Skip to contentSkip to conversation

Build scatterlist covering a process' text segment?

DISCUSSION Arvid Brodin 5 participants 12 messages May 23, 2011 – Jun 17, 2011

lore
Share & tools
lore

Arvid Brodin

author
#0 lore raw

Hi,

How do I build a scatterlist covering the text segment of a process? I have found (struct task_struct).mm->start_code and ->end_code, that gives the address (in some - unknown to me - address space) of the text segment. Since my try to convert these addresses to an sglist:

	sg_init_one(&sg, task->mm->start_code, <length>);

failed miserably, I'm guessing these addresses need to be converted to some address space compatible with sg_init_one()/sg_set_buf() before use? (I probably need to search for pages containing these addresses?)

Thanks,
Arvid Brodin
Enea Services Stockholm AB

author
#1 lore raw
Quoted 7 lines · Arvid Brodin

Hi,

How do I build a scatterlist covering the text segment of a process? I have found (struct task_struct).mm->start_code and ->end_code, that gives the address (in some - unknown to me - address space) of the text segment. Since my try to convert these addresses to an sglist:

Quoted 6 lines
	sg_init_one(&sg, task->mm->start_code, <length>);

failed miserably, I'm guessing these addresses need to be converted to some address space compatible with sg_init_one()/sg_set_buf() before use? (I probably need to search for pages containing these addresses?)

Am I correct in that the addresses in (struct task_struct).mm->start_code and ->end_code belong to the address space of the process whose task_struct I'm looking at?

Signature · 2 lines
Arvid Brodin
Enea Services Stockholm AB
#2 lore raw
Quoted 5 lines · Arvid Brodin

Am I correct in that the addresses in (struct task_struct).mm->start_code and
->end_code belong to the address space of the process whose task_struct I'm looking
at?

I believe yes....

Signature · 7 lines
regards,

Mulyadi Santosa
Freelance Linux trainer and consultant

blog: the-hydra.blogspot.com
training: mulyaditraining.blogspot.com
author
#3 lore raw
Quoted 10 lines · Mulyadi Santosa
Quoted 5 lines · Arvid Brodin

Am I correct in that the addresses in (struct task_struct).mm->start_code and
->end_code belong to the address space of the process whose task_struct I'm looking
at?

I believe yes....

Ok. And looking at e.g. sg_set_buf(), the scatterlist expects a kernel virtual address (it uses virt_to_page() on its "buf" parameter internally, which requires a kernel virtual adress, if I understand correctly).

There seems to be no way to map process adresses to kernel addresses. (Well I guess one could follow the page tables to get the physical page, and then map back to kernel space, but this only works as long as the memory is paged in.) Please correct me if I'm wrong.

Perhaps a better idea is to do_mmap() the "exe_file", (struct mm_struct).exe_file? I'm still confused about this though:

1) do_mmap() returns unsigned long - not the void * expected by sg_set_buf(). Is this a sign that I'm again mixing different address spaces, or is it ok just to cast this?

2) I believe do_mmap() maps a file into the address space of the "current process", and that this address space is also the one used in kernel space at any moment. Is this correct, or is there an additional, kernel-only memory space that I should map the file to instead? If so, how do I do that?

Thanks,
Arvid Brodin
Enea Services Stockholm AB

#4 lore raw

Hi...

Quoted 8 lines · Arvid Brodin

Ok. And looking at e.g. sg_set_buf(), the scatterlist expects a kernel virtual address (it uses virt_to_page() on its "buf" parameter internally, which requires a kernel virtual adress, if I understand correctly).

There seems to be no way to map process adresses to kernel addresses. (Well I guess one could follow the page tables to get the physical page, and then map back to kernel space, but this only works as long as the memory is paged in.) Please correct me if I'm wrong.

I am not good at it, but I think at the first place, you might use get_user_pages() (take a look here :http://lxr.linux.no/#linux+v2.6.39/mm/memory.c#L1703)

then once you get the pointer to the pages (and making sure they're pinned by get_user_pages), I think you just need to use kmap().

I suggest to really observe that scatter gather function and see if address in kernel address space is really needed.... if not, you can avoid using kmap() completely.

