32-bit MEX file (Fortran) works correctly, the same 64-bit MEX file causes "Segmentation Fault"

Hello everyone

I am working with a Fortran compiled MEX file in a Linux based 32-bit environment, and it works properly (*.mexglx file).

I need to work with that same code in a 64-bit environment, so I just re-compile the Fortran code in that machine, getting a *.mexa64 file. Well, this file gets a "Segmentation Fault". The Fortran code is the same, and Matlab did not complain when compiling with 'mex'.

The original code is just ordinary code, with some variable inputs and some variable outputs. It does not "interact" with anything else.

Does anyone know where the problem might be?

Thank you in advance.

Carlos

[Kaustubha Govind]

It might be worthwhile to debug the file to find the exact line causing the issue: http://www.mathworks.com/help/techdoc/matlab_external/f32489.html#bq3wlv9-1

[James Tursa]

Please post your code, or a simplified version of your code that reproduces the problem. You likely have a variable type mismatch somewhere with either mwSize or with mwPointer.

[Carlos Casanueva]

@James Tursa

I'm not an expert on compilation, but if the code worked correctly for 32-bit, why should it be a coding issue? The original Fortran code is exactly the same. If you still think it could be some code-related issue I can post it here.

@Kaustubha Govind

I had some issues debugging, but finally I solved them, I will write a proper answer and not here in the comments.

[Jan]

@Carlos: It is really likly that there is a problem in the code. So please show us the relevant part of it.

Answer 1

[Friedrich]

You have to be carefull with simply recompile the Fortran code on 64bit. The FORTRAN pre-compiler interprets the mwPointer differnt. The FORTRAN preprocessor converts mwPointer to INTEGER*4 when building binary MEX-files on 32-bit platforms and to INTEGER*8 on 64-bit platforms. So make sure you are using mwSize instead of some hardcoded INTEGER*4.

Without any code its hard to tell whats causes this crash. Please post some code which crashes if you need further help with this.

Answer 2

[Carlos Casanueva]

I finally managed to debug the MEX file. First I add a breakpoint at line 1, then I start to debug.

The debugger stops at line 1 and already has the SIGSEV segmentation fault. Anyone knows why?

Then I continue debugging (no more breakpoints, it is the rest of the file) and the crash occurs.

    (gdb) break fsimwear_rev5
    Breakpoint 1 at 0x2aaaab50aaea: file fsimwear_rev5.f, line 1.
    (gdb) continue
    Continuing.

    Program received signal SIGSEGV, Segmentation fault.
    0x000000370407b59b in memcpy () from /lib64/libc.so.6

    Continuing.
    Unable to create crash dump log file

    ------------------------------------------------------------------------
           Segmentation violation detected at Fri Jul  1 14:00:08 2011
    ------------------------------------------------------------------------

    Configuration:
      MATLAB Version:   7.7.0.471 (R2008b)
      MATLAB License:   308809
      Operating System: Linux 2.6.18-128.7.1.el5.pdc5noib #1 SMP Thu Sep 16 19:17:28 CEST 2010 x86_64
      GNU C Library:    2.5 stable
      Window System:    No active display
      Current Visual:   None
      Processor ID:     x86 Family 6 Model 7 Stepping 6, GenuineIntel
      Virtual Machine:  Java is not enabled
      Default Encoding:  US-ASCII

    Fault Count: 1

    Register State:
      rax = 0000000041f46840   rbx = 0000000000000000
      rcx = 0000000000000800   rdx = 00000008000000d0
      rbp = 0000000041f46a10   rsi = 0000000108363798
      rdi = 0000000041f46840   rsp = 0000000041f46758
       r8 = 00000000088240b0    r9 = 0000000000000000
      r10 = 0000000000000000   r11 = 0000000000004000
      r12 = 0000000000000001   r13 = 0000000000000001
      r14 = 000000000897a300   r15 = 0000000041f47290
      rip = 000000370407b59b   flg = 0000000000010206

