Message-ID: <df001b6fbe2a4bdc86999c78933dab7f@huawei.com>
Date:   Sat, 18 May 2019 07:05:07 +0000
From:   "chenjianhong (A)" <chenjianhong2@...wei.com>
To:     Michel Lespinasse <walken@...gle.com>
CC:     Greg Kroah-Hartman <gregkh@...uxfoundation.org>,
        Andrew Morton <akpm@...ux-foundation.org>,
        "mhocko@...e.com" <mhocko@...e.com>,
        "Vlastimil Babka" <vbabka@...e.cz>,
        "Kirill A. Shutemov" <kirill.shutemov@...ux.intel.com>,
        Yang Shi <yang.shi@...ux.alibaba.com>,
        "jannh@...gle.com" <jannh@...gle.com>,
        "steve.capper@....com" <steve.capper@....com>,
        "tiny.windzz@...il.com" <tiny.windzz@...il.com>,
        LKML <linux-kernel@...r.kernel.org>,
        linux-mm <linux-mm@...ck.org>,
        "stable@...r.kernel.org" <stable@...r.kernel.org>,
        "willy@...radead.org" <willy@...radead.org>
Subject: RE: [PATCH] mm/mmap: fix the adjusted length error

I explain my test code and the problem in detail. This problem is found in 
32-bit user process, because its virtual is limited, 3G or 4G. 

First, I explain the bug I found. Function unmapped_area and 
unmapped_area_topdowns adjust search length to account for worst 
case alignment overhead, the code is ' length = info->length + info->align_mask; '.
The variable info->length is the length we allocate and the variable 
info->align_mask accounts for the alignment, because the gap_start  or gap_end 
value also should be an alignment address, but we can't know the alignment offset.
So in the current algorithm, it uses the max alignment offset, this value maybe zero
or other(0x1ff000 for shmat function). 
Is it reasonable way? The required value is longer than what I allocate.
What's more,  why for the first time I can allocate the memory successfully
Via shmat, but after releasing the memory via shmdt and I want to attach
again, it fails. This is not acceptable for many people.

Second, I explain my test code. The code I have sent an email. The following is
the step. I don't think it's something unusual or unreasonable, because the virtual
memory space is enough, but the process can allocate from it. And we can't pass
explicit addresses to function mmap or shmat, the address is getting from the left
vma gap.
 1, we allocat large virtual memory;
 2, we allocate hugepage memory via shmat, and release one
 of the hugepage memory block;
 3, we allocate hugepage memory by shmat again, this will fail.

Third, I want to introduce my change in the current algorithm. I don't change the
current algorithm. Also, I think there maybe a better way to fix this error. Nowadays,
I can just adjust the gap_start value. Take unmapped_area function an example:
The current algorithm:
1, Adjust search length to account for worst case alignment overhead
|----------------------------------------- -----------|----------------------------------------|
gap_start                                                                                                                             gap_end
|____________info->length_____________|___ info->align_mask _________|    
|______________________________length_______________________  ____|                 
2, search for the suitable gap, the required length is the sum of info->length and
info->align_mask
3, Adjust gap address to the desired alignment
My changed:
1, Adjust the gap_start address to the desired alignment
|---------------------------------|--------------------- -----------|--------------------------|
gap_start                            gap_start'                                                                        gap_end
|____ gap_start_offset____|
                                                    |__________________ info->length ___________|
2, search for the suitable gap, from gap_start' to gap_end. the required length is
info->length.

Test code:
#include <stdio.h>
#include <stdlib.h>
#include <sys/shm.h>
#include <errno.h>

int size = 0x4000000;
int shmid[5];
void *shm[5];
int key[5] = {234, 235, 236, 237, 238};
unsigned long seg_size[5] = {0x200000, 0x4000000, 0x1000000,
                        0x400000, 0xa00000};

int init_memory(void)
{
        int i,j;
        for (i = 0; i < 5; i++) {
                shmid[i] = shmget((key_t)key[i], seg_size[i], 0666
                                | IPC_CREAT | SHM_HUGETLB);
                if (shmid[i] == -1) {
                        fprintf(stderr, "shmget[%d] error(%d)\n",
                                i, errno);
                        goto failed;
                }
        }
        return 0;
failed:
        for (j = 0; j < i; j++) {
                shmctl(shmid[j], IPC_RMID, 0);
        }
        return -1;
}

