In one case, we eliminate a local variable; in the other a strlen()
call and some .text.
Signed-off-by: Rasmus Villemoes <[email protected]>
Cc: "Theodore Ts'o" <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
There's no point in starting over when we meet a '/'. This also
eliminates a stack variable and a little .text.
Signed-off-by: Rasmus Villemoes <[email protected]>
Acked-by: NeilBrown <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
This eliminates a little .text and avoids repeating the strchr call when
we meet a '!' (which will happen at least once).
Signed-off-by: Rasmus Villemoes <[email protected]>
Cc: Greg Kroah-Hartman <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
There's probably not many slashes in the name, but starting over when
we see one feels wrong.
Signed-off-by: Rasmus Villemoes <[email protected]>
Cc: Greg Kroah-Hartman <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Part of the disassembly of do_blk_trace_setup:
231b: e8 00 00 00 00 callq 2320 <do_blk_trace_setup+0x50>
231c: R_X86_64_PC32 strlen+0xfffffffffffffffc
2320: eb 0a jmp 232c <do_blk_trace_setup+0x5c>
2322: 66 0f 1f 44 00 00 nopw 0x0(%rax,%rax,1)
2328: 48 83 c3 01 add $0x1,%rbx
232c: 48 39 d8 cmp %rbx,%rax
232f: 76 47 jbe 2378 <do_blk_trace_setup+0xa8>
2331: 41 80 3c 1c 2f cmpb $0x2f,(%r12,%rbx,1)
2336: 75 f0 jne 2328 <do_blk_trace_setup+0x58>
2338: 41 c6 04 1c 5f movb $0x5f,(%r12,%rbx,1)
233d: 4c 89 e7 mov %r12,%rdi
2340: e8 00 00 00 00 callq 2345 <do_blk_trace_setup+0x75>
2341: R_X86_64_PC32 strlen+0xfffffffffffffffc
2345: eb e1 jmp 2328 <do_blk_trace_setup+0x58>
Yep, that's right: gcc isn't smart enough to realize that replacing '/' by
'_' cannot change the strlen(), so we call it again and again (at least
when a '/' is found). Even if gcc were that smart, this construction
would still loop over the string twice, once for the initial strlen() call
and then the open-coded loop.
Let's simply use strreplace() instead.
Signed-off-by: Rasmus Villemoes <[email protected]>
Acked-by: Steven Rostedt <[email protected]>
Liked-by: Jens Axboe <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Strings are sometimes sanitized by replacing a certain character (often
'/') by another (often '!'). In a few places, this is done the same way
Schlemiel the Painter would do it. Others are slightly smarter but still
do multiple strchr() calls. Introduce strreplace() to do this using a
single function call and a single pass over the string.
One would expect the return value to be one of three things: void, s, or
the number of replacements made. I chose the fourth, returning a pointer
to the end of the string. This is more likely to be useful (for example
allowing the caller to avoid a strlen call).
Signed-off-by: Rasmus Villemoes <[email protected]>
Cc: "Theodore Ts'o" <[email protected]>
Cc: Greg Kroah-Hartman <[email protected]>
Cc: Neil Brown <[email protected]>
Cc: Steven Rostedt <[email protected]>
Cc: Joe Perches <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
For !CONFIG_NUMA, hashdist will always be 0, since it's setter is
otherwise compiled out. So we can save 4 bytes of data and some .text
(although mostly in __init functions) by only defining it for
CONFIG_NUMA.
Signed-off-by: Rasmus Villemoes <[email protected]>
Acked-by: David Rientjes <[email protected]>
Reviewed-by: Michal Hocko <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
The first is a keyboard-off-by-one, the other two the ordinary mathy kind.
Signed-off-by: Rasmus Villemoes <[email protected]>
Acked-by: Christoph Lameter <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
According to C99, %2.s means 'print two spaces' (a precision of
. without following digits or * means 0). The kernel's printf
implementation, however, treats that case as if no precision was
given, but relying on that quirk is rather silly. Also, since no -
(aka left-justify) flag is given, the field with of 2 would then cause
the alarm->enabled case to come out as "o n". Deobfuscate it.
Signed-off-by: Rasmus Villemoes <[email protected]>
Signed-off-by: Alexandre Belloni <[email protected]>
list_entry is just a wrapper for container_of, but it is arguably
wrong (and slightly confusing) to use it when the pointed-to struct
member is not a struct list_head. Use container_of directly instead.
Signed-off-by: Rasmus Villemoes <[email protected]>
Signed-off-by: Al Viro <[email protected]>
Some tiny improvements, cutting 180 bytes off the generated code.
- use strchr() for single-character needle
- compute index using pointer subtraction instead of two strlen()
calls
- factor out the common check for whether the initial part of
kctl->id.name (before the space) is identical to w->name.
