Re: [PATCH 2/3] usb: mtu3: add memory barrier before set GPD's HWO

On Mon, 2021-12-13 at 15:18 +0100, Greg Kroah-Hartman wrote:

On Thu, Dec 09, 2021 at 11:14:23AM +0800, Chunfeng Yun wrote:

quoted

There is a seldom issue that the controller access invalid address
and trigger devapc or emimpu violation. That is due to memory
access
is out of order and cause gpd data is not correct.
Make sure GPD is fully written before giving it to HW by setting
its
HWO.

Fixes: 48e0d3735aa5 ("usb: mtu3: supports new QMU format")
Cc: stable@vger.kernel.org
Reported-by: Eddie Hung <redacted>
Signed-off-by: Chunfeng Yun <chunfeng.yun@mediatek.com>
---
 drivers/usb/mtu3/mtu3_qmu.c | 7 ++++++-
 1 file changed, 6 insertions(+), 1 deletion(-)

diff --git a/drivers/usb/mtu3/mtu3_qmu.c

b/drivers/usb/mtu3/mtu3_qmu.c
index 3f414f91b589..34bb5ac67efe 100644

--- a/drivers/usb/mtu3/mtu3_qmu.c
+++ b/drivers/usb/mtu3/mtu3_qmu.c

@@ -273,6 +273,8 @@ static int mtu3_prepare_tx_gpd(struct mtu3_ep

*mep, struct mtu3_request *mreq)
 			gpd->dw3_info |= cpu_to_le32(GPD_EXT_FLAG_ZLP);
 	}
 
+	/* make sure GPD is fully written before giving it to HW */
+	mb();

So this means you are using mmio for this structure? 

No, it's a noncached memory.

 If so, shouldn't
you be using normal io memory read/write calls as well and not just
"raw" pointers like this:

quoted

 	gpd->dw0_info |= cpu_to_le32(GPD_FLAGS_IOC | GPD_FLAGS_HWO);

Are you sure this is ok?

Sprinkling around mb() calls is almost never the correct solution.

If you need to ensure that a write succeeds, shouldn't you do a read
from it afterward?  Many busses require this, doesn't yours?

It works for register access.
Here is noncache memory access, add mb(), just want to prohibite both
the compiler and CPU from reordering read/writes.

quoted

 
 	mreq->gpd = gpd;

@@ -306,6 +308,8 @@ static int mtu3_prepare_rx_gpd(struct mtu3_ep

*mep, struct mtu3_request *mreq)
 	gpd->next_gpd = cpu_to_le32(lower_32_bits(enq_dma));
 	ext_addr |= GPD_EXT_NGP(mtu, upper_32_bits(enq_dma));
 	gpd->dw3_info = cpu_to_le32(ext_addr);
+	/* make sure GPD is fully written before giving it to HW */
+	mb();

Again, mb(); does not ensure that memory-mapped i/o actually hits the
HW.  Or if it does on your platform, how?

Maybe the comment is misleading, I'll change it, here just want to
prevent reordering of compiler and cpu.

mb() is a compiler barrier, not a memory write to a bus
barrier.  Please
read Documentation/memory-barriers.txt for more details.

Ok, I'll do.

Thanks a lot

thanks,

greg k-h