2020-04-15 11:45:00

by Christoph Hellwig

[permalink] [raw]
Subject: [PATCH 8/8] exec: open code copy_string_kernel

Currently copy_string_kernel is just a wrapper around copy_strings that
simplifies the calling conventions and uses set_fs to allow passing a
kernel pointer. But due to the fact the we only need to handle a single
kernel argument pointer, the logic can be sigificantly simplified while
getting rid of the set_fs.

Signed-off-by: Christoph Hellwig <[email protected]>
---
fs/exec.c | 43 ++++++++++++++++++++++++++++++++++---------
1 file changed, 34 insertions(+), 9 deletions(-)

diff --git a/fs/exec.c b/fs/exec.c
index b2a77d5acede..ea90af1fb236 100644
--- a/fs/exec.c
+++ b/fs/exec.c
@@ -592,17 +592,42 @@ static int copy_strings(int argc, struct user_arg_ptr argv,
*/
int copy_string_kernel(const char *arg, struct linux_binprm *bprm)
{
- int r;
- mm_segment_t oldfs = get_fs();
- struct user_arg_ptr argv = {
- .ptr.native = (const char __user *const __user *)&arg,
- };
+ int len = strnlen(arg, MAX_ARG_STRLEN) + 1 /* terminating NUL */;
+ unsigned long pos = bprm->p;
+
+ if (len == 0)
+ return -EFAULT;
+ if (!valid_arg_len(bprm, len))
+ return -E2BIG;
+
+ /* We're going to work our way backwards. */
+ arg += len;
+ bprm->p -= len;
+ if (IS_ENABLED(CONFIG_MMU) && bprm->p < bprm->argmin)
+ return -E2BIG;
+
+ while (len > 0) {
+ unsigned int bytes_to_copy = min_t(unsigned int, len,
+ min_not_zero(offset_in_page(pos), PAGE_SIZE));
+ struct page *page;
+ char *kaddr;

- set_fs(KERNEL_DS);
- r = copy_strings(1, argv, bprm);
- set_fs(oldfs);
+ pos -= bytes_to_copy;
+ arg -= bytes_to_copy;
+ len -= bytes_to_copy;

- return r;
+ page = get_arg_page(bprm, pos, 1);
+ if (!page)
+ return -E2BIG;
+ kaddr = kmap_atomic(page);
+ flush_arg_page(bprm, pos & PAGE_MASK, page);
+ memcpy(kaddr + offset_in_page(pos), arg, bytes_to_copy);
+ flush_kernel_dcache_page(page);
+ kunmap_atomic(kaddr);
+ put_arg_page(page);
+ }
+
+ return 0;
}
EXPORT_SYMBOL(copy_string_kernel);

--
2.25.1


2020-04-18 08:19:09

by Christophe Leroy

[permalink] [raw]
Subject: Re: [PATCH 8/8] exec: open code copy_string_kernel



Le 14/04/2020 à 09:01, Christoph Hellwig a écrit :
> Currently copy_string_kernel is just a wrapper around copy_strings that
> simplifies the calling conventions and uses set_fs to allow passing a
> kernel pointer. But due to the fact the we only need to handle a single
> kernel argument pointer, the logic can be sigificantly simplified while
> getting rid of the set_fs.


Instead of duplicating almost identical code, can you write a function
that takes whether the source is from user or from kernel, then you just
do things like:

if (from_user)
len = strnlen_user(str, MAX_ARG_STRLEN);
else
len = strnlen(str, MAX_ARG_STRLEN);


if (from_user)
copy_from_user(kaddr+offset, str, bytes_to_copy);
else
memcpy(kaddr+offset, str, bytes_to_copy);