Signed-off-by: Rasmus Villemoes <[email protected]>
Acked-by: Vinod Koul <[email protected]>
Tested-by: Fang, Yang A <[email protected]>
Signed-off-by: Mark Brown <[email protected]>
Making a function call with 20 arguments is rather expensive in both
stack and .text. In this case, doing the formatting manually doesn't
make it any less readable, so we might as well save 155 bytes of .text
and 112 bytes of stack.
Signed-off-by: Rasmus Villemoes <[email protected]>
With CONFIG_DEBUG_INFO_REDUCED, we do get quite a lot of debug info
(around 22.7 MB for a defconfig+DEBUG_INFO_REDUCED). However, the
"basenames must match" rule used by -femit-struct-debug-baseonly
option means that we miss some core data structures, such as struct
{device, file, inode, mm_struct, page} etc.
We can easily get these included as well, while still getting the
benefits of CONFIG_DEBUG_INFO_REDUCED (faster build times and smaller
individual object files): All it takes is a dummy translation unit
including a few strategic headers and compiled with a flag overriding
-femit-struct-debug-baseonly.
This increases the size of .debug_info by ~0.3%, but these 90 KB
contain some rather useful info.
Signed-off-by: Rasmus Villemoes <[email protected]>
Signed-off-by: Michal Marek <[email protected]>
Just six days after this FIXME was added seven years ago, Sam Ravnborg
added the missing feature (37c514e3df "Add missing init section
definitions"), though it ended up being called __initconst.
Let's use it; better late than never.
Signed-off-by: Rasmus Villemoes <[email protected]>
Signed-off-by: Tejun Heo <[email protected]>
Currently, two of the branches are dead code, since an earlier smaller
substring would have been found ("TP" in the "TP_NW" case and either
of "BNC" and "AUI" in the "BNC_AUI" case). Rearrange the strstr()
calls so that the longer strings are searched for first.
Signed-off-by: Rasmus Villemoes <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
tty_name no longer uses the buf parameter, so remove it along with all
the 64 byte stack buffers that used to be passed in.
Mostly generated by the coccinelle script
@depends on patch@
identifier buf;
constant C;
expression tty;
@@
- char buf[C];
<+...
- tty_name(tty, buf)
+ tty_name(tty)
...+>
allmodconfig compiles, so I'm fairly confident the stack buffers
weren't used for other purposes as well.
Signed-off-by: Rasmus Villemoes <[email protected]>
Reviewed-by: Peter Hurley <[email protected]>
Acked-by: Jesper Nilsson <[email protected]>
Acked-by: Dmitry Torokhov <[email protected]>
Signed-off-by: Greg Kroah-Hartman <[email protected]>
All users of tty_name pass the return value (the provided buffer) to
some printf-like function. We can thus avoid the strcpy and, more
importantly, later remove the buf parameter completely, eliminating
the need for some 64 byte stack buffers.
Signed-off-by: Rasmus Villemoes <[email protected]>
Reviewed-by: Peter Hurley <[email protected]>
Signed-off-by: Greg Kroah-Hartman <[email protected]>
All users of tty_name pass the result directly to a printf-like
function. This means we can actually let tty_name return the literal
"NULL tty" or tty->name directly, avoiding the strcpy and a lot of
medium-sized stack buffers. In preparation for that, make the return
type const char*.
While at it, we can also constify the tty parameter.
Signed-off-by: Rasmus Villemoes <[email protected]>
Reviewed-by: Peter Hurley <[email protected]>
Signed-off-by: Greg Kroah-Hartman <[email protected]>
sprintf() reliably returns the number of characters printed, so we don't
need to ask strlen() where we are. Also replace calling sprintf("%02x")
in a loop with the much simpler bin2hex().
[[email protected]: it's odd to include kernel.h after everything else]
Signed-off-by: Rasmus Villemoes <[email protected]>
Cc: Al Viro <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
I hadn't had enough coffee when I wrote this. Currently, the final
increment of buf depends on the value loaded from the table, and
causes gcc to emit a cmov immediately before the return. It is smarter
to let it depend on r, since the increment can then be computed in
parallel with the final load/store pair. It also shaves 16 bytes of
.text.
Signed-off-by: Rasmus Villemoes <[email protected]>
Cc: Tejun Heo <[email protected]>
Cc: Peter Zijlstra <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
The most expensive part of decimal conversion is the divisions by 10
(albeit done using reciprocal multiplication with appropriately chosen
constants). I decided to see if one could eliminate around half of
these multiplications by emitting two digits at a time, at the cost of a
200 byte lookup table, and it does indeed seem like there is something
to be gained, especially on 64 bits. Microbenchmarking shows
improvements ranging from -50% (for numbers uniformly distributed in [0,
2^64-1]) to -25% (for numbers heavily biased toward the smaller end, a
more realistic distribution).
On a larger scale, perf shows that top, one of the big consumers of /proc
data, uses 0.5-1.0% fewer cpu cycles.
