[RFC PATCH] linux-user/sparc: Implement v8plus signals

[RFC PATCH] linux-user/sparc: Implement v8plus signals

[Richard Henderson]

Sparc v8plus is a sparc64 running a 32-bit ABI.
The significant difference vs sparc32 is that all 64 bits of
the %g and %o registers, plus %xcc, are saved across interrupts,
context switches, and signals.

There's a special marker in the saved %psr value that's used to
indicate that %xcc and the high bits are present in the frame.

Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---

I have been unable to find an extant v8plus distribution with
which to test this beyond compilation.  Thus the RFC.  I know
debian used to have one, but they have moved to pure sparc64 now.

Thoughts?


r~

---
 target/sparc/cpu.h        |  2 -
 linux-user/sparc/signal.c | 96 ++++++++++++++++++++++++++++++++-------
 2 files changed, 80 insertions(+), 18 deletions(-)

diff --git a/target/sparc/cpu.h b/target/sparc/cpu.h
index ff8ae73002..d2c6e2e4ec 100644
--- a/target/sparc/cpu.h
+++ b/target/sparc/cpu.h
@@ -104,14 +104,12 @@ enum {
 #define PSR_CARRY_SHIFT 20
 #define PSR_CARRY (1 << PSR_CARRY_SHIFT)
 #define PSR_ICC   (PSR_NEG|PSR_ZERO|PSR_OVF|PSR_CARRY)
-#if !defined(TARGET_SPARC64)
 #define PSR_EF    (1<<12)
 #define PSR_PIL   0xf00
 #define PSR_S     (1<<7)
 #define PSR_PS    (1<<6)
 #define PSR_ET    (1<<5)
 #define PSR_CWP   0x1f
-#endif
 
 #define CC_SRC (env->cc_src)
 #define CC_SRC2 (env->cc_src2)
diff --git a/linux-user/sparc/signal.c b/linux-user/sparc/signal.c
index 0cc3db5570..d3d699545b 100644
--- a/linux-user/sparc/signal.c
+++ b/linux-user/sparc/signal.c
@@ -61,6 +61,13 @@ struct target_siginfo_fpu {
 #endif
 };
 
+struct target_siginfo_v8plus {
+#if defined(TARGET_SPARC64) && defined(TARGET_ABI32)
+    uint32_t u_regs_upper[16];
+    uint32_t asi;
+#endif
+};
+
 #ifdef TARGET_ARCH_HAS_SETUP_FRAME
 struct target_signal_frame {
     struct target_stackf ss;
@@ -69,7 +76,8 @@ struct target_signal_frame {
     abi_ulong fpu_save;
     uint32_t insns[2] QEMU_ALIGNED(8);
     abi_ulong extramask[TARGET_NSIG_WORDS - 1];
-    abi_ulong extra_size; /* Should be 0 */
+    abi_ulong extra_size; /* Should be sizeof(v8plus) */
+    struct target_siginfo_v8plus v8plus;
     abi_ulong rwin_save;
 };
 #endif
@@ -87,8 +95,9 @@ struct target_rt_signal_frame {
     abi_ulong fpu_save;
     uint32_t insns[2];
     target_stack_t stack;
-    abi_ulong extra_size; /* Should be 0 */
+    abi_ulong extra_size; /* Should be sizeof(v8plus) */
 #endif
+    struct target_siginfo_v8plus v8plus;
     abi_ulong rwin_save;
 };
 
@@ -121,7 +130,34 @@ static abi_ulong get_sigframe(struct target_sigaction *sa,
     return sp;
 }
 
