Skip to contentSkip to conversation

target/ppc: Catch invalid real address accesses

PATCH Superseded Nicholas Piggin 9 participants 31 messages Jun 23, 2023 – Jun 29, 2023

v1 of 3 Superseded Latest is v3 →

lore
Share & tools
lore
#7 lore raw

PATCH 4/4 target/ppc: Make checkstop stop the system

In reply to Nicholas Piggin

checkstop state does not halt the system, interrupts continue to be serviced, and other CPUs run.

Stop the machine with vm_stop(), and print a register dump too.

---
 target/ppc/excp_helper.c | 35 +++++++++++++++++++++--------------
 1 file changed, 21 insertions(+), 14 deletions(-)
target/ppc/excp_helper.c +21 -14
index 4bfcfc3c3d..51e83d7f07 100644
@@ -19,6 +19,7 @@
 #include "qemu/osdep.h"
 #include "qemu/main-loop.h"
 #include "qemu/log.h"
+#include "sysemu/runstate.h"
 #include "cpu.h"
 #include "exec/exec-all.h"
 #include "internal.h"
@@ -165,6 +166,24 @@ static void ppc_excp_debug_sw_tlb(CPUPPCState *env, int excp)
              env->error_code);
 }
 
+static void powerpc_checkstop(PowerPCCPU *cpu, const char *reason)
+{
+    CPUState *cs = CPU(cpu);
+
+    vm_stop(RUN_STATE_GUEST_PANICKED);
+
+    fprintf(stderr, "Entering checkstop state: %s\n", reason);
+    cpu_dump_state(cs, stderr, CPU_DUMP_FPU | CPU_DUMP_CCOP);
+    if (qemu_log_separate()) {
+        FILE *logfile = qemu_log_trylock();
+        if (logfile) {
+            fprintf(logfile, "Entering checkstop state: %s\n", reason);
+            cpu_dump_state(cs, logfile, CPU_DUMP_FPU | CPU_DUMP_CCOP);
+            qemu_log_unlock(logfile);
+        }
+    }
+}
+
 #if defined(TARGET_PPC64)
 static int powerpc_reset_wakeup(CPUState *cs, CPUPPCState *env, int excp,
                                 target_ulong *msr)
@@ -406,21 +425,9 @@ static void powerpc_set_excp_state(PowerPCCPU *cpu, target_ulong vector,
 
 static void powerpc_mcheck_test_and_checkstop(CPUPPCState *env)
 {
-    CPUState *cs = env_cpu(env);
-
-    if (FIELD_EX64(env->msr, MSR, ME)) {
-        return;
-    }
-
-    /* Machine check exception is not enabled. Enter checkstop state. */
-    fprintf(stderr, "Machine check while not allowed. "
-            "Entering checkstop state\n");
-    if (qemu_log_separate()) {
-        qemu_log("Machine check while not allowed. "
-                 "Entering checkstop state\n");
+    if (!FIELD_EX64(env->msr, MSR, ME)) {
+        powerpc_checkstop(env_archcpu(env), "machine check with MSR[ME]=0");
     }
-    cs->halted = 1;
-    cpu_interrupt_exittb(cs);
 }
 
 static void powerpc_excp_40x(PowerPCCPU *cpu, int excp)
Signature · 1 line
2.40.1
#0 lore raw

cover target/ppc: Catch invalid real address accesses

ppc has always silently ignored access to real (physical) addresses with nothing behind it, which can make debugging difficult at times.

It looks like the way to handle this is implement the transaction failed call, which most target architectures do. Notably not x86 though, I wonder why?

Other question is, sometimes I guess it's nice to avoid crashing in order to try to quickly get past some unimplemented MMIO. Maybe a command line option or something could turn it off? It should probably be a QEMU-wide option if so, so that shouldn't hold this series up, I can propose a option for that if anybody is worried about it.

In any case this seems to work. I have only implemented it for books 64-bit for now because I'm not sure what all the others do. I could try to investigate it if there is interest.

I took Zoltan's patch and tweaked it slightly (hopefully that's okay).

Also made a better checkstop while we're here because the old one really doesn't work.

Thanks,
Nick

BALATON Zoltan (1):
target/ppc: Move common check in machne check handlers to a function

Nicholas Piggin (3):
target/ppc: Machine check on invalid real address access target/ppc: Add POWER9/10 invalid-real machine check codes target/ppc: Make checkstop stop the system

 target/ppc/cpu_init.c    |   1 +
 target/ppc/excp_helper.c | 178 +++++++++++++++++++--------------------
 target/ppc/internal.h    |   5 ++
 3 files changed, 95 insertions(+), 89 deletions(-)
Signature · 1 line
2.40.1
#1 lore raw

PATCH 1/4 target/ppc: Machine check on invalid real address access

ppc currently silently accepts invalid real address access. Catch these and turn them into machine checks. POWER9/10 machine check codes for invalid real address access are implemented.

---
 target/ppc/cpu_init.c    |  1 +
 target/ppc/excp_helper.c | 35 +++++++++++++++++++++++++++++++++++
 target/ppc/internal.h    |  5 +++++
 3 files changed, 41 insertions(+)
target/ppc/cpu_init.c +1 -0
index 7bce421a7c..3b23f2d5b6 100644
@@ -7267,6 +7267,7 @@ static const struct TCGCPUOps ppc_tcg_ops = {
   .cpu_exec_enter = ppc_cpu_exec_enter,
   .cpu_exec_exit = ppc_cpu_exec_exit,
   .do_unaligned_access = ppc_cpu_do_unaligned_access,
+  .do_transaction_failed = ppc_cpu_do_transaction_failed,
 #endif /* !CONFIG_USER_ONLY */
 };
 #endif /* CONFIG_TCG */
target/ppc/excp_helper.c +35 -0
index 12d8a7257b..c9bfa3a827 100644
@@ -1402,7 +1402,9 @@ static void powerpc_excp_books(PowerPCCPU *cpu, int excp)
         /* machine check exceptions don't have ME set */
         new_msr &= ~((target_ulong)1 << MSR_ME);
 
+        msr |= env->error_code;
         break;
+
     case POWERPC_EXCP_DSI:       /* Data storage exception                   */
         trace_ppc_excp_dsi(env->spr[SPR_DSISR], env->spr[SPR_DAR]);
         break;
@@ -3123,5 +3125,38 @@ void ppc_cpu_do_unaligned_access(CPUState *cs, vaddr vaddr,
     env->error_code = insn & 0x03FF0000;
     cpu_loop_exit(cs);
 }
+
+void ppc_cpu_do_transaction_failed(CPUState *cs, hwaddr physaddr,
+                                   vaddr vaddr, unsigned size,
+                                   MMUAccessType access_type,
+                                   int mmu_idx, MemTxAttrs attrs,
+                                   MemTxResult response, uintptr_t retaddr)
+{
+    CPUPPCState *env = cs->env_ptr;
+
+    switch (env->excp_model) {
+#if defined(TARGET_PPC64)
+    case POWERPC_EXCP_970:
+    case POWERPC_EXCP_POWER7:
+    case POWERPC_EXCP_POWER8:
+    case POWERPC_EXCP_POWER9:
+    case POWERPC_EXCP_POWER10:
+        /*
+         * TODO: This does not give the correct machine check code but
+         * it will report a NIP and DAR.
+         */
+        if (access_type == MMU_DATA_LOAD || access_type == MMU_DATA_STORE) {
+            env->spr[SPR_DAR] = vaddr;
+        }
+        break;
+#endif
+    default:
+        /* TODO: Check behaviour for other CPUs, for now do nothing. */
+        return;
+    }
+
+    cs->exception_index = POWERPC_EXCP_MCHECK;
+    cpu_loop_exit_restore(cs, retaddr);
+}
 #endif /* CONFIG_TCG */
 #endif /* !CONFIG_USER_ONLY */
target/ppc/internal.h +5 -0
index 901bae6d39..57acb3212c 100644
@@ -296,6 +296,11 @@ bool ppc_cpu_tlb_fill(CPUState *cs, vaddr address, int size,
 G_NORETURN void ppc_cpu_do_unaligned_access(CPUState *cs, vaddr addr,
                                             MMUAccessType access_type, int mmu_idx,
                                             uintptr_t retaddr);
+void ppc_cpu_do_transaction_failed(CPUState *cs, hwaddr physaddr,
+                                   vaddr addr, unsigned size,
+                                   MMUAccessType access_type,
+                                   int mmu_idx, MemTxAttrs attrs,
+                                   MemTxResult response, uintptr_t retaddr);
 #endif
 
 FIELD(GER_MSK, XMSK, 0, 4)
Signature · 1 line
2.40.1
#2 lore raw

PATCH 2/4 target/ppc: Add POWER9/10 invalid-real machine check codes

In reply to Nicholas Piggin

Implement the correct register settings for the invalid-real access machine check for POWER9/10 processors.

---
 target/ppc/excp_helper.c | 26 ++++++++++++++++++++++++--
 1 file changed, 24 insertions(+), 2 deletions(-)
target/ppc/excp_helper.c +24 -2
index c9bfa3a827..1c26828d8b 100644
@@ -3139,8 +3139,6 @@ void ppc_cpu_do_transaction_failed(CPUState *cs, hwaddr physaddr,
     case POWERPC_EXCP_970:
     case POWERPC_EXCP_POWER7:
     case POWERPC_EXCP_POWER8:
-    case POWERPC_EXCP_POWER9:
-    case POWERPC_EXCP_POWER10:
         /*
          * TODO: This does not give the correct machine check code but
          * it will report a NIP and DAR.
@@ -3149,6 +3147,30 @@ void ppc_cpu_do_transaction_failed(CPUState *cs, hwaddr physaddr,
             env->spr[SPR_DAR] = vaddr;
         }
         break;
+    case POWERPC_EXCP_POWER9:
+    case POWERPC_EXCP_POWER10:
+        /*
+         * Machine check codes can be found in User Manual or Linux or
+         * skiboot source.
+         */
+        if (access_type == MMU_DATA_LOAD) {
+            env->spr[SPR_DAR] = vaddr;
+            env->spr[SPR_DSISR] = PPC_BIT(57);
+            env->error_code = PPC_BIT(42);
+
+        } else if (access_type == MMU_DATA_STORE) {
+            /*
+             * MCE for stores in POWER is asynchronous so hardware does
+             * not set DAR, but QEMU can do better.
+             */
+            env->spr[SPR_DAR] = vaddr;
+            env->error_code = PPC_BIT(36) | PPC_BIT(43) | PPC_BIT(45);
+            env->error_code |= PPC_BIT(42);
+        } else { /* Fetch */
+
+            env->error_code = PPC_BIT(36) | PPC_BIT(44) | PPC_BIT(45);
+        }
+        break;
 #endif
     default:
         /* TODO: Check behaviour for other CPUs, for now do nothing. */
Signature · 1 line
2.40.1
#3 lore raw

PATCH 3/4 target/ppc: Move common check in machne check handlers to a function

In reply to Nicholas Piggin

From: BALATON Zoltan <[email protected]>

All powerpc exception handlers share some code when handling machine check exceptions. Move this to a common function.