    Stack Trace:
      [0] libc.so.6:memcpy~(1, 0x088240b0, 0x127f00000000, 0x41f47298) + 347 bytes
      [1] libmex.so:mexRunMexFile(1, 0x41f47290, 0x100000018, 0x41f471d0) + 108 bytes
      [2] libmex.so:Mfh_mex::runMexFileWithSignalProtection(int, mxArray_tag**, int, mxArray_tag**)(0x41f47050, 0x41f46c50, 0x41f471d0, 0x08399ee0) + 137 bytes
      [3] libmex.so:Mfh_mex::dispatch_file(int, mxArray_tag**, int, mxArray_tag**)(0x41f471d0, 0x0897a300, 0x0897a300, 0x41f47dd0) + 245 bytes
      [4] libmwm_dispatcher.so:Mfh_file::dispatch_fh(int, mxArray_tag**, int, mxArray_tag**)(0x41f471c0, 1, 0x41f47100, 0x2b681483f380) + 297 bytes
      [5] libmwm_interpreter.so:inDispatchFromStack(int, char const*, int, int)(0x41f475ac, 0x2a100000001, 0x087fb38c, 0x100000000) + 1062 bytes
      [6] libmwm_interpreter.so:inDispatchCall(char const*, int, long, int, int*, int*)(0, 0x0821cbc0, 6, 0x41f475b0) + 165 bytes
      [7] libmwm_interpreter.so:inInterp(inDebugCheck, int, int, opcodes, inPcodeNest_tag volatile*, long*)(0x41f47a58, 0x0821cba0, 0x100000000, 0x100000000) + 2931 bytes
      [8] libmwm_interpreter.so:protected_inInterp(inDebugCheck, int, int, opcodes, inPcodeNest_tag*, long*)(0x0839a4a0, 0x07fa92d0, 0x0839a4a0, 0x6b9a72a2efb128) + 140 bytes
      [9] libmwm_interpreter.so:inInterPcodeSJ(inDebugCheck, int, int, opcodes, inPcodeNest_tag*, long*)(0, 0x41f47a20, 0x41f47a6c, 0) + 265 bytes
      [10] libmwm_interpreter.so:inExecuteMFunctionOrScript(Mfh_mp*, bool)(0, 0x3700000001, 0x41f47b40, 0x41f47c50) + 680 bytes
      [11] libmwm_interpreter.so:inRunMfile(int, mxArray_tag**, int, mxArray_tag**, Mfh_mp*, inWorkSpace_tag*)(0x41f48b70, 0x0839a4a0, 0x0839a4a0, 0x41f48db0) + 666 bytes
      [12] libmwm_dispatcher.so:Mfh_file::dispatch_fh(int, mxArray_tag**, int, mxArray_tag**)(0x41f47e70, 0, 0x41f48b70, 0x41f489b0) + 297 bytes
      [13] libmwm_interpreter.so:inEvalPcodeHeaderToWord(_memory_context*, int, mxArray_tag**, _pcodeheader*, Mfh_mp*, unsigned int)(0x41f47e90, 0x41f48030, 0x41f48bc0, 0) + 187 bytes
      [14] libmwm_interpreter.so:in_local_call_script_function(_memory_context*, _pcodeheader*, int, mxArray_tag**, unsigned int, bool)(0x41f48890, 0x41f4802c, 0x41f48a50, 0x41f48930) + 103 bytes
      [15] libmwm_interpreter.so:inEvalStringWithIsVarFcn(_memory_context*, char const*, EvalType, int, mxArray_tag**, inDebugCheck, _pcodeheader*, int*, bool (*)(void*, char const*), void*, bool, bool)(0, 0, 0x2b68157a1960, 0) + 2593 bytes
      [16] libmwm_interpreter.so:inEvalCmdWithLocalReturn(char const*, int*, bool, bool, bool (*)(void*, char const*))(0x41f48bd0, 0x08331bb0, 0x41f48d70, 0) + 149 bytes
      [17] libmwbridge.so:evalCommandWithLongjmpSafety(char const*)(0, 0, 0, 0) + 94 bytes
      [18] libmwbridge.so:mnParser(0x07be3fa0, 0x07be3f30, 0x07bdee60, 0x07bdee40) + 291 bytes
      [19] libmwmcr.so:mcrInstance::mnParser()~(0x41f490bf, 0x41f490b0, 0x7fffffaad0d8, 0x241f48eb0) + 82 bytes
      [20] MATLAB:mcrMain(int, char const**)(0x41f491c0, 0x41f49940, 0, 0x00a00000) + 493 bytes
      [21] libmwmcr.so:0x00002b681542cc9c(0, 0x41f49940, 0, 0)
      [22] libpthread.so.0:0x0000003704c06367(0x41f49940, 1, 0, 0xa99ba000008a0600)

    This error was detected while a MEX-file was running. If the MEX-file
    is not an official MathWorks function, please examine its source code
    for errors. Please consult the External Interfaces Guide for information
    on debugging MEX-files.

