Le 22/06/2021 à 08:04, Nicholas Piggin a écrit :
The PPC_RFI_SRR_DEBUG check added by patch "powerpc/64s: avoid reloading
(H)SRR registers if they are still valid" has a few deficiencies. It
does not fix the actual problem, it's not enabled by default, and it
causes a program check interrupt which can cause more difficulties.

However there are a lot of paths which may clobber SRRs or change return
regs, and difficult to have a high confidence that all paths are covered
without wider testing.

Add a relatively low overhead always-enabled check that catches most
such cases, reports once, and fixes it so the kernel can continue.

Signed-off-by: Nicholas Piggin <npig...@gmail.com>
---
  arch/powerpc/kernel/interrupt.c | 58 +++++++++++++++++++++++++++++++++
  1 file changed, 58 insertions(+)

diff --git a/arch/powerpc/kernel/interrupt.c b/arch/powerpc/kernel/interrupt.c
index 05fa3ae56e25..5920a3e8d1d5 100644
--- a/arch/powerpc/kernel/interrupt.c
+++ b/arch/powerpc/kernel/interrupt.c
@@ -231,6 +231,56 @@ static notrace void booke_load_dbcr0(void)
  #endif
  }
+#include <linux/sched/debug.h> /* for show_regs */
+static void check_return_regs_valid(struct pt_regs *regs)
+{
+#ifdef CONFIG_PPC_BOOK3S_64
+       static bool warned = false;
+
+       if (regs->trap == 0x980 || regs->trap == 0xe00 || regs->trap == 0xe20 ||
+           regs->trap == 0xe40 || regs->trap == 0xe60 || regs->trap == 0xe80 ||
+           regs->trap == 0xea0 || regs->trap == 0xf80 || regs->trap == 0x1200 
||
+           regs->trap == 0x1500 || regs->trap == 0x1600 || regs->trap == 
0x1800) {

Can you use names defined in asm/interrupt.h instead of raw values ?
Some are already there, others can be added.


+               if (local_paca->hsrr_valid) {
+                       unsigned long hsrr0 = mfspr(SPRN_HSRR0);
+                       unsigned long hsrr1 = mfspr(SPRN_HSRR1);
+
+                       if (hsrr0 == regs->nip && hsrr1 == regs->msr)
+                               return;
+
+                       if (!warned) {
+                               warned = true;
+                               printk("HSRR0 was: %lx should be: %lx\n",
+                                       hsrr0, regs->nip);
+                               printk("HSRR1 was: %lx should be: %lx\n",
+                                       hsrr1, regs->msr);
+                               show_regs(regs);
+                       }
+                       local_paca->hsrr_valid = 0; /* fixup */
+               }
+
+       } else if (regs->trap != 0x3000) {
+               if (local_paca->srr_valid) {
+                       unsigned long srr0 = mfspr(SPRN_SRR0);
+                       unsigned long srr1 = mfspr(SPRN_SRR1);
+
+                       if (srr0 == regs->nip && srr1 == regs->msr)
+                               return;
+
+                       if (!warned) {
+                               warned = true;
+                               printk("SRR0 was: %lx should be: %lx\n",
+                                       srr0, regs->nip);
+                               printk("SRR1 was: %lx should be: %lx\n",
+                                       srr1, regs->msr);
+                               show_regs(regs);
+                       }
+                       local_paca->srr_valid = 0; /* fixup */
+               }
+       }
+#endif
+}
+
  /*
   * This should be called after a syscall returns, with r3 the return value
   * from the syscall. If this function returns non-zero, the system call
@@ -327,6 +377,8 @@ notrace unsigned long syscall_exit_prepare(unsigned long r3,
                }
        }
+ check_return_regs_valid(regs);
+
        user_enter_irqoff();
/* scv need not set RI=0 because SRRs are not used */
@@ -405,6 +457,8 @@ notrace unsigned long interrupt_exit_user_prepare(struct 
pt_regs *regs)
                }
        }
+ check_return_regs_valid(regs);
+
        user_enter_irqoff();
if (unlikely(!__prep_irq_for_enabled_exit(true))) {
@@ -469,9 +523,13 @@ notrace unsigned long interrupt_exit_kernel_prepare(struct 
pt_regs *regs)
                        }
                }
+ check_return_regs_valid(regs);
+
                if (unlikely(!prep_irq_for_enabled_exit(true, 
!irqs_disabled_flags(flags))))
                        goto again;
        } else {
+               check_return_regs_valid(regs);
+
                /* Returning to a kernel context with local irqs disabled. */
                __hard_EE_RI_disable();
  #ifdef CONFIG_PPC64

Reply via email to