[np: tweak subject, rename function to powerpc_mcheck_test_and_checkstop]
---
 target/ppc/excp_helper.c | 114 +++++++++------------------------------
 1 file changed, 25 insertions(+), 89 deletions(-)
target/ppc/excp_helper.c +25 -89
index 1c26828d8b..4bfcfc3c3d 100644
@@ -404,6 +404,25 @@ static void powerpc_set_excp_state(PowerPCCPU *cpu, target_ulong vector,
     env->reserve_addr = -1;
 }
 
+static void powerpc_mcheck_test_and_checkstop(CPUPPCState *env)
+{
+    CPUState *cs = env_cpu(env);
+
+    if (FIELD_EX64(env->msr, MSR, ME)) {
+        return;
+    }
+
+    /* Machine check exception is not enabled. Enter checkstop state. */
+    fprintf(stderr, "Machine check while not allowed. "
+            "Entering checkstop state\n");
+    if (qemu_log_separate()) {
+        qemu_log("Machine check while not allowed. "
+                 "Entering checkstop state\n");
+    }
+    cs->halted = 1;
+    cpu_interrupt_exittb(cs);
+}
+
 static void powerpc_excp_40x(PowerPCCPU *cpu, int excp)
 {
     CPUState *cs = CPU(cpu);
@@ -446,21 +465,7 @@ static void powerpc_excp_40x(PowerPCCPU *cpu, int excp)
         srr1 = SPR_40x_SRR3;
         break;
     case POWERPC_EXCP_MCHECK:    /* Machine check exception                  */
-        if (!FIELD_EX64(env->msr, MSR, ME)) {
-            /*
-             * Machine check exception is not enabled.  Enter
-             * checkstop state.
-             */
-            fprintf(stderr, "Machine check while not allowed. "
-                    "Entering checkstop state\n");
-            if (qemu_log_separate()) {
-                qemu_log("Machine check while not allowed. "
-                        "Entering checkstop state\n");
-            }
-            cs->halted = 1;
-            cpu_interrupt_exittb(cs);
-        }
-
+        powerpc_mcheck_test_and_checkstop(env);
         /* machine check exceptions don't have ME set */
         new_msr &= ~((target_ulong)1 << MSR_ME);
 
@@ -577,21 +582,7 @@ static void powerpc_excp_6xx(PowerPCCPU *cpu, int excp)
     case POWERPC_EXCP_CRITICAL:    /* Critical input                         */
         break;
     case POWERPC_EXCP_MCHECK:    /* Machine check exception                  */
-        if (!FIELD_EX64(env->msr, MSR, ME)) {
-            /*
-             * Machine check exception is not enabled.  Enter
-             * checkstop state.
-             */
-            fprintf(stderr, "Machine check while not allowed. "
-                    "Entering checkstop state\n");
-            if (qemu_log_separate()) {
-                qemu_log("Machine check while not allowed. "
-                        "Entering checkstop state\n");
-            }
-            cs->halted = 1;
-            cpu_interrupt_exittb(cs);
-        }
-
+        powerpc_mcheck_test_and_checkstop(env);
         /* machine check exceptions don't have ME set */
         new_msr &= ~((target_ulong)1 << MSR_ME);
 
@@ -750,21 +741,7 @@ static void powerpc_excp_7xx(PowerPCCPU *cpu, int excp)
 
     switch (excp) {
     case POWERPC_EXCP_MCHECK:    /* Machine check exception                  */
-        if (!FIELD_EX64(env->msr, MSR, ME)) {
-            /*
-             * Machine check exception is not enabled.  Enter
-             * checkstop state.
-             */
-            fprintf(stderr, "Machine check while not allowed. "
-                    "Entering checkstop state\n");
-            if (qemu_log_separate()) {
-                qemu_log("Machine check while not allowed. "
-                        "Entering checkstop state\n");
-            }
-            cs->halted = 1;
-            cpu_interrupt_exittb(cs);
-        }
-
+        powerpc_mcheck_test_and_checkstop(env);
         /* machine check exceptions don't have ME set */
         new_msr &= ~((target_ulong)1 << MSR_ME);
 
@@ -935,21 +912,7 @@ static void powerpc_excp_74xx(PowerPCCPU *cpu, int excp)
 
     switch (excp) {
     case POWERPC_EXCP_MCHECK:    /* Machine check exception                  */
-        if (!FIELD_EX64(env->msr, MSR, ME)) {
-            /*
-             * Machine check exception is not enabled.  Enter
-             * checkstop state.
-             */
-            fprintf(stderr, "Machine check while not allowed. "
-                    "Entering checkstop state\n");
-            if (qemu_log_separate()) {
-                qemu_log("Machine check while not allowed. "
-                        "Entering checkstop state\n");
-            }
-            cs->halted = 1;
-            cpu_interrupt_exittb(cs);
-        }
-
+        powerpc_mcheck_test_and_checkstop(env);
         /* machine check exceptions don't have ME set */
         new_msr &= ~((target_ulong)1 << MSR_ME);
 
@@ -1130,21 +1093,7 @@ static void powerpc_excp_booke(PowerPCCPU *cpu, int excp)
         srr1 = SPR_BOOKE_CSRR1;
         break;
     case POWERPC_EXCP_MCHECK:    /* Machine check exception                  */
-        if (!FIELD_EX64(env->msr, MSR, ME)) {
-            /*
-             * Machine check exception is not enabled.  Enter
-             * checkstop state.
-             */
-            fprintf(stderr, "Machine check while not allowed. "
-                    "Entering checkstop state\n");
-            if (qemu_log_separate()) {
-                qemu_log("Machine check while not allowed. "
-                        "Entering checkstop state\n");
-            }
-            cs->halted = 1;
-            cpu_interrupt_exittb(cs);
-        }
-
+        powerpc_mcheck_test_and_checkstop(env);
         /* machine check exceptions don't have ME set */
         new_msr &= ~((target_ulong)1 << MSR_ME);
 
@@ -1377,20 +1326,7 @@ static void powerpc_excp_books(PowerPCCPU *cpu, int excp)
 
     switch (excp) {
     case POWERPC_EXCP_MCHECK:    /* Machine check exception                  */
-        if (!FIELD_EX64(env->msr, MSR, ME)) {
-            /*
-             * Machine check exception is not enabled.  Enter
-             * checkstop state.
-             */
-            fprintf(stderr, "Machine check while not allowed. "
-                    "Entering checkstop state\n");
-            if (qemu_log_separate()) {
-                qemu_log("Machine check while not allowed. "
-                        "Entering checkstop state\n");
-            }
-            cs->halted = 1;
-            cpu_interrupt_exittb(cs);
-        }
+        powerpc_mcheck_test_and_checkstop(env);
         if (env->msr_mask & MSR_HVB) {
             /*
              * ISA specifies HV, but can be delivered to guest with HV
Signature · 1 line
2.40.1
#10 lore raw
In reply to Nicholas Piggin
Quoted 7 lines · Nicholas Piggin

ppc has always silently ignored access to real (physical) addresses with nothing behind it, which can make debugging difficult at times.

It looks like the way to handle this is implement the transaction failed call, which most target architectures do. Notably not x86 though, I wonder why?

Much of this is historical legacy. QEMU originally had no concept of "the system outside the CPU returns some kind of bus error and the CPU raises an exception for it". This is turn is (I think) because the x86 PC doesn't do
that: you always get back some kind of response, I think -1 on reads and writes ignored. We added the do_transaction_failed hook largely because we wanted it to give more accurate emulation of this kind of thing on Arm, but as usual with new facilities we left the other architectures to do it themselves if they wanted -- by default the behaviour remained the same. Some architectures have picked it up; some haven't.

The main reason it's a bit of a pain to turn the correct handling on is because often boards don't actually implement all the devices they're supposed to. For a pile of legacy Arm boards, especially where we didn't have good test images, we use the machine flag ignore_memory_transaction_failures to retain the legacy behaviour. (This isn't great because it's pretty much going to mean we have that flag set on those boards forever because nobody is going to care enough to investigate and test.)

Quoted 6 lines

Other question is, sometimes I guess it's nice to avoid crashing in order to try to quickly get past some unimplemented MMIO. Maybe a command line option or something could turn it off? It should probably be a QEMU-wide option if so, so that shouldn't hold this series up, I can propose a option for that if anybody is worried about it.

I would not recommend going any further than maybe setting the ignore_memory_transaction_failures flag for boards you don't care about. (But in an ideal world, don't set it and deal with any bug reports by implementing stub versions of missing devices. Depends how confident you are in your test coverage.)

thanks
-- PMM

#8 lore raw

Re: [PATCH 4/4] target/ppc: Make checkstop stop the system

In reply to Nicholas Piggin
Quoted 55 lines · Nicholas Piggin

checkstop state does not halt the system, interrupts continue to be serviced, and other CPUs run.

Stop the machine with vm_stop(), and print a register dump too.

---
target/ppc/excp_helper.c | 35 +++++++++++++++++++++--------------

1 file changed, 21 insertions(+), 14 deletions(-)
target/ppc/excp_helper.c +19 -4
index 4bfcfc3c3d..51e83d7f07 100644
@@ -19,6 +19,7 @@
 #include "qemu/osdep.h"
 #include "qemu/main-loop.h"
 #include "qemu/log.h"
+#include "sysemu/runstate.h"
 #include "cpu.h"
 #include "exec/exec-all.h"
 #include "internal.h"
@@ -165,6 +166,24 @@ static void ppc_excp_debug_sw_tlb(CPUPPCState *env, int excp)
             env->error_code);
 }
 
+static void powerpc_checkstop(PowerPCCPU *cpu, const char *reason)
+{
+    CPUState *cs = CPU(cpu);
+
+    vm_stop(RUN_STATE_GUEST_PANICKED);
+
+    fprintf(stderr, "Entering checkstop state: %s\n", reason);
+    cpu_dump_state(cs, stderr, CPU_DUMP_FPU | CPU_DUMP_CCOP);
+    if (qemu_log_separate()) {
+        FILE *logfile = qemu_log_trylock();
+        if (logfile) {
+            fprintf(logfile, "Entering checkstop state: %s\n", reason);
+            cpu_dump_state(cs, logfile, CPU_DUMP_FPU | CPU_DUMP_CCOP);
+            qemu_log_unlock(logfile);
+        }
+    }
+}
+
 #if defined(TARGET_PPC64)
 static int powerpc_reset_wakeup(CPUState *cs, CPUPPCState *env, int excp,
                                target_ulong *msr)
@@ -406,21 +425,9 @@ static void powerpc_set_excp_state(PowerPCCPU *cpu, target_ulong vector,
 
 static void powerpc_mcheck_test_and_checkstop(CPUPPCState *env)
 {
-    CPUState *cs = env_cpu(env);
-
-    if (FIELD_EX64(env->msr, MSR, ME)) {
-        return;

target/ppc/excp_helper.c static void powerpc_set_excp_state(PowerPCCPU *cpu, target_ulong vector, · line 429

Quoted 10 lines
target/ppc/excp_helper.c +2 -8
-    }
-
-    /* Machine check exception is not enabled. Enter checkstop state. */
-    fprintf(stderr, "Machine check while not allowed. "
-            "Entering checkstop state\n");
-    if (qemu_log_separate()) {
-        qemu_log("Machine check while not allowed. "
-                 "Entering checkstop state\n");
+    if (!FIELD_EX64(env->msr, MSR, ME)) {
+        powerpc_checkstop(env_archcpu(env), "machine check with MSR[ME]=0");

I don't mind you twaeking the patch and renaming the function but now this has become another one line function which just clutters code. Either keep this together in one function or inline the if at callers, otherwise this will start to look like Forth where every simple operation gets a new name. :-)

Regards,
BALATON Zoltan

Quoted 7 lines
target/ppc/excp_helper.c +0 -2
    }
-    cs->halted = 1;
-    cpu_interrupt_exittb(cs);
 }
 
 static void powerpc_excp_40x(PowerPCCPU *cpu, int excp)
#11 lore raw
In reply to Peter Maydell
Quoted 38 lines · Peter Maydell
Quoted 7 lines · Nicholas Piggin

ppc has always silently ignored access to real (physical) addresses with nothing behind it, which can make debugging difficult at times.

It looks like the way to handle this is implement the transaction failed call, which most target architectures do. Notably not x86 though, I wonder why?

Much of this is historical legacy. QEMU originally had no concept of "the system outside the CPU returns some kind of bus error and the CPU raises an exception for it". This is turn is (I think) because the x86 PC doesn't do
that: you always get back some kind of response, I think -1 on reads and writes ignored. We added the do_transaction_failed hook largely because we wanted it to give more accurate emulation of this kind of thing on Arm, but as usual with new facilities we left the other architectures to do it themselves if they wanted -- by default the behaviour remained the same. Some architectures have picked it up; some haven't.

The main reason it's a bit of a pain to turn the correct handling on is because often boards don't actually implement all the devices they're supposed to. For a pile of legacy Arm boards, especially where we didn't have good test images, we use the machine flag ignore_memory_transaction_failures to retain the legacy behaviour. (This isn't great because it's pretty much going to mean we have that flag set on those boards forever because nobody is going to care enough to investigate and test.)

Quoted 6 lines

Other question is, sometimes I guess it's nice to avoid crashing in order to try to quickly get past some unimplemented MMIO. Maybe a command line option or something could turn it off? It should probably be a QEMU-wide option if so, so that shouldn't hold this series up, I can propose a option for that if anybody is worried about it.

Quoted 5 lines

I would not recommend going any further than maybe setting the ignore_memory_transaction_failures flag for boards you don't care about. (But in an ideal world, don't set it and deal with any bug reports by implementing stub versions of missing devices. Depends how confident you are in your test coverage.)

It seems it broke the "mac99" and powernv10 machines, using the qemu-ppc-boot images which are mostly buildroot. See below for logs.

Adding Mark for further testing on Mac OS.

Thanks,

C.