int del_segmem(void)
{
        int i = 0;
        for (i = 0; i < 5; i++) {
                shmdt(shm[i]);
                shmctl(shmid[i], IPC_RMID, 0);
        }
        return 0;
}

int fun_C(void)
{
        int i = 0;
        printf("-----------------------fun_C shmat-----------------------\n");
        for (i = 0; i < 5; i+=1) {
                shm[i] = shmat(shmid[i], 0, 0);
                if (shm[i] == (void *)-1) {
                        fprintf(stderr, "shmat[%d] failed %d\n", i, errno);
                        return -1;
                }
        }
        sleep(2);
        system("pid=`ps -e | grep memory | awk '{print $1}'`;cat /proc/$pid/maps");
        shmdt(shm[1]);
        printf("-----------------------after fun_C shmdt-----------------------\n");
        system("pid=`ps -e | grep memory | awk '{print $1}'`;cat /proc/$pid/maps");
        printf("-----------------------fun_C ok-----------------------\n");
        return 0;
}

int fun_A(void)
{
        int i = 1;
        shm[i] = shmat(shmid[1], 0, 0);
        printf("-----------------------fun_A shmat-----------------------\n");
        if (shm[i] == (void *)-1) {
                fprintf(stderr, "funa shmat[%d] size(0x%08x)failed %d\n",
                        i, seg_size[i], errno);
                return -1;
        }
        system("pid=`ps -e | grep memory | awk '{print $1}'`;cat /proc/$pid/maps");
        sleep(2);
        shmdt(shm[1]);
        printf("-----------------------fun_A shmdt-----------------------\n");
        system("pid=`ps -e | grep memory | awk '{print $1}'`;cat /proc/$pid/maps");
        printf("-----------------------fun_A ok-----------------------\n");
        return 0;
}

/*
 * first, we allocat large virtual memory;
 * second, we allocate hugepage memory by shmat, and release one
 * of the hugepage memory block;
 * third, we allocate hugepage memory by shmat again, this will fail.
 */

int main(int argc,char * argv[])
{
        int i;
        int ret = 0;
        for (i == 0; i < 52; i++) {
                malloc(size);//first
        }
        if (init_memory() != 0) {
                ret = -1;
                goto failed_memory;
        }
        fun_C();//second
        sleep(5);
        ret = fun_A();//third
        if (ret != 0) {
                goto failed_memory;
        }
        sleep(3);
failed_memory:
        del_segmem();
        return ret;
}


-----Original Message-----
From: Michel Lespinasse [mailto:walken@...gle.com] 
Sent: Saturday, May 18, 2019 8:13 AM
To: chenjianhong (A) <chenjianhong2@...wei.com>
Cc: Greg Kroah-Hartman <gregkh@...uxfoundation.org>; Andrew Morton <akpm@...ux-foundation.org>; mhocko@...e.com; Vlastimil Babka <vbabka@...e.cz>; Kirill A. Shutemov <kirill.shutemov@...ux.intel.com>; Yang Shi <yang.shi@...ux.alibaba.com>; jannh@...gle.com; steve.capper@....com; tiny.windzz@...il.com; LKML <linux-kernel@...r.kernel.org>; linux-mm <linux-mm@...ck.org>; stable@...r.kernel.org
Subject: Re: [PATCH] mm/mmap: fix the adjusted length error

I worry that the proposed change turns the search from an O(log N)
worst case into a O(N) one.

To see why the current search is O(log N), it is easiest to start by
imagining a simplified search algorithm that wouldn't include the low
and high address limits. In that algorithm, backtracking through the
vma tree is never necessary - the tree walk can always know, prior to
going left or right, if a suitable gap will be found in the
corresponding subtree.

The code we have today does have to respect the low and high address
limits, so it does need to implement backtracking - but this
backtracking only occurs to back out of subtrees that include the low
address limit (the search went 'left' into a subtree that has a large
enough gap, but the gap turns out to be below the limit so it can't be
used and the search needs to go 'right' instead). Because of this, the
amount of backtracking that can occur is very limited, and the search
is still O(log N) in the worst case.