I had to jump through some hoops to get the 32 bit code to compile and run
on my 64 bit machine, so I'm not sure how relevant these numbers are, but
just for comparison the microbenchmark showed improvements between -30%
and -10%.
The bloat-o-meter costs are around 150 bytes (the generated code is a
little smaller, so it's not the full 200 bytes) on both 32 and 64 bit.
I'm aware that extra cache misses won't show up in a microbenchmark as
used above, but on the other hand decimal conversions often happen in bulk
(for example in the case of top).
I have of course tested that the new code generates the same output as the
old, for both the first and last 1e10 numbers in [0,2^64-1] and 4e9
'random' numbers in-between.
Test and verification code on github: https://github.com/Villemoes/dec.
Signed-off-by: Rasmus Villemoes <[email protected]>
Tested-by: Jeff Epler <[email protected]>
Cc: "Peter Zijlstra (Intel)" <[email protected]>
Cc: Tejun Heo <[email protected]>
Cc: Joe Perches <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
The macro BITMAP_LAST_WORD_MASK can be implemented without a conditional,
which will generally lead to slightly better generated code (221 bytes
saved for allmodconfig-GCOV_KERNEL, ~2k with GCOV_KERNEL). As a small
bonus, this also ensures that the nbits parameter is expanded exactly
once.
In BITMAP_FIRST_WORD_MASK, if start is signed gcc is technically allowed
to assume it is positive (or divisible by BITS_PER_LONG), and hence just
do the simple mask. It doesn't seem to use this, and even on an
architecture like x86 where the shift only depends on the lower 5 or 6
bits, and these bits are not affected by the signedness of the expression,
gcc still generates code to compute the C99 mandated value of start %
BITS_PER_LONG. So just use a mask explicitly, also for consistency with
BITMAP_LAST_WORD_MASK.
Signed-off-by: Rasmus Villemoes <[email protected]>
Cc: Tejun Heo <[email protected]>
Reviewed-by: George Spelvin <[email protected]>
Cc: Yury Norov <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
The current semantics of string_escape_mem are inadequate for one of its
current users, vsnprintf(). If that is to honour its contract, it must
know how much space would be needed for the entire escaped buffer, and
string_escape_mem provides no way of obtaining that (short of allocating a
large enough buffer (~4 times input string) to let it play with, and
that's definitely a big no-no inside vsnprintf).
So change the semantics for string_escape_mem to be more snprintf-like:
Return the size of the output that would be generated if the destination
buffer was big enough, but of course still only write to the part of dst
it is allowed to, and (contrary to snprintf) don't do '\0'-termination.
It is then up to the caller to detect whether output was truncated and to
append a '\0' if desired. Also, we must output partial escape sequences,
otherwise a call such as snprintf(buf, 3, "%1pE", "\123") would cause
printf to write a \0 to buf[2] but leaving buf[0] and buf[1] with whatever
they previously contained.
This also fixes a bug in the escaped_string() helper function, which used
to unconditionally pass a length of "end-buf" to string_escape_mem();
since the latter doesn't check osz for being insanely large, it would
happily write to dst. For example, kasprintf(GFP_KERNEL, "something and
then %pE", ...); is an easy way to trigger an oops.
In test-string_helpers.c, the -ENOMEM test is replaced with testing for
getting the expected return value even if the buffer is too small. We
also ensure that nothing is written (by relying on a NULL pointer deref)
if the output size is 0 by passing NULL - this has to work for
kasprintf("%pE") to work.
In net/sunrpc/cache.c, I think qword_add still has the same semantics.
Someone should definitely double-check this.
In fs/proc/array.c, I made the minimum possible change, but longer-term it
should stop poking around in seq_file internals.
[[email protected]: simplify qword_add]
[[email protected]: add missed curly braces]
Signed-off-by: Rasmus Villemoes <[email protected]>
Acked-by: Andy Shevchenko <[email protected]>
Signed-off-by: Andy Shevchenko <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
When printf is given the format specifier %pE, it needs a way of obtaining
the total output size that would be generated if the buffer was large
enough, and string_escape_mem doesn't easily provide that. This is a
refactorization of string_escape_mem in preparation of changing its
external API to provide that information.
The somewhat ugly early returns and subsequent seemingly redundant
conditionals are to make the following patch touch as little as possible
in string_helpers.c while still preserving the current behaviour of never
outputting partial escape sequences. That behaviour must also change for
%pE to work as one expects from every other printf specifier.
Signed-off-by: Rasmus Villemoes <[email protected]>
Acked-by: Andy Shevchenko <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
The helper hex_string() is broken in two ways. First, it doesn't
increment buf regardless of whether there is room to print, so callers
such as kasprintf() that try to probe the correct storage to allocate will
get a too small return value. But even worse, kasprintf() (and likely
anyone else trying to find the size of the result) pass NULL for buf and 0
for size, so we also have end == NULL. But this means that the end-1 in
hex_string() is (char*)-1, so buf < end-1 is true and we get a NULL
pointer deref. I double-checked this with a trivial kernel module that
just did a kasprintf(GFP_KERNEL, "%14ph", "CrashBoomBang").