Saving 256 bits of non-creditable seed for next boot
Starting network: [ 3.876834] Disabling lock debugging due to kernel taint

[    3.877351] Oops: Machine check, sig: 7 [#1]
[    3.877697] BE PAGE_SIZE=4K MMU=Hash SMP NR_CPUS=4 PowerMac
[    3.878019] Modules linked in:
[    3.878233] CPU: 0 PID: 93 Comm: ip Tainted: G   M               6.1.14 #1
[    3.878403] Hardware name: PowerMac3,1 PPC970 0x390202 PowerMac
[    3.878554] NIP:  c0000000007f1c28 LR: c0000000007f18d0 CTR: 0000000000000000
[    3.878721] REGS: c00000000ffefd60 TRAP: 0200   Tainted: G   M                (6.1.14)
[    3.878922] MSR:  8000000000009032 <SF,EE,ME,IR,DR,RI>  CR: 44002884  XER: 00000000
[    3.879149] DAR: c0003e0080103010 DSISR: 42000000 IRQMASK: 0
[    3.879149] GPR00: c0000000007f10fc c0000000037f35e0 c000000000d4bc00 0000000000000024
[    3.879149] GPR04: c0003f00000cbdc8 0000000000000000 0000000000000000 c0000000037f35a6
[    3.879149] GPR08: 0000000000000000 c0003e0080103010 0000000000000000 c000000001513c58
[    3.879149] GPR12: 0000000084002884 c0000000015fa000 000000010d8b0460 000000011a042f13
[    3.879149] GPR16: 000000011a042f7f 000000011a042f87 000000011a04052c 000000011a042f0e
[    3.879149] GPR20: 0000000000000000 0000000000000000 0000000000000000 ffffffffffffffff
[    3.879149] GPR24: c0000000036f8e00 0000000000000000 c00000000319af08 c00000000319af10
[    3.879149] GPR28: c000000003797000 c000000003798000 c00000000319a880 c000000003924c00
[    3.880889] NIP [c0000000007f1c28] .gem_reinit_chip+0xc28/0xc70
[    3.881045] LR [c0000000007f18d0] .gem_reinit_chip+0x8d0/0xc70
[    3.881209] Call Trace:
[    3.881274] [c0000000037f35e0] [c0000000007f10fc] .gem_reinit_chip+0xfc/0xc70 (unreliable)
[    3.881484] [c0000000037f3680] [c0000000007f2df4] .gem_do_start+0x34/0x430
[    3.881658] [c0000000037f3730] [c0000000009396d0] .__dev_open+0x150/0x240
[    3.881835] [c0000000037f37e0] [c000000000939d1c] .__dev_change_flags+0x20c/0x2c0
[    3.882015] [c0000000037f38b0] [c000000000939dfc] .dev_change_flags+0x2c/0x90
[    3.882185] [c0000000037f3940] [c000000000a1dc88] .devinet_ioctl+0x2e8/0x9b0
[    3.882361] [c0000000037f3a20] [c000000000a20854] .inet_ioctl+0x184/0x2c0
[    3.882527] [c0000000037f3b30] [c0000000008fa5d8] .sock_do_ioctl+0x58/0x130
[    3.882693] [c0000000037f3c00] [c0000000008fb77c] .sock_ioctl+0x1cc/0x830
[    3.882876] [c0000000037f3ce0] [c0000000002c6bcc] .__se_sys_ioctl+0x10c/0x130
[    3.883066] [c0000000037f3d80] [c0000000000252e8] .system_call_exception+0x158/0x2d0
[    3.883255] [c0000000037f3e10] [c00000000000b2d4] system_call_common+0xf4/0x258
[    3.883436] --- interrupt: c00 at 0x3fff825da858
[    3.883553] NIP:  00003fff825da858 LR: 0000000119f7b3a0 CTR: 0000000000000000
[    3.883721] REGS: c0000000037f3e80 TRAP: 0c00   Tainted: G   M                (6.1.14)
[    3.883910] MSR:  800000000200f032 <SF,VEC,EE,PR,FP,ME,IR,DR,RI>  CR: 28002884  XER: 00000000
[    3.884149] IRQMASK: 0
[    3.884149] GPR00: 0000000000000036 00003fffc530af10 00003fff826b7400 0000000000000003
[    3.884149] GPR04: 0000000000008914 00003fffc530b0a0 000000011a040518 0000000000000000
[    3.884149] GPR08: 00000000803c7414 0000000000000000 0000000000000000 0000000000000000
[    3.884149] GPR12: 0000000000000000 00003fff827377e0 000000010d8b0460 000000011a042f13
[    3.884149] GPR16: 000000011a042f7f 000000011a042f87 000000011a04052c 000000011a042f0e
[    3.884149] GPR20: 0000000000000000 0000000000000000 0000000000000000 ffffffffffffffff
[    3.884149] GPR24: ffffffffffffffff ffffffffffffffff ffffffffffffffff 00003fffc530bf51
[    3.884149] GPR28: 00003fffc530b0a0 0000000000000001 0000000000000001 000000011a040518
[    3.885775] NIP [00003fff825da858] 0x3fff825da858
[    3.885902] LR [0000000119f7b3a0] 0x119f7b3a0
[    3.886006] --- interrupt: c00
[    3.886109] Instruction dump:
[    3.886296] 39290044 55290032 1d09fffe 7d485214 7d295050 915e01b8 913e01bc 4bfff628
[    3.886537] e93e0000 39400000 39293010 7c0004ac <7d404d2c> a14d02f8 5548043e 2c080000
[    3.886968] ---[ end trace 0000000000000000 ]---
[    3.887112]

and

[   34.443113577,5] XIVE: [ IC 00  ] Resetting one xive...
[   34.446516646,5] RESET: Fast reboot disabled: Kernel re-entered OPAL
[   34.447320802,3] OPAL exiting with locks held, pir=0000 token=80 retval=1
[   34.447483518,3]   hw/xive2.c:3955
[   34.447578862,5] CPU ATTEMPT TO RE-ENTER FIRMWARE! PIR=0000 cpu @0x31c10000 -> pir=0000 token=158
[   34.447781018,5] CPU ATTEMPT TO RE-ENTER FIRMWARE! PIR=0000 cpu @0x31c10000 -> pir=0000 token=158
[    0.000000][    C0] Disabling lock debugging due to kernel taint
[    0.000000][    C0] MCE: CPU0: machine check (Severe)  Real address Load (bad) DAR: 0006030201000038 [Not recovered]
[    0.000000][    C0] MCE: CPU0: NIP: [0000000030093290] 0x30093290
[    0.000000][    C0] MCE: CPU0: Initiator CPU
[    0.000000][    C0] MCE: CPU0: Hardware error
[    0.000000][    C0] opal: Hardware platform error: Unrecoverable Machine Check exception
[    0.000000][    C0] CPU: 0 PID: 0 Comm: swapper/0 Tainted: G   M               6.1.14 #1
[    0.000000][    C0] Hardware name: IBM PowerNV (emulated by qemu) POWER10 0x801200 opal:v7.0 PowerNV
[    0.000000][    C0] NIP:  0000000030093290 LR: 0000000030093284 CTR: 0000000030100f20
[    0.000000][    C0] REGS: c00000003df8fd60 TRAP: 0200   Tainted: G   M                (6.1.14)
[    0.000000][    C0] MSR:  9000000002201002 <SF,HV,VEC,ME,RI>  CR: 84022424  XER: 20040000
[    0.000000][    C0] CFAR: 0000000030112dd8 DAR: 0006030201000038 DSISR: 00000040 IRQMASK: 3
[    0.000000][    C0] GPR00: 0000000030093284 0000000031c13bd0 0000000030192600 8000000000000000
[    0.000000][    C0] GPR04: 0000000000030000 0000000000000300 0000000030144f46 0000000030144f38
[    0.000000][    C0] GPR08: 0000000000000010 0006030201000038 0000000000000000 0000000000000001
[    0.000000][    C0] GPR12: 0000000024022424 c000000002580000 0000000000000000 0000000000000000
[    0.000000][    C0] GPR16: 0000000031c10000 0000000000000001 0000000000000008 0000000000800000
[    0.000000][    C0] GPR20: ffffffffffffff80 0000000030608ea8 0000000030145427 00000000301453e8
[    0.000000][    C0] GPR24: 0000000030145312 000000003014548f 00000000301454bd 0000000030608118
[    0.000000][    C0] GPR28: 0000000000000038 0006030201000000 0000000000000003 0000000030608e08
[    0.000000][    C0] NIP [0000000030093290] 0x30093290
[    0.000000][    C0] LR [0000000030093284] 0x30093284
[    0.000000][    C0] Call Trace:
[    0.000000][    C0] Instruction dump:
[    0.000000][    C0] XXXXXXXX XXXXXXXX XXXXXXXX XXXXXXXX XXXXXXXX XXXXXXXX XXXXXXXX XXXXXXXX
[    0.000000][    C0] XXXXXXXX XXXXXXXX XXXXXXXX XXXXXXXX XXXXXXXX XXXXXXXX XXXXXXXX XXXXXXXX
[   34.458079299,0] OPAL: Reboot requested due to Platform error.
[   34.458349474,3] OPAL: Reboot requested due to Platform error.[   34.458589592,5] Unable to log error
[   34.458756741,2] NVRAM: Failed to load
[    0.000000][    C0] opal: Reboot type 1 not supported for Unrecoverable Machine Check exception
[    0.000000][    C0] Kernel panic legoater@virtlab1024:~/qemu/qemu-ppc-boot.git$
#4 lore raw

Re: [PATCH 3/4] target/ppc: Move common check in machne check handlers to a function

In reply to Nicholas Piggin
Quoted 5 lines · Nicholas Piggin

From: BALATON Zoltan <[email protected]>

All powerpc exception handlers share some code when handling machine check exceptions. Move this to a common function.

Maybe Machine Check is simple enough, but this kind of sharing of code has historically caused pain when people want to change something for the modern cpus and end up affecting the old cpus by mistake.

There is also the inverse scenario where someone has access to the old HW and just want to make an one-off contribution, but the community gets insecure about it because it could also affect the new cpus.

Then comes the obvious "solution" which is to bring in an artificial identifier (excp. model) to be able to have conditional code inside the common function. And that causes problems because no one really knows how it maps to actual hardware/ISA.

No objection, just a little cautionary tale. =)

#5 lore raw

Re: [PATCH 3/4] target/ppc: Move common check in machne check handlers to a function

Quoted 8 lines · Fabiano Rosas
Quoted 5 lines · Nicholas Piggin

From: BALATON Zoltan <[email protected]>

All powerpc exception handlers share some code when handling machine check exceptions. Move this to a common function.

Quoted 7 lines

Maybe Machine Check is simple enough, but this kind of sharing of code has historically caused pain when people want to change something for the modern cpus and end up affecting the old cpus by mistake.

There is also the inverse scenario where someone has access to the old HW and just want to make an one-off contribution, but the community gets insecure about it because it could also affect the new cpus.

It's a trade off between making these independent and avoiding code duplication. I think if we have common things that are the same between different CPU models then we should not duplicate those, otherwise they can diverge when a problem is fixed in one copy but not in the others. We just have to remember that if in the future a new CPU defines these differently then we mayy need to add another function implementing that for the new CPU but the old CPUs can still share the same code without duplicating it. So the question is if this machine check is something that is the same across different CPU models or do we ever need to model this differently for different CPU models? At least the current implementation had the same code duplicated everywhere which this patch resolves.

Regards,
BALATON Zoltan

Quoted 6 lines

Then comes the obvious "solution" which is to bring in an artificial identifier (excp. model) to be able to have conditional code inside the common function. And that causes problems because no one really knows how it maps to actual hardware/ISA.

No objection, just a little cautionary tale. =)

#12 lore raw
In reply to Cédric Le Goater
Quoted 39 lines · Cédric Le Goater
Quoted 38 lines · Peter Maydell
Quoted 7 lines · Nicholas Piggin

ppc has always silently ignored access to real (physical) addresses with nothing behind it, which can make debugging difficult at times.

It looks like the way to handle this is implement the transaction failed call, which most target architectures do. Notably not x86 though, I wonder why?

Much of this is historical legacy. QEMU originally had no concept of "the system outside the CPU returns some kind of bus error and the CPU raises an exception for it". This is turn is (I think) because the x86 PC doesn't do
that: you always get back some kind of response, I think -1 on reads and writes ignored. We added the do_transaction_failed hook largely because we wanted it to give more accurate emulation of this kind of thing on Arm, but as usual with new facilities we left the other architectures to do it themselves if they wanted -- by default the behaviour remained the same. Some architectures have picked it up; some haven't.

The main reason it's a bit of a pain to turn the correct handling on is because often boards don't actually implement all the devices they're supposed to. For a pile of legacy Arm boards, especially where we didn't have good test images, we use the machine flag ignore_memory_transaction_failures to retain the legacy behaviour. (This isn't great because it's pretty much going to mean we have that flag set on those boards forever because nobody is going to care enough to investigate and test.)

Quoted 6 lines

Other question is, sometimes I guess it's nice to avoid crashing in order to try to quickly get past some unimplemented MMIO. Maybe a command line option or something could turn it off? It should probably be a QEMU-wide option if so, so that shouldn't hold this series up, I can propose a option for that if anybody is worried about it.

Quoted 8 lines
Quoted 5 lines

I would not recommend going any further than maybe setting the ignore_memory_transaction_failures flag for boards you don't care about. (But in an ideal world, don't set it and deal with any bug reports by implementing stub versions of missing devices. Depends how confident you are in your test coverage.)

It seems it broke the "mac99" and  powernv10 machines, using the qemu-ppc-boot images which are mostly buildroot. See below for logs.

