[comp.sys.sgi] malloc question

XBR2D96D@DDATHD21.BITNET (Knobi der Rechnerschrat) (01/19/90)

Hallo,

  I've just found a behaviour in a program, that I don't understand.
When running the program mt1 (appended to this message) the value returned
from sbrk(0) is only changed for the first malloc(...). In that case the
difference between old and new break-value is the desired 4MB. In the remaining
9 loops, old and new break-value are the same, but as easily can be seen,
in all 10 loops the same pointer is returned. How do I interpret this
behaviour? Does it mean that free(...) frees the memory allocated by the
process for future use, but does not return the virtual memory? I've written
the test program, because we have a programm that does a lot of malloc/free
calls and it seems its virtual memory (as shown by ps -l) never shrinks. It
grows, but only when more memory is needed. Is this the normal behaviour, or
is something wrong (e.g. my understanding of malloc/free and my interpretation
of sbrk())?  I hope somebody can shed some light on that. We are running a
4D-70/GT with 16 MB of memory, ESDI root-disk and SCSI Data-disk. Our IRIX
is Release 3.2 (btw. when will 3.2.1 be shipped ?).


Regards
Martin Knoblauch

TH-Darmstadt
Physical Chemistry 1
Petersenstrasse 20
D-6100 Darmstadt, FRG

BITNET: <XBR2D96D@DDATHD21>

--------------------------mt1.c------------------------------------------
/* compile using   'cc -O mt1.c -o mt1 -lmalloc'            */
/* using the 'normal' malloc package gives the same results */

#include <stdio.h>
#include <malloc.h>

main()
{
int i=0;
char *sbrk(),*obrk,*nbrk,*dum;

for(i=0;i<10;i++) {
  obrk = sbrk(0);
  dum = malloc(4000000);
  nbrk = sbrk(0);
  printf("Memtest: %x %x %x %d\n",dum,obrk,nbrk,nbrk-obrk);
  free(dum);
  }

}
--------------------------end of mt1.c-----------------------------------

mitch@rock.sgi.com (Tom Mitchell) (03/01/90)

In article <9001181348.aa27967@ADM.BRL.MIL> XBR2D96D@DDATHD21.BITNET (Knobi der Rechnerschrat) writes:
*   I've just found a behaviour in a program, that I don't understand.

* .... Does it mean that free(...) frees the memory allocated by the
* process for future use, but does not return the virtual memory?

Yes.

malloc is a memory allocator tool.  It will sbreak from the
system as needed.  free() returns memory to the malloc library
not to the system.

Perhaps the best way to understand malloc is to look inside.
As luck has it there is a 'public' malloc within Larry
Wall's "perl" package.  So unlike  AT&T code which we
cannot look under the hood at, here is malloc.

As an aside lwall's configure scripts within perl are near
AI in searching out the differences of one UNIX and another.
They are worth looking at.


Enjoy
======= snip =========
malloc from the public 'perl' of lwall.
This code has its origns at BSD/CAltech etc.
--- cut ----
#ifndef lint
static char sccsid[] = "@(#)malloc.c	4.3 (Berkeley) 9/16/83";

#define RCHECK
/*
 * malloc.c (Caltech) 2/21/82
 * Chris Kingsley, kingsley@cit-20.
 *
 * This is a very fast storage allocator.  It allocates blocks of a small 
 * number of different sizes, and keeps free lists of each size.  Blocks that
 * don't exactly fit are passed up to the next larger size.  In this 
 * implementation, the available sizes are 2^n-4 (or 2^n-12) bytes long.
 * This is designed for use in a program that uses vast quantities of memory,
 * but bombs when it runs out. 
 */

#include "EXTERN.h"
#include "perl.h"

/* I don't much care whether these are defined in sys/types.h--LAW */

#define u_char unsigned char
#define u_int unsigned int
#define u_short unsigned short

/*
 * The overhead on a block is at least 4 bytes.  When free, this space
 * contains a pointer to the next free block, and the bottom two bits must
 * be zero.  When in use, the first byte is set to MAGIC, and the second
 * byte is the size index.  The remaining bytes are for alignment.
 * If range checking is enabled and the size of the block fits
 * in two bytes, then the top two bytes hold the size of the requested block
 * plus the range checking words, and the header word MINUS ONE.
 */
union	overhead {
	union	overhead *ov_next;	/* when free */
	struct {
		u_char	ovu_magic;	/* magic number */
		u_char	ovu_index;	/* bucket # */
#ifdef RCHECK
		u_short	ovu_size;	/* actual block size */
		u_int	ovu_rmagic;	/* range magic number */
#endif
	} ovu;
#define	ov_magic	ovu.ovu_magic
#define	ov_index	ovu.ovu_index
#define	ov_size		ovu.ovu_size
#define	ov_rmagic	ovu.ovu_rmagic
};