-static void save_pt_regs(struct target_pt_regs *regs, CPUSPARCState *env)
+/* Fake PSR_IMPL | PSR_VER, meaning 64-bit cpu is present. */
+#define PSR_V8PLUS  0xff000000u
+/* If PSR_V8PLUS, this field contains %xcc. */
+#define PSR_XCC     0x000f0000u
+
+#if defined(TARGET_SPARC64) && defined(TARGET_ABI32)
+# define SAVE_REG(X, I)                                 \
+    do {                                                \
+        __put_user(X, &regs->u_regs[I]);                \
+        __put_user(X >> 32, &v8plus->u_regs_upper[I]);  \
+    } while (0)
+# define RESTORE_REG(X, I)                              \
+    do {                                                \
+        uint32_t l_, h_ = 0;                            \
+        __get_user(l_, &regs->u_regs[I]);               \
+        if ((psr & PSR_V8PLUS) == PSR_V8PLUS) {         \
+            __get_user(h_, &v8plus->u_regs_upper[I]);   \
+        }                                               \
+        X = deposit64(l_, 32, 32, h_);                  \
+    } while (0)
+#else
+# define SAVE_REG(X, I)     __put_user(X, &regs->u_regs[I])
+# define RESTORE_REG(X, I)  __get_user(X, &regs->u_regs[I])
+#endif
+
+static void save_pt_regs(struct target_pt_regs *regs,
+                         struct target_siginfo_v8plus *v8plus,
+                         CPUSPARCState *env)
 {
     int i;
 
@@ -130,7 +166,18 @@ static void save_pt_regs(struct target_pt_regs *regs, CPUSPARCState *env)
     /* TODO: magic should contain PT_REG_MAGIC + %tt. */
     __put_user(0, &regs->magic);
 #else
-    __put_user(cpu_get_psr(env), &regs->psr);
+    uint32_t psr;
+# ifdef TARGET_SPARC64
+    /* See linux/arch/sparc/include/uapi/asm/psrcompat.h, tstate_to_psr */
+    uint64_t tstate = sparc64_tstate(env);
+    psr = (tstate & 0x1f) /* TSTATE_CWP */
+        | ((tstate >> 12) & PSR_ICC)
+        | ((tstate >> 20) & PSR_XCC)
+        | PSR_S | PSR_V8PLUS;
+# else
+    psr = cpu_get_psr(env);
+# endif
+    __put_user(psr, &regs->psr);
 #endif
 
     __put_user(env->pc, &regs->pc);
@@ -138,14 +185,20 @@ static void save_pt_regs(struct target_pt_regs *regs, CPUSPARCState *env)
     __put_user(env->y, &regs->y);
 
     for (i = 0; i < 8; i++) {
-        __put_user(env->gregs[i], &regs->u_regs[i]);
+        SAVE_REG(env->gregs[i], i);
     }
     for (i = 0; i < 8; i++) {
-        __put_user(env->regwptr[WREG_O0 + i], &regs->u_regs[i + 8]);
+        SAVE_REG(env->regwptr[WREG_O0 + i], i + 8);
     }
+
+#if defined(TARGET_SPARC64) && defined(TARGET_ABI32)
+    __put_user(env->asi, &v8plus->asi);
+#endif
 }
 
-static void restore_pt_regs(struct target_pt_regs *regs, CPUSPARCState *env)
+static void restore_pt_regs(struct target_pt_regs *regs,
+                            struct target_siginfo_v8plus *v8plus,
+                            CPUSPARCState *env)
 {
     int i;
 
@@ -163,7 +216,15 @@ static void restore_pt_regs(struct target_pt_regs *regs, CPUSPARCState *env)
      */
     uint32_t psr;
     __get_user(psr, &regs->psr);
-    env->psr = (psr & PSR_ICC) | (env->psr & ~PSR_ICC);
+# ifdef TARGET_SPARC64
+    /* Unconditionally restore %icc and %xcc; conditionally restore %asi. */
+    cpu_put_ccr(env, ((psr & PSR_ICC) >> 20) | ((psr & PSR_XCC) >> 12));
+    if ((psr & PSR_V8PLUS) == PSR_V8PLUS) {
+        env->asi = v8plus->asi & 0xff;
+    }
+# else
+    cpu_put_psr(env, (psr & PSR_ICC) | (env->psr & ~PSR_ICC));
+# endif
 #endif
 