Since commit 21786c7e59 ("softmmu/memory: Log invalid memory accesses") you can log the failed transaction with '-d guest_errors'. See for example commit a13bfa5a05 ("hw/mips/jazz: Map the UART devices unconditionally"):

  $ qemu-system-mips64el -M magnum -d guest_errors,unimp -bios NTPROM.RAW
  Invalid access at addr 0x80007004, size 1, region '(null)', reason: 
rejected
  Invalid access at addr 0x80007001, size 1, region '(null)', reason: 
rejected
  Invalid access at addr 0x80007002, size 1, region '(null)', reason: 
rejected
  Invalid access at addr 0x80007003, size 1, region '(null)', reason: 
rejected
  Invalid access at addr 0x80007004, size 1, region '(null)', reason: 
rejected

Boards booting successfully with ignore_memory_transaction_failures set can often remove this flag by mapping missing accessed ranges as TYPE_UNIMPLEMENTED_DEVICE. (You can then log the same accesses using '-d unimp').

#13 lore raw
Quoted 46 lines · Philippe Mathieu-Daudé
Quoted 45 lines · Cédric Le Goater
Quoted 43 lines · Peter Maydell
Quoted 7 lines · Nicholas Piggin

ppc has always silently ignored access to real (physical) addresses with nothing behind it, which can make debugging difficult at times.

It looks like the way to handle this is implement the transaction failed call, which most target architectures do. Notably not x86 though, I wonder why?

Much of this is historical legacy. QEMU originally had no concept of "the system outside the CPU returns some kind of bus error and the CPU raises an exception for it". This is turn is (I think) because the x86 PC doesn't do
that: you always get back some kind of response, I think -1 on reads and writes ignored. We added the do_transaction_failed hook largely because we wanted it to give more accurate emulation of this kind of thing on Arm, but as usual with new facilities we left the other architectures to do it themselves if they wanted -- by default the behaviour remained the same. Some architectures have picked it up; some haven't.

The main reason it's a bit of a pain to turn the correct handling on is because often boards don't actually implement all the devices they're supposed to. For a pile of legacy Arm boards, especially where we didn't have good test images, we use the machine flag ignore_memory_transaction_failures to retain the legacy behaviour. (This isn't great because it's pretty much going to mean we have that flag set on those boards forever because nobody is going to care enough to investigate and test.)

Quoted 6 lines

Other question is, sometimes I guess it's nice to avoid crashing in order to try to quickly get past some unimplemented MMIO. Maybe a command line option or something could turn it off? It should probably be a QEMU-wide option if so, so that shouldn't hold this series up, I can propose a option for that if anybody is worried about it.

I would not recommend going any further than maybe setting the ignore_memory_transaction_failures flag for boards you don't care about. (But in an ideal world, don't set it and deal with any bug reports by implementing stub versions of missing devices. Depends how confident you are in your test coverage.)

Quoted 6 lines
Quoted 2 lines

It seems it broke the "mac99" and  powernv10 machines, using the qemu-ppc-boot images which are mostly buildroot. See below for logs.

Since commit 21786c7e59 ("softmmu/memory: Log invalid memory accesses") you can log the failed transaction with '-d guest_errors'. See for example commit a13bfa5a05 ("hw/mips/jazz: Map the UART devices

This reminds me I'd still want to split this from guest_errors as discussed here:

https://lists.nongnu.org/archive/html/qemu-devel/2023-02/msg08757.html

Can we get a decision on how to call these debug options?

Quoted 14 lines

unconditionally"):

 $ qemu-system-mips64el -M magnum -d guest_errors,unimp -bios NTPROM.RAW
 Invalid access at addr 0x80007004, size 1, region '(null)', reason: 
rejected
 Invalid access at addr 0x80007001, size 1, region '(null)', reason: 
rejected
 Invalid access at addr 0x80007002, size 1, region '(null)', reason: 
rejected
 Invalid access at addr 0x80007003, size 1, region '(null)', reason: 
rejected
 Invalid access at addr 0x80007004, size 1, region '(null)', reason: 
rejected
Quoted 4 lines

Boards booting successfully with ignore_memory_transaction_failures set can often remove this flag by mapping missing accessed ranges as TYPE_UNIMPLEMENTED_DEVICE. (You can then log the same accesses using '-d unimp').

The mac99 may have a lot of unimplemented devices and they are also not quite documented so it may even be difficult to find where they should be.

Regards,
BALATON Zoltan

#9 lore raw

Re: [PATCH 4/4] target/ppc: Make checkstop stop the system

In reply to BALATON Zoltan
Quoted 64 lines · BALATON Zoltan
Quoted 63 lines · Nicholas Piggin

checkstop state does not halt the system, interrupts continue to be serviced, and other CPUs run.

Stop the machine with vm_stop(), and print a register dump too.

---
target/ppc/excp_helper.c | 35 +++++++++++++++++++++--------------

1 file changed, 21 insertions(+), 14 deletions(-)
target/ppc/excp_helper.c +19 -12
index 4bfcfc3c3d..51e83d7f07 100644
@@ -19,6 +19,7 @@
 #include "qemu/osdep.h"
 #include "qemu/main-loop.h"
 #include "qemu/log.h"
+#include "sysemu/runstate.h"
 #include "cpu.h"
 #include "exec/exec-all.h"
 #include "internal.h"
@@ -165,6 +166,24 @@ static void ppc_excp_debug_sw_tlb(CPUPPCState *env, int excp)
             env->error_code);
 }
 
+static void powerpc_checkstop(PowerPCCPU *cpu, const char *reason)
+{
+    CPUState *cs = CPU(cpu);
+
+    vm_stop(RUN_STATE_GUEST_PANICKED);
+
+    fprintf(stderr, "Entering checkstop state: %s\n", reason);
+    cpu_dump_state(cs, stderr, CPU_DUMP_FPU | CPU_DUMP_CCOP);
+    if (qemu_log_separate()) {
+        FILE *logfile = qemu_log_trylock();
+        if (logfile) {
+            fprintf(logfile, "Entering checkstop state: %s\n", reason);
+            cpu_dump_state(cs, logfile, CPU_DUMP_FPU | CPU_DUMP_CCOP);
+            qemu_log_unlock(logfile);
+        }
+    }
+}
+
 #if defined(TARGET_PPC64)
 static int powerpc_reset_wakeup(CPUState *cs, CPUPPCState *env, int excp,
                                target_ulong *msr)
@@ -406,21 +425,9 @@ static void powerpc_set_excp_state(PowerPCCPU *cpu, target_ulong vector,
 
 static void powerpc_mcheck_test_and_checkstop(CPUPPCState *env)
 {
-    CPUState *cs = env_cpu(env);
-
-    if (FIELD_EX64(env->msr, MSR, ME)) {
-        return;
-    }
-
-    /* Machine check exception is not enabled. Enter checkstop state. */
-    fprintf(stderr, "Machine check while not allowed. "
-            "Entering checkstop state\n");
-    if (qemu_log_separate()) {
-        qemu_log("Machine check while not allowed. "
-                 "Entering checkstop state\n");
Quoted 8 lines

target/ppc/excp_helper.c static void powerpc_set_excp_state(PowerPCCPU *cpu, target_ulong vector, · line 429

Quoted 2 lines
target/ppc/excp_helper.c +2 -0
+    if (!FIELD_EX64(env->msr, MSR, ME)) {
+        powerpc_checkstop(env_archcpu(env), "machine check with MSR[ME]=0");

I don't mind you twaeking the patch and renaming the function but now this has become another one line function which just clutters code. Either keep this together in one function or inline the if at callers, otherwise this will start to look like Forth where every simple operation gets a new name. :-)

Yeah good point. I did want to have a powerpc_checkstop function with a reason because other places might start to also call it in future.

As far as the machine check ME test goes... we could re-inline that I suppose.

Thanks,
Nick

#30 lore raw
In reply to Peter Maydell
Quoted 38 lines · Peter Maydell
Quoted 7 lines · Nicholas Piggin

ppc has always silently ignored access to real (physical) addresses with nothing behind it, which can make debugging difficult at times.

It looks like the way to handle this is implement the transaction failed call, which most target architectures do. Notably not x86 though, I wonder why?

Much of this is historical legacy. QEMU originally had no concept of "the system outside the CPU returns some kind of bus error and the CPU raises an exception for it". This is turn is (I think) because the x86 PC doesn't do
that: you always get back some kind of response, I think -1 on reads and writes ignored. We added the do_transaction_failed hook largely because we wanted it to give more accurate emulation of this kind of thing on Arm, but as usual with new facilities we left the other architectures to do it themselves if they wanted -- by default the behaviour remained the same. Some architectures have picked it up; some haven't.

The main reason it's a bit of a pain to turn the correct handling on is because often boards don't actually implement all the devices they're supposed to. For a pile of legacy Arm boards, especially where we didn't have good test images, we use the machine flag ignore_memory_transaction_failures to retain the legacy behaviour. (This isn't great because it's pretty much going to mean we have that flag set on those boards forever because nobody is going to care enough to investigate and test.)

Quoted 6 lines

Other question is, sometimes I guess it's nice to avoid crashing in order to try to quickly get past some unimplemented MMIO. Maybe a command line option or something could turn it off? It should probably be a QEMU-wide option if so, so that shouldn't hold this series up, I can propose a option for that if anybody is worried about it.

Quoted 5 lines

I would not recommend going any further than maybe setting the ignore_memory_transaction_failures flag for boards you don't care about. (But in an ideal world, don't set it and deal with any bug reports by implementing stub versions of missing devices. Depends how confident you are in your test coverage.)

Thanks for the background, interesting and helpful. So I think it is the right place for powerpc BookS 64 to hook into. Point taken about adding a global option for it. Will try to fix the known problems first, maybe it won't be too hard.

Thanks,
Nick

#6 lore raw

Re: [PATCH 3/4] target/ppc: Move common check in machne check handlers to a function

In reply to Fabiano Rosas
Quoted 16 lines · Fabiano Rosas
Quoted 5 lines · Nicholas Piggin

From: BALATON Zoltan <[email protected]>

All powerpc exception handlers share some code when handling machine check exceptions. Move this to a common function.

Maybe Machine Check is simple enough, but this kind of sharing of code has historically caused pain when people want to change something for the modern cpus and end up affecting the old cpus by mistake.

There is also the inverse scenario where someone has access to the old HW and just want to make an one-off contribution, but the community gets insecure about it because it could also affect the new cpus.

Quoted 6 lines

Then comes the obvious "solution" which is to bring in an artificial identifier (excp. model) to be able to have conditional code inside the common function. And that causes problems because no one really knows how it maps to actual hardware/ISA.

No objection, just a little cautionary tale. =)

Thanks Fabiano, good point. I know you spent a lot of work on untangling this mess. I'll think a bit more about it. Seems we need to at least make a few fixes first before we can turn this on for upstream.

Thanks,
Nick

#14 lore raw
In reply to Cédric Le Goater
Quoted 45 lines · Cédric Le Goater
Quoted 38 lines · Peter Maydell
Quoted 7 lines · Nicholas Piggin

ppc has always silently ignored access to real (physical) addresses with nothing behind it, which can make debugging difficult at times.

It looks like the way to handle this is implement the transaction failed call, which most target architectures do. Notably not x86 though, I wonder why?

Much of this is historical legacy. QEMU originally had no concept of "the system outside the CPU returns some kind of bus error and the CPU raises an exception for it". This is turn is (I think) because the x86 PC doesn't do
that: you always get back some kind of response, I think -1 on reads and writes ignored. We added the do_transaction_failed hook largely because we wanted it to give more accurate emulation of this kind of thing on Arm, but as usual with new facilities we left the other architectures to do it themselves if they wanted -- by default the behaviour remained the same. Some architectures have picked it up; some haven't.

The main reason it's a bit of a pain to turn the correct handling on is because often boards don't actually implement all the devices they're supposed to. For a pile of legacy Arm boards, especially where we didn't have good test images, we use the machine flag ignore_memory_transaction_failures to retain the legacy behaviour. (This isn't great because it's pretty much going to mean we have that flag set on those boards forever because nobody is going to care enough to investigate and test.)

Quoted 6 lines

Other question is, sometimes I guess it's nice to avoid crashing in order to try to quickly get past some unimplemented MMIO. Maybe a command line option or something could turn it off? It should probably be a QEMU-wide option if so, so that shouldn't hold this series up, I can propose a option for that if anybody is worried about it.

Quoted 5 lines

I would not recommend going any further than maybe setting the ignore_memory_transaction_failures flag for boards you don't care about. (But in an ideal world, don't set it and deal with any bug reports by implementing stub versions of missing devices. Depends how confident you are in your test coverage.)

Quoted 4 lines

It seems it broke the "mac99" and  powernv10 machines, using the qemu-ppc-boot images which are mostly buildroot. See below for logs.

Adding Mark for further testing on Mac OS.

Mac OS 9.2 fails to boot with a popup saying :

Sorry, a system error occured.
"Sound Manager"
address error
To temporarily turn off extensions, restart and hold down the shift key

Darwin and Mac OSX look OK.

Thanks,

C.