Nobody seems to be using %ph with kasprintf, but we might as well fix it
before it hits someone.
Signed-off-by: Rasmus Villemoes <[email protected]>
Acked-by: Andy Shevchenko <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
gcc doesn't merge or overlap const char[] objects with identical contents
(probably language lawyers would also insist that these things have
different addresses), but there's no reason to have the string
"0123456789ABCDEF" occur in multiple places. hex_asc_upper is declared in
kernel.h and defined in lib/hexdump.c, which is unconditionally compiled
in.
Signed-off-by: Rasmus Villemoes <[email protected]>
Cc: Peter Zijlstra <[email protected]>
Cc: Tejun Heo <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
At least since the initial git commit, when base was passed as a separate
parameter, number() has only been called with bases 8, 10 and 16. I'm
guessing that 66 was to accommodate 64 0/1, a sign and a '\0', but the
buffer is only used for the actual digits. Octal digits carry 3 bits of
information, so 24 is enough. Spell that 3*sizeof(num) so one less place
needs to be changed should long long ever be 128 bits. Also remove the
commented-out code that would handle an arbitrary base.
Signed-off-by: Rasmus Villemoes <[email protected]>
Cc: Peter Zijlstra <[email protected]>
Cc: Tejun Heo <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Since FORMAT_TYPE_INT is simply 1 more than FORMAT_TYPE_UINT, and
similarly for BYTE/UBYTE, SHORT/USHORT, LONG/ULONG, we can eliminate a few
instructions by making SIGN have the value 1 instead of 2, and then use
arithmetic instead of branches for computing the right spec->type. It's a
little hacky, but certainly in the same spirit as SMALL needing to have
the value 0x20. For example for the spec->qualifier == 'l' case, gcc now
generates
75e: 0f b6 53 01 movzbl 0x1(%rbx),%edx
762: 83 e2 01 and $0x1,%edx
765: 83 c2 09 add $0x9,%edx
768: 88 13 mov %dl,(%rbx)
instead of
763: 0f b6 53 01 movzbl 0x1(%rbx),%edx
767: 83 e2 02 and $0x2,%edx
76a: 80 fa 01 cmp $0x1,%dl
76d: 19 d2 sbb %edx,%edx
76f: 83 c2 0a add $0xa,%edx
772: 88 13 mov %dl,(%rbx)
Signed-off-by: Rasmus Villemoes <[email protected]>
Cc: Peter Zijlstra <[email protected]>
Cc: Tejun Heo <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Commit 607ca46e97 ("UAPI: (Scripted) Disintegrate include/linux") left
behind some empty conditional blocks. Since they are useless and may
cause a reader to wonder whether something is missing, remove them.
Signed-off-by: Rasmus Villemoes <[email protected]>
Cc: Geert Uytterhoeven <[email protected]>
Cc: David Howells <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
According to C99, %*.s means the same as %*.0s, in other words, print as
many spaces as the field width argument says and effectively ignore the
string argument. That is certainly not what was meant here. The kernel's
printf implementation, however, treats it as if the . was not there,
i.e. as %*s. I don't know if de->name is nul-terminated or not, but in
any case I'm guessing the intention was to use de->name_len as precision
instead of field width.
[ Note: this is debugging code which is commented out, so this is not
security issue; a developer would have to explicitly enable
INLINE_DIR_DEBUG before this would be an issue. ]
Signed-off-by: Rasmus Villemoes <[email protected]>
Signed-off-by: Theodore Ts'o <[email protected]>
The tty_name member of struct ser_device is never set or used, so it
can be removed. (The definition of struct ser_device is private to
this .c file, and the identifier tty_name only occurs in this one
place.)
Signed-off-by: Rasmus Villemoes <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
The Subtlety (1) referred to vanished with 6ba2ef7baa ("cpumask:
Move deprecated functions to end of header."). That used to mention
some suboptimal code generation by a, by now, rather ancient gcc. With
gcc 4.7, I don't see any change in the generated code by making it a
static inline, so let's add type checking and get rid of the ghost
reference.
Signed-off-by: Rasmus Villemoes <[email protected]>
Signed-off-by: Rusty Russell <[email protected]>
There's no reason to allocate the dec{32,64}table on the stack; it
just wastes a bunch of instructions setting them up and, of course,
also consumes quite a bit of stack. Using size_t for such small
integers is a little excessive.