     /* Note that pc and npc are handled in the caller. */
@@ -171,13 +232,16 @@ static void restore_pt_regs(struct target_pt_regs *regs, CPUSPARCState *env)
     __get_user(env->y, &regs->y);
 
     for (i = 0; i < 8; i++) {
-        __get_user(env->gregs[i], &regs->u_regs[i]);
+        RESTORE_REG(env->gregs[i], i);
     }
     for (i = 0; i < 8; i++) {
-        __get_user(env->regwptr[WREG_O0 + i], &regs->u_regs[i + 8]);
+        RESTORE_REG(env->regwptr[WREG_O0 + i], i + 8);
     }
 }
 
+#undef SAVE_REG
+#undef RESTORE_REG
+
 static void save_reg_win(struct target_reg_window *win, CPUSPARCState *env)
 {
     int i;
@@ -259,8 +323,8 @@ void setup_frame(int sig, struct target_sigaction *ka,
     }
 
     /* 2. Save the current process state */
-    save_pt_regs(&sf->regs, env);
-    __put_user(0, &sf->extra_size);
+    save_pt_regs(&sf->regs, &sf->v8plus, env);
+    __put_user(sizeof(sf->v8plus), &sf->extra_size);
 
     save_fpu((struct target_siginfo_fpu *)(sf + 1), env);
     __put_user(sf_addr + sizeof(*sf), &sf->fpu_save);
@@ -321,7 +385,7 @@ void setup_rt_frame(int sig, struct target_sigaction *ka,
 
     /* 2. Save the current process state */
     save_reg_win(&sf->ss.win, env);
-    save_pt_regs(&sf->regs, env);
+    save_pt_regs(&sf->regs, &sf->v8plus, env);
 
     save_fpu((struct target_siginfo_fpu *)(sf + 1), env);
     __put_user(sf_addr + sizeof(*sf), &sf->fpu_save);
@@ -333,7 +397,7 @@ void setup_rt_frame(int sig, struct target_sigaction *ka,
     target_save_altstack(&sf->stack, env);
 
 #ifdef TARGET_ABI32
-    __put_user(0, &sf->extra_size);
+    __put_user(sizeof(sf->v8plus), &sf->extra_size);
 #endif
 
     /* 3. signal handler back-trampoline and parameters */
@@ -404,7 +468,7 @@ long do_sigreturn(CPUSPARCState *env)
     }
 
     /* 2. Restore the state */
-    restore_pt_regs(&sf->regs, env);
+    restore_pt_regs(&sf->regs, &sf->v8plus, env);
     env->pc = pc;
     env->npc = npc;
 
@@ -471,7 +535,7 @@ long do_rt_sigreturn(CPUSPARCState *env)
     }
 
     /* 2. Restore the state */
-    restore_pt_regs(&sf->regs, env);
+    restore_pt_regs(&sf->regs, &sf->v8plus, env);
 
     __get_user(ptr, &sf->fpu_save);
     if (ptr) {
-- 
2.25.1

Re: [RFC PATCH] linux-user/sparc: Implement v8plus signals

[Laurent Vivier]

Le 26/05/2021 à 03:13, Richard Henderson a écrit :
> Sparc v8plus is a sparc64 running a 32-bit ABI.
> The significant difference vs sparc32 is that all 64 bits of
> the %g and %o registers, plus %xcc, are saved across interrupts,
> context switches, and signals.
> 
> There's a special marker in the saved %psr value that's used to
> indicate that %xcc and the high bits are present in the frame.
> 
> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
> ---
> 
> I have been unable to find an extant v8plus distribution with
> which to test this beyond compilation.  Thus the RFC.  I know
> debian used to have one, but they have moved to pure sparc64 now.
> 
> Thoughts?

