Re: [Qemu-devel] [PATCH v2 3/4] linux-user: Tidy and enforce reserved_va initialization

[Alistair Francis]

On Thu, Oct 5, 2017 at 7:36 AM, Richard Henderson
<address@hidden> wrote:
> From: Richard Henderson <address@hidden>
>
> We had a check using TARGET_VIRT_ADDR_SPACE_BITS to make sure
> that the allocation coming in from the command-line option was
> not too large, but that didn't include target-specific knowledge
> about other restrictions on user-space.
>
> Remove several target-specific hacks in linux-user/main.c.
>
> For MIPS and Nios, we can replace them with proper adjustments
> to the respective target's TARGET_VIRT_ADDR_SPACE_BITS definition.
>
> For ARM, we had no existing ifdef but I suspect that the current
> default value of 0xf7000000 was chosen with this in mind.  Define
> a workable value in linux-user/arm/, and also document why the
> special case is required.
>
> Signed-off-by: Richard Henderson <address@hidden>
> Reviewed-by: Peter Maydell <address@hidden>
> Message-Id: <address@hidden>

Hey,

I just had the arduous task of bisecting QEMU 2.10 and 2.11 to figure
out that somehow this patch causes the Yocto webkitgtk build for x86
to hang.

Any ideas?

My host is 64-bit x86 and I'm building for 32-bit x86. I'm going to
test now to see if I see the same hang building for 64-bit. I am not
sure yet where QEMU usermode is used, but apparently it is.

Thanks,
Alistair

> ---
>  linux-user/arm/target_cpu.h |  4 ++++
>  target/mips/mips-defs.h     |  6 +++++-
>  target/nios2/cpu.h          |  6 +++++-
>  linux-user/main.c           | 38 +++++++++++++++++++++++++-------------
>  4 files changed, 39 insertions(+), 15 deletions(-)
>
> diff --git a/linux-user/arm/target_cpu.h b/linux-user/arm/target_cpu.h
> index d888219150..c3eb4b243d 100644
> --- a/linux-user/arm/target_cpu.h
> +++ b/linux-user/arm/target_cpu.h
> @@ -19,6 +19,10 @@
>  #ifndef ARM_TARGET_CPU_H
>  #define ARM_TARGET_CPU_H
>
> +/* We need to be able to map the commpage.
> +   See validate_guest_space in linux-user/elfload.c.  */
> +#define MAX_RESERVED_VA  0xffff0000ul
> +
>  static inline void cpu_clone_regs(CPUARMState *env, target_ulong newsp)
>  {
>      if (newsp) {
> diff --git a/target/mips/mips-defs.h b/target/mips/mips-defs.h
> index 047554ee45..d239069975 100644
> --- a/target/mips/mips-defs.h
> +++ b/target/mips/mips-defs.h
> @@ -15,7 +15,11 @@
>  #else
>  #define TARGET_LONG_BITS 32
>  #define TARGET_PHYS_ADDR_SPACE_BITS 40
> -#define TARGET_VIRT_ADDR_SPACE_BITS 32
> +# ifdef CONFIG_USER_ONLY
> +#  define TARGET_VIRT_ADDR_SPACE_BITS 31
> +# else
> +#  define TARGET_VIRT_ADDR_SPACE_BITS 32
> +#endif
>  #endif
>
>  /* Masks used to mark instructions to indicate which ISA level they
> diff --git a/target/nios2/cpu.h b/target/nios2/cpu.h
> index 50d803a217..9119eee587 100644
> --- a/target/nios2/cpu.h
> +++ b/target/nios2/cpu.h
> @@ -226,7 +226,11 @@ qemu_irq *nios2_cpu_pic_init(Nios2CPU *cpu);
>  void nios2_check_interrupts(CPUNios2State *env);
>
>  #define TARGET_PHYS_ADDR_SPACE_BITS 32
> -#define TARGET_VIRT_ADDR_SPACE_BITS 32
> +#ifdef CONFIG_USER_ONLY
> +# define TARGET_VIRT_ADDR_SPACE_BITS 31
> +#else
> +# define TARGET_VIRT_ADDR_SPACE_BITS 32
> +#endif
>
>  #define cpu_init(cpu_model) cpu_generic_init(TYPE_NIOS2_CPU, cpu_model)
>
> diff --git a/linux-user/main.c b/linux-user/main.c
> index 829f974662..fd54d344bb 100644
> --- a/linux-user/main.c
> +++ b/linux-user/main.c
> @@ -60,23 +60,38 @@ do {                                                      
>               \
>      }                                                                   \
>  } while (0)
>
> -#if (TARGET_LONG_BITS == 32) && (HOST_LONG_BITS == 64)
>  /*
>   * When running 32-on-64 we should make sure we can fit all of the possible
>   * guest address space into a contiguous chunk of virtual host memory.
>   *
>   * This way we will never overlap with our own libraries or binaries or stack
>   * or anything else that QEMU maps.
> + *
> + * Many cpus reserve the high bit (or more than one for some 64-bit cpus)
> + * of the address for the kernel.  Some cpus rely on this and user space
> + * uses the high bit(s) for pointer tagging and the like.  For them, we
> + * must preserve the expected address space.
>   */
> -# if defined(TARGET_MIPS) || defined(TARGET_NIOS2)
> -/*
> - * MIPS only supports 31 bits of virtual address space for user space.
> - * Nios2 also only supports 31 bits.
> - */
> -unsigned long reserved_va = 0x77000000;
> +#ifndef MAX_RESERVED_VA
> +# if HOST_LONG_BITS > TARGET_VIRT_ADDR_SPACE_BITS
> +#  if TARGET_VIRT_ADDR_SPACE_BITS == 32 && \
> +      (TARGET_LONG_BITS == 32 || defined(TARGET_ABI32))
> +/* There are a number of places where we assign reserved_va to a variable
> +   of type abi_ulong and expect it to fit.  Avoid the last page.  */
> +#   define MAX_RESERVED_VA  (0xfffffffful & TARGET_PAGE_MASK)
> +#  else
> +#   define MAX_RESERVED_VA  (1ul << TARGET_VIRT_ADDR_SPACE_BITS)
> +#  endif
>  # else
> -unsigned long reserved_va = 0xf7000000;
> +#  define MAX_RESERVED_VA  0
>  # endif
> +#endif
> +
> +/* That said, reserving *too* much vm space via mmap can run into problems
> +   with rlimits, oom due to page table creation, etc.  We will still try it,
> +   if directed by the command-line option, but not by default.  */
> +#if HOST_LONG_BITS == 64 && TARGET_VIRT_ADDR_SPACE_BITS <= 32
> +unsigned long reserved_va = MAX_RESERVED_VA;
>  #else
>  unsigned long reserved_va;
>  #endif
> @@ -3978,11 +3993,8 @@ static void handle_arg_reserved_va(const char *arg)
>          unsigned long unshifted = reserved_va;
>          p++;
>          reserved_va <<= shift;
> -        if (((reserved_va >> shift) != unshifted)
> -#if HOST_LONG_BITS > TARGET_VIRT_ADDR_SPACE_BITS
> -            || (reserved_va > (1ul << TARGET_VIRT_ADDR_SPACE_BITS))
> -#endif
> -            ) {
> +        if (reserved_va >> shift != unshifted
> +            || (MAX_RESERVED_VA && reserved_va > MAX_RESERVED_VA)) {
>              fprintf(stderr, "Reserved virtual address too big\n");
>              exit(EXIT_FAILURE);
>          }
> --
> 2.13.6
>
>