alpha fp-emu vs module refcounting
arch/alpha/math-emu/math.c still uses MOD_{INC,DEC}_USE_COUNT to protect from unloading which is
a) unsafe
b) in the way of nuking them as soon as 2.7 is out
any chance we could either do a try_module_get on a struct module *fp_emul_module in alpha core code before or completely disallow module unloading for it (by removing the cleanup_module() callback)?
Quoted 3 lines · Christoph Hellwig
any chance we could either do a
try_module_geton a struct module*fp_emul_modulein alpha core code before or completely disallow module unloading for it (by removing thecleanup_module()callback)?
How about this (also fixes modular build of math-emu)?
Ivan.
arch/alpha/math-emu/math.c +8 -6
@@ -51,7 +51,9 @@ extern void alpha_write_fp_reg_s (unsign#ifdef MODULE +MODULE_AUTHOR ("Richard Henderson <[email protected]>" );MODULE_DESCRIPTION ("FP Software completion module" ); +MODULE_LICENSE ("GPL" );extern long (*alpha_fp_emul_imprecise)(struct pt_regs *,unsigned long );extern long (*alpha_fp_emul) (unsigned long pc);
@@ -106,7 +108,7 @@ alpha_fp_emul (unsigned long pc)__u32 insn;long si_code; -MOD_INC_USE_COUNT ; +BUG_ON (!try_module_get (THIS_MODULE ));get_user (insn, (__u32 *)pc); fc = (insn >>0 ) &0x1f ;/* destination register */
@@ -320,7 +322,7 @@ done:if (_fex &IEEE_TRAP_ENABLE_INV ) si_code =FPE_FLTINV ; } -MOD_DEC_USE_COUNT ; +module_put (THIS_MODULE );return si_code; }
@@ -328,13 +330,13 @@ done:requires that the result *always* be written... so we do the write immediately after the operations above. */ -MOD_DEC_USE_COUNT ; +module_put (THIS_MODULE );return 0 ; bad_insn:printk (KERN_ERR "alpha_fp_emul: Invalid FP insn %#x at %#lx\n" , insn, pc); -MOD_DEC_USE_COUNT ; +module_put (THIS_MODULE );return -1 ; }
@@ -344,7 +346,7 @@ alpha_fp_emul_imprecise (struct pt_regsunsigned long trigger_pc = regs->pc -4 ;unsigned long insn, opcode, rc, si_code =0 ; -MOD_INC_USE_COUNT ; +BUG_ON (!try_module_get (THIS_MODULE ));/* * Turn off the bits corresponding to registers that are the
@@ -403,6 +405,6 @@ alpha_fp_emul_imprecise (struct pt_regs } egress: -MOD_DEC_USE_COUNT ; +module_put (THIS_MODULE );return si_code; }
arch/alpha/math-emu/Makefile +3 -1
@@ -4,4 +4,6 @@ EXTRA_CFLAGS := -w -obj-$(CONFIG_MATHEMU) += math.o qrnnd.o +obj-$(CONFIG_MATHEMU) += math-emu.o + +math-emu-objs := math.o qrnnd.o
Review
Quoted 2 lines · Ivan Kokshaysky
(quoted diff) +1
-1- MOD_INC_USE_COUNT; + BUG_ON(!try_module_get(THIS_MODULE));
This is broken. If you know you already have a reference somewhere - and I can't find how that would work - you'd have to use __module_get, but if it's not you can get a failure from try_module_get legally and you should do the try_module_get outside the module, before it's code is exectuted.
Quoted 9 lines · Christoph Hellwig
Review
Quoted 2 lines · Ivan Kokshaysky
(quoted diff) +1
-1- MOD_INC_USE_COUNT; + BUG_ON(!try_module_get(THIS_MODULE));This is broken. If you know you already have a reference somewhere - and I can't find how that would work - you'd have to use
__module_get, but if it's not you can get a failure fromtry_module_getlegally and you should do thetry_module_getoutside the module, before it's code is exectuted.
Ok, I realize - this seems to be confusing. I'll try to clarify:
- First of all, mere mortals are _not_ allowed to compile mandatory alpha IEEE fp emulation code as a module. Which is documented in arch/alpha/Kconfig.
- Roughly speaking, the fp emu _module_ code intercepts the fp traps, so races vs module loading/unloading are fundamentally unavoidable. These refcounting attempts just narrow the window.
And no, try_module_get should never fail here.
Alternatively, we could just drop _all_ module related stuff from alpha/math-emu...
Ivan.
Quoted 2 lines · Christoph Hellwig
either that or just marking it unloadable by removing the
cleanup_modulehandler sound like the simplest solution I guess.
Or leave it as it is for now - 'CONFIG_MATHEMU=m' won't compile. ;-)
Ivan.
