Date: Tue, 2 Jun 2015 21:59:08 +0200 From: Sebastian Gottschall <s.gottschall@...wrt.com> To: Rich Felker <dalias@...c.org> Cc: musl@...ts.openwall.com Subject: Re: stable 1.1.9 & current GIT broken on mips Am 02.06.2015 um 21:11 schrieb Rich Felker: > On Tue, Jun 02, 2015 at 07:52:15PM +0200, Sebastian Gottschall wrote: >> Am 02.06.2015 um 19:19 schrieb Rich Felker: >>> On Tue, Jun 02, 2015 at 05:57:23PM +0200, Sebastian Gottschall wrote: >>>> Hello >>>> >>>> i tested today the current 1.1.9 (and later also current git so see >>>> if its the same behaviour) >>>> build on a mipsr2 big endian target (atheros ar7xxx) based on my >>>> dd-wrt firmware. >>>> i found out that mips seems to be broken on musl right now. the >>>> behaviour is >>>> that a call using execvp will not result in calling the desired >>>> application. >>>> on a second call and a following return call, the userspace will >>>> lock up with no way todo anything anymore. > What do you mean by "the userspace will lock up"? The process hangs? > Or the whole system? Is there any way to observe what's going on with > strace? no serial input anymore, looks like system hangs. but sometimes a kernel message still may race up. ([81.730000] random: nonblocking pool is initialized for instance) so its not completelly dead, but its not possible to interface from userspace anymore with serial console etc. > > And can you clarify what you mean about execvp? Are you saying the > first call to execvp returns with an error? I don't get what you mean > about "a second call and a following return call". On success execvp > does not return. i never checked for a return value. i just can say that the programm was never called. see the function i provided _eval("devinit"); //returns, but app never gets called _eval("sysinit"); //app never gets called, _eval hangs complete system on return call of _eval > > If you're trying to start a dynamic-linked application, it's likely > that something is going wrong in the dynamic linker after execvp > succeeds but before execution passes into the main program. You could > try running a program with global constructors and see if they run. > There's a jump at the end of dynamic linking which is not compatible > with calling into mips16 code, but as long as crt1.o is not mips16 > (and it shouldn't be; on mips it's still built from a .s file, and you > said you're not using mips16 in libc) this should not be a problem. > > Another possibility I should mention since this is DD-WRT is that > you've got an ancient kernel that's not compatible with TLS. As of > 1.1.0 musl deprecated running without a valid thread pointer, but > still worked if you happened not to invoke code that needs it. 1.1.9 > removed the last remnants of support for no-thread-pointer and now > aborts with SIGSEGV or SIGILL in the startup code if setting the > thread pointer fails, which will be the case on 2.4 kernels. its kernel 3.18 and kernel 3.10 which is mainly used. in the testcase 3.18 was used dont think that dd-wrt still used 2.4 (this is only the case for old wrt54g devices, but these devices are uclibc based) musl is the standard for all mips and armbe/le based devices in dd-wrt only y86,x64,powerpc and mips64 devices are still using uclibs since mips64 doesnt work with musl and power wasnt working months ago. may have changed in between. never tested so far. but target is to use musl on all cpu architectures in future, once its working on all. Sebastian > > Please let me know if you think this might be the case and if it's > blocking an important usage case for you. Since the rdhwr to get the > thread pointer is potentially very expensive on old hardware > (MIPS-I/II) I'm already considering making musl use a non-TLS global > for the thread pointer on mips in single-threaded programs, and this > same mechanism could make it possible to run (without threads, of > course) on kernels without rdhwr emulation or set_thread_area syscall. > But if possible it's really best to just move away from 2.4. It has > lots of bugs and cannot provide a fully-working environment for musl. > > Rich >
Powered by blists - more mailing lists
Confused about mailing lists and their use? Read about mailing lists on Wikipedia and check out these guidelines on proper formatting of your messages.