[Qemu-devel] [PATCH] accel/tcg/user-exec: Don't parse aarch64 insns to test for read vs write

[Qemu-devel] [PATCH] accel/tcg/user-exec: Don't parse aarch64 insns to test for read vs write

[Peter Maydell]

In cpu_signal_handler() for aarch64 hosts, currently we parse
the faulting instruction to see if it is a load or a store.
Since the 3.16 kernel (~2014), the kernel has provided us with
the syndrome register for a fault, which includes the WnR bit.
Use this instead if it is present, only falling back to
instruction parsing if not.

Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
---
Since I originally asked the kernel folks to add the ESR context
so we could use it in QEMU, I figured that it was about time
(five years later...) to write the code to make use of it.

I wanted to say "everybody surely has at least a 3.16
kernel for aarch64 machines" and delete the fallback code,
but it turns out that the gcc compile farm box has 3.13.0...
---
 accel/tcg/user-exec.c | 66 ++++++++++++++++++++++++++++++++++---------
 1 file changed, 52 insertions(+), 14 deletions(-)

diff --git a/accel/tcg/user-exec.c b/accel/tcg/user-exec.c
index 941295ea49b..66cc818e3f3 100644
--- a/accel/tcg/user-exec.c
+++ b/accel/tcg/user-exec.c
@@ -479,28 +479,66 @@ int cpu_signal_handler(int host_signum, void *pinfo,
 
 #elif defined(__aarch64__)
 
+#ifndef ESR_MAGIC
+/* Pre-3.16 kernel headers don't have these, so provide fallback definitions */
+#define ESR_MAGIC 0x45535201
+struct esr_context {
+    struct _aarch64_ctx head;
+    uint64_t esr;
+};
+#endif
+
+static inline struct _aarch64_ctx *first_ctx(ucontext_t *uc)
+{
+    return (struct _aarch64_ctx *)&uc->uc_mcontext.__reserved;
+}
+
+static inline struct _aarch64_ctx *next_ctx(struct _aarch64_ctx *hdr)
+{
+    return (struct _aarch64_ctx *)((char *)hdr + hdr->size);
+}
+
 int cpu_signal_handler(int host_signum, void *pinfo, void *puc)
 {
     siginfo_t *info = pinfo;
     ucontext_t *uc = puc;
     uintptr_t pc = uc->uc_mcontext.pc;
-    uint32_t insn = *(uint32_t *)pc;
     bool is_write;
+    struct _aarch64_ctx *hdr;
+    struct esr_context const *esrctx = NULL;
 
-    /* XXX: need kernel patch to get write flag faster.  */
-    is_write = (   (insn & 0xbfff0000) == 0x0c000000   /* C3.3.1 */
-                || (insn & 0xbfe00000) == 0x0c800000   /* C3.3.2 */
-                || (insn & 0xbfdf0000) == 0x0d000000   /* C3.3.3 */
-                || (insn & 0xbfc00000) == 0x0d800000   /* C3.3.4 */
-                || (insn & 0x3f400000) == 0x08000000   /* C3.3.6 */
-                || (insn & 0x3bc00000) == 0x39000000   /* C3.3.13 */
-                || (insn & 0x3fc00000) == 0x3d800000   /* ... 128bit */
-                /* Ingore bits 10, 11 & 21, controlling indexing.  */
-                || (insn & 0x3bc00000) == 0x38000000   /* C3.3.8-12 */
-                || (insn & 0x3fe00000) == 0x3c800000   /* ... 128bit */
-                /* Ignore bits 23 & 24, controlling indexing.  */
-                || (insn & 0x3a400000) == 0x28000000); /* C3.3.7,14-16 */
+    /* Find the esr_context, which has the WnR bit in it */
+    for (hdr = first_ctx(uc); hdr->magic; hdr = next_ctx(hdr)) {
+        if (hdr->magic == ESR_MAGIC) {
+            esrctx = (struct esr_context const *)hdr;
+            break;
+        }
+    }
 