Answer 3

[Carlos Casanueva]

OK, thanks for your help. Here is the code with some mathematics removed:

          subroutine fsimwear_rev5(dz,m,n,a,b,ffn,cx,cy,cz,upsx,upsy,phi,
       &                         fric,V,k1,k2,k3,k4,kp_lim,kv_lim_low,
       &                         kv_lim_high,H,ga,elastic,psimp,fricwear,
       &                         D_wheel,dz_WLD,dA_PSH)

        implicit  none

        integer   m,n
        real*8    dz(m,n),a,b,ffn,cx,cy,cz,upsx,upsy,phi,fric,V,k1,k2,k3,
       &          k4,kp_lim,kv_lim_low,kv_lim_high,H,ga,D_wheel,
       &          dz_WLD(m,n),dA_PSH
        logical   elastic,psimp,fricwear

        integer   j,k
        real*8    x,y,dx,dxf,dy,t,xle,xx,yy
        real*8    lx,ly,lz,uxl,uyl,uzl,vslipx,vslipy,vslip,s,kalle
        real*8    f,ph_max,px,py,pz,pbound,para,q
        real*8    fn,fx,fy,amz
        real*8    fricwk,dfric,Tgamma_A(m,n),kfric,kfric_g
        logical   debug/.false./,forces/.false./

        real*8    pi/3.14159265358979/,ro/7850/   ! pi, steel density

c     Initiate variables

        if (debug) then
          print *, '   m=',   m,'      n=',n
          print *, '   a=',   a,'      b=',b
          print *, ' ffn=', ffn
          print *, '  cx=',  cx,'     cy=',  cy,'     cz=',  cz
          print *, 'upsx=',upsx,'   upsy=',upsy,'    phi=', phi
          print *, 'fric=',fric,'      V=',   V
          print *, '  k1=',  k1,'     k2=',  k2
          print *, '  k3=',  k3,'     k4=',  k4
        end if

        fn=ffn/fric
        uxl=upsx/lx
        uyl=upsy/ly
        uzl=phi/lz
        if (debug) print *, '  lx=',  lx,'     ly=',  ly,'     lz=', lz
        if (debug) print *, ' uxl=', uxl,'    uyl=', uyl,'    uzl=',uzl

        f=2*ffn/(pi*a*b)
        if (.not.psimp) then
          ph_max=1.5*fn/(pi*a*b)
        end if
        fx=0
        dy=2*b/n
        y=-b+dy/2

c     
      t=dx/V
      if (debug) write(*,*) (t)

c --- Outer loop 

        if (debug) then
          write(*,'(a)') ' '
          write(*,'(a)') '      x            y            q    '//
       &                '        fx           fy           fn'
        end if

        do 20 j=1,n
          yy=(y/b)**2
          xle=a*sqrt(1-yy)
  c r3    x=a-dx/2
          py=0
    
  c --- Inner loop

          if (debug) write(*,'(a)') ' '
          do 10 k=1,m

c --- 

            if (abs(x).lt.xle) then
        xx=(x/a)**2
            dxf=xle-x                      ! r5 distance to front edge
            if (dxf.lt.dx) then            ! r5
              px=px-(uxl-uzl*y)*dxf        ! r5
              py=py-(uyl+(uzl*(x+dxf)+uzl*x)/2)*dxf  ! r5 Heun
            else                           ! r5            
              px=px-(uxl-uzl*y)*dx
            end if                         ! r5
        para=1-yy-xx
        pbound=f*para
              if (debug) print *, 'px=',px,'   py=',py
              if (debug) print *, 'pbound=',pbound
              q=sqrt(px**2+py**2+1e-20)/pbound 
        
  c           Normal stress distribution     
              if (psimp) then             ! Kalker's simplified
          pz=pbound/fric
              else                        ! Hertz
                pz=ph_max*sqrt(para)
              end if

