* Re: [PATCH v2] efifb: prevent null dereferences by removing unused array indices from dmi_list
[not found] ` <1376684395.6529.5.camel@hermes>
@ 2013-08-16 20:36 ` David Herrmann
2013-09-05 8:12 ` Tomi Valkeinen
1 sibling, 0 replies; 4+ messages in thread
From: David Herrmann @ 2013-08-16 20:36 UTC (permalink / raw)
To: James Bates
Cc: Peter Jones, Jean-Christophe Plagniol-Villard, Tomi Valkeinen,
linux-fbdev@vger.kernel.org, linux-kernel, H. Peter Anvin
Hi
On Fri, Aug 16, 2013 at 10:19 PM, James Bates <james.h.bates@gmail.com> wrote:
> On Fri, 2013-08-16 at 15:37 +0200, David Herrmann wrote:
>
> Hi
> (CC'ing hpa)
>
> Yepp, this patch looks good:
> Reviewed-by: David Herrmann <dh.herrmann@gmail.com>
>
> However, I'd like to see some code to prevent this from happening
> again. It isn't really obvious that removing an entry will result in a
> NULL-deref. I am not the maintainer of this code, but I'd really like
> to see a "(dmi_list[i].optname && !strcmp())" check in efifb_setup().
> Otherwise we _will_ run into this again.
>
> Side note: this code got moved to arch/x86/kernel/sysfb_efi.c in the
> x86 tree. I am adding hpa here so he will remember this once Linus
> gets a merge conflict (iff the sysfb changes get merged through the
> x86 tree).
>
> Thanks
> David
>
> Sure thing, here is the patch again, with the extra check in efifb_setup()
> (it turns out one can simply
> switch around the order of the checks: implicitly initialized dmi_list items
> will have base == 0):
You can put such comments below the "---" in your patch so they will
not show up in the commit-message (between commit message and the
diffstat).
> Full patch v2:
>
> the dmi_list array is initialized using gnu designated initializers, and
> therefore
> contains fewer explicitly defined entries as there are elements in it. This
> is because the enum above with M_blabla constants contains more items than
> the designated
> initializer. Those elements not explicitly initialized are implicitly set to
> 0.
>
> Now efifb_setup(), L.322 & L.323, loops through all these array elements,
> and performs
> a strcmp o a field (optname) in each item. For non explicitly initialized
> elements this
> will be a null pointer:
>
> for (i = 0; i < M_UNKNOWN; i++) {
> if (!strcmp(this_opt, dmi_list[i].optname)
> &&
>
> On my macbook6,1 the predefined values are for some reason incorrect, and
> most parameters
> are preset correctly by my efi bootloader (elilo). but stride/line_length is
> not detected
> correctly and so I wish to set it explicitly using a
> "video=efifb:stride:2048" command-line
> argument. Because of the above null dereference, an exception (presumably)
> occurs before
> the parsing code (L.333) is ever reached. I say presumably since the mac
> hangs on boot
> without a console, and I can therefore not see any output.
>
> By removing the unused values from the enum, and thus preventing implicitly
> initialized items
> in the dmi_list array, the null dereference does not occur, my customer
> command-line arg is
> parsed correctly, and my console displays correctly.
>
> This patch removes the unused enum values, and also guards against any
> future implicit
> initializing by inverting the check order in the if statement, and checking
> first whether
> dmi_list[i].base is null.
>
> Signed-off-by: James Bates <james.h.bates@yahoo.com>
> ---
> drivers/video/efifb.c | 7 ++-----
> 1 file changed, 2 insertions(+), 5 deletions(-)
>
> diff --git a/drivers/video/efifb.c b/drivers/video/efifb.c
> index 50fe668..161757b 100644
>
> --- a/drivers/video/efifb.c
> +++ b/drivers/video/efifb.c
> @@ -50,12 +50,9 @@ enum {
> M_MINI_3_1, /* Mac Mini, 3,1th gen */
> M_MINI_4_1, /* Mac Mini, 4,1th gen */
> M_MB, /* MacBook */
> - M_MB_2, /* MacBook, 2nd rev. */
> - M_MB_3, /* MacBook, 3rd rev. */
> M_MB_5_1, /* MacBook, 5th rev. */
> M_MB_6_1, /* MacBook, 6th rev. */
> M_MB_7_1, /* MacBook, 7th rev. */
> - M_MB_SR, /* MacBook, 2nd gen, (Santa Rosa) */
> M_MBA, /* MacBook Air */
> M_MBA_3, /* Macbook Air, 3rd rev */
> M_MBP, /* MacBook Pro */
> @@ -323,8 +320,8 @@ static int __init efifb_setup(char *options)
> if (!*this_opt) continue;
>
>
> for (i = 0; i < M_UNKNOWN; i++) {
> - if (!strcmp(this_opt, dmi_list[i].optname) &&
> - dmi_list[i].base != 0) {
> + if (dmi_list[i].base != 0 &&
>From a quick glance, M_MBA_3 has "base == 0" but is a valid entry. Hm,
I think we can be sure that there will never be a framebuffer at phys
0, so yeah, I am fine with this.