+    if (esrctx) {
+        /* For data aborts ESR.EC is 0b10010x: then bit 6 is the WnR bit */
+        uint64_t esr = esrctx->esr;
+        is_write = extract32(esr, 27, 5) == 0x12 && extract32(esr, 6, 1) == 1;
+    } else {
+        /*
+         * Fall back to parsing instructions; will only be needed
+         * for really ancient (pre-3.16) kernels.
+         */
+        uint32_t insn = *(uint32_t *)pc;
+
+        is_write = ((insn & 0xbfff0000) == 0x0c000000   /* C3.3.1 */
+                    || (insn & 0xbfe00000) == 0x0c800000   /* C3.3.2 */
+                    || (insn & 0xbfdf0000) == 0x0d000000   /* C3.3.3 */
+                    || (insn & 0xbfc00000) == 0x0d800000   /* C3.3.4 */
+                    || (insn & 0x3f400000) == 0x08000000   /* C3.3.6 */
+                    || (insn & 0x3bc00000) == 0x39000000   /* C3.3.13 */
+                    || (insn & 0x3fc00000) == 0x3d800000   /* ... 128bit */
+                    /* Ignore bits 10, 11 & 21, controlling indexing.  */
+                    || (insn & 0x3bc00000) == 0x38000000   /* C3.3.8-12 */
+                    || (insn & 0x3fe00000) == 0x3c800000   /* ... 128bit */
+                    /* Ignore bits 23 & 24, controlling indexing.  */
+                    || (insn & 0x3a400000) == 0x28000000); /* C3.3.7,14-16 */
+    }
     return handle_cpu_signal(pc, info, is_write, &uc->uc_sigmask);
 }
 
-- 
2.20.1

Re: [Qemu-devel] [PATCH] accel/tcg/user-exec: Don't parse aarch64 insns to test for read vs write

[Peter Maydell]

Ping for review?