In my test, I use debian wheezy, and it seems to be sparc32plus v8+:

$ file chroot/sparc/wheezy/bin/ls
chroot/sparc/wheezy/bin/ls: ELF 32-bit MSB executable, SPARC32PLUS, V8+ Required, total store
ordering, version 1 (SYSV), dynamically linked, interpreter /lib/ld-linux.so.2, for GNU/Linux
2.6.26, BuildID[sha1]=3fdfb5b9f829824f6110b7a2d91efd9947614263, stripped

I get wheezy from http://archive.debian.org/debian

If you want I can make a try with your patch.

Thanks,
Laurent

Re: [RFC PATCH] linux-user/sparc: Implement v8plus signals

[Richard Henderson]

On 6/14/21 11:48 PM, Laurent Vivier wrote:
> Le 26/05/2021 à 03:13, Richard Henderson a écrit :
>> Sparc v8plus is a sparc64 running a 32-bit ABI.
>> The significant difference vs sparc32 is that all 64 bits of
>> the %g and %o registers, plus %xcc, are saved across interrupts,
>> context switches, and signals.
>>
>> There's a special marker in the saved %psr value that's used to
>> indicate that %xcc and the high bits are present in the frame.
>>
>> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
>> ---
>>
>> I have been unable to find an extant v8plus distribution with
>> which to test this beyond compilation.  Thus the RFC.  I know
>> debian used to have one, but they have moved to pure sparc64 now.
>>
>> Thoughts?
> 
> In my test, I use debian wheezy, and it seems to be sparc32plus v8+:
> 
> $ file chroot/sparc/wheezy/bin/ls
> chroot/sparc/wheezy/bin/ls: ELF 32-bit MSB executable, SPARC32PLUS, V8+ Required, total store
> ordering, version 1 (SYSV), dynamically linked, interpreter /lib/ld-linux.so.2, for GNU/Linux
> 2.6.26, BuildID[sha1]=3fdfb5b9f829824f6110b7a2d91efd9947614263, stripped
> 
> I get wheezy from http://archive.debian.org/debian
> 
> If you want I can make a try with your patch.

Yes, please.


r~

Re: [RFC PATCH] linux-user/sparc: Implement v8plus signals

[Laurent Vivier]

On 15/06/2021 16:44, Richard Henderson wrote:
> On 6/14/21 11:48 PM, Laurent Vivier wrote:
>> Le 26/05/2021 à 03:13, Richard Henderson a écrit :
>>> Sparc v8plus is a sparc64 running a 32-bit ABI.
>>> The significant difference vs sparc32 is that all 64 bits of
>>> the %g and %o registers, plus %xcc, are saved across interrupts,
>>> context switches, and signals.
>>>
>>> There's a special marker in the saved %psr value that's used to
>>> indicate that %xcc and the high bits are present in the frame.
>>>
>>> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
>>> ---
>>>
>>> I have been unable to find an extant v8plus distribution with
>>> which to test this beyond compilation.  Thus the RFC.  I know
>>> debian used to have one, but they have moved to pure sparc64 now.
>>>
>>> Thoughts?
>>
>> In my test, I use debian wheezy, and it seems to be sparc32plus v8+:
>>
>> $ file chroot/sparc/wheezy/bin/ls
>> chroot/sparc/wheezy/bin/ls: ELF 32-bit MSB executable, SPARC32PLUS, V8+ Required, total
>> store
>> ordering, version 1 (SYSV), dynamically linked, interpreter /lib/ld-linux.so.2, for
>> GNU/Linux
>> 2.6.26, BuildID[sha1]=3fdfb5b9f829824f6110b7a2d91efd9947614263, stripped
>>
>> I get wheezy from http://archive.debian.org/debian
>>
>> If you want I can make a try with your patch.
> 
> Yes, please.
> 

This breaks two tests in LTP: sigaction01 and rt_sigaction01 (they hang).

Tested with debian wheezy and LTP 20200930.

Note: even with master, debootstrap is broken and needs manual hacks to have a working chroot.

Thanks,
Laurent