> On Aug 5, 2015, at 05:18, Dave Hansen <d...@sr71.net> wrote:
>
> On 08/03/2015 08:37 PM, yalin wang wrote:
>> This change kcore_read() to use __copy_from_user_inatomic() to
>> copy data from kernel address, because kern_addr_valid() just make sure
>> page table is valid during call it, whne it return, the page table may
>> change, for example, like set_fixmap() function will change kernel page
>> table, then maybe trigger kernel crash if encounter this unluckily.
>
> I don't see any cases at the moment that will crash. set_fixmap()
> doesn't ever clear out any ptes, right?
>
> I guess the root problem here is that we don't have any good (generic)
> locking of kernel page tables inside the linear map. Can you come up
> with a case where this will _actually_ crash?
>
Thanks for your comments.
i don’t have crash for this, but when i read code, i see this part not safe,
so i make this patch :).
>> fs/proc/kcore.c | 30 ++++++++++++++++++++++++------
>> 1 file changed, 24 insertions(+), 6 deletions(-)
>>
>> diff --git a/fs/proc/kcore.c b/fs/proc/kcore.c
>> index 92e6726..b085fde 100644
>> --- a/fs/proc/kcore.c
>> +++ b/fs/proc/kcore.c
>> @@ -86,8 +86,8 @@ static size_t get_kcore_size(int *nphdr, size_t
>> *elf_buflen)
>> size = try;
>> *nphdr = *nphdr + 1;
>> }
>> - *elf_buflen = sizeof(struct elfhdr) +
>> - (*nphdr + 2)*sizeof(struct elf_phdr) +
>> + *elf_buflen = sizeof(struct elfhdr) +
>> + (*nphdr + 2)*sizeof(struct elf_phdr) +
>
> I'm having a hard time spotting the change here. Whitespace?
i will seperate in another patch for format correctness.
>
>> 3 * ((sizeof(struct elf_note)) +
>> roundup(sizeof(CORE_STR), 4)) +
>> roundup(sizeof(struct elf_prstatus), 4) +
>> @@ -435,6 +435,7 @@ read_kcore(struct file *file, char __user *buffer,
>> size_t buflen, loff_t *fpos)
>> size_t elf_buflen;
>> int nphdr;
>> unsigned long start;
>> + unsigned long page = 0;
>>
>> read_lock(&kclist_lock);
>> size = get_kcore_size(&nphdr, &elf_buflen);
>> @@ -485,7 +486,7 @@ read_kcore(struct file *file, char __user *buffer,
>> size_t buflen, loff_t *fpos)
>> start = kc_offset_to_vaddr(*fpos - elf_buflen);
>> if ((tsz = (PAGE_SIZE - (start & ~PAGE_MASK))) > buflen)
>> tsz = buflen;
>> -
>> +
>
> Please keep the unnecessary whitespace changes for another patch.
>
>> while (buflen) {
>> struct kcore_list *m;
>>
>> @@ -515,15 +516,32 @@ read_kcore(struct file *file, char __user *buffer,
>> size_t buflen, loff_t *fpos)
>> } else {
>> if (kern_addr_valid(start)) {
>> unsigned long n;
>> + mm_segment_t old_fs = get_fs();
>> +
>> + if (page == 0) {
>> + page = __get_free_page(GFP_KERNEL);
>> + if (page == 0)
>> + return -ENOMEM;
>
> FWIW, we usually code this as "!page" instead of "page == 0". I also
> wouldn't call it 'page'.
>
> Also, why is this using a raw __get_free_page() while the code above it
> uses a kmalloc()?
>
because i am using a page size buffer, more efficient to use __get_free_page()
than kmalloc() here .
>> - n = copy_to_user(buffer, (char *)start, tsz);
>> + }
>> + set_fs(KERNEL_DS);
>> + pagefault_disable();
>> + n = __copy_from_user_inatomic((void *)page,
>> + (__force const void __user *)start,
>> + tsz);
>> + pagefault_enable();
>> + set_fs(old_fs);
>> + if (n)
>> + memset((void *)page + tsz - n, 0, n);
>> +
>> + n = copy_to_user(buffer, (char *)page, tsz);
>
> So, first of all, we are using __copy_from_user_inatomic() to copy to
> and from a *kernel* addresses, and it doesn't even get a comment? :)
>
i will add comment.
> Fundamentally, we're trying to be able to safely survive faults in the
> kernel linear map here. I think we've got to get a better handle on
> when that happens rather than just paper over it when it does. (Aside:
> There might actually be a missing use of get_online_mems() here.)
>
ok.
> Maybe we should just be walking the kernel page tables ourselves and do
> a kmap(). We might have a stale pte but we don't have to worry about
> actual racy updates while we are doing the copy.
>
so if do like this, we can remove kern_addr_valid() function, and i just walk
pte and use get_page_unelss_zero()
to grab the valid page ?
>> /*
>> * We cannot distinguish between fault on source
>> * and fault on destination. When this happens
>> * we clear too and hope it will trigger the
>> * EFAULT again.
>> */
>
> This comment seems wrong after the patch.
Ok.--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