Quoted 8 lines · Ivan Kokshaysky
Ok, I realize - this seems to be confusing. I'll try to clarify:
- First of all, mere mortals are _not_ allowed to compile mandatory alpha IEEE fp emulation code as a module. Which is documented inarch/alpha/Kconfig.
- Roughly speaking, the fp emu _module_ code intercepts the fp traps, so races vs module loading/unloading are fundamentally unavoidable. These refcounting attempts just narrow the window.
Quoted 4 lines
And no,
try_module_getshould never fail here.Alternatively, we could just drop _all_ module related stuff from alpha/math-emu...
either that or just marking it unloadable by removing the cleanup_module handler sound like the simplest solution I guess.
Quoted 5 lines · Ivan Kokshaysky
Quoted 2 lines · Christoph Hellwig
either that or just marking it unloadable by removing the
cleanup_modulehandler sound like the simplest solution I guess.Or leave it as it is for now - '
CONFIG_MATHEMU=m' won't compile. ;-)
Well, still false positives in grep. What about this patch to simply remove any traces of CONFIG_MATHEMU and modular math emulation?
arch/alpha/Kconfig +0 -8
@@ -625,14 +625,6 @@ Say Y here if you are developing drivers or trying to debug and identify kernel problems. -config MATHEMU - tristate "Kernel FP software completion" if DEBUG_KERNEL - default y if !DEBUG_KERNEL - help - This option is required for IEEE compliant floating point arithmetic - on the Alpha. The only time you would ever not say Y is to say M in - order to debug the code. Say Y unless you know what you are doing. - config DEBUG_SLAB bool "Debug memory allocations" depends on DEBUG_KERNEL
===== arch/alpha/kernel/alpha_ksyms.c 1.37 vs edited =====
arch/alpha/kernel/alpha_ksyms.c +0 -7
@@ -169,13 +169,6 @@EXPORT_SYMBOL (csum_partial_copy_from_user);EXPORT_SYMBOL (csum_ipv6_magic); -#ifdef CONFIG_MATHEMU_MODULE -extern long (*alpha_fp_emul_imprecise)(struct pt_regs *,unsigned long ); -extern long (*alpha_fp_emul) (unsigned long pc); -EXPORT_SYMBOL (alpha_fp_emul_imprecise); -EXPORT_SYMBOL (alpha_fp_emul); -#endif -#ifdef CONFIG_ALPHA_BROKEN_IRQ_MASK EXPORT_SYMBOL (__min_ipl);#endif
===== arch/alpha/kernel/traps.c 1.29 vs edited =====
arch/alpha/kernel/traps.c +0 -8
@@ -191,16 +191,8 @@do_exit (SIGSEGV ); } -#ifndef CONFIG_MATHEMU -static long dummy_emul (void ) {return 0 ; } -long (*alpha_fp_emul_imprecise)(struct pt_regs *regs,unsigned long writemask) - = (void *)dummy_emul; -long (*alpha_fp_emul) (unsigned long pc) - = (void *)dummy_emul; -#else long alpha_fp_emul_imprecise (struct pt_regs *regs,unsigned long writemask);long alpha_fp_emul (unsigned long pc); -#endif asmlinkage void do_entArith(unsigned long summary, unsigned long write_mask,
===== arch/alpha/math-emu/Makefile 1.9 vs edited =====
arch/alpha/math-emu/Makefile +1 -1
@@ -4,4 +4,4 @@ EXTRA_CFLAGS := -w -obj-$(CONFIG_MATHEMU) += math.o qrnnd.o +obj-y += math.o qrnnd.o
===== arch/alpha/math-emu/math.c 1.5 vs edited =====
arch/alpha/math-emu/math.c +0 -45
85 lines not on this page · raw message
Quoted 2 lines · Christoph Hellwig
Well, still false positives in grep. What about this patch to simply remove any traces of
CONFIG_MATHEMUand modular math emulation?
Personally, I'm fine with it.
Yet another simple solution would be just to remove refcounting and only allow modular build of math-emu when CONFIG_SMP=n, which is safe vs module unload and still fine for debugging.
Richard?
Ivan.
Quoted 9 lines · Ivan Kokshaysky
Quoted 2 lines · Christoph Hellwig
Well, still false positives in grep. What about this patch to simply remove any traces of
CONFIG_MATHEMUand modular math emulation?Personally, I'm fine with it.
Yet another simple solution would be just to remove refcounting and only allow modular build of math-emu when
CONFIG_SMP=n, which is safe vs module unload and still fine for debugging.
UP is not safe vs module unloading per se. Especially not with CONFIG_PREEMPT. I'd say just apply the patch and if someone badly needs modular mathemu for debugging he/she can apply some local hack.
Quoted 5 lines · Ivan Kokshaysky
Yet another simple solution would be just to remove refcounting and only allow modular build of math-emu when
CONFIG_SMP=n, which is safe vs module unload and still fine for debugging.Richard?
Seems reasonable. We've already got it under the debugging config section, right?
r~