Signature · 7 lines
regards,

Mulyadi Santosa
Freelance Linux trainer and consultant

blog: the-hydra.blogspot.com
training: mulyaditraining.blogspot.com
author
#5 lore raw
Quoted 16 lines · Mulyadi Santosa

Hi...

Quoted 8 lines · Arvid Brodin

Ok. And looking at e.g. sg_set_buf(), the scatterlist expects a kernel virtual address (it uses virt_to_page() on its "buf" parameter internally, which requires a kernel virtual adress, if I understand correctly).

There seems to be no way to map process adresses to kernel addresses. (Well I guess one could follow the page tables to get the physical page, and then map back to kernel space, but this only works as long as the memory is paged in.) Please correct me if I'm wrong.

I am not good at it, but I think at the first place, you might use get_user_pages() (take a look here :http://lxr.linux.no/#linux+v2.6.39/mm/memory.c#L1703)

Quoted 7 lines

then once you get the pointer to the pages (and making sure they're pinned by get_user_pages), I think you just need to use kmap().

I suggest to really observe that scatter gather function and see if address in kernel address space is really needed.... if not, you can avoid using kmap() completely.

Many thanks for that tip! I've been reading LDD3 ("Linux Device Drivers 3rd edition") and "Understanding the Linux Kernel" as well as searching google and of course reading code for many hours, and nowhere has there been a mention of get_user_pages()! Obviously I haven't been using the right search terms...

(Also, for some reason your mail didn't reach my inbox, so I only saw it a few days ago when looking at a mail list archive.)

Anyway, the code looks like this now (below), and seems to work on 2.7.37.6. It will probably have problems on systems with > 896 MiB and high memory enabled though; I'm thinking the scatterlist functions won't handle virtual kernel addresses?

Any ideas for improvements are welcome; especially if I'm doing something stupid that risks a kernel panic.

static int proc_pid_text_checksum(struct seq_file *m, struct pid_namespace *ns,
struct pid *pid, struct task_struct *task)

{
	int retval;

	int text_size;
	int nr_pages, nr_pages_mapped;
	int i;
	struct page **pages;

	struct scatterlist *sgl, *sg;

	u8 result[MD5_DIGEST_SIZE + 2];
	struct crypto_hash *tfm;
	struct hash_desc desc;

	retval = 0;

	if (!task->mm)
		return -EINVAL;

	text_size = task->mm->end_code - task->mm->start_code;
	nr_pages = (text_size + PAGE_SIZE - 1) >> PAGE_SHIFT;


	/**** User page code ****/

	pages = kmalloc(nr_pages * sizeof(*pages), GFP_KERNEL);
	if (!pages) {
		retval = -ENOMEM;
		goto err_pages;
	}

	down_read(&task->mm->mmap_sem);
	nr_pages_mapped = get_user_pages(current, task->mm,
			task->mm->start_code, nr_pages, 0, 0, pages, NULL);
	up_read(&task->mm->mmap_sem);

	if (nr_pages_mapped < nr_pages) {
		retval = -EBUSY; /* Weird error code for this,
		                    couldn't find any better */
		goto err_mapped;
	}


	/**** Scatterlist code ****/

	sgl = kmalloc(nr_pages_mapped * sizeof(*sgl), GFP_KERNEL);
	if (!sgl) {
		retval = -ENOMEM;
		goto err_sg;
	}

	sg_init_table(sgl, nr_pages_mapped);

	for_each_sg(sgl, sg, nr_pages_mapped, i)
		sg_set_page(sg, pages[i], (i < nr_pages_mapped) ? PAGE_SIZE :
						text_size & ~PAGE_MASK, 0);


	/**** Crypto code ****/

	tfm = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
	if (IS_ERR(tfm)) {
		retval = -ENOMEM;
		goto err_crypto;
	}

	desc.tfm = tfm;
	desc.flags = CRYPTO_TFM_REQ_MAY_SLEEP;

	memset(result, 0, MD5_DIGEST_SIZE + 2);
	retval = crypto_hash_digest(&desc, sgl, text_size, result);
	if (retval)
		goto err_digest;

	for (i = 0; i < MD5_DIGEST_SIZE; i++)
		seq_printf(m, "%02x", result[i]);
	seq_printf(m, "\n");

err_digest:

	crypto_free_hash(tfm);

err_crypto:

	kfree(sgl);
	for (i = 0; i < nr_pages_mapped; i++)
		put_page(pages[i]);

err_sg:
err_mapped:

	kfree(pages);

err_pages:

	return retval;
}
Signature · 2 lines
Arvid Brodin
Enea Services Stockholm AB
#6 lore raw

On Fri, Jun 17, 2011 at 02:47:25AM +0200, Arvid Brodin wrote:
[...]

