move a0, sp /* pt_regs */
        move a1, s0 /* current_thread_info->flags */
-       tail do_notify_resume
-work_resched:
-       tail schedule
+       call do_work_pending
+       j resume_userspace
 
 /* Slow paths for ptrace. */
 handle_syscall_trace_enter:
arch/riscv/kernel/signal.c +12 -13
index 5c591123c440..bfb2afa4135f 100644
@@ -313,19 +313,27 @@ static void do_signal(struct pt_regs *regs)
 }
 
 /*
- * notification of userspace execution resumption
- * - triggered by the _TIF_WORK_MASK flags
+ * Handle any pending work on the resume-to-userspace path, as indicated by
+ * _TIF_WORK_MASK. Entered from assembly with IRQs off.
  */
-asmlinkage __visible void do_notify_resume(struct pt_regs *regs,
-                                          unsigned long thread_info_flags)
+asmlinkage __visible void do_work_pending(struct pt_regs *regs,
+                                         unsigned long thread_info_flags)
 {
-       if (thread_info_flags & _TIF_UPROBE)
-               uprobe_notify_resume(regs);
-
-       /* Handle pending signal delivery */
-       if (thread_info_flags & (_TIF_SIGPENDING | _TIF_NOTIFY_SIGNAL))
-               do_signal(regs);
-
-       if (thread_info_flags & _TIF_NOTIFY_RESUME)
-               resume_user_mode_work(regs);
+       do {
+               if (thread_info_flags & _TIF_NEED_RESCHED) {
+                       schedule();
+               } else {
+                       local_irq_enable();
+                       if (thread_info_flags & _TIF_UPROBE)
+                               uprobe_notify_resume(regs);
+                       /* Handle pending signal delivery */

arch/riscv/kernel/signal.c static void do_signal(struct pt_regs *regs) · line 339

Quoted 10 lines
arch/riscv/kernel/signal.c +9 -0
+                       if (thread_info_flags & (_TIF_SIGPENDING |
+                                                _TIF_NOTIFY_SIGNAL))
+                               do_signal(regs);
+                       if (thread_info_flags & _TIF_NOTIFY_RESUME)
+                               resume_user_mode_work(regs);
+               }
+               local_irq_disable();
+               thread_info_flags = read_thread_flags();
+       } while (thread_info_flags & _TIF_WORK_MASK);
 }

The more graceful code has been written in kernel/entry/common.c. Let's base it on that:

static unsigned long exit_to_user_mode_loop(struct pt_regs *regs,
unsigned long ti_work)

{
        /*
         * Before returning to user space ensure that all pending work
         * items have been completed.
         */
        while (ti_work & EXIT_TO_USER_MODE_WORK) {

                local_irq_enable_exit_to_user(ti_work);

                if (ti_work & _TIF_NEED_RESCHED)
                        schedule();

                if (ti_work & _TIF_UPROBE)
                        uprobe_notify_resume(regs);

                if (ti_work & _TIF_PATCH_PENDING)
                        klp_update_patch_state(current);

                if (ti_work & (_TIF_SIGPENDING | _TIF_NOTIFY_SIGNAL))
                        arch_do_signal_or_restart(regs);

                if (ti_work & _TIF_NOTIFY_RESUME)
                        resume_user_mode_work(regs);

                /* Architecture specific TIF work */
                arch_exit_to_user_mode_work(regs, ti_work);

                /*
                 * Disable interrupts and reevaluate the work flags as they
                 * might have changed while interrupts and preemption was
                 * enabled above.
                 */
                local_irq_disable_exit_to_user();

                /* Check if any of the above work has queued a
deferred wakeup */
                tick_nohz_user_enter_prepare();

                ti_work = read_thread_flags();
        }

        /* Return the latest work state for arch_exit_to_user_mode() */
        return ti_work;
}
Quoted 8 lines
Quoted 3 lines

--
2.25.1

--
Best Regards
Guo Ren

_______________________________________________
linux-riscv mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-riscv

#6 lore raw
Quoted 16 lines · Conor Dooley
Quoted 14 lines · Andrew Bresticker
Quoted 12 lines · Guo Ren
Quoted 8 lines · Andrew Bresticker

The return to userspace path in entry.S may enable interrupts without the corresponding lockdep annotation, producing a splat[0] when DEBUG_LOCKDEP is enabled. Simply calling __trace_hardirqs_on() here gets a bit messy due to the use of RA to point back to ret_from_exception, so just move the whole slow-path loop into C. It's more readable and it lets us use local_irq_{enable,disable}(), avoiding the need for manual annotations altogether.