c --- 

              if (q.gt.1) then
      
  c             Sliding velocity
                if (elastic) then  
            vslipx=-V*lx*px*(1-1/q)/dx    ! S
            vslipy=-V*ly*py*(1-1/q)/dx    ! eq. after 3.38b in 
                else
                  vslipy=V*(upsy+phi*x)
                end if
          vslip=sqrt(vslipx**2+vslipy**2)

c                

c             
              px=px/q
              py=py/q

c --- 

                if(pz.gt.kp_lim) then             
                  kalle=k1
                else 
                  if(vslip.lt.kv_lim_high) then
                    if(vslip.lt.kv_lim_low) then
                      kalle=k2
                    end if
                    if(vslip.gt.kv_lim_low) then
                      kalle=k3
                    end if
                  else
                    kalle=k4
                  end if      
                end if

c             
              dz(j,k)=kalle*s*pz/H

c --- 

              else
                dz(j,k)=0
              end if  ! slip zone

    
c --- 

              if (forces) then
                fx=fx+px*dx*dy
                amz=amz+(py*x-px*y)*dx*dy
              end if                                                       ! r5
              if (forces.or.fricwear) then                                 ! r5                              ! r5
  c r5          fricwk=fricwk+((upsx-phi*y)*px+(upsy+phi*x)*py)*dx*dy
              end if

c --- 

              if (fricwear) then
  c r5          dfric=(dabs((upsx-phi*y)*px)+dabs((upsy+phi*x)*py))*dx*dy
                Tgamma_A(j,k)=dabs(dfric)/(dx*dy)
  c r5          fricwk=fricwk+dfric

                if (Tgamma_A(j,k).lt.10.4E6) then
                  kfric=Tgamma_A(j,k)*5.3E-15
                elseif (Tgamma_A(j,k).lt.77.2E6) then
                else
                  kfric=Tgamma_A(j,k)*61.9*1E-15-4724*1E-9
                end if

c             
              dz_WLD(j,k)=kfric*Tgamma_A(j,k)*dx/ro  
            end if

c --- 

            else
              dz(j,k)=0  
            end if  ! contact patch
            x=x-dx

            if (debug) then
              write(*,'(1p7e13.5)') x, y, q, px*dx*dy, py*dx*dy, 
       &                            (1-yy-xx)*f, dx*dy
            end if
            if (debug) print *,'dz(',j,k,')=',dz(j,k)

   10     continue  ! inner loop
          y=y+dy

   20   continue  ! outer loop

c --- 

        if (fricwear) then

c     
        if (fricwk.lt.100) then
        elseif (fricwk.lt.200) then
          kfric_g=25/(D_wheel*1000)
        else
          kfric_g=(1.19*fricwk-154)/(D_wheel*1000)
        end if

c     
        dA_PSH=kfric_g*1E-6*pi*D_wheel/1000
      end if

        return
        end

c =============================================================

        subroutine mexFunction(nlhs,plhs,nrhs,prhs)

        implicit none
        integer  plhs(*),prhs(*),alen,nlhs,nrhs

        parameter(alen=26)

        integer in_ptr,out_ptr,m,n,pflag,elflag,fricflag,
       &        mxGetPr,mxCreateDoubleMatrix,mxREAL
        real*8  dz(100,100),a,b,ffn,cx,cy,cz,upsx,upsy,phi,fric,V,
       &        k1,k2,k3,k4,kp_lim,kv_lim_low,kv_lim_high,H,ga,
       &        D_wheel,dz_WLD(100,100),dA_PSH,fsin(alen)
        logical psimp,elastic,fricwear

c     Extract input data from pointer prhs
      in_ptr=mxGetPr(prhs(1))
      call mxCopyPtrToReal8(in_ptr, fsin, alen)

        m=fsin(1)
        n=fsin(2)
        a=fsin(3)
        b=fsin(4)
        ffn=fsin(5)
        cx=fsin(6)
        cy=fsin(7)
        cz=fsin(8)
        upsx=fsin(9)
        upsy=fsin(10)
        phi=fsin(11)
        fric=fsin(12)
        V=fsin(13)
        k1=fsin(14)
        k2=fsin(15)
        k3=fsin(16)
        k4=fsin(17)
        kp_lim=fsin(18)
        kv_lim_low=fsin(19)
        kv_lim_high=fsin(20)
        H=fsin(21)
        ga=fsin(22)
        elflag=fsin(23)
        elastic=.true.
        if (elflag.eq.0) elastic=.false.
        pflag=fsin(24)
        psimp=.true.
        if (pflag.eq.0) psimp=.false.
        fricflag=fsin(25)
        fricwear=.true.
        if (fricflag.eq.0) fricwear=.false.
        D_wheel=fsin(26)