thanks
-- PMM

On Tue, 8 Jan 2019 at 18:00, Peter Maydell <peter.maydell@linaro.org> wrote:
>
> In cpu_signal_handler() for aarch64 hosts, currently we parse
> the faulting instruction to see if it is a load or a store.
> Since the 3.16 kernel (~2014), the kernel has provided us with
> the syndrome register for a fault, which includes the WnR bit.
> Use this instead if it is present, only falling back to
> instruction parsing if not.
>
> Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
> ---
> Since I originally asked the kernel folks to add the ESR context
> so we could use it in QEMU, I figured that it was about time
> (five years later...) to write the code to make use of it.
>
> I wanted to say "everybody surely has at least a 3.16
> kernel for aarch64 machines" and delete the fallback code,
> but it turns out that the gcc compile farm box has 3.13.0...
> ---
>  accel/tcg/user-exec.c | 66 ++++++++++++++++++++++++++++++++++---------
>  1 file changed, 52 insertions(+), 14 deletions(-)
>
> diff --git a/accel/tcg/user-exec.c b/accel/tcg/user-exec.c
> index 941295ea49b..66cc818e3f3 100644
> --- a/accel/tcg/user-exec.c
> +++ b/accel/tcg/user-exec.c
> @@ -479,28 +479,66 @@ int cpu_signal_handler(int host_signum, void *pinfo,
>
>  #elif defined(__aarch64__)
>
> +#ifndef ESR_MAGIC
> +/* Pre-3.16 kernel headers don't have these, so provide fallback definitions */
> +#define ESR_MAGIC 0x45535201
> +struct esr_context {
> +    struct _aarch64_ctx head;
> +    uint64_t esr;
> +};
> +#endif
> +
> +static inline struct _aarch64_ctx *first_ctx(ucontext_t *uc)
> +{
> +    return (struct _aarch64_ctx *)&uc->uc_mcontext.__reserved;
> +}
> +
> +static inline struct _aarch64_ctx *next_ctx(struct _aarch64_ctx *hdr)
> +{
> +    return (struct _aarch64_ctx *)((char *)hdr + hdr->size);
> +}
> +
>  int cpu_signal_handler(int host_signum, void *pinfo, void *puc)
>  {
>      siginfo_t *info = pinfo;
>      ucontext_t *uc = puc;
>      uintptr_t pc = uc->uc_mcontext.pc;
> -    uint32_t insn = *(uint32_t *)pc;
>      bool is_write;
> +    struct _aarch64_ctx *hdr;
> +    struct esr_context const *esrctx = NULL;
>
> -    /* XXX: need kernel patch to get write flag faster.  */
> -    is_write = (   (insn & 0xbfff0000) == 0x0c000000   /* C3.3.1 */
> -                || (insn & 0xbfe00000) == 0x0c800000   /* C3.3.2 */
> -                || (insn & 0xbfdf0000) == 0x0d000000   /* C3.3.3 */
> -                || (insn & 0xbfc00000) == 0x0d800000   /* C3.3.4 */
> -                || (insn & 0x3f400000) == 0x08000000   /* C3.3.6 */
> -                || (insn & 0x3bc00000) == 0x39000000   /* C3.3.13 */
> -                || (insn & 0x3fc00000) == 0x3d800000   /* ... 128bit */
> -                /* Ingore bits 10, 11 & 21, controlling indexing.  */
> -                || (insn & 0x3bc00000) == 0x38000000   /* C3.3.8-12 */
> -                || (insn & 0x3fe00000) == 0x3c800000   /* ... 128bit */
> -                /* Ignore bits 23 & 24, controlling indexing.  */
> -                || (insn & 0x3a400000) == 0x28000000); /* C3.3.7,14-16 */
> +    /* Find the esr_context, which has the WnR bit in it */
> +    for (hdr = first_ctx(uc); hdr->magic; hdr = next_ctx(hdr)) {
> +        if (hdr->magic == ESR_MAGIC) {
> +            esrctx = (struct esr_context const *)hdr;
> +            break;
> +        }
> +    }
>
> +    if (esrctx) {
> +        /* For data aborts ESR.EC is 0b10010x: then bit 6 is the WnR bit */
> +        uint64_t esr = esrctx->esr;
> +        is_write = extract32(esr, 27, 5) == 0x12 && extract32(esr, 6, 1) == 1;
> +    } else {
> +        /*
> +         * Fall back to parsing instructions; will only be needed
> +         * for really ancient (pre-3.16) kernels.
> +         */
> +        uint32_t insn = *(uint32_t *)pc;
> +
> +        is_write = ((insn & 0xbfff0000) == 0x0c000000   /* C3.3.1 */
> +                    || (insn & 0xbfe00000) == 0x0c800000   /* C3.3.2 */
> +                    || (insn & 0xbfdf0000) == 0x0d000000   /* C3.3.3 */
> +                    || (insn & 0xbfc00000) == 0x0d800000   /* C3.3.4 */
> +                    || (insn & 0x3f400000) == 0x08000000   /* C3.3.6 */
> +                    || (insn & 0x3bc00000) == 0x39000000   /* C3.3.13 */
> +                    || (insn & 0x3fc00000) == 0x3d800000   /* ... 128bit */
> +                    /* Ignore bits 10, 11 & 21, controlling indexing.  */
> +                    || (insn & 0x3bc00000) == 0x38000000   /* C3.3.8-12 */
> +                    || (insn & 0x3fe00000) == 0x3c800000   /* ... 128bit */
> +                    /* Ignore bits 23 & 24, controlling indexing.  */
> +                    || (insn & 0x3a400000) == 0x28000000); /* C3.3.7,14-16 */
> +    }
>      return handle_cpu_signal(pc, info, is_write, &uc->uc_sigmask);
>  }
>
> --
> 2.20.1
>

Re: [Qemu-devel] [PATCH] accel/tcg/user-exec: Don't parse aarch64 insns to test for read vs write

[Richard Henderson]

On 1/9/19 5:00 AM, Peter Maydell wrote:
> In cpu_signal_handler() for aarch64 hosts, currently we parse
> the faulting instruction to see if it is a load or a store.
> Since the 3.16 kernel (~2014), the kernel has provided us with
> the syndrome register for a fault, which includes the WnR bit.
> Use this instead if it is present, only falling back to
> instruction parsing if not.
> 
> Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
> ---
> Since I originally asked the kernel folks to add the ESR context
> so we could use it in QEMU, I figured that it was about time
> (five years later...) to write the code to make use of it.
> 
> I wanted to say "everybody surely has at least a 3.16
> kernel for aarch64 machines" and delete the fallback code,
> but it turns out that the gcc compile farm box has 3.13.0...
> ---
>  accel/tcg/user-exec.c | 66 ++++++++++++++++++++++++++++++++++---------
>  1 file changed, 52 insertions(+), 14 deletions(-)

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

I should note that we fail to generate esr_context from aarch64-linux-user.


r~