• Florian Fainelli's avatar
    MIPS: Fix code generation for non-DSP capable CPUs · 63c2b681
    Florian Fainelli authored
    Commit 32a7ede6 (MIPS: dsp: Add assembler support for DSP ASEs) has
    enabled the use of DSP ASE specific instructions such as rddsp and wrdsp
    under the idea that all code path that will make use of these two
    instructions are properly checking for cpu_has_dsp to ensure that the
    particular CPU we are running on *actually* supports DSP ASE.
    
    This commit actually causes the following oops on QEMU Malta emulating a
    MIPS 24Kc without the DSP ASE implemented:
    
    [    7.960000] Reserved instruction in kernel
    [    7.960000] Cpu 0
    [    7.960000] $ 0   : 00000000 00000000 00000014 00000005
    [    7.960000] $ 4   : 8fc2de48 00000001 00000000 8f59ddb0
    [    7.960000] $ 8   : 8f5ceec4 00000018 00000c00 00800000
    [    7.960000] $12   : 00000100 00000200 00000000 00457b84
    [    7.960000] $16   : 00000000 8fc2ba78 8f4ec980 00000001
    [    7.960000] $20   : 80418f90 00000000 00000000 000002dd
    [    7.960000] $24   : 0000009c 7730d7b8
    [    7.960000] $28   : 8f59c000 8f59dd38 00000001 80104248
    [    7.960000] Hi    : 0000001d
    [    7.960000] Lo    : 0000000b
    [    7.960000] epc   : 801041ec thread_saved_pc+0x2c/0x38
    [    7.960000]     Not tainted
    [    7.960000] ra    : 80104248 get_wchan+0x48/0xac
    [    7.960000] Status: 1000b703    KERNEL EXL IE
    [    7.960000] Cause : 10800028
    [    7.960000] PrId  : 00019300 (MIPS 24Kc)
    [    7.960000] Modules linked in:
    [    7.960000] Process killall (pid: 1574, threadinfo=8f59c000,
    task=8fd14558, tls=773aa440)
    [    7.960000] Stack : 8fc2ba78 8012b008 0000000c 0000001d 00000000
    00000000 8f58a380
                      8f58a380 8fc2ba78 80202668 8f59de78 8f468600 8f59de28
    801b2a3c 8f59df00 8f98ba20 74696e69
                      8f468600 8f59de28 801b7308 0081c007 00000000 00000000
    00000000 00000000 00000000 00000000
                      00000000 8fc2bbb4 00000001 0000001d 0000000b 77f038cc
    7fe80648 ffffffff ffffffff 00000000
                      00000001 0016e000 00000000 ...
    [    7.960000] Call Trace:
    [    7.960000] [<801041ec>] thread_saved_pc+0x2c/0x38
    [    7.960000] [<80104248>] get_wchan+0x48/0xac
    
    The disassembly of thread_saved_pc points to the following:
    000006d0 <thread_saved_pc>:
     6d0:   8c820208        lw      v0,520(a0)
     6d4:   3c030000        lui     v1,0x0
     6d8:   24630000        addiu   v1,v1,0
     6dc:   10430008        beq     v0,v1,700 <thread_saved_pc+0x30>
     6e0:   00000000        nop
     6e4:   3c020000        lui     v0,0x0
     6e8:   8c43000c        lw      v1,12(v0)
     6ec:   04620004        bltzl   v1,700 <thread_saved_pc+0x30>
     6f0:   00001021        move    v0,zero
     6f4:   8c840200        lw      a0,512(a0)
     6f8:   00031080        sll     v0,v1,0x2
     6fc:   7c44100a        lwx     v0,a0(v0)   <------------
     700:   03e00008        jr      ra
     704:   00000000        nop
    
    If we specifically disable -mdsp/-mdspr2 for arch/mips/kernel/process.o,
    we get the following (non-crashing) assembly:
    
    00000708 <thread_saved_pc>:
     708:   8c820208        lw      v0,520(a0)
     70c:   3c030000        lui     v1,0x0
     710:   24630000        addiu   v1,v1,0
     714:   10430009        beq     v0,v1,73c <thread_saved_pc+0x34>
     718:   00000000        nop
     71c:   3c020000        lui     v0,0x0
     720:   8c42000c        lw      v0,12(v0)
     724:   04420005        bltzl   v0,73c <thread_saved_pc+0x34>
     728:   00001021        move    v0,zero
     72c:   8c830200        lw      v1,512(a0)
     730:   00021080        sll     v0,v0,0x2
     734:   00431021        addu    v0,v0,v1
     738:   8c420000        lw      v0,0(v0)
     73c:   03e00008        jr      ra
     740:   00000000        nop
    
    The specific line that leads a different assembly being produced is:
    
    unsigned long thread_saved_pc(struct task_struct *tsk)
    ...
    	return ((unsigned long *)t->reg29)[schedule_mfi.pc_offset]; <---
    
    The problem here is that the compiler was given the right to use DSP
    instructions with the -mdsp / -mdspr2 command-line switches and
    performed some optimization for us and used DSP ASE instructions where
    we are not checking that the running CPU actually supports DSP ASE.
    
    This patch fixes the issue by partially reverting commit 32a7ede6 for
    arch/mips/kernel/Makefile in order to remove the -mdsp / -mdspr2
    compiler command-line switches such that we are now guaranteed that the
    compiler will not optimize using DSP ASE reserved instructions. We also
    need to fixup the rddsp/wrdsp and m{t,h}{hi,lo}{0,1,2,3} macros in
    arch/mips/include/asm/mipsregs.h to tell the assembler that we are going
    to explicitely use DSP ASE reserved instructions. The comment in
    arch/mips/kernel/Makefile is also updated to reflect that.
    Signed-off-by: default avatarFlorian Fainelli <florian@openwrt.org>
    Acked-by: default avatarSteven J. Hill <Steven.Hill@imgtec.com>
    Cc: linux-mips@linux-mips.org
    Cc: blogic@openwrt.org
    Signed-off-by: default avatarRalf Baechle <ralf@linux-mips.org>
    63c2b681
Makefile 3.92 KB