Russel was out of the loop, re-adding him as he may have interesting thoughts about this. Sorry for the double e-mail.
Miquel Raynal miquel.raynal@bootlin.com wrote on Tue, 22 Oct 2019 09:44:51 +0200:
Hello,
Boris Brezillon boris.brezillon@collabora.com wrote on Mon, 21 Oct 2019 10:01:05 +0200:
On Fri, 18 Oct 2019 16:36:43 +0200 Miquel Raynal miquel.raynal@bootlin.com wrote:
Any write with either dd or flashcp to a device driven by the spear_smi.c driver will pass through the spear_smi_cpy_toio() function. This function will get called for chunks of up to 256 bytes. If the amount of data is smaller, we may have a problem if the data length is not 4-byte aligned. In this situation, the kernel panics during the memcpy:
# dd if=/dev/urandom bs=1001 count=1 of=/dev/mtd6 spear_smi_cpy_toio [620] dest c9070000, src c7be8800, len 256 spear_smi_cpy_toio [620] dest c9070100, src c7be8900, len 256 spear_smi_cpy_toio [620] dest c9070200, src c7be8a00, len 256 spear_smi_cpy_toio [620] dest c9070300, src c7be8b00, len 233 Unhandled fault: external abort on non-linefetch (0x808) at 0xc90703e8 [...] PC is at memcpy+0xcc/0x330Can you find out which instruction is at memcpy+0xcc/0x330? For the record, the assembly is here [1].
The full disassembled file is attached, here is the failing part:
7: ldmfd sp!, {r5 - r8} b8: e8bd01e0 pop {r5, r6, r7, r8} UNWIND( .fnend ) @ end of second stmfd block
UNWIND( .fnstart ) usave r4, lr @ still in first stmdb block 8: movs r2, r2, lsl #31 bc: e1b02f82 lsls r2, r2, #31 ldr1b r1, r3, ne, abort=21f c0: 14d13001 ldrbne r3, [r1], #1 ldr1b r1, r4, cs, abort=21f c4: 24d14001 ldrbcs r4, [r1], #1 ldr1b r1, ip, cs, abort=21f c8: 24d1c001 ldrbcs ip, [r1], #1 str1b r0, r3, ne, abort=21f cc: 14c03001 strbne r3, [r0], #1 str1b r0, r4, cs, abort=21f d0: 24c04001 strbcs r4, [r0], #1 str1b r0, ip, cs, abort=21f d4: 24c0c001 strbcs ip, [r0], #1
exit r4, pcd8: e8bd8011 pop {r0, r4, pc}
So the fault is triggered on a strbne instruction.
Workaround this issue by using the alternate _memcpy_toio() method which at least does not present the same problem.
Fixes: f18dbbb1bfe0 ("mtd: ST SPEAr: Add SMI driver for serial NOR flash") Cc: stable@vger.kernel.org Suggested-by: Boris Brezillon boris.brezillon@collabora.com
I don't remember suggesting that as a final solution. I probably suggested to test with _memcpy_toio() to see if using a byte accessor was fixing the problem, but it's definitely not the right solution (using byte access with a memory barrier for 256 bytes buffers is likely to cause a huge perf penalty).
Signed-off-by: Miquel Raynal miquel.raynal@bootlin.com
Hello,
This patch could not be tested with a mainline kernel (only compiled) but was tested with a stable 4.14.x kernel. I have really no idea why memcpy fails in this situation that's why I propose this workaround but I bet there is something deeper not working.
Thanks, Miquèl
drivers/mtd/devices/spear_smi.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/drivers/mtd/devices/spear_smi.c b/drivers/mtd/devices/spear_smi.c index 986f81d2f93e..d888625a3244 100644 --- a/drivers/mtd/devices/spear_smi.c +++ b/drivers/mtd/devices/spear_smi.c @@ -614,7 +614,7 @@ static inline int spear_smi_cpy_toio(struct spear_smi *dev, u32 bank, ctrlreg1 = readl(dev->io_base + SMI_CR1); writel((ctrlreg1 | WB_MODE) & ~SW_MODE, dev->io_base + SMI_CR1);
- memcpy_toio(dest, src, len);
- _memcpy_toio(dest, src, len);
writel(ctrlreg1, dev->io_base + SMI_CR1);
[1]https://elixir.bootlin.com/linux/v5.4-rc2/source/arch/arm/lib/memcpy.S
Thanks, Miquèl