Build / dev/kmem to enumerate all Linux kernel modules (including hidden ones)

In fact, Linux is a later thing. In the initial UNIX era, enumeration processes such as ps were scanned from / dev/kmem, which is all files Later, the Linux kernel expanded procfs inappropriately, and put all kinds of messy things into it, such as modules, filesystems, vmallocinfo, etc., which are obviously not processes, all of them are thrown in, which is not appropriate, but because this is the Linux kernel, so everything is right!

Of course, Linux also retains / dev/mem and / dev/kmem, two extremely special and fun files:

• /dev/mem: map all the physical memory of the system.
• /dev/kmem: map all kernel state virtual memory of the system.

Later, because the permissions exposed by / dev/kmem are too large and there are security risks, the general kernel blocks this character device, only retains / dev/mem, and it is still limited:

# CONFIG_DEVKMEM is not set CONFIG_STRICT_DEVMEM=y

This article shows how to enumerate all kernel modules by scanning / dev/kmem.

what? Isn't it true that config? Devkmem is disabled? Easy to do! This is the case when I replant it. In addition, the version I replant can support the mapping of vmalloc space.

The code is as follows:

// kmem.c #include <linux/module.h> #include <linux/mm.h> #include <linux/kallsyms.h> #include <linux/cdev.h> #include <linux/fs.h> pgprot_t (*_phys_mem_access_prot)(struct file *, unsigned long, unsigned long, pgprot_t); phys_addr_t (*_slow_virt_to_phys)(void *); pte_t *(*_lookup_address)(unsigned long, unsigned int *); static const struct vm_operations_struct mmap_mem_ops = { .access = generic_access_phys }; static int mmap_kmem(struct file *file, struct vm_area_struct *vma) { unsigned long pfn; pte_t *pte; unsigned int level = 0; size_t size; // This is what I added. Because vmalloc space is included, it is necessary to prevent crash caused by the virtual address value of unmapped page. pte = _lookup_address((u64)vma->vm_pgoff << PAGE_SHIFT, &level); if (!pte || !pte_present(*pte)) return -EIO; // pfn is obtained by a general method instead of only considering linear mapping. pfn = _slow_virt_to_phys((void *)(vma->vm_pgoff << PAGE_SHIFT)) >> PAGE_SHIFT; if (!pfn_valid(pfn)) return -EIO; vma->vm_pgoff = pfn; size = vma->vm_end - vma->vm_start; vma->vm_page_prot = _phys_mem_access_prot(file, vma->vm_pgoff, size, vma->vm_page_prot); vma->vm_ops = &mmap_mem_ops; if (remap_pfn_range(vma, vma->vm_start, vma->vm_pgoff, size, vma->vm_page_prot)) { return -EAGAIN; } return 0; } static const struct file_operations kmem_fops = { .mmap = mmap_kmem, }; dev_t dev = 0; static struct cdev kmem_cdev; static int __init devkmem_init(void) { _phys_mem_access_prot = (void *)kallsyms_lookup_name("phys_mem_access_prot"); _slow_virt_to_phys = (void *)kallsyms_lookup_name("slow_virt_to_phys"); _lookup_address = (void *)kallsyms_lookup_name("lookup_address"); if((alloc_chrdev_region(&dev, 0, 1, "test_dev")) <0){ printk("alloc failed\n"); return -1; } printk("major=%d minor=%d \n",MAJOR(dev), MINOR(dev)); cdev_init(&kmem_cdev, &kmem_fops); if ((cdev_add(&kmem_cdev, dev, 1)) < 0) { printk("add failed\n"); goto out; } return 0; out: unregister_chrdev_region(dev,1); return -1; } void __exit devkmem_exit(void) { cdev_del(&kmem_cdev); unregister_chrdev_region(dev, 1); } module_init(devkmem_init); module_exit(devkmem_exit); MODULE_LICENSE("GPL");

OK, we compile and load it, and create character device:

insmod ./kmem.ko mknod /dev/kmem c 248 0

Then, the following script shows how to enumerate all the modules. Because I map and parse address by address, and my bash level is very low, neither python nor go, so the efficiency of the script is very low, and the operation is very slow. If all the memory from 0xffffffa00000000 to 0xffffffff000000 is mapped at one time, it will be much faster. Although it is slow, it is absolutely detailed enough. Let's see:

#!/bin/bash # mlist.sh start='' end='' base='' moktype=$(cat /proc/kallsyms|grep module_ktype|awk '') # For pattern matching moktype=$(echo $moktype|tr 'a-z' 'A-Z') for line in $(cat /proc/vmallocinfo |grep 0xffffffffa|awk '') do start=$(echo $line|awk -F '-' ''|awk -F '0x' '') start=$(echo $start|tr 'a-z' 'A-Z') end=$(echo $line|awk -F '-' ''|awk -F '0x' '') end=$(echo $end|tr 'a-z' 'A-Z') base=$start next=$base while true; do val=$(./a.out $next); if [ $? -ne 0 ]; then break; fi if [ $val == $base ]; then mod=$(echo "ibase=16;$next-138"|bc) mod=$(echo "obase=16;$mod"|bc) state=$(./a.out $mod) if [ $? -ne 0 ] || [ $state != '0' ]; then next=$(echo "ibase=16;$next+8"|bc) next=$(echo "obase=16;$next"|bc) continue; fi ktype=$(echo "ibase=16;$mod+78"|bc) ktype=$(echo "obase=16;$ktype"|bc) type=$(./a.out $ktype) if [ $? -ne 0 ] || [ $type != $moktype ]; then next=$(echo "ibase=16;$next+8"|bc) next=$(echo "obase=16;$next"|bc) continue; fi namea=$(echo "ibase=16;$mod+18"|bc) namea=$(echo "obase=16;$namea"|bc) name=$(./a.out $namea) # Only the first eight characters of the name are truncated name=$(echo -n $name|sed 's/\([0-9A-F]\{2\}\)/\\\\\\x\1/gI' | xargs printf) name=$(echo $name|rev 2>/dev/null) if [ $? -eq 0 ]; then echo name-- $name fi fi next=$(echo "ibase=16;$next+8"|bc) next=$(echo "obase=16;$next"|bc) done done;

The results are as follows:

[root@localhost test]# ./mlist.sh name-- dm_mod name-- dm_regio name-- serio_ra name-- dm_log name-- dm_mirro name-- libata name-- ahci name-- ata_gene name-- crct10di name-- e1000 name-- pata_acp name-- cdrom name-- sr_mod name-- crc_t10d name-- sd_mod name-- ablk_hel name-- libcrc32 name-- ip_table name-- video name-- i2c_piix name-- parport name-- cryptd name-- parport_ name-- ata_piix name-- kmem name-- i2c_core ...

Here is a brief explanation.

We know that the address mapping space of modules is from 0xffffffffa00000000 to 0xffffffff000000, so we need to find them in this space. As long as the modules are loaded through init module system call in a normal way, they must be in this space, so we only need to scan the memory of this space to match the two key features:

• The self referential feature of module core.
• ktype feature of kobject of module.

OK, let's look at how to find the hidden module in this address space, that is, the module of chain removal:

• Scan the gaps!!

Come on:

#!/bin/bash start='' end='' base='' moktype=$(cat /proc/kallsyms|grep module_ktype|awk '') moktype=$(echo $moktype|tr 'a-z' 'A-Z') for line in $(cat /proc/vmallocinfo |grep 0xffffffffa|awk '') do start=$(echo $line|awk -F '-' ''|awk -F '0x' '') start=$(echo $start|tr 'a-z' 'A-Z') if [ $start == 'FFFFFFFFA0000000' ]; then end=$(echo $line|awk -F '-' ''|awk -F '0x' '') end=$(echo $end|tr 'a-z' 'A-Z') continue; fi if [ $start == $end ];then end=$(echo $line|awk -F '-' ''|awk -F '0x' '') end=$(echo $end|tr 'a-z' 'A-Z') continue; fi base=$end next=$base end=$(echo $line|awk -F '-' ''|awk -F '0x' '') end=$(echo $end|tr 'a-z' 'A-Z') while true; do val=$(./a.out $next); if [ $? -ne 0 ]; then break; fi if [ $val == $base ]; then mod=$(echo "ibase=16;$next-138"|bc) mod=$(echo "obase=16;$mod"|bc) state=$(./a.out $mod) if [ $? -ne 0 ] || [ $state != '0' ]; then next=$(echo "ibase=16;$next+8"|bc) next=$(echo "obase=16;$next"|bc) continue; fi ktype=$(echo "ibase=16;$mod+78"|bc) ktype=$(echo "obase=16;$ktype"|bc) type=$(./a.out $ktype) if [ $? -ne 0 ] || [ $type != $moktype ]; then next=$(echo "ibase=16;$next+8"|bc) next=$(echo "obase=16;$next"|bc) continue; fi namea=$(echo "ibase=16;$mod+18"|bc) namea=$(echo "obase=16;$namea"|bc) name=$(./a.out $namea) name=$(echo -n $name|sed 's/\([0-9A-F]\{2\}\)/\\\\\\x\1/gI' | xargs printf) name=$(echo $name|rev 2>/dev/null) if [ $? -eq 0 ]; then echo name-- $name fi fi next=$(echo "ibase=16;$next+8"|bc) next=$(echo "obase=16;$next"|bc) done done;

Try it. As long as the module is moved after the insmod command is loaded, it's easy to find out the hidden module, only for a long time.

It's a long time because my script is very inefficient. In fact, if every page is mapped once, it will be much better. But because I can't program, I have to do this first.

As mentioned earlier, as long as the modules loaded through the insmod command, i.e. init module system call, can be found. What if you don't load the module in this normal way?

In fact, think about the number of entries into the kernel, especially the entry of code into the kernel

• init_module
• ptrace
• ftrace
• eBPF
• ...

There are not many in nature. Init module is the most commonly used and easiest. If you don't use init module, what else can you use?

Wenzhou leather shoes in Zhejiang Province are wet, and they will not be fat if it rains and floods.

16 May 2020, 23:53 | Views: 2642