>
> Signed-off-by: Christoph Hellwig <[email protected]>
> ---
> fs/exec.c | 43 ++++++++++++++++++++++++++++++++++---------
> 1 file changed, 34 insertions(+), 9 deletions(-)
>
> diff --git a/fs/exec.c b/fs/exec.c
> index b2a77d5acede..ea90af1fb236 100644
> --- a/fs/exec.c
> +++ b/fs/exec.c
> @@ -592,17 +592,42 @@ static int copy_strings(int argc, struct user_arg_ptr argv,
> */
> int copy_string_kernel(const char *arg, struct linux_binprm *bprm)
> {
> - int r;
> - mm_segment_t oldfs = get_fs();
> - struct user_arg_ptr argv = {
> - .ptr.native = (const char __user *const __user *)&arg,
> - };
> + int len = strnlen(arg, MAX_ARG_STRLEN) + 1 /* terminating NUL */;
> + unsigned long pos = bprm->p;
> +
> + if (len == 0)
> + return -EFAULT;
> + if (!valid_arg_len(bprm, len))
> + return -E2BIG;
> +
> + /* We're going to work our way backwards. */
> + arg += len;
> + bprm->p -= len;
> + if (IS_ENABLED(CONFIG_MMU) && bprm->p < bprm->argmin)
> + return -E2BIG;
> +
> + while (len > 0) {
> + unsigned int bytes_to_copy = min_t(unsigned int, len,
> + min_not_zero(offset_in_page(pos), PAGE_SIZE));
> + struct page *page;
> + char *kaddr;
>
> - set_fs(KERNEL_DS);
> - r = copy_strings(1, argv, bprm);
> - set_fs(oldfs);
> + pos -= bytes_to_copy;
> + arg -= bytes_to_copy;
> + len -= bytes_to_copy;
>
> - return r;
> + page = get_arg_page(bprm, pos, 1);
> + if (!page)
> + return -E2BIG;
> + kaddr = kmap_atomic(page);
> + flush_arg_page(bprm, pos & PAGE_MASK, page);
> + memcpy(kaddr + offset_in_page(pos), arg, bytes_to_copy);
> + flush_kernel_dcache_page(page);
> + kunmap_atomic(kaddr);
> + put_arg_page(page);
> + }
> +
> + return 0;
> }
> EXPORT_SYMBOL(copy_string_kernel);
>
>

Christophe

2020-04-19 08:10:34

by Christoph Hellwig

[permalink] [raw]
Subject: Re: [PATCH 8/8] exec: open code copy_string_kernel

On Sat, Apr 18, 2020 at 10:15:42AM +0200, Christophe Leroy wrote:
>
>
> Le 14/04/2020 ? 09:01, Christoph Hellwig a ?crit?:
>> Currently copy_string_kernel is just a wrapper around copy_strings that
>> simplifies the calling conventions and uses set_fs to allow passing a
>> kernel pointer. But due to the fact the we only need to handle a single
>> kernel argument pointer, the logic can be sigificantly simplified while
>> getting rid of the set_fs.
>
>
> Instead of duplicating almost identical code, can you write a function that
> takes whether the source is from user or from kernel, then you just do
> things like:
>
> if (from_user)
> len = strnlen_user(str, MAX_ARG_STRLEN);
> else
> len = strnlen(str, MAX_ARG_STRLEN);
>
>
> if (from_user)
> copy_from_user(kaddr+offset, str, bytes_to_copy);
> else
> memcpy(kaddr+offset, str, bytes_to_copy);

We'll need two different str variables then with and without __user
annotations to keep type safety. And introduce a branch-y and unreadable
mess in the exec fast path instead of adding a simple and well understood
function for the kernel case that just deals with the much simpler case
of just copying a single arg vector from a kernel address.

2020-04-19 09:46:11

by Christophe Leroy

[permalink] [raw]
Subject: Re: [PATCH 8/8] exec: open code copy_string_kernel



Le 19/04/2020 à 10:06, Christoph Hellwig a écrit :
> On Sat, Apr 18, 2020 at 10:15:42AM +0200, Christophe Leroy wrote:
>>
>>
>> Le 14/04/2020 à 09:01, Christoph Hellwig a écrit :
>>> Currently copy_string_kernel is just a wrapper around copy_strings that
>>> simplifies the calling conventions and uses set_fs to allow passing a
>>> kernel pointer. But due to the fact the we only need to handle a single
>>> kernel argument pointer, the logic can be sigificantly simplified while
>>> getting rid of the set_fs.
>>
>>
>> Instead of duplicating almost identical code, can you write a function that
>> takes whether the source is from user or from kernel, then you just do
>> things like:
>>
>> if (from_user)
>> len = strnlen_user(str, MAX_ARG_STRLEN);
>> else
>> len = strnlen(str, MAX_ARG_STRLEN);
>>
>>
>> if (from_user)
>> copy_from_user(kaddr+offset, str, bytes_to_copy);
>> else
>> memcpy(kaddr+offset, str, bytes_to_copy);
>
> We'll need two different str variables then with and without __user
> annotations to keep type safety. And introduce a branch-y and unreadable
> mess in the exec fast path instead of adding a simple and well understood
> function for the kernel case that just deals with the much simpler case
> of just copying a single arg vector from a kernel address.
>

About the branch, I was expecting GCC to inline and eliminate the unused
branch.