Reviewed-by: David Herrmann <dh.herrmann@gmail.com>
Regards
David
> + !strcmp(this_opt, dmi_list[i].optname)) {
> screen_info.lfb_base = dmi_list[i].base;
> screen_info.lfb_linelength = dmi_list[i].stride;
> screen_info.lfb_width = dmi_list[i].width;
> --
>
^ permalink raw reply [flat|nested] 4+ messages in thread* Re: [PATCH v2] efifb: prevent null dereferences by removing unused array indices from dmi_list
[not found] ` <1376684395.6529.5.camel@hermes>
2013-08-16 20:36 ` [PATCH v2] " David Herrmann
@ 2013-09-05 8:12 ` Tomi Valkeinen
1 sibling, 0 replies; 4+ messages in thread
From: Tomi Valkeinen @ 2013-09-05 8:12 UTC (permalink / raw)
To: James Bates
Cc: David Herrmann, Peter Jones, Jean-Christophe Plagniol-Villard,
linux-fbdev, linux-kernel, H. Peter Anvin
[-- Attachment #1: Type: text/plain, Size: 4245 bytes --]
Hi James,
Can you resend this? I don't know how I could apply the patch from this
mail without re-creating it manually.
Tomi
On 16/08/13 23:19, James Bates wrote:
> On Fri, 2013-08-16 at 15:37 +0200, David Herrmann wrote:
>> Hi
>> (CC'ing hpa)
>>
>> Yepp, this patch looks good:
>> Reviewed-by: David Herrmann <dh.herrmann@gmail.com <mailto:dh.herrmann@gmail.com>>
>>
>> However, I'd like to see some code to prevent this from happening
>> again. It isn't really obvious that removing an entry will result in a
>> NULL-deref. I am not the maintainer of this code, but I'd really like
>> to see a "(dmi_list[i].optname && !strcmp())" check in efifb_setup().
>> Otherwise we _will_ run into this again.
>>
>> Side note: this code got moved to arch/x86/kernel/sysfb_efi.c in the
>> x86 tree. I am adding hpa here so he will remember this once Linus
>> gets a merge conflict (iff the sysfb changes get merged through the
>> x86 tree).
>>
>> Thanks
>> David
> Sure thing, here is the patch again, with the extra check in
> efifb_setup() (it turns out one can simply
> switch around the order of the checks: implicitly initialized dmi_list
> items will have base == 0):
>
> Full patch v2:
> the dmi_list array is initialized using gnu designated initializers, and
> therefore
> contains fewer explicitly defined entries as there are elements in it. This
> is because the enum above with M_blabla constants contains more items
> than the designated
> initializer. Those elements not explicitly initialized are implicitly
> set to 0.
>
> Now efifb_setup(), L.322 & L.323, loops through all these array
> elements, and performs
> a strcmp o a field (optname) in each item. For non explicitly
> initialized elements this
> will be a null pointer:
>
> for (i = 0; i < M_UNKNOWN; i++) {
> if (!strcmp(this_opt,
> dmi_list[i].optname) &&
>
> On my macbook6,1 the predefined values are for some reason incorrect,
> and most parameters
> are preset correctly by my efi bootloader (elilo). but
> stride/line_length is not detected
> correctly and so I wish to set it explicitly using a
> "video=efifb:stride:2048" command-line
> argument. Because of the above null dereference, an exception
> (presumably) occurs before
> the parsing code (L.333) is ever reached. I say presumably since the mac
> hangs on boot
> without a console, and I can therefore not see any output.
>
> By removing the unused values from the enum, and thus preventing
> implicitly initialized items
> in the dmi_list array, the null dereference does not occur, my customer
> command-line arg is
> parsed correctly, and my console displays correctly.
>
> This patch removes the unused enum values, and also guards against any
> future implicit
> initializing by inverting the check order in the if statement, and
> checking first whether
> dmi_list[i].base is null.
>
> Signed-off-by: James Bates <james.h.bates@yahoo.com
> <mailto:james.h.bates@yahoo.com>>
> ---
> drivers/video/efifb.c | 7 ++-----
> 1 file changed, 2 insertions(+), 5 deletions(-)
>
> diff --git a/drivers/video/efifb.c b/drivers/video/efifb.c
> index 50fe668..161757b 100644
> --- a/drivers/video/efifb.c
> +++ b/drivers/video/efifb.c
> @@ -50,12 +50,9 @@ enum {
> M_MINI_3_1, /* Mac Mini, 3,1th gen */
> M_MINI_4_1, /* Mac Mini, 4,1th gen */
> M_MB, /* MacBook */
> - M_MB_2, /* MacBook, 2nd rev. */
> - M_MB_3, /* MacBook, 3rd rev. */
> M_MB_5_1, /* MacBook, 5th rev. */
> M_MB_6_1, /* MacBook, 6th rev. */
> M_MB_7_1, /* MacBook, 7th rev. */
> - M_MB_SR, /* MacBook, 2nd gen, (Santa Rosa) */
> M_MBA, /* MacBook Air */
> M_MBA_3, /* Macbook Air, 3rd rev */
> M_MBP, /* MacBook Pro */
> @@ -323,8 +320,8 @@ static int __init efifb_setup(char *options)
> if (!*this_opt) continue;
>
> for (i = 0; i < M_UNKNOWN; i++) {
> - if (!strcmp(this_opt, dmi_list[i].optname) &&
> - dmi_list[i].base != 0) {
> + if (dmi_list[i].base != 0 &&
> + !strcmp(this_opt, dmi_list[i].optname)) {
> screen_info.lfb_base = dmi_list[i].base;
> screen_info.lfb_linelength = dmi_list[i].stride;
> screen_info.lfb_width = dmi_list[i].width;
> --
>
[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 901 bytes --]
^ permalink raw reply [flat|nested] 4+ messages in thread