$ scripts/bloat-o-meter /tmp/built-in.o lib/built-in.o
add/remove: 2/2 grow/shrink: 2/0 up/down: 1304/-1548 (-244)
function old new delta
lz4_decompress_unknownoutputsize 55 718 +663
lz4_decompress 55 632 +577
dec64table - 32 +32
dec32table - 32 +32
lz4_uncompress 747 - -747
lz4_uncompress_unknownoutputsize 801 - -801
The now inlined lz4_uncompress functions used to have a stack
footprint of 176 bytes (according to -fstack-usage); their inlinees
have increased their stack use from 32 bytes to 48 and 80 bytes,
respectively.
Signed-off-by: Rasmus Villemoes <[email protected]>
Signed-off-by: Greg Kroah-Hartman <[email protected]>
The documentation and the code disagrees; fix the former.
Signed-off-by: Rasmus Villemoes <[email protected]>
Signed-off-by: Jonathan Corbet <[email protected]>
Not everybody uses a utf8 locale (unfortunately), so let's avoid
non-ascii characters in the kernel log. Replace the 3-byte utf8
sequence with a 3-byte ascii equivalent.
Signed-off-by: Rasmus Villemoes <[email protected]>
Signed-off-by: Felipe Balbi <[email protected]>
The only difference between the three power_sensor_name_templates is
whether there is a suffix of "", "_lowest" or "_highest". We might as
well pull those into an array and use a literal format string,
allowing gcc to do type checking of the arguments to
sprintf. Incidentially, the same three suffixes are used in the
temp_sensor_name_templates case, so we end up eliminating one static
array.
Signed-off-by: Rasmus Villemoes <[email protected]>
[Guenter Roeck: Fixed line length over 80 characters]
Signed-off-by: Guenter Roeck <[email protected]>
By extracting the only part that differs we can allow static checking
of the format string, and possibly save a little .rodata.
Signed-off-by: Rasmus Villemoes <[email protected]>
[Guenter Roeck: continuation line alignment]
Signed-off-by: Guenter Roeck <[email protected]>
Commit fab4c256a5 ("PCI/AER: Add a TLP header print helper") introduced
the helper function __print_tlp_header(), but contrary to the intention,
the behaviour did change: Since we're taking the address of the parameter
t, the first 4 or 8 bytes printed will be the value of the pointer t
itself, and the remaining 12 or 8 bytes will be who-knows-what (something
from the stack).
We want to show the values of the four members of the struct
aer_header_log_regs; that can be done without ugly and error-prone casts.
On little-endian this should produce the same output as originally
intended, and since no-one has complained about getting garbage output so
far, I think big-endian should be ok too.
Fixes: fab4c256a5 ("PCI/AER: Add a TLP header print helper")
Signed-off-by: Rasmus Villemoes <[email protected]>
Signed-off-by: Bjorn Helgaas <[email protected]>
Acked-by: Borislav Petkov <[email protected]>
CC: [email protected] # v3.14+
The # flag in %X means print a 0X prefix. Remove the extra 0x prefix.
Signed-off-by: Rasmus Villemoes <[email protected]>
Signed-off-by: Kalle Valo <[email protected]>
C99 says that a precision which is simply '.' with no following digits
or * should be interpreted as 0, which means that these format strings
actually mean 'print 16 spaces'. However, the kernel's printf
implementation treats this case as if the precision was omitted. Don't
rely on that quirk.
Signed-off-by: Rasmus Villemoes <[email protected]>
Signed-off-by: Greg Kroah-Hartman <[email protected]>
In the expressions idx/32 and idx%32, both idx and 32 have signed
type, and unfortunately the C standard prescribes rounding to 0, so
unless gcc can prove that idx is non-negative, these cannot be
implemented as simple shift respectively mask operations. Help gcc by
changing the type of idx to unsigned - this cuts another few
instructions from the generated code.
Signed-off-by: Rasmus Villemoes <[email protected]>
Acked-by: Selvin Xavier <[email protected]>
Signed-off-by: Roland Dreier <[email protected]>
gcc emits a surprising amount of code in order to flip a bit. One
would think that a single instruction is enough.
$ scripts/bloat-o-meter /tmp/ocrdma_verbs.o drivers/infiniband/hw/ocrdma/ocrdma_verbs.o
add/remove: 0/0 grow/shrink: 0/3 up/down: 0/-142 (-142)
function old new delta
ocrdma_post_srq_recv 498 460 -38
ocrdma_poll_cq 2010 1962 -48
ocrdma_discard_cqes 495 439 -56
All three calls of ocrdma_srq_toggle_bit happen within spinlocks, so
saving a few useless instructions might be worthwhile.
Signed-off-by: Rasmus Villemoes <[email protected]>
Acked-by: Selvin Xavier <[email protected]>
Signed-off-by: Roland Dreier <[email protected]>
Instead of potentially passing over the string twice in case c is not
found, just keep track of the last occurrence. According to
bloat-o-meter, this also cuts the generated code by a third (54 vs 36
bytes). Oh, and we get rid of those 7-space indented lines.
