Message ID | 10e97a04980d933b2cfecb6b124bf9046b6e4f16.1582216144.git.leonard.crestez@nxp.com |
---|---|
State | New |
Headers | show |
Series | firmware: imx: Align imx SC msg structs to 4 | expand |
On Thu, Feb 20, 2020 at 06:29:32PM +0200, Leonard Crestez wrote: > The imx SC api strongly assumes that messages are composed out of > 4-bytes words but some of our message structs have odd sizeofs. > > This produces many oopses with CONFIG_KASAN=y. > > Fix by marking with __aligned(4). > > Fixes: fe37b4820417 ("clk: imx: add scu clock common part") > Signed-off-by: Leonard Crestez <leonard.crestez@nxp.com> Applied, thanks.
Quoting Leonard Crestez (2020-02-20 08:29:32) > The imx SC api strongly assumes that messages are composed out of > 4-bytes words but some of our message structs have odd sizeofs. > > This produces many oopses with CONFIG_KASAN=y. > > Fix by marking with __aligned(4). > > Fixes: fe37b4820417 ("clk: imx: add scu clock common part") > Signed-off-by: Leonard Crestez <leonard.crestez@nxp.com> > --- > drivers/clk/imx/clk-scu.c | 6 +++--- > 1 file changed, 3 insertions(+), 3 deletions(-) > > diff --git a/drivers/clk/imx/clk-scu.c b/drivers/clk/imx/clk-scu.c > index fbef740704d0..3c5c42d8833e 100644 > --- a/drivers/clk/imx/clk-scu.c > +++ b/drivers/clk/imx/clk-scu.c > @@ -41,16 +41,16 @@ struct clk_scu { > struct imx_sc_msg_req_set_clock_rate { > struct imx_sc_rpc_msg hdr; > __le32 rate; > __le16 resource; > u8 clk; > -} __packed; > +} __packed __aligned(4); Sorry, this still doesn't make sense to me. Having __aligned(4) means that the struct is placed on the stack at some alignment, great, but it still has __packed so the sizeof this struct is some odd number like 11. If this struct is the last element on the stack it will end at some unaligned address and the mailbox code will read a few bytes beyond the end of the stack. I see that the calling code puts 3 as the 'size' for this struct in clk_scu_set_rate(). hdr->size = 3; That seems to say that the struct is 3 words long, or 12 bytes. Then we call imx_scu_call_rpc(), passing the pointer to this struct on the stack and that eventually gets into imx_scu_ipc_write() calling mbox_send_message() with u32 pointers. for (i = 0; i < hdr->size; i++) { sc_chan = &sc_ipc->chans[i % 4]; ret = mbox_send_message(sc_chan->ch, &data[i]); So we've taken the 11 byte struct (data in this case) and casted it to a u32 array with 3 elements, which is bad. This is what kasan is warning about. Adding aligned sometimes fixes it because the compiler will place the next stack variable at the naturally aligned location and thus we get the one byte padding but I don't see how that works when it's the last stack element. The stack will end at some unaligned address. The better solution would be to drop __aligned(4) and make a union of the struct with whatever size number of words the message is or do a copy of the struct into a u32 array that is passed to imx_scu_call_rpc(). For example: struct imx_sc_msg_req_set_clock_rate { union { struct packed_message { struct imx_sc_rpc_msg hdr; __le32 rate; __le16 resource; u8 clk; } __packed; u32 data[3]; }; }; If the union approach was used then each time imx_scu_call_rpc() is called we can simply pass the 'data' member and make the second argument 'msg' strongly typed to be a u32 pointer. kasan should be happy too.