#15 lore raw
Quoted 54 lines · Cédric Le Goater
Quoted 45 lines · Cédric Le Goater
Quoted 38 lines · Peter Maydell
Quoted 7 lines · Nicholas Piggin

ppc has always silently ignored access to real (physical) addresses with nothing behind it, which can make debugging difficult at times.

It looks like the way to handle this is implement the transaction failed call, which most target architectures do. Notably not x86 though, I wonder why?

Much of this is historical legacy. QEMU originally had no concept of "the system outside the CPU returns some kind of bus error and the CPU raises an exception for it". This is turn is (I think) because the x86 PC doesn't do
that: you always get back some kind of response, I think -1 on reads and writes ignored. We added the do_transaction_failed hook largely because we wanted it to give more accurate emulation of this kind of thing on Arm, but as usual with new facilities we left the other architectures to do it themselves if they wanted -- by default the behaviour remained the same. Some architectures have picked it up; some haven't.

The main reason it's a bit of a pain to turn the correct handling on is because often boards don't actually implement all the devices they're supposed to. For a pile of legacy Arm boards, especially where we didn't have good test images, we use the machine flag ignore_memory_transaction_failures to retain the legacy behaviour. (This isn't great because it's pretty much going to mean we have that flag set on those boards forever because nobody is going to care enough to investigate and test.)

Quoted 6 lines

Other question is, sometimes I guess it's nice to avoid crashing in order to try to quickly get past some unimplemented MMIO. Maybe a command line option or something could turn it off? It should probably be a QEMU-wide option if so, so that shouldn't hold this series up, I can propose a option for that if anybody is worried about it.

Quoted 5 lines

I would not recommend going any further than maybe setting the ignore_memory_transaction_failures flag for boards you don't care about. (But in an ideal world, don't set it and deal with any bug reports by implementing stub versions of missing devices. Depends how confident you are in your test coverage.)

Quoted 4 lines

It seems it broke the "mac99" and  powernv10 machines, using the qemu-ppc-boot images which are mostly buildroot. See below for logs.

Adding Mark for further testing on Mac OS.

Mac OS 9.2 fails to boot with a popup saying :

Quoted 8 lines

Sorry, a system error occured.
"Sound Manager"
address error
To temporarily turn off extensions, restart and hold down the shift key

Darwin and Mac OSX look OK.

Might have to to restrict it to POWER machines for now then. Seems like it will break working systems.

We could just log a guest error for the others.

Thanks,
Nick

#16 lore raw
Quoted 15 lines
Quoted 15 lines
Quoted 4 lines

It seems it broke the "mac99" and  powernv10 machines, using the qemu-ppc-boot images which are mostly buildroot. See below for logs.

Adding Mark for further testing on Mac OS.

Mac OS 9.2 fails to boot with a popup saying :

Sorry, a system error occured.
"Sound Manager"
address error
To temporarily turn off extensions, restart and hold down the shift key

Quoted 6 lines
Quoted 1 line

Darwin and Mac OSX look OK.

Might have to to restrict it to POWER machines for now then. Seems like it will break working systems.

We could just log a guest error for the others.

The "mac99" and powernv machines run fine now with the patches you and Fred just sent. There is a "fix" window after soft freeze when we can set the 'ignore_memory_transaction_failures' flag as Peter suggested.

Thanks,

C.

#17 lore raw
In reply to Cédric Le Goater
Quoted 60 lines · Cédric Le Goater
Quoted 45 lines · Cédric Le Goater
Quoted 38 lines · Peter Maydell
Quoted 7 lines · Nicholas Piggin

ppc has always silently ignored access to real (physical) addresses with nothing behind it, which can make debugging difficult at times.

It looks like the way to handle this is implement the transaction failed call, which most target architectures do. Notably not x86 though, I wonder why?

Much of this is historical legacy. QEMU originally had no concept of "the system outside the CPU returns some kind of bus error and the CPU raises an exception for it". This is turn is (I think) because the x86 PC doesn't do
that: you always get back some kind of response, I think -1 on reads and writes ignored. We added the do_transaction_failed hook largely because we wanted it to give more accurate emulation of this kind of thing on Arm, but as usual with new facilities we left the other architectures to do it themselves if they wanted -- by default the behaviour remained the same. Some architectures have picked it up; some haven't.

The main reason it's a bit of a pain to turn the correct handling on is because often boards don't actually implement all the devices they're supposed to. For a pile of legacy Arm boards, especially where we didn't have good test images, we use the machine flag ignore_memory_transaction_failures to retain the legacy behaviour. (This isn't great because it's pretty much going to mean we have that flag set on those boards forever because nobody is going to care enough to investigate and test.)

Quoted 6 lines

Other question is, sometimes I guess it's nice to avoid crashing in order to try to quickly get past some unimplemented MMIO. Maybe a command line option or something could turn it off? It should probably be a QEMU-wide option if so, so that shouldn't hold this series up, I can propose a option for that if anybody is worried about it.

Quoted 5 lines

I would not recommend going any further than maybe setting the ignore_memory_transaction_failures flag for boards you don't care about. (But in an ideal world, don't set it and deal with any bug reports by implementing stub versions of missing devices. Depends how confident you are in your test coverage.)

Quoted 4 lines

It seems it broke the "mac99" and  powernv10 machines, using the qemu-ppc-boot images which are mostly buildroot. See below for logs.

Adding Mark for further testing on Mac OS.

Mac OS 9.2 fails to boot with a popup saying :
        Sorry, a system error occured.
        "Sound Manager"
          address error
        To temporarily turn off extensions, restart and hold down the shift key

Quoted 1 line

Darwin and Mac OSX look OK.

My guess would be that MacOS 9.2 is trying to access the sound chip registers which isn't implemented in QEMU for the moment (I have a separate screamer branch available, but it's not ready for primetime yet). In theory they shouldn't be accessed at all because the sound device isn't present in the OpenBIOS device tree, but this is all fairly old stuff.

Does implementing the sound registers using a dummy device help at all?

hw/misc/macio/macio.c +10 -0
index 265c0bbd8d..e55f938da7 100644
@@ -26,6 +26,7 @@
 #include "qemu/osdep.h"
 #include "qapi/error.h"
 #include "qemu/module.h"
+#include "hw/misc/unimp.h"
 #include "hw/misc/macio/cuda.h"
 #include "hw/pci/pci.h"
 #include "hw/ppc/mac_dbdma.h"
@@ -94,6 +95,7 @@ static bool macio_common_realize(PCIDevice *d, Error **errp)
 {
     MacIOState *s = MACIO(d);
     SysBusDevice *sbd;
+    DeviceState *dev;
 
     if (!qdev_realize(DEVICE(&s->dbdma), BUS(&s->macio_bus), errp)) {
         return false;
@@ -102,6 +104,14 @@ static bool macio_common_realize(PCIDevice *d, Error **errp)
     memory_region_add_subregion(&s->bar, 0x08000,
                                 sysbus_mmio_get_region(sbd, 0));
 
+    dev = qdev_new(TYPE_UNIMPLEMENTED_DEVICE);
+    qdev_prop_set_string(dev, "name", "screamer");
+    qdev_prop_set_uint64(dev, "size", 0x1000);
+    sysbus_realize_and_unref(SYS_BUS_DEVICE(dev), &error_fatal);
+    sbd = SYS_BUS_DEVICE(dev);
+    memory_region_add_subregion(&s->bar, 0x14000,
+                                sysbus_mmio_get_region(sbd, 0));
+
     qdev_prop_set_uint32(DEVICE(&s->escc), "disabled", 0);
     qdev_prop_set_uint32(DEVICE(&s->escc), "frequency", ESCC_CLOCK);
     qdev_prop_set_uint32(DEVICE(&s->escc), "it_shift", 4);
include/hw/misc/macio/macio.h +1 -0
index 86df2c2b60..1894178a68 100644
@@ -109,6 +109,7 @@ struct MacIOState {
     PMUState pmu;
     DBDMAState dbdma;
     ESCCState escc;
+    MemoryRegion screamer;
     uint64_t frequency;
 };
 

ATB,

Mark.

#18 lore raw
Quoted 75 lines · Mark Cave-Ayland
Quoted 3 lines · Cédric Le Goater
Quoted 2 lines · Cédric Le Goater
Quoted 1 line · Peter Maydell

On Fri, 23 Jun 2023 at 09:21, Nicholas Piggin <[email protected]>

Quoted 57 lines · wrote:
Quoted 43 lines
Quoted 37 lines
Quoted 7 lines

ppc has always silently ignored access to real (physical) addresses with nothing behind it, which can make debugging difficult at times.

It looks like the way to handle this is implement the transaction failed call, which most target architectures do. Notably not x86 though, I wonder why?

Much of this is historical legacy. QEMU originally had no concept of "the system outside the CPU returns some kind of bus error and the CPU raises an exception for it". This is turn is (I think) because the x86 PC doesn't do
that: you always get back some kind of response, I think -1 on reads and writes ignored. We added the do_transaction_failed hook largely because we wanted it to give more accurate emulation of this kind of thing on Arm, but as usual with new facilities we left the other architectures to do it themselves if they wanted -- by default the behaviour remained the same. Some architectures have picked it up; some haven't.

The main reason it's a bit of a pain to turn the correct handling on is because often boards don't actually implement all the devices they're supposed to. For a pile of legacy Arm boards, especially where we didn't have good test images, we use the machine flag ignore_memory_transaction_failures to retain the legacy behaviour. (This isn't great because it's pretty much going to mean we have that flag set on those boards forever because nobody is going to care enough to investigate and test.)

Quoted 6 lines

Other question is, sometimes I guess it's nice to avoid crashing in order to try to quickly get past some unimplemented MMIO. Maybe a command line option or something could turn it off? It should probably be a QEMU-wide option if so, so that shouldn't hold this series up, I can propose a option for that if anybody is worried about it.

Quoted 5 lines

I would not recommend going any further than maybe setting the ignore_memory_transaction_failures flag for boards you don't care about. (But in an ideal world, don't set it and deal with any bug reports by implementing stub versions of missing devices. Depends how confident you are in your test coverage.)

Quoted 4 lines

It seems it broke the "mac99" and powernv10 machines, using the qemu-ppc-boot images which are mostly buildroot. See below for logs.

Adding Mark for further testing on Mac OS.

Mac OS 9.2 fails to boot with a popup saying :
Sorry, a system error occured.
"Sound Manager"
address error
To temporarily turn off extensions, restart and hold down the shift key

Quoted 1 line

Darwin and Mac OSX look OK.

My guess would be that MacOS 9.2 is trying to access the sound chip registers which
isn't implemented in QEMU for the moment (I have a separate screamer branch
available, but it's not ready for primetime yet). In theory they shouldn't be
accessed at all because the sound device isn't present in the OpenBIOS device tree,
but this is all fairly old stuff.

Quoted 3 lines

Does implementing the sound registers using a dummy device help at all?

My uneducated guess is that you stumbled on a longstanding, but intermittently occurring, issue specific to Mac OS 9.2 related to sound support over USB in Apple monitors.
I believe It is not fixed by the patch set from the 23 of june, I still get system errors when running Mac OS 9.2 with the mac99 machine after applying them.
Mac OS 9.2 has required mac99,via=pmu for a long time now to always boot successfully. (while 9.0.4 requires mac99 to boot, due to an undiagnosed OHCI USB problem with the specific drivers that ship with it.) ;-)

Best,
Howard

Quoted 57 lines
hw/misc/macio/macio.c +2 -0
index 265c0bbd8d..e55f938da7 100644
@@ -26,6 +26,7 @@
  #include "qemu/osdep.h"
  #include "qapi/error.h"
  #include "qemu/module.h"
+#include "hw/misc/unimp.h"
  #include "hw/misc/macio/cuda.h"
  #include "hw/pci/pci.h"
  #include "hw/ppc/mac_dbdma.h"
@@ -94,6 +95,7 @@ static bool macio_common_realize(PCIDevice *d, Error
 **errp)
  {
      MacIOState *s = MACIO(d);
      SysBusDevice *sbd;
+    DeviceState *dev;
 
      if (!qdev_realize(DEVICE(&s->dbdma), BUS(&s->macio_bus), errp)) {
          return false;
hw/misc/macio/macio.c +8 -0
@@ -102,6 +104,14 @@ static bool macio_common_realize(PCIDevice *d, Error
 **errp)
      memory_region_add_subregion(&s->bar, 0x08000,
                                  sysbus_mmio_get_region(sbd, 0));
 
+    dev = qdev_new(TYPE_UNIMPLEMENTED_DEVICE);
+    qdev_prop_set_string(dev, "name", "screamer");
+    qdev_prop_set_uint64(dev, "size", 0x1000);
+    sysbus_realize_and_unref(SYS_BUS_DEVICE(dev), &error_fatal);
+    sbd = SYS_BUS_DEVICE(dev);
+    memory_region_add_subregion(&s->bar, 0x14000,
+                                sysbus_mmio_get_region(sbd, 0));
+
      qdev_prop_set_uint32(DEVICE(&s->escc), "disabled", 0);
      qdev_prop_set_uint32(DEVICE(&s->escc), "frequency", ESCC_CLOCK);
      qdev_prop_set_uint32(DEVICE(&s->escc), "it_shift", 4);
include/hw/misc/macio/macio.h +1 -0
index 86df2c2b60..1894178a68 100644
@@ -109,6 +109,7 @@ struct MacIOState {
      PMUState pmu;
      DBDMAState dbdma;
      ESCCState escc;
+    MemoryRegion screamer;
      uint64_t frequency;
  };
 

ATB,

Mark.

