perf tools: fix type mismatch - long vs __statfs_word
PATCH v2 1/1 perf tools: fix type mismatch - long vs __statfs_word
In case of compilation with "-Wextra" following breakage happens:
--->---fs.c: In function ‘fs__valid_mount’:
fs.c:76:24: error: comparison between signed and unsigned integer expressions [-Werror=sign-compare]
else if (st_fs.f_type != magic)
^
cc1: all warnings being treated as errors
--->---
Note that now when fs.c is in "lib/api/fs" and in "tools/lib/api/Makefile" CFLAGS has hard-coded "-Wextra" this is inevitable even if one builds "perf" with "WERROR=0".
Even though "debugfs.c" gets compiled successfully because DEBUGFS_MAGIC is definitely positive (MSB bit is not set, so compiler doesn't care about explisit cast to "long") it's good to cast "st_fs.f_type" to "long" as well.
And here's an explanation of observed issue.
Perf tools use libc headers instead of ones from kernel source tree. This is because "perf" is just a user-space tool even thought its sources are merged in Linux kernel source tree.
And so "statfs.h" comes from either uClibc, glibc or musl.
In case of uClibc most of architectures use "legacy" version from here:
http://git.uclibc.org/uClibc/tree/libc/sysdeps/linux/common/bits/statfs.h
New UAPI compliant architectures (looks like for now it's only ARC) uses UAPI version of "statfs.h" from here:
http://git.uclibc.org/uClibc/tree/libc/sysdeps/linux/common-generic/bits/statfs.h
In this particular situation most interesting difference is in data type used for "struct statfs" member "f_type".
In "legacy" version it's:
--->---__SWORD_TYPE f_type;
--->---
In UAPI version it's:
--->---__U32_TYPE f_type;
--->---
In its turn mentioned types are defined in
http://git.uclibc.org/uClibc/tree/libc/sysdeps/linux/common/bits/types.h
this way:
--->---
...
--->---
And for 32-bit architectures "int" is of the same length as "long" so compiler is happy on data type check.
But if "f_type" is "unsigned int" it's definitely not an unsigned "long".
Essential fix is explicit casting of "f_type" to "long".
---
Changes in v2:
* Added type cast to DEBUGFS_MAGIC in "debugfs.c"
* Added verbose explanation of root cause
Thanks a lot to Arnaldo for mention of another instance of hidden missing cast in "debugfs.c".
--- tools/lib/api/fs/debugfs.c | 2 +- tools/lib/api/fs/fs.c | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-)
tools/lib/api/fs/debugfs.c +1 -1
@@ -67,7 +67,7 @@ int debugfs_valid_mountpoint(const char *debugfs)if (statfs (debugfs, &st_fs) <0 )return -ENOENT ; -else if (st_fs.f_type != (long )DEBUGFS_MAGIC ) +else if ((long ) st_fs.f_type != (long )DEBUGFS_MAGIC )return -ENOENT ;return 0 ;
tools/lib/api/fs/fs.c +1 -1
@@ -75,7 +75,7 @@ static int fs__valid_mount(const char *fs, long magic)if (statfs (fs, &st_fs) <0 )return -ENOENT ; -else if (st_fs.f_type != magic) +else if ((long ) st_fs.f_type != magic)return -ENOENT ;return 0 ;
Signature · 1 line
1.9.3
Quoted 85 lines · Alexey Brodkin
In case of compilation with "-Wextra" following breakage happens:
--->---fs.c: In function ‘fs__valid_mount’:
fs.c:76:24: error: comparison between signed and unsigned integer expressions [-Werror=sign-compare]
else if (st_fs.f_type != magic)^cc1: all warnings being treated as errors
--->---Note that now when
fs.cis in "lib/api/fs" and in "tools/lib/api/Makefile" CFLAGS has hard-coded "-Wextra" this is inevitable even if one builds "perf" with "WERROR=0".Even though "
debugfs.c" gets compiled successfully becauseDEBUGFS_MAGICis definitely positive (MSB bit is not set, so compiler doesn't care about explisit cast to "long") it's good to cast "st_fs.f_type" to "long" as well.And here's an explanation of observed issue.
Perf tools use libc headers instead of ones from kernel source tree. This is because "perf" is just a user-space tool even thought its sources are merged in Linux kernel source tree.
And so "statfs.h" comes from either uClibc, glibc or musl.In case of uClibc most of architectures use "legacy" version from here:
http://git.uclibc.org/uClibc/tree/libc/sysdeps/linux/common/bits/statfs.hNew UAPI compliant architectures (looks like for now it's only ARC) uses UAPI version of "
statfs.h" from here:
http://git.uclibc.org/uClibc/tree/libc/sysdeps/linux/common-generic/bits/statfs.hIn this particular situation most interesting difference is in data type used for "struct statfs" member "
f_type".In "legacy" version it's:
--->---__SWORD_TYPEf_type;
--->---In UAPI version it's:
--->---__U32_TYPEf_type;
--->---In its turn mentioned types are defined in
http://git.uclibc.org/uClibc/tree/libc/sysdeps/linux/common/bits/types.h
this way:
--->---
...
--->---And for 32-bit architectures "int" is of the same length as "long" so compiler is happy on data type check.
But if "
f_type" is "unsigned int" it's definitely not an unsigned "long".Essential fix is explicit casting of "
f_type" to "long".Signed-off-by Alexey Brodkin <[email protected]>Cc Vineet Gupta <[email protected]>Cc Borislav Petkov <[email protected]>Cc Jiri Olsa <[email protected]>Cc Cody P Schafer <[email protected]>Cc Arnaldo Carvalho de Melo <[email protected]>---
Changes in v2:
* Added type cast to
DEBUGFS_MAGICin "debugfs.c"
* Added verbose explanation of root causeThanks a lot to Arnaldo for mention of another instance of hidden missing cast in "
debugfs.c".--- tools/lib/api/fs/debugfs.c | 2 +- tools/lib/api/fs/fs.c | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-)tools/lib/api/fs/debugfs.c +0
-0
tools/lib/api/fs/debugfs.c int debugfs_valid_mountpoint(const char *debugfs) · line 70
Quoted 5 lines
tools/lib/api/fs/debugfs.c +1
-1@@ -67,7 +67,7 @@ int debugfs_valid_mountpoint(const char *debugfs)if (statfs (debugfs, &st_fs) <0 )return -ENOENT ; -else if (st_fs.f_type != (long )DEBUGFS_MAGIC ) +else if ((long ) st_fs.f_type != (long )DEBUGFS_MAGIC )
Hi Arnaldo,
Does this fix look good to you. We need something like this to enable upstream ARC kernels to build perf.
OTOH do we really need the cast on both sides for this 2nd hunk ?
Thx,
-Vineet
Quoted 16 lines
tools/lib/api/fs/debugfs.c +0
-0return -ENOENT ;return 0 ;tools/lib/api/fs/fs.c +1
-1@@ -75,7 +75,7 @@ static int fs__valid_mount(const char *fs, long magic)if (statfs (fs, &st_fs) <0 )return -ENOENT ; -else if (st_fs.f_type != magic) +else if ((long ) st_fs.f_type != magic)return -ENOENT ;return 0 ;