#define	MAGIC		0xff		/* magic # on accounting info */
#define OLDMAGIC	0x7f		/* same after a free() */
#define RMAGIC		0x55555555	/* magic # on range info */
#ifdef RCHECK
#define	RSLOP		sizeof (u_int)
#else
#define	RSLOP		0
#endif

/*
 * nextf[i] is the pointer to the next free block of size 2^(i+3).  The
 * smallest allocatable block is 8 bytes.  The overhead information
 * precedes the data area returned to the user.
 */
#define	NBUCKETS 30
static	union overhead *nextf[NBUCKETS];
extern	char *sbrk();

#ifdef MSTATS
/*
 * nmalloc[i] is the difference between the number of mallocs and frees
 * for a given block size.
 */
static	u_int nmalloc[NBUCKETS];
#include <stdio.h>
#endif

#ifdef debug
#define	ASSERT(p)   if (!(p)) botch("p"); else
static
botch(s)
	char *s;
{

	printf("assertion botched: %s\n", s);
	abort();
}
#else
#define	ASSERT(p)
#endif

char *
malloc(nbytes)
	register unsigned nbytes;
{
  	register union overhead *p;
  	register int bucket = 0;
  	register unsigned shiftr;

	/*
	 * Convert amount of memory requested into
	 * closest block size stored in hash buckets
	 * which satisfies request.  Account for
	 * space used per block for accounting.
	 */
  	nbytes += sizeof (union overhead) + RSLOP;
  	nbytes = (nbytes + 3) &~ 3; 
  	shiftr = (nbytes - 1) >> 2;
	/* apart from this loop, this is O(1) */
  	while (shiftr >>= 1)
  		bucket++;
	/*
	 * If nothing in hash bucket right now,
	 * request more memory from the system.
	 */
  	if (nextf[bucket] == NULL)    
  		morecore(bucket);
  	if ((p = (union overhead *)nextf[bucket]) == NULL)
  		return (NULL);
	/* remove from linked list */
	if (*((int*)p) > 0x10000000)
#ifndef I286
	    fprintf(stderr,"Corrupt malloc ptr 0x%x at 0x%x\n",*((int*)p),p);
#else
	    fprintf(stderr,"Corrupt malloc ptr 0x%lx at 0x%lx\n",*((int*)p),p);
#endif
  	nextf[bucket] = nextf[bucket]->ov_next;
	p->ov_magic = MAGIC;
	p->ov_index= bucket;
#ifdef MSTATS
  	nmalloc[bucket]++;
#endif
#ifdef RCHECK
	/*
	 * Record allocated size of block and
	 * bound space with magic numbers.
	 */
  	if (nbytes <= 0x10000)
		p->ov_size = nbytes - 1;
	p->ov_rmagic = RMAGIC;
  	*((u_int *)((caddr_t)p + nbytes - RSLOP)) = RMAGIC;
#endif
  	return ((char *)(p + 1));
}

/*
 * Allocate more memory to the indicated bucket.
 */
static
morecore(bucket)
	register int bucket;
{
  	register union overhead *op;
  	register int rnu;       /* 2^rnu bytes will be requested */
  	register int nblks;     /* become nblks blocks of the desired size */
	register int siz;

  	if (nextf[bucket])
  		return;
	/*
	 * Insure memory is allocated
	 * on a page boundary.  Should
	 * make getpageize call?
	 */
  	op = (union overhead *)sbrk(0);
#ifndef I286
  	if ((int)op & 0x3ff)
  		(void)sbrk(1024 - ((int)op & 0x3ff));
#else
	/* The sbrk(0) call on the I286 always returns the next segment */
#endif

#ifndef I286
	/* take 2k unless the block is bigger than that */
  	rnu = (bucket <= 8) ? 11 : bucket + 3;
#else
	/* take 16k unless the block is bigger than that 
	   (80286s like large segments!)		*/
  	rnu = (bucket <= 11) ? 14 : bucket + 3;
#endif
  	nblks = 1 << (rnu - (bucket + 3));  /* how many blocks to get */
  	if (rnu < bucket)
		rnu = bucket;
	op = (union overhead *)sbrk(1 << rnu);
	/* no more room! */
  	if ((int)op == -1)
  		return;
	/*
	 * Round up to minimum allocation size boundary
	 * and deduct from block count to reflect.
	 */
#ifndef I286
  	if ((int)op & 7) {
  		op = (union overhead *)(((int)op + 8) &~ 7);
  		nblks--;
  	}
#else
	/* Again, this should always be ok on an 80286 */
#endif
	/*
	 * Add new memory allocated to that on
	 * free list for this hash bucket.
	 */
  	nextf[bucket] = op;
  	siz = 1 << (bucket + 3);
  	while (--nblks > 0) {
		op->ov_next = (union overhead *)((caddr_t)op + siz);
		op = (union overhead *)((caddr_t)op + siz);
  	}
}

free(cp)
	char *cp;
{   
  	register int size;
	register union overhead *op;