With your proposed change, backtracking could occur not only around
the low address limit, but also at any node within the search tree,
when it turns out that a gap that seemed large enough actually isn't
due to alignment constraints. So, the code should still work, but it
could backtrack more in the worst case, turning the worst case search
into an O(N) thing.

I am not sure what to do about this. First I would want to understand
more about your test case; is this something that you stumbled upon
without expecting it or was it an artificially constructed case to
show the limitations of the current search algorithm ? Also, if your
process does something unusual and expects to be able to map (close
to) the entirety of its address space, would it be reasonable for it
to manually manage the address space and pass explicit addresses to
mmap / shmat ?

On Thu, May 16, 2019 at 11:02 PM jianhong chen <chenjianhong2@...wei.com> wrote:
> In linux version 4.4, a 32-bit process may fail to allocate 64M hugepage
> memory by function shmat even though there is a 64M memory gap in
> the process.
>
> It is the adjusted length that causes the problem, introduced from
> commit db4fbfb9523c935 ("mm: vm_unmapped_area() lookup function").
> Accounting for the worst case alignment overhead, function unmapped_area
> and unmapped_area_topdown adjust the search length before searching
> for available vma gap. This is an estimated length, sum of the desired
> length and the longest alignment offset, which can cause misjudgement
> if the system has very few virtual memory left. For example, if the
> longest memory gap available is 64M, we can’t get it from the system
> by allocating 64M hugepage memory via shmat function. The reason is
> that it requires a longger length, the sum of the desired length(64M)
> and the longest alignment offset.
>
> To fix this error ,we can calculate the alignment offset of
> gap_start or gap_end to get a desired gap_start or gap_end value,
> before searching for the available gap. In this way, we don't
> need to adjust the search length.
>
> Problem reproduces procedure:
> 1. allocate a lot of virtual memory segments via shmat and malloc
> 2. release one of the biggest memory segment via shmdt
> 3. attach the biggest memory segment via shmat
>
> e.g.
> process maps:
> 00008000-00009000 r-xp 00000000 00:12 3385    /tmp/memory_mmap
> 00011000-00012000 rw-p 00001000 00:12 3385    /tmp/memory_mmap
> 27536000-f756a000 rw-p 00000000 00:00 0
> f756a000-f7691000 r-xp 00000000 01:00 560     /lib/libc-2.11.1.so
> f7691000-f7699000 ---p 00127000 01:00 560     /lib/libc-2.11.1.so
> f7699000-f769b000 r--p 00127000 01:00 560     /lib/libc-2.11.1.so
> f769b000-f769c000 rw-p 00129000 01:00 560     /lib/libc-2.11.1.so
> f769c000-f769f000 rw-p 00000000 00:00 0
> f769f000-f76c0000 r-xp 00000000 01:00 583     /lib/libgcc_s.so.1
> f76c0000-f76c7000 ---p 00021000 01:00 583     /lib/libgcc_s.so.1
> f76c7000-f76c8000 rw-p 00020000 01:00 583     /lib/libgcc_s.so.1
> f76c8000-f76e5000 r-xp 00000000 01:00 543     /lib/ld-2.11.1.so
> f76e9000-f76ea000 rw-p 00000000 00:00 0
> f76ea000-f76ec000 rw-p 00000000 00:00 0
> f76ec000-f76ed000 r--p 0001c000 01:00 543     /lib/ld-2.11.1.so
> f76ed000-f76ee000 rw-p 0001d000 01:00 543     /lib/ld-2.11.1.so
> f7800000-f7a00000 rw-s 00000000 00:0e 0       /SYSV000000ea (deleted)
> fba00000-fca00000 rw-s 00000000 00:0e 65538   /SYSV000000ec (deleted)
> fca00000-fce00000 rw-s 00000000 00:0e 98307   /SYSV000000ed (deleted)
> fce00000-fd800000 rw-s 00000000 00:0e 131076  /SYSV000000ee (deleted)
> ff913000-ff934000 rw-p 00000000 00:00 0       [stack]
> ffff0000-ffff1000 r-xp 00000000 00:00 0       [vectors]
>
> from 0xf7a00000 to fba00000, it has 64M memory gap, but we can't get
> it from kernel.

Powered by blists - more mailing lists