On 25.02.2020 18:52, Stephen Boyd wrote: > Quoting Leonard Crestez (2020-02-20 08:29:32) >> The imx SC api strongly assumes that messages are composed out of >> 4-bytes words but some of our message structs have odd sizeofs. >> >> This produces many oopses with CONFIG_KASAN=y. >> >> Fix by marking with __aligned(4). >> >> Fixes: fe37b4820417 ("clk: imx: add scu clock common part") >> Signed-off-by: Leonard Crestez <leonard.crestez@nxp.com> >> --- >> drivers/clk/imx/clk-scu.c | 6 +++--- >> 1 file changed, 3 insertions(+), 3 deletions(-) >> >> diff --git a/drivers/clk/imx/clk-scu.c b/drivers/clk/imx/clk-scu.c >> index fbef740704d0..3c5c42d8833e 100644 >> --- a/drivers/clk/imx/clk-scu.c >> +++ b/drivers/clk/imx/clk-scu.c >> @@ -41,16 +41,16 @@ struct clk_scu { >> struct imx_sc_msg_req_set_clock_rate { >> struct imx_sc_rpc_msg hdr; >> __le32 rate; >> __le16 resource; >> u8 clk; >> -} __packed; >> +} __packed __aligned(4); > > Sorry, this still doesn't make sense to me. Having __aligned(4) means > that the struct is placed on the stack at some alignment, great, but it > still has __packed so the sizeof this struct is some odd number like 11. > If this struct is the last element on the stack it will end at some > unaligned address and the mailbox code will read a few bytes beyond the > end of the stack. I checked again and marking the struct with __aligned(4) makes it have sizeof == 12 as intended. It was 11 before. static_assert(sizeof(struct imx_sc_msg_req_set_clock_rate) == 12); After reading through your email and gcc docs again I'm not sure if this portable/reliable this is but as far as I understand "sizeof" needs to account for alignment. Or is this just an accident with my compiler? Marking a structure both __packed and __aligned(4) means that __packed only affects internal struct member layout but sizeof is still rounded up to a multiple of 4: struct test { u8 a; u16 b; } __packed __aligned(4); static_assert(sizeof(struct test) == 4); static_assert(offsetof(struct test, a) == 0); static_assert(offsetof(struct test, b) == 1); This test is not realistic because I don't think SCU messages have any such oddly-aligned members. > > I see that the calling code puts 3 as the 'size' for this struct in > clk_scu_set_rate(). > > hdr->size = 3; > > That seems to say that the struct is 3 words long, or 12 bytes. Then we > call imx_scu_call_rpc(), passing the pointer to this struct on the stack > and that eventually gets into imx_scu_ipc_write() calling > mbox_send_message() with u32 pointers. > > for (i = 0; i < hdr->size; i++) { > sc_chan = &sc_ipc->chans[i % 4]; > ret = mbox_send_message(sc_chan->ch, &data[i]); > > So we've taken the 11 byte struct (data in this case) and casted it to a > u32 array with 3 elements, which is bad. This is what kasan is warning > about. Adding aligned sometimes fixes it because the compiler will place > the next stack variable at the naturally aligned location and thus we > get the one byte padding but I don't see how that works when it's the > last stack element. The stack will end at some unaligned address. > > The better solution would be to drop __aligned(4) and make a union of > the struct with whatever size number of words the message is or do a > copy of the struct into a u32 array that is passed to > imx_scu_call_rpc(). > > For example: > > struct imx_sc_msg_req_set_clock_rate { > union { > struct packed_message { > struct imx_sc_rpc_msg hdr; > __le32 rate; > __le16 resource; > u8 clk; > } __packed; > u32 data[3]; > }; > }; > > If the union approach was used then each time imx_scu_call_rpc() is > called we can simply pass the 'data' member and make the second argument > 'msg' strongly typed to be a u32 pointer. kasan should be happy too. >
On Mon, Feb 24, 2020 at 03:22:18PM +0800, Shawn Guo wrote: > On Thu, Feb 20, 2020 at 06:29:32PM +0200, Leonard Crestez wrote: > > The imx SC api strongly assumes that messages are composed out of > > 4-bytes words but some of our message structs have odd sizeofs. > > > > This produces many oopses with CONFIG_KASAN=y. > > > > Fix by marking with __aligned(4). > > > > Fixes: fe37b4820417 ("clk: imx: add scu clock common part") > > Signed-off-by: Leonard Crestez <leonard.crestez@nxp.com> > > Applied, thanks. Patch #1 and #2 were dropped from my clk queue, as Stephen hasn't been convinced by this change. Shawn