Quoted 3 lines
	/**** User page code ****/

	pages = kmalloc(nr_pages * sizeof(*pages), GFP_KERNEL);

kcalloc?

You don't need the zeroing but the overflow check might be useful.

Quoted 4 lines
	if (!pages) {
		retval = -ENOMEM;
		goto err_pages;
	}

Greetings,
Jonathan Neusch?fer

#7 lore raw

Jonathan, it's kmalloc(), not kcalloc().

Quoted 22 lines · Jonathan Neusch?fer

On Fri, Jun 17, 2011 at 02:47:25AM +0200, Arvid Brodin wrote:
[...]

Quoted 3 lines
   /**** User page code ****/

   pages = kmalloc(nr_pages * sizeof(*pages), GFP_KERNEL);

kcalloc?

You don't need the zeroing but the overflow check might be useful.

Quoted 4 lines
   if (!pages) {
       retval = -ENOMEM;
       goto err_pages;
   }

Greetings,
Jonathan Neusch?fer

_______________________________________________
Kernelnewbies mailing list
Kernelnewbies at kernelnewbies.org
http://lists.kernelnewbies.org/mailman/listinfo/kernelnewbies

#8 lore raw

Hi,

Quoted 1 line · Metin KAYA

Jonathan, it's kmalloc(), not kcalloc().

The kernel's kmalloc doesn't initialize the memory. There is a function called kzalloc which will allocate and zero the memory, which is sort of similar to the user-space calloc function, although calloc takes num-of-items and size-of-one-item, whereas kzalloc takes "size-of-allocation".

Signature · 3 lines
Dave Hylands
Shuswap, BC, Canada
http://www.davehylands.com
#9 lore raw
In reply to Metin KAYA
Quoted 1 line · Metin KAYA

Jonathan, it's kmalloc(), not kcalloc().

No, I was trying to suggest using kcalloc(), which behaves roughly like the userspace calloc, and implements overflow checking, which you may need and zeroing, which you don't seem to need.

(see include/linux/slab.h, line 223 for reference)

Quoted 13 lines · Jonathan Neusch?fer
Quoted 11 lines · Arvid Brodin

[...]

Quoted 3 lines
   /**** User page code ****/

   pages = kmalloc(nr_pages * sizeof(*pages), GFP_KERNEL);

kcalloc?

You don't need the zeroing but the overflow check might be useful.

Quoted 7 lines
Quoted 7 lines
Quoted 4 lines
   if (!pages) {
       retval = -ENOMEM;
       goto err_pages;
   }

Greetings,
Jonathan Neusch?fer

HTH,
Jonathan Neusch?fer

#10 lore raw

Hey Jonathan,

Quoted 6 lines · Jonathan Neusch?fer
Quoted 1 line · Metin KAYA

Jonathan, it's kmalloc(), not kcalloc().

No, I was trying to suggest using kcalloc(), which behaves roughly like the userspace calloc, and implements overflow checking, which you may need and zeroing, which you don't seem to need.

Cool - I missed that somehow. So we now have kmalloc, kcalloc, and kzalloc, and I see krealloc too.

Signature · 3 lines
Dave Hylands
Shuswap, BC, Canada
http://www.davehylands.com
#11 lore raw
In reply to Jonathan Neuschäfer
Quoted 6 lines · Jonathan Neusch?fer
Quoted 1 line · Metin KAYA

Jonathan, it's kmalloc(), not kcalloc().

No, I was trying to suggest using kcalloc(), which behaves roughly like the userspace calloc, and implements overflow checking, which you may need and zeroing, which you don't seem to need.

Arvid, not you, sorry.

Jonathan Neusch?fer