sprofil(2)

NAME

sprofil() — execution time profile for disjointed text spaces

SYNOPSIS

#include <sys/time.h> 
#include <sys/profil.h> 

int sprofil( 
    struct prof     *profp, 
    int             profcnt, 
    struct timeval  *tvp, 
    unsigned int    flags 
); 

DESCRIPTION

sprofil() controls profiling, by which the system maintains estimates of the amount of time the calling program spends executing at various places in its address space. It differs from its predecessor profil() in that it allows simultaneous profiling of many disjointed regions of memory.

profp must point to an ordered array of prof structures.

The prof structure is defined as:

struct prof { 
       void          *pr_base;      /* buffer base */ 
       unsigned int  pr_size;       /* buffer size */ 
       void          *pr_offset;    /* pc offset */ 
       unsigned int  pr_scale;      /* pc scaling */ 
}; 

profcnt specifies the number of prof structures in the profp array. If profcnt is 0, profiling is disabled.

If non-NULL, tvp points to a struct timeval which on return contains the time value corresponding to one clock tick.

The only valid flags values are PROF_USHORT, PROF_UINT, and PROF_UINT64. The flags parameter must be set to one of these values.

When profiling is enabled, each thread's program counter (pc) is examined each clock tick (CLK_TICK times per second), and the system searches the profp array for a region that includes pc.

Each prof structure is interpreted as follows:

pr_base must point to an area of memory whose length (in bytes) is given by pr_size. This memory must not be modified or freed while profiling is enabled. This memory is treated as an array of 2, 4 or 8-byte elements corresponding to the PROF_USHORT, PROF_UINT, or PROF_UINT64 flag specified by the flags parameter.
pr_size must be a non-zero multiple of the element size.
pr_offset is the start address of the region to be profiled. Note that it is always interpreted as a text address and never as the address of a function descriptor.
pr_scale is interpreted as an unsigned, fixed-point value with sixteen bits to the right of the binary point.

The pc value falls within a region if (offset=((pc-pr_offset) * pr_scale) / 65536) < pr_size.

If a matching region is found, offset is rounded down to a 2, 4 or 8-byte boundary corresponding to the element size specified by the flags parameter. The element at pr_base + offset is incremented by 1, unless that would result in overflow.

If the calculation above results in an invalid address, profiling is silently disabled.

The profp array must be sorted in ascending order by pr_offset value. The profiling regions must not overlap.

If a pr_scale field is set to 0 or 1, the corresponding prof array entry is ignored.

The last entry in the profp array is special. If its pr_offset field is set to 0 and its pr_scale field is set to 2, it is treated as an overflow descriptor. In this case, its pr_size field must be set to the size of one element. If the pc value did not fall within the range of one of the preceding descriptors, the single element specified by the pr_base field is incremented, unless that would result in overflow.

Each call to sprofil() disables profiling if it was enabled and then enables profiling of the regions specified by the prof array entries.

Due to variations in instruction length on different architectures, the pr_scale setting results in different entry-to-instruction ratios. The following tables show the number of instructions that map into each element of the array that pr_base points to based on instruction size, pr_scale and flags.

On PA-RISC machines, instructions are 4 bytes each.

Scale	PROF_USHORT	PROF_UINT	PROF_UINT64
2	16384	32768	65536
4	8192	16384	32768
8	4096	8192	16384
16	2048	4096	8192
32	1024	2048	4096
64	512	1024	2048
128	256	512	1024
256	128	256	512
512	64	128	256
1024	32	64	128
2048	16	32	64
4096	8	16	32
8192	4	8	16
16384	2	4	8
32768	1	2	4
65536	0.5	1	2
131072	0.25	0.5	1

On Itanium-based machines, pc points to a 16-byte bundle containing three instructions. It is not possible to profile individual instructions.

Scale	PROF_USHORT	PROF_UINT	PROF_UINT64
2	4096	8192	16384
4	2048	4096	8192
8	1024	2048	4096
16	512	1024	2048
32	256	512	1024
64	128	256	512
128	64	128	256
256	32	64	128
512	16	32	64
1024	8	16	32
2048	4	8	16
4096	2	4	8
8192	1	2	4
16384	0.5	1	2
32768	0.25	0.5	1

Note that a 1-1 mapping of PROF_UINT counters consumes twice as much space as for PROF_USHORT counters, and should only be used in those cases where you expect a 16-bit counter to overflow. Likewise, a 1-1 mapping of PROF_UINT64 counters consumes twice as much space as for PROF_UINT counters, and should only be used in those cases where you expect a 32-bit counter to overflow.

The results of specifying a pr_scale that corresponds to an instruction-to-profiling element ratio smaller than 1-1 mapping produces undefined behavior.

If overlapping text regions are specified, then the pr_base associated with the largest offset less than the value of the program counter (pc) gets incremented. If more than one text region has the same offset, the first one is chosen.

A text region with the pr_offset of 0 and pr_scale of 0x0002 acts as an overflow bin. Any pc value not previously matched or outside the regions of text space specified is considered to match this text region and the first counter of the specified pr_base is incremented. If specified, such an overflow bin should always be last in the list of specified regions.

Notes

Profiling is turned off when an exec() is executed, but remains on in both child and parent processes after a fork().

A call to profil() has the same effect as calling sprofil() with just one member in the profp array with the fields profcnt, tvp, and flags taking on the values of 1, NULL, and PROF_USHORT respectively.

RETURN VALUE

On success, sprofil() returns 0. Otherwise, it returns -1 and sets the global variable errno to indicate the error.

ERRORS

The sprofil() function fails if:

E2BIG: profcnt was less than or equal to 0, or greater than sysconf(_SC_PROFIL_MAX).
EFAULT: Either profp or tvp points to an illegal address.
EINVAL: The flags parameter is not valid or a pr_size field is zero or greater than (2^46 * pr_scale) / 65536 or is not a multiple of the element size or the overflow descriptor is not last in the list of specified regions or is more than one element long or the regions are not sorted in ascending order by pr_offset or the regions overlap.

sprofil(2)

Technical documentation

» Table of Contents

» Index