linux-user: Fix the computation of the requested heap size
commit4d1de87c75007ee7e29dd271ebb4afdcf01ad7aa
authorvincent <cedric.vincent@st.com>
Tue, 14 Jun 2011 21:56:33 +0000 (14 21:56 +0000)
committerRiku Voipio <riku.voipio@iki.fi>
Tue, 21 Jun 2011 17:30:09 +0000 (21 20:30 +0300)
tree49f6e0735bdbb78292f7a8953c6a4592885f0dfa
parent5382a012e8ce7cf5ea612d291286be827574c181
linux-user: Fix the computation of the requested heap size

There were several remaining bugs in the previous implementation of
do_brk():

    1. the value of "new_alloc_size" was one page too large when the
       requested brk was aligned on a host page boundary.

    2. no new pages should be (re-)allocated when the requested brk is
       in the range of the pages that were already allocated
       previsouly (for the same purpose).  Technically these pages are
       never unmapped in the current implementation.

The problem/fix can be reproduced/validated with the test-suite above:

    #include <unistd.h>       /* syscall(2),      */
    #include <sys/syscall.h>  /* SYS_brk,         */
    #include <stdio.h>        /* puts(3),         */
    #include <stdlib.h>       /* exit(3), EXIT_*, */
    #include <stdint.h>       /* uint*_t,         */
    #include <sys/mman.h>     /* mmap(2), MAP_*,  */
    #include <string.h>       /* memset(3), */

    int main()
    {
        int exit_status = EXIT_SUCCESS;
        uint8_t *current_brk = 0;
        uint8_t *initial_brk;
        uint8_t *new_brk;
        uint8_t *old_brk;
        int failure = 0;
        int i;

        void test_brk(int increment, int expected_result) {
            new_brk = (uint8_t *)syscall(SYS_brk, current_brk + increment);
            if ((new_brk == current_brk) == expected_result)
                failure = 1;
            current_brk = (uint8_t *)syscall(SYS_brk, 0);
        }

        void test_result() {
            if (!failure)
                puts("OK");
            else {
                puts("failure");
                exit_status = EXIT_FAILURE;
            }
        }

        void test_title(const char *title) {
            failure = 0;
            printf("%-45s : ", title);
            fflush(stdout);
        }

        test_title("Initialization");
        test_brk(0, 1);
        initial_brk = current_brk;
        test_result();

        test_title("Don't overlap \"brk\" pages");
        test_brk(HOST_PAGE_SIZE, 1);
        test_brk(HOST_PAGE_SIZE, 1);
        test_result();

        /* Preparation for the test "Re-allocated heap is initialized".  */
        old_brk = current_brk - HOST_PAGE_SIZE;
        memset(old_brk, 0xFF, HOST_PAGE_SIZE);

        test_title("Don't allocate the same \"brk\" page twice");
        test_brk(-HOST_PAGE_SIZE, 1);
        test_brk(HOST_PAGE_SIZE, 1);
        test_result();

        test_title("Re-allocated \"brk\" pages are initialized");
        for (i = 0; i < HOST_PAGE_SIZE; i++) {
            if (old_brk[i] != 0) {
                printf("(index = %d, value = 0x%x) ", i, old_brk[i]);
                failure = 1;
                break;
            }
        }
        test_result();

        test_title("Don't allocate \"brk\" pages over \"mmap\" pages");
        new_brk = mmap(current_brk, HOST_PAGE_SIZE / 2, PROT_READ, MAP_PRIVATE | MAP_ANONYMOUS | MAP_FIXED, -1, 0);
        if (new_brk == (void *) -1)
            puts("unknown");
        else {
            test_brk(HOST_PAGE_SIZE, 0);
            test_result();
        }

        test_title("All \"brk\" pages are writable (please wait)");
        if (munmap(current_brk, HOST_PAGE_SIZE / 2) != 0)
            puts("unknown");
        else {
            while (current_brk - initial_brk < 2*1024*1024*1024UL) {
                old_brk = current_brk;

                test_brk(HOST_PAGE_SIZE, -1);
                if (old_brk == current_brk)
                    break;

                for (i = 0; i < HOST_PAGE_SIZE; i++)
                    old_brk[i] = 0xAA;
            }
            puts("OK");
        }

        test_title("Maximum size of the heap > 16MB");
        failure = (current_brk - initial_brk) < 16*1024*1024;
        test_result();

        exit(exit_status);
    }

Changes introduced in patch v2:

    * extend the "brk" test-suite embedded within the commit message;

    * heap contents have to be initialized to zero, this bug was
      exposed by "tst-calloc.c" from the GNU C library;

    * don't [try to] allocate a new host page if the new "brk" is
      equal to the latest allocated host page ("brk_page"); and

    * print some debug information when DEBUGF_BRK is defined.

Signed-off-by: Cédric VINCENT <cedric.vincent@st.com>
Reviewed-by: Christophe Guillon <christophe.guillon@st.com>
Cc: Riku Voipio <riku.voipio@iki.fi>
Signed-off-by: Riku Voipio <riku.voipio@iki.fi>
linux-user/syscall.c