Could generic_entry solve your problem? please try:
https://lore.kernel.org/linux-riscv/[email protected]/

Quoted 8 lines
Quoted 3 lines

Indeed it does, as I mentioned below, however it wasn't clear to me how close your series was to landing so I'v emailed this small fix for the existing bug in case your series does not make it into 6.2

The backportablilty of this 26+, 26- change has to be considered too. The generic entry patchset has 20x the changes and would not ordinarily be stable material. How much of that series would be required to solve the problem?

We'd need up to patch 6, "riscv: convert to generic entry":
https://lore.kernel.org/linux-riscv/[email protected]/

-Andrew

Quoted 187 lines
Quoted 187 lines
Quoted 174 lines
Quoted 114 lines

[0]:

  ------------[ cut here ]------------
  DEBUG_LOCKS_WARN_ON(!lockdep_hardirqs_enabled())
  WARNING: CPU: 2 PID: 1 at kernel/locking/lockdep.c:5512 check_flags+0x10a/0x1e0
  Modules linked in:
  CPU: 2 PID: 1 Comm: init Not tainted 6.1.0-rc4-00160-gb56b6e2b4f31 #53
  Hardware name: riscv-virtio,qemu (DT)
  epc : check_flags+0x10a/0x1e0
  ra : check_flags+0x10a/0x1e0
  <snip>
   status: 0000000200000100 badaddr: 0000000000000000 cause: 0000000000000003
  [<ffffffff808edb90>] lock_is_held_type+0x78/0x14e
  [<ffffffff8003dae2>] __might_resched+0x26/0x22c
  [<ffffffff8003dd24>] __might_sleep+0x3c/0x66
  [<ffffffff80022c60>] get_signal+0x9e/0xa70
  [<ffffffff800054a2>] do_notify_resume+0x6e/0x422
  [<ffffffff80003c68>] ret_from_exception+0x0/0x10
  irq event stamp: 44512
  hardirqs last  enabled at (44511): [<ffffffff808f901c>] _raw_spin_unlock_irqrestore+0x54/0x62
  hardirqs last disabled at (44512): [<ffffffff80008200>] __trace_hardirqs_off+0xc/0x14
  softirqs last  enabled at (44472): [<ffffffff808f9fbe>] __do_softirq+0x3de/0x51e
  softirqs last disabled at (44467): [<ffffffff80017760>] irq_exit+0xd6/0x104
  ---[ end trace 0000000000000000 ]---
  possible reason: unannotated irqs-on.

---
This should also theoretically be fixed by the conversion to generic entry, but it's not clear how far away that series is from landing.

---
 arch/riscv/kernel/entry.S  | 18 +++++-------------
 arch/riscv/kernel/signal.c | 34 +++++++++++++++++++++-------------
 2 files changed, 26 insertions(+), 26 deletions(-)
arch/riscv/kernel/entry.S +5 -13
index b9eda3fcbd6d..58dfa8595e19 100644
@@ -263,12 +263,11 @@ ret_from_exception:
 #endif
        bnez s0, resume_kernel
 
-resume_userspace:
        /* Interrupts must be disabled here so flags are checked atomically */
        REG_L s0, TASK_TI_FLAGS(tp) /* current_thread_info->flags */
        andi s1, s0, _TIF_WORK_MASK
-       bnez s1, work_pending
-
+       bnez s1, resume_userspace_slow
+resume_userspace:
 #ifdef CONFIG_CONTEXT_TRACKING_USER
        call user_enter_callable
 #endif
@@ -368,19 +367,12 @@ resume_kernel:
        j restore_all
 #endif
 