#19 lore raw
Quoted 73 lines · Howard Spoelstra

On Tue, Jun 27, 2023 at 10:15 AM Mark Cave-Ayland <[email protected] <mailto:[email protected]>> wrote:

On 26/06/2023 14:35, Cédric Le Goater wrote:

> On 6/23/23 14:37, Cédric Le Goater wrote:
>> On 6/23/23 11:10, Peter Maydell wrote:
>>> On Fri, 23 Jun 2023 at 09:21, Nicholas Piggin <[email protected]
<mailto:[email protected]>> wrote:
>>>>
>>>> ppc has always silently ignored access to real (physical) addresses >>>> with nothing behind it, which can make debugging difficult at times. >>>>
>>>> It looks like the way to handle this is implement the transaction >>>> failed call, which most target architectures do. Notably not x86 >>>> though, I wonder why?
>>>
>>> Much of this is historical legacy. QEMU originally had no >>> concept of "the system outside the CPU returns some kind >>> of bus error and the CPU raises an exception for it". >>> This is turn is (I think) because the x86 PC doesn't do >>> that: you always get back some kind of response, I think >>> -1 on reads and writes ignored. We added the do_transaction_failed >>> hook largely because we wanted it to give more accurate >>> emulation of this kind of thing on Arm, but as usual with new >>> facilities we left the other architectures to do it themselves >>> if they wanted -- by default the behaviour remained the same. >>> Some architectures have picked it up; some haven't.
>>>
>>> The main reason it's a bit of a pain to turn the correct >>> handling on is because often boards don't actually implement >>> all the devices they're supposed to. For a pile of legacy Arm >>> boards, especially where we didn't have good test images, >>> we use the machine flag ignore_memory_transaction_failures to >>> retain the legacy behaviour. (This isn't great because it's >>> pretty much going to mean we have that flag set on those >>> boards forever because nobody is going to care enough to >>> investigate and test.)
>>>
>>>> Other question is, sometimes I guess it's nice to avoid crashing in >>>> order to try to quickly get past some unimplemented MMIO. Maybe a >>>> command line option or something could turn it off? It should >>>> probably be a QEMU-wide option if so, so that shouldn't hold this >>>> series up, I can propose a option for that if anybody is worried >>>> about it.
>>>
>>> I would not recommend going any further than maybe setting the >>> ignore_memory_transaction_failures flag for boards you don't >>> care about. (But in an ideal world, don't set it and deal with >>> any bug reports by implementing stub versions of missing devices. >>> Depends how confident you are in your test coverage.)
>>
>> It seems it broke the "mac99" and  powernv10 machines, using the >> qemu-ppc-boot images which are mostly buildroot. See below for logs. >>
>> Adding Mark for further testing on Mac OS.
>
>
> Mac OS 9.2 fails to boot with a popup saying :
>          Sorry, a system error occured.
>          "Sound Manager"
>            address error
>          To temporarily turn off extensions, restart and
>          hold down the shift key
>
>
> Darwin and Mac OSX look OK.

My guess would be that MacOS 9.2 is trying to access the sound chip registers which isn't implemented in QEMU for the moment (I have a separate screamer branch available, but it's not ready for primetime yet). In theory they shouldn't be accessed at all because the sound device isn't present in the OpenBIOS device tree, but this is all fairly old stuff.

Quoted 4 lines

Does implementing the sound registers using a dummy device help at all?

My uneducated guess is that you stumbled on a longstanding, but intermittently occurring, issue specific to Mac OS 9.2 related to sound support over USB in Apple monitors.

I'm not sure I understand this: are there non-standard command line options being used here other than "qemu-system-ppc -M mac99 -cdrom macos92.iso -boot d"?

Quoted 2 lines

I believe It is not fixed by the patch set from the 23 of june, I still get system errors when running Mac OS 9.2 with the mac99 machine after applying them.
Mac OS 9.2 has required mac99,via=pmu for a long time now to always boot successfully. (while 9.0.4 requires mac99 to boot, due to an undiagnosed OHCI USB problem with the specific drivers that ship with it.)  ;-)

I always test MacOS 9.2 boot both with and without via=pmu for my OpenBIOS tests, so I'd expect this to work unless a regression has slipped in?

ATB,

Mark.

#28 lore raw
In reply to Mark Cave-Ayland
Quoted 8 lines
Quoted 8 lines

Mac OS 9.2 fails to boot with a popup saying :
         Sorry, a system error occured.
         "Sound Manager"
           address error
         To temporarily turn off extensions, restart and hold down the shift key

Quoted 5 lines
Quoted 1 line

Darwin and Mac OSX look OK.

My guess would be that MacOS 9.2 is trying to access the sound chip registers which isn't implemented in QEMU for the moment (I have a separate screamer branch available, but it's not ready for primetime yet). In theory they shouldn't be accessed at all because the sound device isn't present in the OpenBIOS device tree, but this is all fairly old stuff.

Does implementing the sound registers using a dummy device help at all?

Nope. OS 9 loops earlier (little black/white disk spinning).

Thanks,

C.

Quoted 54 lines
hw/misc/macio/macio.c +10 -0
index 265c0bbd8d..e55f938da7 100644
@@ -26,6 +26,7 @@
  #include "qemu/osdep.h"
  #include "qapi/error.h"
  #include "qemu/module.h"
+#include "hw/misc/unimp.h"
  #include "hw/misc/macio/cuda.h"
  #include "hw/pci/pci.h"
  #include "hw/ppc/mac_dbdma.h"
@@ -94,6 +95,7 @@ static bool macio_common_realize(PCIDevice *d, Error **errp)
  {
      MacIOState *s = MACIO(d);
      SysBusDevice *sbd;
+    DeviceState *dev;
 
      if (!qdev_realize(DEVICE(&s->dbdma), BUS(&s->macio_bus), errp)) {
          return false;
@@ -102,6 +104,14 @@ static bool macio_common_realize(PCIDevice *d, Error **errp)
      memory_region_add_subregion(&s->bar, 0x08000,
                                  sysbus_mmio_get_region(sbd, 0));
 
+    dev = qdev_new(TYPE_UNIMPLEMENTED_DEVICE);
+    qdev_prop_set_string(dev, "name", "screamer");
+    qdev_prop_set_uint64(dev, "size", 0x1000);
+    sysbus_realize_and_unref(SYS_BUS_DEVICE(dev), &error_fatal);
+    sbd = SYS_BUS_DEVICE(dev);
+    memory_region_add_subregion(&s->bar, 0x14000,
+                                sysbus_mmio_get_region(sbd, 0));
+
      qdev_prop_set_uint32(DEVICE(&s->escc), "disabled", 0);
      qdev_prop_set_uint32(DEVICE(&s->escc), "frequency", ESCC_CLOCK);
      qdev_prop_set_uint32(DEVICE(&s->escc), "it_shift", 4);
include/hw/misc/macio/macio.h +1 -0
index 86df2c2b60..1894178a68 100644
@@ -109,6 +109,7 @@ struct MacIOState {
      PMUState pmu;
      DBDMAState dbdma;
      ESCCState escc;
+    MemoryRegion screamer;
      uint64_t frequency;
  };
 

ATB,

Mark.

#20 lore raw
In reply to Mark Cave-Ayland
Quoted 98 lines · Mark Cave-Ayland
Quoted 1 line · Howard Spoelstra

On Tue, Jun 27, 2023 at 10:15 AM Mark Cave-Ayland <

[email protected]

Quoted 7 lines

<mailto:[email protected]>> wrote:

    On 26/06/2023 14:35, Cédric Le Goater wrote:

     > On 6/23/23 14:37, Cédric Le Goater wrote:
     >> On 6/23/23 11:10, Peter Maydell wrote:
     >>> On Fri, 23 Jun 2023 at 09:21, Nicholas Piggin <

[email protected]

Quoted 3 lines

<mailto:[email protected]>> wrote:
>>>>
>>>> ppc has always silently ignored access to real (physical)

addresses

Quoted 1 line

>>>> with nothing behind it, which can make debugging difficult at

times.

Quoted 2 lines

>>>>
>>>> It looks like the way to handle this is implement the

transaction

Quoted 1 line

>>>> failed call, which most target architectures do. Notably not

x86

Quoted 8 lines

>>>> though, I wonder why?
>>>
>>> Much of this is historical legacy. QEMU originally had no >>> concept of "the system outside the CPU returns some kind >>> of bus error and the CPU raises an exception for it". >>> This is turn is (I think) because the x86 PC doesn't do >>> that: you always get back some kind of response, I think >>> -1 on reads and writes ignored. We added the

do_transaction_failed

Quoted 9 lines

>>> hook largely because we wanted it to give more accurate >>> emulation of this kind of thing on Arm, but as usual with new >>> facilities we left the other architectures to do it themselves >>> if they wanted -- by default the behaviour remained the same. >>> Some architectures have picked it up; some haven't. >>>
>>> The main reason it's a bit of a pain to turn the correct >>> handling on is because often boards don't actually implement >>> all the devices they're supposed to. For a pile of legacy Arm

Quoted 8 lines

>>> boards, especially where we didn't have good test images, >>> we use the machine flag ignore_memory_transaction_failures to >>> retain the legacy behaviour. (This isn't great because it's >>> pretty much going to mean we have that flag set on those >>> boards forever because nobody is going to care enough to >>> investigate and test.)
>>>
>>>> Other question is, sometimes I guess it's nice to avoid

crashing in

Quoted 1 line

>>>> order to try to quickly get past some unimplemented MMIO.

Maybe a

Quoted 2 lines

>>>> command line option or something could turn it off? It should >>>> probably be a QEMU-wide option if so, so that shouldn't hold

this

Quoted 1 line

>>>> series up, I can propose a option for that if anybody is

worried

Quoted 6 lines

>>>> about it.
>>>
>>> I would not recommend going any further than maybe setting the >>> ignore_memory_transaction_failures flag for boards you don't >>> care about. (But in an ideal world, don't set it and deal with >>> any bug reports by implementing stub versions of missing

devices.

Quoted 4 lines

>>> Depends how confident you are in your test coverage.) >>
>> It seems it broke the "mac99" and powernv10 machines, using the >> qemu-ppc-boot images which are mostly buildroot. See below for

logs.

Quoted 14 lines

>>
>> Adding Mark for further testing on Mac OS. >
>
> Mac OS 9.2 fails to boot with a popup saying :
> Sorry, a system error occured.
> "Sound Manager"
> address error
> To temporarily turn off extensions, restart and
> hold down the shift key
>
>
> Darwin and Mac OSX look OK.

Quoted 1 line

My guess would be that MacOS 9.2 is trying to access the sound chip

registers which

Quoted 1 line

isn't implemented in QEMU for the moment (I have a separate screamer

branch

Quoted 1 line

available, but it's not ready for primetime yet). In theory they

shouldn't be

Quoted 1 line

accessed at all because the sound device isn't present in the

OpenBIOS device tree,

Quoted 3 lines

but this is all fairly old stuff.

Does implementing the sound registers using a dummy device help at

all?

Quoted 3 lines

My uneducated guess is that you stumbled on a longstanding, but

Quoted 9 lines

intermittently

Quoted 1 line

occurring, issue specific to Mac OS 9.2 related to sound support over

USB in Apple

Quoted 1 line

monitors.

I'm not sure I understand this: are there non-standard command line options being
used here other than "qemu-system-ppc -M mac99 -cdrom macos92.iso -boot d"?

It must be my windows host ;-)

qemu-system-ppc.exe -M mac99,via=pmu -cdrom C:\mac-iso\9.2.2.iso -boot d -L pc-bios
crashes Mac OS with an address error. (with unpatched and patched builds).

qemu-system-ppc.exe -M mac99 -hda C:\mac-hd\9.2.2-clean.img -boot c -L pc-bios sometimes crashes with an illegal instruction.

qemu-system-ppc.exe -M mac99,via=pmu -hda C:\mac-hd\9.2.2-clean.img -boot c -L pc-bios sometimes crashes with Sound manager address error. (with both patched and non-patched versions).

Best,
Howard

Quoted 20 lines
Quoted 2 lines

I believe It is not fixed by the patch set from the 23 of june, I still

get system

Quoted 1 line

errors when running Mac OS 9.2 with the mac99 machine after applying

them.

Quoted 2 lines

Mac OS 9.2 has required mac99,via=pmu for a long time now to always boot successfully. (while 9.0.4 requires mac99 to boot, due to an undiagnosed

OHCI USB

Quoted 1 line

problem with the specific drivers that ship with it.) ;-)

I always test MacOS 9.2 boot both with and without via=pmu for my OpenBIOS tests, so
I'd expect this to work unless a regression has slipped in?