  	if (cp == NULL)
  		return;
	op = (union overhead *)((caddr_t)cp - sizeof (union overhead));
#ifdef debug
  	ASSERT(op->ov_magic == MAGIC);		/* make sure it was in use */
#else
	if (op->ov_magic != MAGIC) {
		fprintf(stderr,"%s free() ignored\n",
		    op->ov_magic == OLDMAGIC ? "Duplicate" : "Bad");
		return;				/* sanity */
	}
	op->ov_magic = OLDMAGIC;
#endif
#ifdef RCHECK
  	ASSERT(op->ov_rmagic == RMAGIC);
	if (op->ov_index <= 13)
		ASSERT(*(u_int *)((caddr_t)op + op->ov_size + 1 - RSLOP) == RMAGIC);
#endif
  	ASSERT(op->ov_index < NBUCKETS);
  	size = op->ov_index;
	op->ov_next = nextf[size];
  	nextf[size] = op;
#ifdef MSTATS
  	nmalloc[size]--;
#endif
}

/*
 * When a program attempts "storage compaction" as mentioned in the
 * old malloc man page, it realloc's an already freed block.  Usually
 * this is the last block it freed; occasionally it might be farther
 * back.  We have to search all the free lists for the block in order
 * to determine its bucket: 1st we make one pass thru the lists
 * checking only the first block in each; if that fails we search
 * ``reall_srchlen'' blocks in each list for a match (the variable
 * is extern so the caller can modify it).  If that fails we just copy
 * however many bytes was given to realloc() and hope it's not huge.
 */
int reall_srchlen = 4;	/* 4 should be plenty, -1 =>'s whole list */

char *
realloc(cp, nbytes)
	char *cp; 
	unsigned nbytes;
{   
  	register u_int onb;
	union overhead *op;
  	char *res;
	register int i;
	int was_alloced = 0;

  	if (cp == NULL)
  		return (malloc(nbytes));
	op = (union overhead *)((caddr_t)cp - sizeof (union overhead));
	if (op->ov_magic == MAGIC) {
		was_alloced++;
		i = op->ov_index;
	} else {
		/*
		 * Already free, doing "compaction".
		 *
		 * Search for the old block of memory on the
		 * free list.  First, check the most common
		 * case (last element free'd), then (this failing)
		 * the last ``reall_srchlen'' items free'd.
		 * If all lookups fail, then assume the size of
		 * the memory block being realloc'd is the
		 * smallest possible.
		 */
		if ((i = findbucket(op, 1)) < 0 &&
		    (i = findbucket(op, reall_srchlen)) < 0)
			i = 0;
	}
	onb = (1 << (i + 3)) - sizeof (*op) - RSLOP;
	/* avoid the copy if same size block */
	if (was_alloced &&
	    nbytes <= onb && nbytes > (onb >> 1) - sizeof(*op) - RSLOP)
		return(cp);
  	if ((res = malloc(nbytes)) == NULL)
  		return (NULL);
  	if (cp != res)			/* common optimization */
		(void)bcopy(cp, res, (int)((nbytes < onb) ? nbytes : onb));
  	if (was_alloced)
		free(cp);
  	return (res);
}

/*
 * Search ``srchlen'' elements of each free list for a block whose
 * header starts at ``freep''.  If srchlen is -1 search the whole list.
 * Return bucket number, or -1 if not found.
 */
static
findbucket(freep, srchlen)
	union overhead *freep;
	int srchlen;
{
	register union overhead *p;
	register int i, j;

	for (i = 0; i < NBUCKETS; i++) {
		j = 0;
		for (p = nextf[i]; p && j != srchlen; p = p->ov_next) {
			if (p == freep)
				return (i);
			j++;
		}
	}
	return (-1);
}

#ifdef MSTATS
/*
 * mstats - print out statistics about malloc
 * 
 * Prints two lines of numbers, one showing the length of the free list
 * for each size category, the second showing the number of mallocs -
 * frees for each size category.
 */
mstats(s)
	char *s;
{
  	register int i, j;
  	register union overhead *p;
  	int totfree = 0,
  	totused = 0;

  	fprintf(stderr, "Memory allocation statistics %s\nfree:\t", s);
  	for (i = 0; i < NBUCKETS; i++) {
  		for (j = 0, p = nextf[i]; p; p = p->ov_next, j++)
  			;
  		fprintf(stderr, " %d", j);
  		totfree += j * (1 << (i + 3));
  	}
  	fprintf(stderr, "\nused:\t");
  	for (i = 0; i < NBUCKETS; i++) {
  		fprintf(stderr, " %d", nmalloc[i]);
  		totused += nmalloc[i] * (1 << (i + 3));
  	}
  	fprintf(stderr, "\n\tTotal in use: %d, total free: %d\n",
	    totused, totfree);
}
#endif
#endif /* lint */

  Thomas P. Mitchell   --  mitch@sgi.com
	"All things in moderation; including moderation."