-work_pending:
+resume_userspace_slow:
        /* Enter slow path for supplementary processing */
-       la ra, ret_from_exception
-       andi s1, s0, _TIF_NEED_RESCHED
-       bnez s1, work_resched
-work_notifysig:
-       /* Handle pending signals and notify-resume requests */
-       csrs CSR_STATUS, SR_IE /* Enable interrupts for do_notify_resume() */
        move a0, sp /* pt_regs */
        move a1, s0 /* current_thread_info->flags */
-       tail do_notify_resume
-work_resched:
-       tail schedule
+       call do_work_pending
+       j resume_userspace
 
 /* Slow paths for ptrace. */
 handle_syscall_trace_enter:
arch/riscv/kernel/signal.c +12 -13
index 5c591123c440..bfb2afa4135f 100644
@@ -313,19 +313,27 @@ static void do_signal(struct pt_regs *regs)
 }
 
 /*
- * notification of userspace execution resumption
- * - triggered by the _TIF_WORK_MASK flags
+ * Handle any pending work on the resume-to-userspace path, as indicated by
+ * _TIF_WORK_MASK. Entered from assembly with IRQs off.
  */
-asmlinkage __visible void do_notify_resume(struct pt_regs *regs,
-                                          unsigned long thread_info_flags)
+asmlinkage __visible void do_work_pending(struct pt_regs *regs,
+                                         unsigned long thread_info_flags)
 {
-       if (thread_info_flags & _TIF_UPROBE)
-               uprobe_notify_resume(regs);
-
-       /* Handle pending signal delivery */
-       if (thread_info_flags & (_TIF_SIGPENDING | _TIF_NOTIFY_SIGNAL))
-               do_signal(regs);
-
-       if (thread_info_flags & _TIF_NOTIFY_RESUME)
-               resume_user_mode_work(regs);
+       do {
+               if (thread_info_flags & _TIF_NEED_RESCHED) {
+                       schedule();
+               } else {
+                       local_irq_enable();
+                       if (thread_info_flags & _TIF_UPROBE)
+                               uprobe_notify_resume(regs);
+                       /* Handle pending signal delivery */

arch/riscv/kernel/signal.c static void do_signal(struct pt_regs *regs) · line 339

Quoted 10 lines
arch/riscv/kernel/signal.c +9 -0
+                       if (thread_info_flags & (_TIF_SIGPENDING |
+                                                _TIF_NOTIFY_SIGNAL))
+                               do_signal(regs);
+                       if (thread_info_flags & _TIF_NOTIFY_RESUME)
+                               resume_user_mode_work(regs);
+               }
+               local_irq_disable();
+               thread_info_flags = read_thread_flags();
+       } while (thread_info_flags & _TIF_WORK_MASK);
 }

The more graceful code has been written in kernel/entry/common.c. Let's base it on that:

static unsigned long exit_to_user_mode_loop(struct pt_regs *regs,
unsigned long ti_work)

{
        /*
         * Before returning to user space ensure that all pending work
         * items have been completed.
         */
        while (ti_work & EXIT_TO_USER_MODE_WORK) {

                local_irq_enable_exit_to_user(ti_work);

                if (ti_work & _TIF_NEED_RESCHED)
                        schedule();

                if (ti_work & _TIF_UPROBE)
                        uprobe_notify_resume(regs);

                if (ti_work & _TIF_PATCH_PENDING)
                        klp_update_patch_state(current);

                if (ti_work & (_TIF_SIGPENDING | _TIF_NOTIFY_SIGNAL))
                        arch_do_signal_or_restart(regs);

                if (ti_work & _TIF_NOTIFY_RESUME)
                        resume_user_mode_work(regs);

                /* Architecture specific TIF work */
                arch_exit_to_user_mode_work(regs, ti_work);

                /*
                 * Disable interrupts and reevaluate the work flags as they
                 * might have changed while interrupts and preemption was
                 * enabled above.
                 */
                local_irq_disable_exit_to_user();