> From: Stephen Boyd <sboyd@kernel.org> > Sent: Wednesday, February 26, 2020 12:52 AM > Quoting Leonard Crestez (2020-02-20 08:29:32) > > The imx SC api strongly assumes that messages are composed out of > > 4-bytes words but some of our message structs have odd sizeofs. > > > > This produces many oopses with CONFIG_KASAN=y. > > > > Fix by marking with __aligned(4). > > > > Fixes: fe37b4820417 ("clk: imx: add scu clock common part") > > Signed-off-by: Leonard Crestez <leonard.crestez@nxp.com> > > --- > > drivers/clk/imx/clk-scu.c | 6 +++--- > > 1 file changed, 3 insertions(+), 3 deletions(-) > > > > diff --git a/drivers/clk/imx/clk-scu.c b/drivers/clk/imx/clk-scu.c > > index fbef740704d0..3c5c42d8833e 100644 > > --- a/drivers/clk/imx/clk-scu.c > > +++ b/drivers/clk/imx/clk-scu.c > > @@ -41,16 +41,16 @@ struct clk_scu { > > struct imx_sc_msg_req_set_clock_rate { > > struct imx_sc_rpc_msg hdr; > > __le32 rate; > > __le16 resource; > > u8 clk; > > -} __packed; > > +} __packed __aligned(4); > > Sorry, this still doesn't make sense to me. Having __aligned(4) means that the > struct is placed on the stack at some alignment, great, but it still has __packed so > the sizeof this struct is some odd number like 11 > If this struct is the last element on the stack it will end at some unaligned address > and the mailbox code will read a few bytes beyond the end of the stack. Hi Leonard, Can you construct this case to see if we can reproduce the issue as pointed by Stephen? Regards Aisheng > > I see that the calling code puts 3 as the 'size' for this struct in clk_scu_set_rate(). > > hdr->size = 3; > > That seems to say that the struct is 3 words long, or 12 bytes. Then we call > imx_scu_call_rpc(), passing the pointer to this struct on the stack and that > eventually gets into imx_scu_ipc_write() calling > mbox_send_message() with u32 pointers. > > for (i = 0; i < hdr->size; i++) { > sc_chan = &sc_ipc->chans[i % 4]; > ret = mbox_send_message(sc_chan->ch, &data[i]); > > So we've taken the 11 byte struct (data in this case) and casted it to a > u32 array with 3 elements, which is bad. This is what kasan is warning about. > Adding aligned sometimes fixes it because the compiler will place the next stack > variable at the naturally aligned location and thus we get the one byte padding > but I don't see how that works when it's the last stack element. The stack will > end at some unaligned address. > > The better solution would be to drop __aligned(4) and make a union of the > struct with whatever size number of words the message is or do a copy of the > struct into a u32 array that is passed to imx_scu_call_rpc(). > > For example: > > struct imx_sc_msg_req_set_clock_rate { > union { > struct packed_message { > struct imx_sc_rpc_msg hdr; > __le32 rate; > __le16 resource; > u8 clk; > } __packed; > u32 data[3]; > }; > }; > > If the union approach was used then each time imx_scu_call_rpc() is called we > can simply pass the 'data' member and make the second argument 'msg' > strongly typed to be a u32 pointer. kasan should be happy too.
diff --git a/drivers/clk/imx/clk-scu.c b/drivers/clk/imx/clk-scu.c index fbef740704d0..3c5c42d8833e 100644 --- a/drivers/clk/imx/clk-scu.c +++ b/drivers/clk/imx/clk-scu.c @@ -41,16 +41,16 @@ struct clk_scu { struct imx_sc_msg_req_set_clock_rate { struct imx_sc_rpc_msg hdr; __le32 rate; __le16 resource; u8 clk; -} __packed; +} __packed __aligned(4); struct req_get_clock_rate { __le16 resource; u8 clk; -} __packed; +} __packed __aligned(4); struct resp_get_clock_rate { __le32 rate; }; @@ -119,11 +119,11 @@ struct imx_sc_msg_req_clock_enable { struct imx_sc_rpc_msg hdr; __le16 resource; u8 clk; u8 enable; u8 autog; -} __packed; +} __packed __aligned(4); static inline struct clk_scu *to_clk_scu(struct clk_hw *hw) { return container_of(hw, struct clk_scu, hw); }
The imx SC api strongly assumes that messages are composed out of 4-bytes words but some of our message structs have odd sizeofs. This produces many oopses with CONFIG_KASAN=y. Fix by marking with __aligned(4). Fixes: fe37b4820417 ("clk: imx: add scu clock common part") Signed-off-by: Leonard Crestez <leonard.crestez@nxp.com> --- drivers/clk/imx/clk-scu.c | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-)