Signed-off-by: Rasmus Villemoes <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
The first of these conditionals is completely redundant: If k == lim-1, we
must have off==0, so the second conditional will also trigger and then it
wouldn't matter if upper had some high bits set. But the second
conditional is in fact also redundant, since it only serves to clear out
some high-order "don't care" bits of dst, about which no guarantee is
made.
Signed-off-by: Rasmus Villemoes <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
We can shift the bits from lower and upper into place before assembling
dst[k + off]; moving the shift of lower into the branch where we already
know that rem is non-zero allows us to remove a conditional.
Signed-off-by: Rasmus Villemoes <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
gcc can generate slightly better code for stuff like "nbits %
BITS_PER_LONG" when it knows nbits is not negative. Since negative size
bitmaps or shift amounts don't make sense, change these parameters of
bitmap_shift_right to unsigned.
If off >= lim (which requires shift >= nbits), k is initialized with a
large positive value, but since I've let k continue to be signed, the loop
will never run and dst will be zeroed as expected. Inside the loop, k is
guaranteed to be non-negative, so the fact that it is promoted to unsigned
in the various expressions it appears in is harmless.
Also use "shift" and "nbits" consistently for the parameter names.
Signed-off-by: Rasmus Villemoes <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
If left is 0, we can just let mask be ~0UL, so that anding with it is a
no-op. Conveniently, BITMAP_LAST_WORD_MASK provides precisely what we
need, and we can eliminate left.
Signed-off-by: Rasmus Villemoes <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
If the condition k==lim-1 is true, we must have off == 0 (otherwise, k
could never become that big). But in that case we have upper == 0 and
hence dst[k] == (src[k] & mask) >> rem. Since mask consists of a
consecutive range of bits starting from the LSB, anding dst[k] with mask
is a no-op.
Signed-off-by: Rasmus Villemoes <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
We can shift the bits from lower and upper into place before assembling
dst[k]; moving the shift of upper into the branch where we already know
that rem is non-zero allows us to remove a conditional.
Signed-off-by: Rasmus Villemoes <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
I've previously changed the nbits parameter of most bitmap_* functions to
unsigned; now it is bitmap_shift_{left,right}'s turn. This alone saves
some .text, but while at it I found that there were a few other things one
could do. The end result of these seven patches is
$ scripts/bloat-o-meter /tmp/bitmap.o.{old,new}
add/remove: 0/0 grow/shrink: 0/2 up/down: 0/-328 (-328)
function old new delta
__bitmap_shift_right 384 226 -158
__bitmap_shift_left 306 136 -170
and less importantly also a smaller stack footprint
$ stack-o-meter.pl master bitmap
file function old new delta
lib/bitmap.o __bitmap_shift_right 24 8 -16
lib/bitmap.o __bitmap_shift_left 24 0 -24
For each pair of 0 <= shift <= nbits <= 256 I've tested the end result
with a few randomly filled src buffers (including garbage beyond nbits),
in each case verifying that the shift {left,right}-most bits of dst are
zero and the remaining nbits-shift bits correspond to src, so I'm fairly
confident I didn't screw up. That hasn't stopped me from being wrong
before, though.
This patch (of 7):
gcc can generate slightly better code for stuff like "nbits %
BITS_PER_LONG" when it knows nbits is not negative. Since negative size
bitmaps or shift amounts don't make sense, change these parameters of
bitmap_shift_right to unsigned.
The expressions involving "lim - 1" are still ok, since if lim is 0 the
loop is never executed.
Also use "shift" and "nbits" consistently for the parameter names.
Signed-off-by: Rasmus Villemoes <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
On little-endian, there's no reason to have an extra, presumably less
efficient, way of copying a bitmap.
Signed-off-by: Rasmus Villemoes <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Make the prototype of bitmap_copy_le the same as bitmap_copy's. All other
bitmap_* functions take unsigned long* parameters; there's no reason this
should be special.
The only current user is the static inline uwb_mas_bm_copy_le, which
already does the void* laundering, so the end users can pass their u8 or
__le32 buffers without a cast.
Furthermore, this allows us to simply let bitmap_copy_le be an alias for
bitmap_copy on little-endian; see next patch.
Signed-off-by: Rasmus Villemoes <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
We don't need all the stuff kernel.h pulls in; just compiler.h since
export.h doesn't do necessary #includes. This removes more than 100
dependencies.
Signed-off-by: Rasmus Villemoes <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
These three #includes seem to be completely redundant: Removing them
yields identical objdump -d output for each of {allyes,allno,def}config,
and neither included file end up in the generated dependency file through
some recursive include. In total, about 50 lines are eliminated from
.percpu.o.cmd.
Signed-off-by: Rasmus Villemoes <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
strncpy_from_user.c only needs EXPORT_SYMBOL, so just include compiler.h
and export.h instead of the whole module.h machinery.