                /* Check if any of the above work has queued a
deferred wakeup */
                tick_nohz_user_enter_prepare();

                ti_work = read_thread_flags();
        }

        /* Return the latest work state for arch_exit_to_user_mode() */
        return ti_work;
}
Quoted 8 lines
Quoted 3 lines

--
2.25.1

--
Best Regards
Guo Ren

_______________________________________________
linux-riscv mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-riscv

#7 lore raw
Quoted 23 lines · Andrew Bresticker
Quoted 20 lines · Conor Dooley
Quoted 17 lines · Andrew Bresticker
Quoted 10 lines · Guo Ren
Quoted 8 lines · Andrew Bresticker

The return to userspace path in entry.S may enable interrupts without the corresponding lockdep annotation, producing a splat[0] when DEBUG_LOCKDEP is enabled. Simply calling __trace_hardirqs_on() here gets a bit messy due to the use of RA to point back to ret_from_exception, so just move the whole slow-path loop into C. It's more readable and it lets us use local_irq_{enable,disable}(), avoiding the need for manual annotations altogether.

Quoted 2 lines

Could generic_entry solve your problem? please try:
https://lore.kernel.org/linux-riscv/[email protected]/

Indeed it does, as I mentioned below, however it wasn't clear to me how close your series was to landing so I'v emailed this small fix for the existing bug in case your series does not make it into 6.2

Quoted 7 lines
Quoted 4 lines

The backportablilty of this 26+, 26- change has to be considered too. The generic entry patchset has 20x the changes and would not ordinarily be stable material. How much of that series would be required to solve the problem?

We'd need up to patch 6, "riscv: convert to generic entry":
https://lore.kernel.org/linux-riscv/[email protected]/

Yes, but 6 needs 1 and 5, and they are very small patches.

Quoted 190 lines

-Andrew

Quoted 187 lines
Quoted 187 lines
Quoted 174 lines
Quoted 114 lines

[0]:

  ------------[ cut here ]------------
  DEBUG_LOCKS_WARN_ON(!lockdep_hardirqs_enabled())
  WARNING: CPU: 2 PID: 1 at kernel/locking/lockdep.c:5512 check_flags+0x10a/0x1e0
  Modules linked in:
  CPU: 2 PID: 1 Comm: init Not tainted 6.1.0-rc4-00160-gb56b6e2b4f31 #53
  Hardware name: riscv-virtio,qemu (DT)
  epc : check_flags+0x10a/0x1e0
  ra : check_flags+0x10a/0x1e0
  <snip>
   status: 0000000200000100 badaddr: 0000000000000000 cause: 0000000000000003
  [<ffffffff808edb90>] lock_is_held_type+0x78/0x14e
  [<ffffffff8003dae2>] __might_resched+0x26/0x22c
  [<ffffffff8003dd24>] __might_sleep+0x3c/0x66
  [<ffffffff80022c60>] get_signal+0x9e/0xa70
  [<ffffffff800054a2>] do_notify_resume+0x6e/0x422
  [<ffffffff80003c68>] ret_from_exception+0x0/0x10
  irq event stamp: 44512
  hardirqs last  enabled at (44511): [<ffffffff808f901c>] _raw_spin_unlock_irqrestore+0x54/0x62
  hardirqs last disabled at (44512): [<ffffffff80008200>] __trace_hardirqs_off+0xc/0x14
  softirqs last  enabled at (44472): [<ffffffff808f9fbe>] __do_softirq+0x3de/0x51e
  softirqs last disabled at (44467): [<ffffffff80017760>] irq_exit+0xd6/0x104
  ---[ end trace 0000000000000000 ]---
  possible reason: unannotated irqs-on.

---
This should also theoretically be fixed by the conversion to generic entry, but it's not clear how far away that series is from landing.

