Re: [RFC PATCH glibc 11/12] hurd, htl: Add some x86_64-specific code

[Samuel Thibault <samuel.thibault@gnu.org>]

Sergey Bugaev, le dim. 12 févr. 2023 14:10:42 +0300, a ecrit:
> It seems that GCC expects TLS on x86_64 to be done relative to %fs, not %gs, so
> that's what I attempted to do in tls.h. The main thing missing there is the
> ability to actually set (and read) the %fs base address of a thread. It is my
> understanding (but note that I have no idea what I'm talking about) that on
> x86_64 the segment descriptors (as in GDT/LDT) are not used for this,

segmentation has somewhat disappeared in x86_64, yes.

> and instead the address can be set by writing to a MSR. Linux exposes
> the arch_prctl (ARCH_[GS]ET_[FG]S) syscall for this; so maybe GNU Mach
> could also have an explicit routine for this, perhaps like this:
> 
> routine i386_set_fgs_base (
> 	target_thread: thread_t;
> 	which: int;
> 	value: rpc_vm_address_t);

Indeed.

> We should not need a getter routine, because one can simply inspect the target
> thread's state (unless, again, I misunderstand things horribly).

For 16bit fs/gs values we could read them from userland yes. But for
fs/gs base, the FSGSBASE instruction is not available on all 64bit
processors. And ATM in THREAD_TCB we want to be able to get the base of
another thread.

> diff --git a/sysdeps/mach/hurd/x86_64/static-start.S b/sysdeps/mach/hurd/x86_64/static-start.S
> new file mode 100644
> index 00000000..982d3d52
> --- /dev/null
> +++ b/sysdeps/mach/hurd/x86_64/static-start.S
> @@ -0,0 +1,27 @@
> +/* Type of the TCB.  */
> +typedef struct
> +{
> +  void *tcb;			/* Points to this structure.  */
> +  dtv_t *dtv;			/* Vector of pointers to TLS data.  */
> +  thread_t self;		/* This thread's control port.  */
> +  int __glibc_padding1;
> +  int multiple_threads;
> +  int gscope_flag;
> +  uintptr_t sysinfo;
> +  uintptr_t stack_guard;
> +  uintptr_t pointer_guard;
> +  long __glibc_padding2[2];
> +  int private_futex;

? Isn't that rather feature_1 ?

> +  int __glibc_padding3;
> +  /* Reservation of some values for the TM ABI.  */
> +  void *__private_tm[4];
> +  /* GCC split stack support.  */
> +  void *__private_ss;
> +  /* The lowest address of shadow stack.  */
> +  unsigned long long int ssp_base;
> +
> +  /* Keep these fields last, so offsets of fields above can continue being
> +     compatible with the x86_64 NPTL version.  */
> +  mach_port_t reply_port;      /* This thread's reply port.  */
> +  struct hurd_sigstate *_hurd_sigstate;
> +
> +  /* Used by the exception handling implementation in the dynamic loader.  */
> +  struct rtld_catch *rtld_catch;
> +} tcbhead_t;
> +


> +/* GCC generates %fs:0x28 to access the stack guard.  */
> +_Static_assert (offsetof (tcbhead_t, stack_guard) == 0x28,
> +                "stack guard offset");
> +/* libgcc uses %fs:0x70 to access the split stack pointer.  */
> +_Static_assert (offsetof (tcbhead_t, __private_ss) == 0x70,
> +                "split stack pointer offset");

Indeed. Could you perhaps also add them to the i386 tls.h?

> +/* FIXME */
> +# define __LIBC_NO_TLS() 0

We'll want an efficient way to know whether we have configured TLS
indeed. At worse we can make it a global variable.

> +/* The TCB can have any size and the memory following the address the
> +   thread pointer points to is unspecified.  Allocate the TCB there.  */
> +# define TLS_TCB_AT_TP	1
> +# define TLS_DTV_AT_TP	0
> +

Also copy the comment above TCB_ALIGNMENT.

> +/* Install new dtv for current thread.  */
> +# define INSTALL_NEW_DTV(dtvp) THREAD_SETMEM (THREAD_SELF, dtv, dtvp)
> +/* Return the address of the dtv for the current thread.  */
> +# define THREAD_DTV() THREAD_GETMEM (THREAD_SELF, dtv)

While at it, try to make the i386 version use that too?

Samuel