Signed-off-by: Rasmus Villemoes <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
stmp_device.c only needs EXPORT_SYMBOL, so just include compiler.h and
export.h instead of the whole module.h machinery.
Signed-off-by: Rasmus Villemoes <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
The sort function and its helpers don't do memory allocation, so the
slab.h include is redundant. Move it inside the #if 0 protecting the
self-test, similar to how it is done in lib/list_sort.c. This removes
over 450 lines from the generated dependency file.
Signed-off-by: Rasmus Villemoes <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
show_mem.c doesn't use anything from nmi.h. Removing it yields identical
objdump -d output for each of {allyes,allno,def}config and eliminates more
than 100 lines in the dependency file.
Signed-off-by: Rasmus Villemoes <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
The comment helpfully explains why hardirq.h is included, but since
commit 2d4b84739f ("hardirq: Split preempt count mask definitions")
in_interrupt() has been provided by preempt_mask.h. Use that instead,
saving around 40 lines in the generated dependency file.
Signed-off-by: Rasmus Villemoes <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Removing the include of linux/spinlock.h produces byte-identical output
for {allno,def}config, and identical objdump -d output for allyesconfig.
In the former two cases, more than a 100 lines are eliminated from the
generated dependency file.
Signed-off-by: Rasmus Villemoes <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
nlattr.c doesn't seem to rely on anything from netdevice.h. Removing it
yields identical objdump -d output for each of {allyes,allno,def}config,
and eliminates more than 200 lines from the generated dependency file.
Signed-off-by: Rasmus Villemoes <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
The file doesn't seem to use anything from linux/user_namespace.h, and
removing it yields byte-identical object code and strictly fewer
dependencies in the .cmd file.
Signed-off-by: Rasmus Villemoes <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
This file doesn't seem to use anything provided by linux/interrupt.h or
anything recursively included through that. Removing it produces
byte-identical output, while reducing .llist.o.cmd from 541 to 156 lines.
Signed-off-by: Rasmus Villemoes <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
md5.c doesn't use anything from kernel.h, except that that pulls in
compiler.h, which is needed for the export.h to work.
Signed-off-by: Rasmus Villemoes <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Memory allocation only happens in the self test, just as random numbers
are only used there. So move the inclusion of slab.h inside the
CONFIG_TEST_LIST_SORT.
We don't need module.h and all of the stuff it carries with it, so replace
with export.h and compiler.h. Unfortunately, the ARRAY_SIZE macro from
kernel.h requires the user to ensure bug.h is also included (for
BUILD_BUG_ON_ZERO, used by __must_be_array). We used to get that through
some maze of nested includes, but just include it explicitly.
linux/string.h is then only included implicitly through
kernel.h->printk.h->dynamic_debug.h, but only if !CONFIG_DYNAMIC_DEBUG, so
just include it explicitly (for memset).
objdump -d says the generated code is the same, and wc -l says that
lib/.list_sort.o.cmd went from 579 to 165 lines.
Signed-off-by: Rasmus Villemoes <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
idr.c doesn't seem to use anything from hardirq.h (or anything included
from that). Removing it produces identical objdump -d output, and gives
44 fewer lines in the .idr.o.cmd dependency file.
Signed-off-by: Rasmus Villemoes <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
We only need EXPORT_SYMBOL, so compiler.h and export.h suffice. This
means linux/types.h is no longer implicitly included, so add an include of
uapi/linux/types.h to linux/cryptohash.h for __u32. Other users of
cryptohash.h cannot be affected, since they must already have been
including uapi/linux/types.h in order for gcc not to complain about
unknown types.
Signed-off-by: Rasmus Villemoes <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
The file doesn't use anything from ctype.h. Instead of module.h, just use
export.h for EXPORT_SYMBOL. The latter requires the user to include
compiler.h, so do that explicitly instead of relying on some other header
pulling it in.
Signed-off-by: Rasmus Villemoes <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
sort.c doesn't use facilities from kernel.h, but does use some types
defined in linux/types.h. Include the latter directly instead of relying
on some other header doing it. Similarly, include linux/export.h directly
instead of through module.h. This removes 80 lines from the dependency
file .sort.o.cmd.
Signed-off-by: Rasmus Villemoes <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
The file uses nothing from init.h, and also doesn't need the full module.h
machinery; export.h is sufficient. The latter requires the user to ensure
compiler.h is included, so do that explicitly instead of relying on some
other header pulling it in.
Signed-off-by: Rasmus Villemoes <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Also, rename bits to nbits. Both changes for consistency with other
bitmap_* functions.
Signed-off-by: Rasmus Villemoes <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Make the return value and the ord and nbits parameters of
bitmap_ord_to_pos unsigned.
Also, simplify the implementation and as a side effect make the result
fully defined, returning nbits for ord >= weight, in analogy with what
find_{first,next}_bit does. This is a better sentinel than the former
("unofficial") 0. No current users are affected by this change.