---
 arch/riscv/kernel/entry.S  | 18 +++++-------------
 arch/riscv/kernel/signal.c | 34 +++++++++++++++++++++-------------
 2 files changed, 26 insertions(+), 26 deletions(-)
arch/riscv/kernel/entry.S +5 -13
index b9eda3fcbd6d..58dfa8595e19 100644
@@ -263,12 +263,11 @@ ret_from_exception:
 #endif
        bnez s0, resume_kernel
 
-resume_userspace:
        /* Interrupts must be disabled here so flags are checked atomically */
        REG_L s0, TASK_TI_FLAGS(tp) /* current_thread_info->flags */
        andi s1, s0, _TIF_WORK_MASK
-       bnez s1, work_pending
-
+       bnez s1, resume_userspace_slow
+resume_userspace:
 #ifdef CONFIG_CONTEXT_TRACKING_USER
        call user_enter_callable
 #endif
@@ -368,19 +367,12 @@ resume_kernel:
        j restore_all
 #endif
 
-work_pending:
+resume_userspace_slow:
        /* Enter slow path for supplementary processing */
-       la ra, ret_from_exception
-       andi s1, s0, _TIF_NEED_RESCHED
-       bnez s1, work_resched
-work_notifysig:
-       /* Handle pending signals and notify-resume requests */
-       csrs CSR_STATUS, SR_IE /* Enable interrupts for do_notify_resume() */
        move a0, sp /* pt_regs */
        move a1, s0 /* current_thread_info->flags */
-       tail do_notify_resume
-work_resched:
-       tail schedule
+       call do_work_pending
+       j resume_userspace
 
 /* Slow paths for ptrace. */
 handle_syscall_trace_enter:
arch/riscv/kernel/signal.c +12 -13
index 5c591123c440..bfb2afa4135f 100644
@@ -313,19 +313,27 @@ static void do_signal(struct pt_regs *regs)
 }
 
 /*
- * notification of userspace execution resumption
- * - triggered by the _TIF_WORK_MASK flags
+ * Handle any pending work on the resume-to-userspace path, as indicated by
+ * _TIF_WORK_MASK. Entered from assembly with IRQs off.
  */
-asmlinkage __visible void do_notify_resume(struct pt_regs *regs,
-                                          unsigned long thread_info_flags)
+asmlinkage __visible void do_work_pending(struct pt_regs *regs,
+                                         unsigned long thread_info_flags)
 {
-       if (thread_info_flags & _TIF_UPROBE)
-               uprobe_notify_resume(regs);
-
-       /* Handle pending signal delivery */
-       if (thread_info_flags & (_TIF_SIGPENDING | _TIF_NOTIFY_SIGNAL))
-               do_signal(regs);
-
-       if (thread_info_flags & _TIF_NOTIFY_RESUME)
-               resume_user_mode_work(regs);
+       do {
+               if (thread_info_flags & _TIF_NEED_RESCHED) {
+                       schedule();
+               } else {
+                       local_irq_enable();
+                       if (thread_info_flags & _TIF_UPROBE)
+                               uprobe_notify_resume(regs);
+                       /* Handle pending signal delivery */

arch/riscv/kernel/signal.c static void do_signal(struct pt_regs *regs) · line 339

Quoted 10 lines
arch/riscv/kernel/signal.c +9 -0
+                       if (thread_info_flags & (_TIF_SIGPENDING |
+                                                _TIF_NOTIFY_SIGNAL))
+                               do_signal(regs);
+                       if (thread_info_flags & _TIF_NOTIFY_RESUME)
+                               resume_user_mode_work(regs);
+               }
+               local_irq_disable();
+               thread_info_flags = read_thread_flags();
+       } while (thread_info_flags & _TIF_WORK_MASK);
 }

The more graceful code has been written in kernel/entry/common.c. Let's base it on that:

static unsigned long exit_to_user_mode_loop(struct pt_regs *regs,
unsigned long ti_work)

{
        /*
         * Before returning to user space ensure that all pending work
         * items have been completed.
         */
        while (ti_work & EXIT_TO_USER_MODE_WORK) {

                local_irq_enable_exit_to_user(ti_work);

                if (ti_work & _TIF_NEED_RESCHED)
                        schedule();

                if (ti_work & _TIF_UPROBE)
                        uprobe_notify_resume(regs);

                if (ti_work & _TIF_PATCH_PENDING)
                        klp_update_patch_state(current);

                if (ti_work & (_TIF_SIGPENDING | _TIF_NOTIFY_SIGNAL))
                        arch_do_signal_or_restart(regs);

                if (ti_work & _TIF_NOTIFY_RESUME)
                        resume_user_mode_work(regs);

