See this conversation here ( https://lore.kernel.org/io-uring/CAHk-=wh0KoEZXPYMGkfkeVEerSCEF1AiCZSvz9TRrx=Kj74D+Q@mail.gmail.com/ ). The problem seems to be that gdb picks the wrong thread to select 32-bit/64-bit property for a process. AFAICT, this is done here in x86_linux_nat_target::read_description: ... /* GNU/Linux LWP ID's are process ID's. */ tid = inferior_ptid.lwp (); if (tid == 0) tid = inferior_ptid.pid (); /* Not a threaded program. */ ... It's not obvious to me from the discussion what is wrong with that.
Anyway, I've tested this patch and no regressions: ... diff --git a/gdb/x86-linux-nat.c b/gdb/x86-linux-nat.c index 85c7f0ddc94..bdf381f1430 100644 --- a/gdb/x86-linux-nat.c +++ b/gdb/x86-linux-nat.c @@ -114,9 +114,7 @@ x86_linux_nat_target::read_description () uint64_t xcr0_features_bits; /* GNU/Linux LWP ID's are process ID's. */ - tid = inferior_ptid.lwp (); - if (tid == 0) - tid = inferior_ptid.pid (); /* Not a threaded program. */ + tid = inferior_ptid.pid (); #ifdef __x86_64__ { ... so I wonder if this fixes the observed problems.
Posted patch: https://sourceware.org/pipermail/gdb-patches/2021-May/178596.html
The master branch has been updated by Tom de Vries <vries@sourceware.org>: https://sourceware.org/git/gitweb.cgi?p=binutils-gdb.git;h=fbf3c4b97907cb198168f58e7a22d497868e5926 commit fbf3c4b97907cb198168f58e7a22d497868e5926 Author: Tom de Vries <tdevries@suse.de> Date: Sun May 23 10:08:45 2021 +0200 [gdb/tdep] Use pid to choose process 64/32-bitness In a linux kernel mailing list discussion, it was mentioned that "gdb has this odd thing where it takes the 64-bit vs 32-bit data for the whole process from one thread, and picks the worst possible thread to do it (ie explicitly not even the main thread, ...)" [1]. The picking of the thread is done here in x86_linux_nat_target::read_description: ... /* GNU/Linux LWP ID's are process ID's. */ tid = inferior_ptid.lwp (); if (tid == 0) tid = inferior_ptid.pid (); /* Not a threaded program. */ ... To understand what this code does, let's investigate a scenario in which inferior_ptid.lwp () != inferior_ptid.pid (). Say we start exec jit-attach-pie, identified with pid x. The main thread starts another thread that sleeps, and then the main thread waits for the sleeping thread. So we have two threads, identified with LWP IDs x and x+1: ... PID LWP CMD x x ./jit-attach-pie x x+1 ./jit-attach-pie ... [ The thread with LWP x is known as the thread group leader. ] When attaching to this exec using the pid, gdb does a stop_all_threads which iterates over all the threads, first LWP x, and then LWP x+1. So the state we arrive with at x86_linux_nat_target::read_description is: ... (gdb) p inferior_ptid $1 = {m_pid = x, m_lwp = x+1, m_tid = 0} ... and consequently we probe 64/32-bitness from thread LWP x+1. [ Note that this is different from when gdb doesn't attach but instead launches the exec itself, in which case there's just one thread to begin with, and consequently the probed thread is LWP x. ] According to aforementioned remark, a better choice would have been the main thread, that is, LWP x. This patch implement that choice, by simply doing: ... tid = inferior_ptid.pid (); ... The fact that gdb makes a per-process permanent choice for 64/32-bitness is a problem in itself: each thread can be in either 64 or 32 bit mode, and change forth and back. That is a problem that this patch doesn't fix. Now finally: why does this matter in the context of the linux kernel discussion? The discussion was related to a patch that exposed io_uring threads to user-space. This made it possible that one of those threads would be picked out to select 64/32-bitness. Given that such threads are atypical user-space threads in the sense that they don't return to user-space and don't have a userspace register state, reading their registers returns garbage, and so it could f.i. occur that in a 64-bit process with all normal user-space threads in 64-bit mode, the probing would return 32-bit. It may be that this is worked-around on the kernel side by providing userspace register state in those threads such that current gdb is happy. Nevertheless, it seems prudent to fix this on the gdb size as well. Tested on x86_64-linux. [1] https://lore.kernel.org/io-uring/CAHk-=wh0KoEZXPYMGkfkeVEerSCEF1AiCZSvz9TRrx=Kj74D+Q@mail.gmail.com/ gdb/ChangeLog: 2021-05-23 Tom de Vries <tdevries@suse.de> PR tdep/27822 * target.h (struct target_ops): Mention target_thread_architecture in read_description comment. * x86-linux-nat.c (x86_linux_nat_target::read_description): Use pid to determine if process is 64-bit or 32-bit. * aarch64-linux-nat.c (aarch64_linux_nat_target::read_description): Same. * ppc-linux-nat.c (ppc_linux_nat_target::read_description): Same. * riscv-linux-nat.c (riscv_linux_nat_target::read_description): Same. * s390-linux-nat.c (s390_linux_nat_target::read_description): Same. * arm-linux-nat.c (arm_linux_nat_target::read_description): Same. Likewise, use pid to determine if kernel supports reading VFP registers.
Patch committed, marking resolved-fixed.