Signed-off-by: Rasmus Villemoes <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
The ordinal of a set bit is simply the number of set bits before it;
counting those doesn't need to be done one bit at a time. While at it,
update the parameters to unsigned int.
It is not completely unthinkable that gcc would see pos as compile-time
constant 0 in one of the uses of bitmap_pos_to_ord. Since the static
inline frontend bitmap_weight doesn't handle nbits==0 correctly (it would
behave exactly as if nbits==BITS_PER_LONG), use __bitmap_weight.
Alternatively, the last line could be spelled bitmap_weight(buf, pos+1)-1,
but this is simpler.
Signed-off-by: Rasmus Villemoes <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Change the sz and nbits parameters of bitmap_fold to unsigned int for
consistency with other bitmap_* functions, and to save another few bytes
in the generated code.
[[email protected]: fix kerneldoc]
Signed-off-by: Rasmus Villemoes <[email protected]>
Cc: Wu Fengguang <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Change the nbits parameter of bitmap_onto to unsigned int for consistency
with other bitmap_* functions.
[[email protected]: coding-style fixes]
Signed-off-by: Rasmus Villemoes <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Since the various bitmap_* functions now take an unsigned int as nbits
parameter, it makes sense to also update the various wrappers, even though
they're marked as obsolete.
Signed-off-by: Rasmus Villemoes <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Since the various bitmap_* functions now take an unsigned int as nbits
parameter, it makes sense to also update the various wrappers.
Signed-off-by: Rasmus Villemoes <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
For consistency with the other bitmap_* functions, also make the nbits
parameter of bitmap_zero, bitmap_fill and bitmap_copy unsigned.
Signed-off-by: Rasmus Villemoes <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
string_get_size() was documented to return an error, but in fact always
returned 0. Since the output always fits in 9 bytes, just document that
and let callers do what they do now: pass a small stack buffer and ignore
the return value.
Signed-off-by: Rasmus Villemoes <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
The remainder from do_div is always a u32, and after size has been reduced
to be below 1000 (or 1024), it certainly fits in u32. So both remainder
and sf_cap can be made u32s, the format specifiers can be simplified (%lld
wasn't the right thing to use for _unsigned_ long long anyway), and we can
replace a do_div with an ordinary 32/32 bit division.
Signed-off-by: Rasmus Villemoes <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
While commit 3c9f3681d0 ("[SCSI] lib: add generic helper to print
sizes rounded to the correct SI range") says that Z and Y are included
in preparation for 128 bit computers, they just waste .text currently.
If and when we get u128, string_get_size needs updating anyway (and ISO
needs to come up with four more prefixes).
Also there's no need to include and test for the NULL sentinel; once we
reach "E" size is at most 18. [The test is also wrong; it should be
units_str[units][i+1]; if we've reached NULL we're already doomed.]
Signed-off-by: Rasmus Villemoes <[email protected]>
Cc: James Bottomley <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
All callers of skip_atoi have already checked for the first character
being a digit. In this case, gcc generates simpler code for a do
while-loop.
Signed-off-by: Rasmus Villemoes <[email protected]>
Cc: Jiri Kosina <[email protected]>
Cc: Randy Dunlap <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
On 64 bit, size may very well be huge even if bit 31 happens to be 0.
Somehow it doesn't feel right that one can pass a 5 GiB buffer but not a
3 GiB one. So cap at INT_MAX as was probably the intention all along.
This is also the made-up value passed by sprintf and vsprintf.
Signed-off-by: Rasmus Villemoes <[email protected]>
Cc: Jiri Kosina <[email protected]>
Cc: Randy Dunlap <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
It seems a little simpler to consume the p from a %p specifier in
format_decode, just as it is done for the surrounding %c, %s and %% cases.
While there, delete a redundant and misplaced comment.
Signed-off-by: Rasmus Villemoes <[email protected]>
Cc: Jiri Kosina <[email protected]>
Cc: Randy Dunlap <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
__FUNCTION__ hasn't been treated as a string literal since gcc 3.4, so
this only helps people who only test-compile using 3.3 (compiler-gcc3.h
barks at anything older than that). Besides, there are almost no
occurrences of __FUNCTION__ left in the tree.
[[email protected]: convert remaining __FUNCTION__ references]
Signed-off-by: Rasmus Villemoes <[email protected]>
Cc: Michal Nazarewicz <[email protected]>
Cc: Joe Perches <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
src_ip is a pointer to a union vxlan_addr, one member of which is a
struct sockaddr. Passing a pointer to src_ip is wrong; one should pass
the value of src_ip itself. Since %pIS formally expects something of
type struct sockaddr*, let's pass a pointer to the appropriate union
member, though this of course doesn't change the generated code.
Fixes: e4c7ed4153 ("vxlan: add ipv6 support")
Signed-off-by: Rasmus Villemoes <[email protected]>
Acked-by: Cong Wang <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
Rasmus Villemoes