                /* Architecture specific TIF work */
                arch_exit_to_user_mode_work(regs, ti_work);

                /*
                 * Disable interrupts and reevaluate the work flags as they
                 * might have changed while interrupts and preemption was
                 * enabled above.
                 */
                local_irq_disable_exit_to_user();

                /* Check if any of the above work has queued a
deferred wakeup */
                tick_nohz_user_enter_prepare();

                ti_work = read_thread_flags();
        }

        /* Return the latest work state for arch_exit_to_user_mode() */
        return ti_work;
}
Quoted 8 lines
Quoted 3 lines

--
2.25.1

--
Best Regards
Guo Ren

_______________________________________________
linux-riscv mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-riscv

Signature · 2 lines
Best Regards
 Guo Ren
maintainer
#8 lore raw
Quoted 25 lines · On Mon, 14 Nov 2022 07:54:47 PST (-0800), [email protected] wrote:
Quoted 23 lines · Andrew Bresticker
Quoted 20 lines · Conor Dooley
Quoted 17 lines · Andrew Bresticker
Quoted 10 lines · Guo Ren
Quoted 8 lines · Andrew Bresticker

The return to userspace path in entry.S may enable interrupts without the corresponding lockdep annotation, producing a splat[0] when DEBUG_LOCKDEP is enabled. Simply calling __trace_hardirqs_on() here gets a bit messy due to the use of RA to point back to ret_from_exception, so just move the whole slow-path loop into C. It's more readable and it lets us use local_irq_{enable,disable}(), avoiding the need for manual annotations altogether.

Quoted 2 lines

Could generic_entry solve your problem? please try:
https://lore.kernel.org/linux-riscv/[email protected]/

Indeed it does, as I mentioned below, however it wasn't clear to me how close your series was to landing so I'v emailed this small fix for the existing bug in case your series does not make it into 6.2

Quoted 8 lines
Quoted 7 lines
Quoted 4 lines

The backportablilty of this 26+, 26- change has to be considered too. The generic entry patchset has 20x the changes and would not ordinarily be stable material. How much of that series would be required to solve the problem?

We'd need up to patch 6, "riscv: convert to generic entry":
https://lore.kernel.org/linux-riscv/[email protected]/

Yes, but 6 needs 1 and 5, and they are very small patches.

I'm just going to take this one, for the ease of backporting. The generic entry one keeps getting re-spun and I'm not sure it'll be in time for 6.2, so at least this way we can get the fix in now.

I've merge th is just on top of 6.1-rc1, so it's easier to deal with any conflicts later.

Thanks!

Quoted 183 lines
Quoted 183 lines

-Andrew

Quoted 179 lines
Quoted 179 lines
Quoted 179 lines
Quoted 114 lines

[0]:

  ------------[ cut here ]------------
  DEBUG_LOCKS_WARN_ON(!lockdep_hardirqs_enabled())
  WARNING: CPU: 2 PID: 1 at kernel/locking/lockdep.c:5512 check_flags+0x10a/0x1e0
  Modules linked in:
  CPU: 2 PID: 1 Comm: init Not tainted 6.1.0-rc4-00160-gb56b6e2b4f31 #53
  Hardware name: riscv-virtio,qemu (DT)
  epc : check_flags+0x10a/0x1e0
  ra : check_flags+0x10a/0x1e0
  <snip>
   status: 0000000200000100 badaddr: 0000000000000000 cause: 0000000000000003
  [<ffffffff808edb90>] lock_is_held_type+0x78/0x14e
  [<ffffffff8003dae2>] __might_resched+0x26/0x22c
  [<ffffffff8003dd24>] __might_sleep+0x3c/0x66
  [<ffffffff80022c60>] get_signal+0x9e/0xa70
  [<ffffffff800054a2>] do_notify_resume+0x6e/0x422
  [<ffffffff80003c68>] ret_from_exception+0x0/0x10
  irq event stamp: 44512
  hardirqs last  enabled at (44511): [<ffffffff808f901c>] _raw_spin_unlock_irqrestore+0x54/0x62
  hardirqs last disabled at (44512): [<ffffffff80008200>] __trace_hardirqs_off+0xc/0x14
  softirqs last  enabled at (44472): [<ffffffff808f9fbe>] __do_softirq+0x3de/0x51e
  softirqs last disabled at (44467): [<ffffffff80017760>] irq_exit+0xd6/0x104
  ---[ end trace 0000000000000000 ]---
  possible reason: unannotated irqs-on.