ATB,

Mark.

#21 lore raw
Quoted 92 lines · Howard Spoelstra

On Tue, Jun 27, 2023 at 1:24 PM Mark Cave-Ayland <[email protected] <mailto:[email protected]>> wrote:

On 27/06/2023 11:28, Howard Spoelstra wrote:

> On Tue, Jun 27, 2023 at 10:15 AM Mark Cave-Ayland <[email protected] <mailto:[email protected]> > <mailto:[email protected] <mailto:[email protected]>>> wrote:
>
>     On 26/06/2023 14:35, Cédric Le Goater wrote:
>
>      > On 6/23/23 14:37, Cédric Le Goater wrote:
>      >> On 6/23/23 11:10, Peter Maydell wrote:
>      >>> On Fri, 23 Jun 2023 at 09:21, Nicholas Piggin <[email protected] <mailto:[email protected]> >     <mailto:[email protected] <mailto:[email protected]>>> wrote:
>      >>>>
>      >>>> ppc has always silently ignored access to real (physical) addresses
>      >>>> with nothing behind it, which can make debugging difficult at times.
>      >>>>
>      >>>> It looks like the way to handle this is implement the transaction
>      >>>> failed call, which most target architectures do. Notably not x86
>      >>>> though, I wonder why?
>      >>>
>      >>> Much of this is historical legacy. QEMU originally had no
>      >>> concept of "the system outside the CPU returns some kind
>      >>> of bus error and the CPU raises an exception for it".
>      >>> This is turn is (I think) because the x86 PC doesn't do
>      >>> that: you always get back some kind of response, I think
>      >>> -1 on reads and writes ignored. We added the do_transaction_failed
>      >>> hook largely because we wanted it to give more accurate
>      >>> emulation of this kind of thing on Arm, but as usual with new
>      >>> facilities we left the other architectures to do it themselves
>      >>> if they wanted -- by default the behaviour remained the same.
>      >>> Some architectures have picked it up; some haven't.
>      >>>
>      >>> The main reason it's a bit of a pain to turn the correct
>      >>> handling on is because often boards don't actually implement
>      >>> all the devices they're supposed to. For a pile of legacy Arm
>      >>> boards, especially where we didn't have good test images,
>      >>> we use the machine flag ignore_memory_transaction_failures to
>      >>> retain the legacy behaviour. (This isn't great because it's
>      >>> pretty much going to mean we have that flag set on those
>      >>> boards forever because nobody is going to care enough to
>      >>> investigate and test.)
>      >>>
>      >>>> Other question is, sometimes I guess it's nice to avoid crashing in
>      >>>> order to try to quickly get past some unimplemented MMIO. Maybe a
>      >>>> command line option or something could turn it off? It should
>      >>>> probably be a QEMU-wide option if so, so that shouldn't hold this
>      >>>> series up, I can propose a option for that if anybody is worried
>      >>>> about it.
>      >>>
>      >>> I would not recommend going any further than maybe setting the
>      >>> ignore_memory_transaction_failures flag for boards you don't
>      >>> care about. (But in an ideal world, don't set it and deal with
>      >>> any bug reports by implementing stub versions of missing devices.
>      >>> Depends how confident you are in your test coverage.)
>      >>
>      >> It seems it broke the "mac99" and  powernv10 machines, using the
>      >> qemu-ppc-boot images which are mostly buildroot. See below for logs.
>      >>
>      >> Adding Mark for further testing on Mac OS.
>      >
>      >
>      > Mac OS 9.2 fails to boot with a popup saying :
>      >          Sorry, a system error occured.
>      >          "Sound Manager"
>      >            address error
>      >          To temporarily turn off extensions, restart and
>      >          hold down the shift key
>      >
>      >
>      > Darwin and Mac OSX look OK.
>
>     My guess would be that MacOS 9.2 is trying to access the sound chip registers which
>     isn't implemented in QEMU for the moment (I have a separate screamer branch
>     available, but it's not ready for primetime yet). In theory they shouldn't be
>     accessed at all because the sound device isn't present in the OpenBIOS device tree,
>     but this is all fairly old stuff.
>
>     Does implementing the sound registers using a dummy device help at all?
>
>
> My uneducated guess is that you stumbled on a longstanding, but intermittently
> occurring, issue specific to Mac OS 9.2 related to sound support over USB in Apple
> monitors.

I'm not sure I understand this: are there non-standard command line options being used here other than "qemu-system-ppc -M mac99 -cdrom macos92.iso -boot d"?

Quoted 4 lines

It must be my windows host ;-)

qemu-system-ppc.exe -M mac99,via=pmu -cdrom C:\mac-iso\9.2.2.iso -boot d -L pc-bios crashes Mac OS with an address error. (with unpatched and patched builds).

Same on Linux. I get an invalid opcode. QEMU 7.2 work fine though.

C.

Quoted 24 lines

qemu-system-ppc.exe -M mac99 -hda C:\mac-hd\9.2.2-clean.img -boot c -L pc-bios sometimes crashes with an illegal instruction.

qemu-system-ppc.exe -M mac99,via=pmu -hda C:\mac-hd\9.2.2-clean.img -boot c -L pc-bios sometimes crashes with Sound manager address error. (with both patched and non-patched versions).

Best,
Howard

> I believe It is not fixed by the patch set from the 23 of june, I still get system > errors when running Mac OS 9.2 with the mac99 machine after applying them. > Mac OS 9.2 has required mac99,via=pmu for a long time now to always boot > successfully. (while 9.0.4 requires mac99 to boot, due to an undiagnosed OHCI USB > problem with the specific drivers that ship with it.)  ;-)

I always test MacOS 9.2 boot both with and without via=pmu for my OpenBIOS tests, so I'd expect this to work unless a regression has slipped in?

ATB,

Mark.

#29 lore raw
In reply to Cédric Le Goater
Quoted 12 lines · Cédric Le Goater
Quoted 12 lines
Quoted 9 lines

Mac OS 9.2 fails to boot with a popup saying :
         Sorry, a system error occured.
         "Sound Manager"
           address error
         To temporarily turn off extensions, restart and hold down the shift key

Darwin and Mac OSX look OK.

My guess would be that MacOS 9.2 is trying to access the sound chip registers which isn't implemented in QEMU for the moment (I have a separate screamer branch available, but it's not ready for primetime yet). In theory they shouldn't be accessed at all because the sound device isn't present in the OpenBIOS device tree, but this is all fairly old stuff.

Quoted 7 lines
Quoted 1 line

Does implementing the sound registers using a dummy device help at all?

Nope. OS 9 loops earlier (little black/white disk spinning).

Thanks,

C.

Hmmm that's annoying. Another one to add to the TODO list then...

ATB,

Mark.

#22 lore raw
In reply to Cédric Le Goater
Quoted 93 lines · Cédric Le Goater
Quoted 92 lines · Howard Spoelstra

On Tue, Jun 27, 2023 at 1:24 PM Mark Cave-Ayland <[email protected] <mailto:[email protected]>> wrote:

    On 27/06/2023 11:28, Howard Spoelstra wrote:

     > On Tue, Jun 27, 2023 at 10:15 AM Mark Cave-Ayland <[email protected] <mailto:[email protected]> > <mailto:[email protected] <mailto:[email protected]>>> wrote:
     >
     >     On 26/06/2023 14:35, Cédric Le Goater wrote:
     >
     >      > On 6/23/23 14:37, Cédric Le Goater wrote:
     >      >> On 6/23/23 11:10, Peter Maydell wrote:
     >      >>> On Fri, 23 Jun 2023 at 09:21, Nicholas Piggin <[email protected] <mailto:[email protected]> >     <mailto:[email protected] <mailto:[email protected]>>> wrote:
     >      >>>>
     >      >>>> ppc has always silently ignored access to real (physical) addresses
     >      >>>> with nothing behind it, which can make debugging difficult at times.
     >      >>>>
     >      >>>> It looks like the way to handle this is implement the transaction
     >      >>>> failed call, which most target architectures do. Notably not x86
     >      >>>> though, I wonder why?
     >      >>>
     >      >>> Much of this is historical legacy. QEMU originally had no
     >      >>> concept of "the system outside the CPU returns some kind
     >      >>> of bus error and the CPU raises an exception for it".
     >      >>> This is turn is (I think) because the x86 PC doesn't do
     >      >>> that: you always get back some kind of response, I think
     >      >>> -1 on reads and writes ignored. We added the do_transaction_failed
     >      >>> hook largely because we wanted it to give more accurate
     >      >>> emulation of this kind of thing on Arm, but as usual with new
     >      >>> facilities we left the other architectures to do it themselves
     >      >>> if they wanted -- by default the behaviour remained the same.
     >      >>> Some architectures have picked it up; some haven't.
     >      >>>
     >      >>> The main reason it's a bit of a pain to turn the correct
     >      >>> handling on is because often boards don't actually implement
     >      >>> all the devices they're supposed to. For a pile of legacy Arm
     >      >>> boards, especially where we didn't have good test images,
     >      >>> we use the machine flag ignore_memory_transaction_failures to
     >      >>> retain the legacy behaviour. (This isn't great because it's
     >      >>> pretty much going to mean we have that flag set on those
     >      >>> boards forever because nobody is going to care enough to
     >      >>> investigate and test.)
     >      >>>
     >      >>>> Other question is, sometimes I guess it's nice to avoid crashing in
     >      >>>> order to try to quickly get past some unimplemented MMIO. Maybe a
     >      >>>> command line option or something could turn it off? It should
     >      >>>> probably be a QEMU-wide option if so, so that shouldn't hold this
     >      >>>> series up, I can propose a option for that if anybody is worried
     >      >>>> about it.
     >      >>>
     >      >>> I would not recommend going any further than maybe setting the
     >      >>> ignore_memory_transaction_failures flag for boards you don't
     >      >>> care about. (But in an ideal world, don't set it and deal with
     >      >>> any bug reports by implementing stub versions of missing devices.
     >      >>> Depends how confident you are in your test coverage.)
     >      >>
     >      >> It seems it broke the "mac99" and  powernv10 machines, using the
     >      >> qemu-ppc-boot images which are mostly buildroot. See below for logs.
     >      >>
     >      >> Adding Mark for further testing on Mac OS.
     >      >
     >      >
     >      > Mac OS 9.2 fails to boot with a popup saying :
     >      >          Sorry, a system error occured.
     >      >          "Sound Manager"
     >      >            address error
     >      >          To temporarily turn off extensions, restart and
     >      >          hold down the shift key
     >      >
     >      >
     >      > Darwin and Mac OSX look OK.
     >
     >     My guess would be that MacOS 9.2 is trying to access the sound chip registers which
     >     isn't implemented in QEMU for the moment (I have a separate screamer branch
     >     available, but it's not ready for primetime yet). In theory they shouldn't be
     >     accessed at all because the sound device isn't present in the OpenBIOS device tree,
     >     but this is all fairly old stuff.
     >
     >     Does implementing the sound registers using a dummy device help at all?
     >
     >
     > My uneducated guess is that you stumbled on a longstanding, but intermittently
     > occurring, issue specific to Mac OS 9.2 related to sound support over USB in Apple
     > monitors.

    I'm not sure I understand this: are there non-standard command line options being used here other than "qemu-system-ppc -M mac99 -cdrom macos92.iso -boot d"?

Quoted 8 lines
Quoted 4 lines

It must be my windows host ;-)

qemu-system-ppc.exe -M mac99,via=pmu -cdrom C:\mac-iso\9.2.2.iso -boot d -L pc-bios crashes Mac OS with an address error. (with unpatched and patched builds).

Same on Linux. I get an invalid opcode. QEMU 7.2 work fine though.

C.

That certainly shouldn't happen, and if it worked in 7.2 then there's definitely a regression which has crept in there somewhere. I'll try and bisect this at some point soon, but feel free to try and beat me ;)

ATB,

Mark.

