Writing a byte to a pl011 SFR overwrites the whole SFR

[Daniels Umanovskis]

The bug is present in QEMU 2.8.1 and, if my analysis is correct, also on master.

I first noticed that a PL011 UART driver, which is fine on real hardware, fails to enable the RX interrupt in the IMSC register when running in QEMU. However, the problem only comes up if the code is compiled without optimizations. I think I've narrowed it down to a minimal example that will exhibit the problem if run as a bare-metal application.

Given:

pl011_addr: .word 0x10009000

The following snippet will be problematic:

     ldr r3, pl011_addr
     ldrb r2, [r3, #0x38] // IMSC
     mov r2, #0
     orr r2, r2, #0x10 // R2 == 0x10
     strb r2, [r3, #0x38] // Whole word reads correctly after this
     ldrb r2, [r3, #0x39]
     mov r2, #0
     strb r2, [r3, #0x39] // Problem here! Overwrites offset 0x38 as well

After the first strb instruction, which writes to 0x10009038, everything is fine. It can be seen in the QEMU monitor:

(qemu) xp 0x10009038
0000000010009038: 0x00000010

After the second strb instruction, the write to 0x10009039 clears the entire word:

(qemu) xp 0x10009038
0000000010009038: 0x00000000

QEMU command-line, using the vexpress-a9 which has the PL011 at 0x10009000:

qemu-system-arm -S -M vexpress-a9 -m 32M -no-reboot -nographic -monitor telnet:127.0.0.1:1234,server,nowait -kernel pl011-sfr.bin -gdb tcp::2159 -serial mon:stdio

Compiling the original C code with optimizations makes the driver work. It compiles down to assembly that only does a single write:

    ldr r3, pl011_addr
    mov r2, #0x10
    str r2, [r3, #0x38]

Attached is the an assembly file, and linkscript, that shows the problem, and also includes the working code.

I haven't debugged inside of QEMU itself but it seems to me that the problem is in pl011_write in pl011.c - the functions looks at which offset is being written, and then writes the entire SFR that offset falls under, which means that changing a single byte will change the whole SFR.

[Daniels Umanovskis (dumanovskis)]

Adding the link script.

[Peter Maydell (pmaydell)]

Yes, our PL011 implementation assumes that you only ever access the 32-bit registers with full width 32-bit word reads and writes. Don't try to do byte accesses to them. The PL011 data sheet doesn't specifically say that partial-width accesses to registers are permitted, so I think that trying to access offset 0x39 falls under the general note in section 3.1 that attempting to access reserved or unused address locations can result in unpredictable behaviour.

You need to make sure you write your C code in a manner which enforces that accesses to device registers are done as single 32-bit accesses, and the compiler does not silently break them down into multiple reads and writes, or you will be in for a lot of pain trying to figure out what is going on if the compiler ever does it with registers that are write-to-clear or similar behaviour. Linux, for instance, does this by having readl() and writel() functions that end up doing inline asm of ldr/str instructions.

[Daniels Umanovskis (dumanovskis)]

Thanks for the response.

I don't think section 3.1 applies to 8-bit accesses. That is specifically about reserved locations, and neither offset 0x38 nor 0x39 are reserved, so I think it's a matter of whether 32-bit access is required or not.

From what I usually see in ARM documentation, 32-bit access is explicitly mentioned when required. For the PL011, it's mentioned for the UARTPeriphID_n registers, for instance. In many other cases access size depends on the implementation and the corresponding memory mapping of that implementation.

I understand that *in practice* you should ensure single-access writes unless doing otherwise is explicitly allowed. However, in cases like the PL011 it seems ambiguous whether that is actually required, so it seems like the best choice would be to explicitly document it for the QEMU implementation. That would save some guesswork.

[Thomas Huth (th-huth)]

The QEMU project is currently considering to move its bug tracking to
another system. For this we need to know which bugs are still valid
and which could be closed already. Thus we are setting older bugs to
"Incomplete" now.

If you still think this bug report here is valid, then please switch
the state back to "New" within the next 60 days, otherwise this report
will be marked as "Expired". Or please mark it as "Fix Released" if
the problem has been solved with a newer version of QEMU already.

Thank you and sorry for the inconvenience.

[Launchpad Janitor (janitor)]

[Expired for QEMU because there has been no activity for 60 days.]