---
This should also theoretically be fixed by the conversion to generic entry, but it's not clear how far away that series is from landing.

---
 arch/riscv/kernel/entry.S  | 18 +++++-------------
 arch/riscv/kernel/signal.c | 34 +++++++++++++++++++++-------------
 2 files changed, 26 insertions(+), 26 deletions(-)
arch/riscv/kernel/entry.S +5 -13
index b9eda3fcbd6d..58dfa8595e19 100644
@@ -263,12 +263,11 @@ ret_from_exception:
 #endif
        bnez s0, resume_kernel
 
-resume_userspace:
        /* Interrupts must be disabled here so flags are checked atomically */
        REG_L s0, TASK_TI_FLAGS(tp) /* current_thread_info->flags */
        andi s1, s0, _TIF_WORK_MASK
-       bnez s1, work_pending
-
+       bnez s1, resume_userspace_slow
+resume_userspace:
 #ifdef CONFIG_CONTEXT_TRACKING_USER
        call user_enter_callable
 #endif
@@ -368,19 +367,12 @@ resume_kernel:
        j restore_all
 #endif
 
-work_pending:
+resume_userspace_slow:
        /* Enter slow path for supplementary processing */
-       la ra, ret_from_exception
-       andi s1, s0, _TIF_NEED_RESCHED
-       bnez s1, work_resched
-work_notifysig:
-       /* Handle pending signals and notify-resume requests */
-       csrs CSR_STATUS, SR_IE /* Enable interrupts for do_notify_resume() */
        move a0, sp /* pt_regs */
        move a1, s0 /* current_thread_info->flags */
-       tail do_notify_resume
-work_resched:
-       tail schedule
+       call do_work_pending
+       j resume_userspace
 
 /* Slow paths for ptrace. */
 handle_syscall_trace_enter:
arch/riscv/kernel/signal.c +12 -13
index 5c591123c440..bfb2afa4135f 100644
@@ -313,19 +313,27 @@ static void do_signal(struct pt_regs *regs)
 }
 
 /*
- * notification of userspace execution resumption
- * - triggered by the _TIF_WORK_MASK flags
+ * Handle any pending work on the resume-to-userspace path, as indicated by
+ * _TIF_WORK_MASK. Entered from assembly with IRQs off.
  */
-asmlinkage __visible void do_notify_resume(struct pt_regs *regs,
-                                          unsigned long thread_info_flags)
+asmlinkage __visible void do_work_pending(struct pt_regs *regs,
+                                         unsigned long thread_info_flags)
 {
-       if (thread_info_flags & _TIF_UPROBE)
-               uprobe_notify_resume(regs);
-
-       /* Handle pending signal delivery */
-       if (thread_info_flags & (_TIF_SIGPENDING | _TIF_NOTIFY_SIGNAL))
-               do_signal(regs);
-
-       if (thread_info_flags & _TIF_NOTIFY_RESUME)
-               resume_user_mode_work(regs);
+       do {
+               if (thread_info_flags & _TIF_NEED_RESCHED) {
+                       schedule();
+               } else {
+                       local_irq_enable();
+                       if (thread_info_flags & _TIF_UPROBE)
+                               uprobe_notify_resume(regs);
+                       /* Handle pending signal delivery */

arch/riscv/kernel/signal.c static void do_signal(struct pt_regs *regs) · line 339