Reviewed-by Philippe Mathieu-Daudé
#23 lore raw
Quoted 97 lines · Mark Cave-Ayland
Quoted 95 lines · Cédric Le Goater
Quoted 94 lines · Howard Spoelstra

On Tue, Jun 27, 2023 at 1:24 PM Mark Cave-Ayland <[email protected] <mailto:[email protected]>> wrote:

    On 27/06/2023 11:28, Howard Spoelstra wrote:

     > On Tue, Jun 27, 2023 at 10:15 AM Mark Cave-Ayland <[email protected] <mailto:[email protected]> > <mailto:[email protected] <mailto:[email protected]>>> wrote:
     >
     >     On 26/06/2023 14:35, Cédric Le Goater wrote:
     >
     >      > On 6/23/23 14:37, Cédric Le Goater wrote:
     >      >> On 6/23/23 11:10, Peter Maydell wrote:
     >      >>> On Fri, 23 Jun 2023 at 09:21, Nicholas Piggin <[email protected] <mailto:[email protected]> >     <mailto:[email protected] <mailto:[email protected]>>> wrote:
     >      >>>>
     >      >>>> ppc has always silently ignored access to real (physical) addresses
     >      >>>> with nothing behind it, which can make debugging difficult at times.
     >      >>>>
     >      >>>> It looks like the way to handle this is implement the transaction
     >      >>>> failed call, which most target architectures do. Notably not x86
     >      >>>> though, I wonder why?
     >      >>>
     >      >>> Much of this is historical legacy. QEMU originally had no
     >      >>> concept of "the system outside the CPU returns some kind
     >      >>> of bus error and the CPU raises an exception for it".
     >      >>> This is turn is (I think) because the x86 PC doesn't do
     >      >>> that: you always get back some kind of response, I think
     >      >>> -1 on reads and writes ignored. We added the do_transaction_failed
     >      >>> hook largely because we wanted it to give more accurate
     >      >>> emulation of this kind of thing on Arm, but as usual with new
     >      >>> facilities we left the other architectures to do it themselves
     >      >>> if they wanted -- by default the behaviour remained the same.
     >      >>> Some architectures have picked it up; some haven't.
     >      >>>
     >      >>> The main reason it's a bit of a pain to turn the correct
     >      >>> handling on is because often boards don't actually implement
     >      >>> all the devices they're supposed to. For a pile of legacy Arm
     >      >>> boards, especially where we didn't have good test images,
     >      >>> we use the machine flag ignore_memory_transaction_failures to
     >      >>> retain the legacy behaviour. (This isn't great because it's
     >      >>> pretty much going to mean we have that flag set on those
     >      >>> boards forever because nobody is going to care enough to
     >      >>> investigate and test.)
     >      >>>
     >      >>>> Other question is, sometimes I guess it's nice to avoid crashing in
     >      >>>> order to try to quickly get past some unimplemented MMIO. Maybe a
     >      >>>> command line option or something could turn it off? It should
     >      >>>> probably be a QEMU-wide option if so, so that shouldn't hold this
     >      >>>> series up, I can propose a option for that if anybody is worried
     >      >>>> about it.
     >      >>>
     >      >>> I would not recommend going any further than maybe setting the
     >      >>> ignore_memory_transaction_failures flag for boards you don't
     >      >>> care about. (But in an ideal world, don't set it and deal with
     >      >>> any bug reports by implementing stub versions of missing devices.
     >      >>> Depends how confident you are in your test coverage.)
     >      >>
     >      >> It seems it broke the "mac99" and  powernv10 machines, using the
     >      >> qemu-ppc-boot images which are mostly buildroot. See below for logs.
     >      >>
     >      >> Adding Mark for further testing on Mac OS.
     >      >
     >      >
     >      > Mac OS 9.2 fails to boot with a popup saying :
     >      >          Sorry, a system error occured.
     >      >          "Sound Manager"
     >      >            address error
     >      >          To temporarily turn off extensions, restart and
     >      >          hold down the shift key
     >      >
     >      >
     >      > Darwin and Mac OSX look OK.
     >
     >     My guess would be that MacOS 9.2 is trying to access the sound chip registers which
     >     isn't implemented in QEMU for the moment (I have a separate screamer branch
     >     available, but it's not ready for primetime yet). In theory they shouldn't be
     >     accessed at all because the sound device isn't present in the OpenBIOS device tree,
     >     but this is all fairly old stuff.
     >
     >     Does implementing the sound registers using a dummy device help at all?
     >
     >
     > My uneducated guess is that you stumbled on a longstanding, but intermittently
     > occurring, issue specific to Mac OS 9.2 related to sound support over USB in Apple
     > monitors.

    I'm not sure I understand this: are there non-standard command line options being used here other than "qemu-system-ppc -M mac99 -cdrom macos92.iso -boot d"?

It must be my windows host ;-)

Quoted 8 lines
Quoted 6 lines
Quoted 2 lines

qemu-system-ppc.exe -M mac99,via=pmu -cdrom C:\mac-iso\9.2.2.iso -boot d -L pc-bios crashes Mac OS with an address error. (with unpatched and patched builds).

Same on Linux. I get an invalid opcode. QEMU 7.2 work fine though.

C.

That certainly shouldn't happen, and if it worked in 7.2 then there's definitely a regression which has crept in there somewhere. I'll try and bisect this at some point soon, but feel free to try and beat me ;)

bisect points to :

commit e506ad6a05c806bbef460a7d014a184ff8d707a6

Author: Richard Henderson <[email protected]>
Date:   Mon Mar 6 04:30:11 2023 +0300

accel/tcg: Pass last not end to tb_invalidate_phys_range

Pass the address of the last byte to be changed, rather than the first address past the last byte. This avoids overflow when the last page of the address space is involved.

 include/exec/exec-all.h   |  2 +-
 accel/tcg/tb-maint.c      | 31 ++++++++++++++++---------------
 accel/tcg/translate-all.c |  2 +-
 accel/tcg/user-exec.c     |  2 +-
 softmmu/physmem.c         |  2 +-
 5 files changed, 20 insertions(+), 19 deletions(-)

I think the instruction is fnmadds. Needs more digging.

Thanks,

C.

#24 lore raw
Quoted 30 lines
Quoted 8 lines
Quoted 6 lines
Quoted 2 lines

qemu-system-ppc.exe -M mac99,via=pmu -cdrom C:\mac-iso\9.2.2.iso -boot d -L pc-bios crashes Mac OS with an address error. (with unpatched and patched builds).

Same on Linux. I get an invalid opcode. QEMU 7.2 work fine though.

C.

That certainly shouldn't happen, and if it worked in 7.2 then there's definitely a regression which has crept in there somewhere. I'll try and bisect this at some point soon, but feel free to try and beat me ;)

bisect points to :

commit e506ad6a05c806bbef460a7d014a184ff8d707a6
Author: Richard Henderson <[email protected]> Date:   Mon Mar 6 04:30:11 2023 +0300

    accel/tcg: Pass last not end to tb_invalidate_phys_range Pass the address of the last byte to be changed, rather than the first address past the last byte.  This avoids overflow when the last page of the address space is involved.

  include/exec/exec-all.h   |  2 +-
  accel/tcg/tb-maint.c      | 31 ++++++++++++++++---------------
  accel/tcg/translate-all.c |  2 +-
  accel/tcg/user-exec.c     |  2 +-
  softmmu/physmem.c         |  2 +-
  5 files changed, 20 insertions(+), 19 deletions(-)
Quoted 1 line

I think the instruction is fnmadds. Needs more digging.

the invalid opcode is just a symptom of something bad happening.

C.

#25 lore raw
Quoted 30 lines · Cédric Le Goater
Quoted 30 lines
Quoted 8 lines
Quoted 6 lines
Quoted 2 lines

qemu-system-ppc.exe -M mac99,via=pmu -cdrom C:\mac-iso\9.2.2.iso -boot d -L pc-bios crashes Mac OS with an address error. (with unpatched and patched builds).

Same on Linux. I get an invalid opcode. QEMU 7.2 work fine though.

C.

That certainly shouldn't happen, and if it worked in 7.2 then there's definitely a regression which has crept in there somewhere. I'll try and bisect this at some point soon, but feel free to try and beat me ;)

bisect points to :

commit e506ad6a05c806bbef460a7d014a184ff8d707a6
Author: Richard Henderson <[email protected]> Date:   Mon Mar 6 04:30:11 2023 +0300

     accel/tcg: Pass last not end to tb_invalidate_phys_range Pass the address of the last byte to be changed, rather than the first address past the last byte.  This avoids overflow when the last page of the address space is involved.

  include/exec/exec-all.h   |  2 +- accel/tcg/tb-maint.c      | 31 ++++++++++++++++--------------- accel/tcg/translate-all.c |  2 +- accel/tcg/user-exec.c     |  2 +-
  softmmu/physmem.c         |  2 +-

  5 files changed, 20 insertions(+), 19 deletions(-)
Quoted 5 lines
Quoted 1 line

I think the instruction is fnmadds. Needs more digging.

the invalid opcode is just a symptom of something bad happening.

C.

Indeed, it appears to be a copy/paste error within that commit. I've just posted a proposed fix for this: https://lore.kernel.org/qemu-devel/[email protected]/T/.

Nick: you may wish to try your series again with this fix applied to see if there are still problems with the CPUs used in the Mac machines.

ATB,

Mark.

#26 lore raw
Quoted 32 lines · Mark Cave-Ayland
Quoted 30 lines · Cédric Le Goater
Quoted 30 lines
Quoted 8 lines
Quoted 6 lines
Quoted 2 lines

qemu-system-ppc.exe -M mac99,via=pmu -cdrom C:\mac-iso\9.2.2.iso -boot d -L pc-bios crashes Mac OS with an address error. (with unpatched and patched builds).

Same on Linux. I get an invalid opcode. QEMU 7.2 work fine though.

C.

That certainly shouldn't happen, and if it worked in 7.2 then there's definitely a regression which has crept in there somewhere. I'll try and bisect this at some point soon, but feel free to try and beat me ;)

bisect points to :

commit e506ad6a05c806bbef460a7d014a184ff8d707a6
Author: Richard Henderson <[email protected]> Date:   Mon Mar 6 04:30:11 2023 +0300

     accel/tcg: Pass last not end to tb_invalidate_phys_range Pass the address of the last byte to be changed, rather than the first address past the last byte.  This avoids overflow when the last page of the address space is involved.

  include/exec/exec-all.h   |  2 +- accel/tcg/tb-maint.c      | 31 ++++++++++++++++--------------- accel/tcg/translate-all.c |  2 +- accel/tcg/user-exec.c     |  2 +-
  softmmu/physmem.c         |  2 +-

  5 files changed, 20 insertions(+), 19 deletions(-)
Quoted 7 lines
Quoted 5 lines
Quoted 1 line

I think the instruction is fnmadds. Needs more digging.

the invalid opcode is just a symptom of something bad happening.

C.

Indeed, it appears to be a copy/paste error within that commit. I've just posted a proposed fix for this: https://lore.kernel.org/qemu-devel/[email protected]/T/.

Looks good. I could boot macos 9.2.1 from an iso and 9.2.2 from disk.

Quoted 1 line

Nick: you may wish to try your series again with this fix applied to see if there are still problems with the CPUs used in the Mac machines.

Looks good too, with these :

[PATCH v2 1/4] target/ppc: Machine check on invalid real address [PATCH v2 2/4] target/ppc: Move common check in machine check [PATCH v2 3/4] target/ppc: Make checkstop actually stop the system

Patch 4 (attn) doesn't compile but it's an extra behavior on top of checkstop.

Thanks,

C.

#27 lore raw
Quoted 43 lines · Cédric Le Goater
Quoted 32 lines · Mark Cave-Ayland
Quoted 30 lines · Cédric Le Goater
Quoted 30 lines
Quoted 8 lines
Quoted 6 lines
Quoted 2 lines

qemu-system-ppc.exe -M mac99,via=pmu -cdrom C:\mac-iso\9.2.2.iso -boot d -L pc-bios crashes Mac OS with an address error. (with unpatched and patched builds).

Same on Linux. I get an invalid opcode. QEMU 7.2 work fine though.

C.

That certainly shouldn't happen, and if it worked in 7.2 then there's definitely a regression which has crept in there somewhere. I'll try and bisect this at some point soon, but feel free to try and beat me ;)

bisect points to :

commit e506ad6a05c806bbef460a7d014a184ff8d707a6
Author: Richard Henderson <[email protected]> Date:   Mon Mar 6 04:30:11 2023 +0300

     accel/tcg: Pass last not end to tb_invalidate_phys_range Pass the address of the last byte to be changed, rather than the first address past the last byte.  This avoids overflow when the last page of the address space is involved.

  include/exec/exec-all.h   |  2 +- accel/tcg/tb-maint.c      | 31 ++++++++++++++++--------------- accel/tcg/translate-all.c |  2 +- accel/tcg/user-exec.c     |  2 +-
  softmmu/physmem.c         |  2 +-

  5 files changed, 20 insertions(+), 19 deletions(-)
Quoted 7 lines
Quoted 5 lines
Quoted 1 line

I think the instruction is fnmadds. Needs more digging.

the invalid opcode is just a symptom of something bad happening.

C.

Indeed, it appears to be a copy/paste error within that commit. I've just posted a proposed fix for this: https://lore.kernel.org/qemu-devel/[email protected]/T/.

Looks good. I could boot macos 9.2.1 from an iso and 9.2.2 from disk.

Quoted 7 lines
Quoted 1 line

Nick: you may wish to try your series again with this fix applied to see if there are still problems with the CPUs used in the Mac machines.

Looks good too, with these :

[PATCH v2 1/4] target/ppc: Machine check on invalid real address [PATCH v2 2/4] target/ppc: Move common check in machine check [PATCH v2 3/4] target/ppc: Make checkstop actually stop the system

In the v2 series I removed the machine check for 970 btw so there might still be invalid memory access.

Thanks,
Nick