Re: [RFC PATCH] glibc: Perform rseq(2) registration at nptl init and thread creation

[Mathieu Desnoyers <mathieu dot desnoyers at efficios dot com>]

----- On Sep 20, 2018, at 4:20 PM, Joseph Myers joseph@codesourcery.com wrote:

> On Thu, 20 Sep 2018, Mathieu Desnoyers wrote:
> 
>> Are you saying glibc has an explicit check for the kernel version visible
>> from /proc before using specific features ? If so, how can this work with
>> the variety of feature backports we find in the distribution kernels out
>> there ?
> 
> See sysdeps/unix/sysv/linux/dl-sysdep.c and
> sysdeps/unix/sysv/linux/dl-osinfo.h.  As I said, Carlos has proposed
> removing that check.

For the system calls I implement and maintain, I typically ensure there is
a set of parameters that can be used when issuing the system call so it
can either succeed or fail with ENOSYS without having side-effects. It's
specifically meant to be used for feature discovery in a library
initialization phase. It's especially useful if the application needs to
keep state around related to the system call across its execution, e.g.
robust futexes.

> 
>> For too-old headers at compile time, one possibility is that we don't event
>> expose the __rseq_abi TLS symbol. OTOH, if we need to keep exposing it anyway
>> for ABI consistency purposes, then we'd leave its cpu_id field at the initial
>> value (-1). But that would require that we copy linux/rseq.h into the glibc
>> source tree.
> 
> The ABI needs to be independent of the kernel headers used.  I don't think
> you need to copy linux/rseq.h; all you should need is to e.g. define an
> array of suitable size and alignment with the relevant member initialized
> and a suitable explanatory comment.

In that case, I'm thinking declaring a minimal structure in glibc code may be
clearer than the array, e.g.:

[pthreadP.h]

enum libc_rseq_cpu_id_state {
  LIBC_RSEQ_CPU_ID_UNINITIALIZED = -1,
  LIBC_RSEQ_CPU_ID_REGISTRATION_FAILED = -2,
};

/* linux/rseq.h defines struct rseq as aligned on 32 bytes. The kernel ABI
   size is 20 bytes. For support of multiple rseq users within a process,
   user-space defines an extra 4 bytes field as a reference count, for a
   total of 24 bytes.  */
struct libc_rseq {
  /* kernel-userspace ABI. */
  uint32_t cpu_id_start;
  uint32_t cpu_id;
  uint64_t rseq_cs;
  uint32_t flags;
  /* user-space ABI. */
  uint32_t refcount;
} __attribute__((aligned(4 * sizeof(uint64_t))));

[pthread_create.h]

__thread volatile struct libc_rseq __rseq_abi = {
  .cpu_id = LIBC_RSEQ_CPU_ID_UNINITIALIZED,
};

Thanks,

Mathieu

-- 
Mathieu Desnoyers
EfficiOS Inc.
http://www.efficios.com