Quoted 10 lines
arch/riscv/kernel/signal.c +9 -0
+                       if (thread_info_flags & (_TIF_SIGPENDING |
+                                                _TIF_NOTIFY_SIGNAL))
+                               do_signal(regs);
+                       if (thread_info_flags & _TIF_NOTIFY_RESUME)
+                               resume_user_mode_work(regs);
+               }
+               local_irq_disable();
+               thread_info_flags = read_thread_flags();
+       } while (thread_info_flags & _TIF_WORK_MASK);
 }

The more graceful code has been written in kernel/entry/common.c. Let's base it on that:

static unsigned long exit_to_user_mode_loop(struct pt_regs *regs,
unsigned long ti_work)

{
        /*
         * Before returning to user space ensure that all pending work
         * items have been completed.
         */
        while (ti_work & EXIT_TO_USER_MODE_WORK) {

                local_irq_enable_exit_to_user(ti_work);

                if (ti_work & _TIF_NEED_RESCHED)
                        schedule();

                if (ti_work & _TIF_UPROBE)
                        uprobe_notify_resume(regs);

                if (ti_work & _TIF_PATCH_PENDING)
                        klp_update_patch_state(current);

                if (ti_work & (_TIF_SIGPENDING | _TIF_NOTIFY_SIGNAL))
                        arch_do_signal_or_restart(regs);

                if (ti_work & _TIF_NOTIFY_RESUME)
                        resume_user_mode_work(regs);

                /* Architecture specific TIF work */
                arch_exit_to_user_mode_work(regs, ti_work);

                /*
                 * Disable interrupts and reevaluate the work flags as they
                 * might have changed while interrupts and preemption was
                 * enabled above.
                 */
                local_irq_disable_exit_to_user();

                /* Check if any of the above work has queued a
deferred wakeup */
                tick_nohz_user_enter_prepare();

                ti_work = read_thread_flags();
        }

        /* Return the latest work state for arch_exit_to_user_mode() */
        return ti_work;
}
Quoted 3 lines

--
2.25.1

Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 3 lines

--
Best Regards
Guo Ren

_______________________________________________
linux-riscv mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-riscv

_______________________________________________
linux-riscv mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-riscv

#9 lore raw
In reply to Andrew Bresticker
Quoted 9 lines · Andrew Bresticker

The return to userspace path in entry.S may enable interrupts without the corresponding lockdep annotation, producing a splat[0] when DEBUG_LOCKDEP is enabled. Simply calling __trace_hardirqs_on() here gets a bit messy due to the use of RA to point back to ret_from_exception, so just move the whole slow-path loop into C. It's more readable and it lets us use local_irq_{enable,disable}(), avoiding the need for manual annotations altogether.

[...]

Applied, thanks!

[1/1] RISC-V: Fix unannoted hardirqs-on in return to userspace slow-path

      https://git.kernel.org/palmer/c/e1ceb0964163

Best regards,

Signature · 6 lines
Palmer Dabbelt <[email protected]>

_______________________________________________
linux-riscv mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-riscv
#10 lore raw
In reply to Andrew Bresticker
This patch was applied to riscv/linux.git (for-next) by Palmer Dabbelt <[email protected]>:

Hello:

This patch was applied to riscv/linux.git (for-next) by Palmer Dabbelt <[email protected]>:

Quoted 9 lines · On Fri, 11 Nov 2022 17:31:08 -0500 you wrote:

The return to userspace path in entry.S may enable interrupts without the corresponding lockdep annotation, producing a splat[0] when DEBUG_LOCKDEP is enabled. Simply calling __trace_hardirqs_on() here gets a bit messy due to the use of RA to point back to ret_from_exception, so just move the whole slow-path loop into C. It's more readable and it lets us use local_irq_{enable,disable}(), avoiding the need for manual annotations altogether.

[...]

Here is the summary with links:
- RISC-V: Fix unannoted hardirqs-on in return to userspace slow-path

    https://git.kernel.org/riscv/c/b0f4c74eadbf

You are awesome, thank you!

Signature · 2 lines
Deet-doot-dot, I am a bot.
https://korg.docs.kernel.org/patchwork/pwbot.html