Messages in this thread | | | Date | Sun, 10 May 2015 16:52:03 +0300 | From | Alexey Dobriyan <> | Subject | Re: [PATCH 02/12] Add parse_integer() (replacement for simple_strto*()) |
| |
On Fri, May 08, 2015 at 01:46:46PM -0700, Andrew Morton wrote: > On Fri, 8 May 2015 21:30:29 +0300 Alexey Dobriyan <adobriyan@gmail.com> wrote: > > > kstrto*() and kstrto*_from_user() family of functions were added > > to help with parsing one integer written as string to proc/sysfs/debugfs > > files. But they have a limitation: string passed must end with \0 or \n\0. > > There are enough places where kstrto*() functions can't be used because of > > this limitation. Trivial example: major:minor "%u:%u". > > > > Currently the only way to parse everything is simple_strto*() functions. > > But they are suboptimal: > > * they do not detect overflow (can be fixed, but no one bothered since ~0.99.11), > > * there are only 4 of them -- long and "long long" versions, > > This leads to silent truncation in the most simple case: > > > > val = strtoul(s, NULL, 0); > > > > * half of the people think that "char **endp" argument is necessary and > > add unnecessary variable. > > > > OpenBSD people, fed up with how complex correct integer parsing is, added > > strtonum(3) to fixup for deficiencies of libc-style integer parsing: > > http://www.openbsd.org/cgi-bin/man.cgi/OpenBSD-current/man3/strtonum.3?query=strtonum&arch=i386 > > > > It'd be OK to copy that but it relies on "errno" and fixed strings as > > error reporting channel which I think is not OK for kernel. > > strtonum() also doesn't report number of characted consumed. > > > > What to do? > > > > Enter parse_integer(). > > fs/binfmt_misc.c | 12 > fs/cachefiles/daemon.c | 84 ++-- > fs/dcache.c | 2 > fs/ext2/super.c | 6 > fs/ext3/super.c | 7 > fs/ext4/super.c | 15 > fs/inode.c | 2 > fs/libfs.c | 26 - > fs/namespace.c | 4 > fs/ocfs2/cluster/heartbeat.c | 54 +- > fs/ocfs2/cluster/nodemanager.c | 50 +- > fs/ocfs2/stack_user.c | 52 +- > include/linux/kernel.h | 129 ------- > include/linux/parse-integer.h | 188 ++++++++++ > lib/Kconfig.debug | 3 > lib/Makefile | 2 > lib/cmdline.c | 42 +- > lib/kstrtox.c | 254 ------------- > lib/kstrtox.h | 1 > lib/parse-integer.c | 222 ++++++++++++ > lib/parser.c | 33 - > lib/swiotlb.c | 2 > lib/test-kstrtox.c | 6 > lib/test-parse-integer.c | 563 +++++++++++++++++++++++++++++++ > lib/vsprintf.c | 81 ++-- > mm/memcontrol.c | 19 - > mm/memtest.c | 2 > mm/page_alloc.c | 2 > mm/shmem.c | 14 > 29 files changed, 1242 insertions(+), 635 deletions(-) > > So not counting lib/test-parse-integer.c, it's a net addition of 44 > lines. That's OK. > > My overall reaction to this is "oh god, not again". Is it really worth > it?
I think giving good examples to people is always worth it :-)
> > +#define parse_integer(s, base, val) \ > > +({ \ > > + const char *_s = (s); \ > > + unsigned int _base = (base); \ > > + typeof(&(val)[0]) _val = (val); \ > > + \ > > + __builtin_choose_expr( \ > > + __builtin_types_compatible_p(typeof(_val), signed char *), \ > > + _parse_integer_sc(_s, _base, (void *)_val), \ > > + __builtin_choose_expr( \ > > + __builtin_types_compatible_p(typeof(_val), unsigned char *), \ > > + _parse_integer_uc(_s, _base, (void *)_val), \ > > + __builtin_choose_expr( \ > > + __builtin_types_compatible_p(typeof(_val), short *), \ > > + _parse_integer_s(_s, _base, (void *)_val), \ > > + __builtin_choose_expr( \ > > + __builtin_types_compatible_p(typeof(_val), unsigned short *), \ > > + _parse_integer_us(_s, _base, (void *)_val), \ > > + __builtin_choose_expr( \ > > + __builtin_types_compatible_p(typeof(_val), int *), \ > > + _parse_integer_i(_s, _base, (void *)_val), \ > > + __builtin_choose_expr( \ > > + __builtin_types_compatible_p(typeof(_val), unsigned int *), \ > > + _parse_integer_u(_s, _base, (void *)_val), \ > > + __builtin_choose_expr( \ > > + __builtin_types_compatible_p(typeof(_val), long *) && sizeof(long) == 4,\ > > + _parse_integer_i(_s, _base, (void *)_val), \ > > + __builtin_choose_expr( \ > > + __builtin_types_compatible_p(typeof(_val), long *) && sizeof(long) == 8,\ > > + _parse_integer_ll(_s, _base, (void *)_val), \ > > + __builtin_choose_expr( \ > > + __builtin_types_compatible_p(typeof(_val), unsigned long *) && sizeof(unsigned long) == 4,\ > > + _parse_integer_u(_s, _base, (void *)_val), \ > > + __builtin_choose_expr( \ > > + __builtin_types_compatible_p(typeof(_val), unsigned long *) && sizeof(unsigned long) == 8,\ > > + _parse_integer_ull(_s, _base, (void *)_val), \ > > + __builtin_choose_expr( \ > > + __builtin_types_compatible_p(typeof(_val), long long *), \ > > + _parse_integer_ll(_s, _base, (void *)_val), \ > > + __builtin_choose_expr( \ > > + __builtin_types_compatible_p(typeof(_val), unsigned long long *),\ > > + _parse_integer_ull(_s, _base, (void *)_val), \ > > + _parse_integer_link_time_error())))))))))))); \ > > +}) > > Wow.
Yep, but blame the language not me. Macro would be simpler with _Generic, but it is too early for that.
> > +/* internal, do not use */ > > +int _parse_integer_sc(const char *s, unsigned int base, signed char *val); > > +int _parse_integer_uc(const char *s, unsigned int base, unsigned char *val); > > +int _parse_integer_s(const char *s, unsigned int base, short *val); > > +int _parse_integer_us(const char *s, unsigned int base, unsigned short *val); > > +int _parse_integer_i(const char *s, unsigned int base, int *val); > > +int _parse_integer_u(const char *s, unsigned int base, unsigned int *val); > > +int _parse_integer_ll(const char *s, unsigned int base, long long *val); > > +int _parse_integer_ull(const char *s, unsigned int base, unsigned long long *val); > > These all have fairly lengthy implementations. Could it all be done > with a single function? > > int __parse_integer(const char *s, unsigned int base, unsigned int size, void *val); > > Where "size" is 1,2,4,8 with the top bit set if signed?
The question is why bother.
With smallish VM config I was testing, .text size difference is very small, LOC-wise the difference doesn't exist as well (additional error checking adds lines, not code per se). In the end _parse_integer() and simple_strto*() functions will be removed as well.
With 4-arg dispatch function, every callsite will be more bloated:
simple_strto* (pointer, pointer, int) parse_integer (pointer, int, pointer) 4-arg (pointer, int, int, pointer)
I think this code is way understandable than any mask-shift alternative:
int _parse_integer_sc(const char *s, unsigned int base, signed char *val) { long long tmp; int rv;
rv = _parse_integer_ll(s, base, &tmp); if (rv < 0) return rv; if (tmp != (signed char)tmp) return -ERANGE; *val = tmp; return rv; }
| |