c     Call Fastsim
      call fsimwear_rev5(dz,m,n,a,b,ffn,cx,cy,cz,upsx,upsy,phi,fric,V,
     &                   k1,k2,k3, k4,kp_lim, kv_lim_low,kv_lim_high,H,
     &                   ga,elastic,psimp,fricwear,D_wheel,dz_WLD,
     &                   dA_PSH)

c     Assign output arrays

        plhs(1)=mxCreateDoubleMatrix(m,n,mxREAL)
        out_ptr=mxGetPr(plhs(1))
        call mxCopyReal8ToPtr(dz,out_ptr,m*n)

        plhs(2)=mxCreateDoubleMatrix(m,n,mxREAL)
        out_ptr=mxGetPr(plhs(2))
        call mxCopyReal8ToPtr(dz_WLD,out_ptr,m*n)

        plhs(3)=mxCreateDoubleMatrix(1,1,mxREAL)
        out_ptr=mxGetPr(plhs(3))
        call mxCopyReal8ToPtr(dA_PSH,out_ptr,1)

        return
        end

[Friedrich]

i think for example this can be dangerous:

Extract input data from pointer prhs

in_ptr=mxGetPr(prhs(1))

call mxCopyPtrToReal8(in_ptr, fsin, alen)

Since in_ptr is an integer and not an mwPointer as it should arcording the documentation. Is this integer 32bit or 64bit in the compiled mex file? because when its 32bit it will crash since you call mxCopyPtrToReal8 on it which expects a 64bit variable. please modify your code that it matched the datatypes mentioned in the documentation.

Answer 4

[Jan]

What exactly is "integer" in your 64bit compiler? nrhs and nlhs must be "integer*4" for example. If they are "integer*8", the mex should crash.

[James Tursa]

A Fortran default integer is typically integer*4, even for 64-bit systems. It is likely that this is causing a pointer problem in the code.

Answer 5

[James Tursa]

A Fortran default integer is almost certainly an integer*4, even for 64-bit compilers. Compiler vendors typically do this so that older code that depends on a default integer being an integer*4 have a better chance of compiling and running correctly when they use their code on a 64-bit machine. However, this makes a default integer type unreliable to use for holding a pointer, and using it that way will almost certainly crash MATLAB. This is the EXACT problem that I was expecting to see in the code and why one should always use mwPointer for the pointer types and mwSize for the size types. Your code needs to be completely scrubbed and everything made consistent with regards to the API interface. If you want to see examples of this you can take a look at my Fortran 95 - MATLAB API interface submission on the FEX:

http://www.mathworks.com/matlabcentral/fileexchange/25934-fortran-95-interface-to-matlab-api-with-extras

That being said, I haven't completely updated this submission for 64-bit yet. I have all of the coding done but haven't completed testing yet. Hopefully I will be able to upload it in the next week or so.

Another reason to take a look at my Fortran 95 submission is that it has routines to get Fortran pointers to the mxArray contents, so you don't need to use the clunky mxCopyReal8ToPointer and mxCopyPointerToReal8 routines at all.

Answer 6

[Carlos Casanueva]

Thanks for your help. As I said before, my fortran knowledge is really basic and this code is inherited from a former worker, so I was quite lost. I will try to re-read all the mex file documentation and with the tips you gave me I guess I will repair it.

CC

[Adrianna]

Did you get this fixed? If so, please share your fix.

I'm having a similar difficulty. I know my problem is with the mex wrapper as the fortran code migrated fine to 64bit fortran.

[James Tursa]

@Adrianna: See my comments on your other post.

[Carlos Casanueva]

No, sorry, I am still working on this, but as James Tursa said, check his comments: that is what I am using to try to fix it.

Answer 7

[Helen]

For more information, see Upgrade MEX-Files to Use 64-Bit API. In particular, see the section Update Fortran Source Code.