I'm running into a problem that seems to be caused by this really old
changeset:
http://linux.bkbits.net:8080/linux-2.5/cset@3d9f67f2BWvXiLsZCFwD-8s_E9AN6A
When I run 'setserial /dev/ttyS1 uart 16450' on an ia64 system w/ MMIO
UARTs, I get a NAT consumption oops from the kernel. The problem is
that this code path calls serial8250_release_port() where the membase
gets cleared. However, the subsequent call to serial8250_request_port()
doesn't restore membase, causing a read from a bad address. I don't see
many users of the UPF_IOREMAP flag, so I think the solution is to simply
make the remap case symmetric to the unmap case. Patch below. Thanks,
Alex
--
Signed-off-by: Alex Williamson <[email protected]>
===== drivers/serial/8250.c 1.67 vs edited =====
--- 1.67/drivers/serial/8250.c 2004-09-13 18:23:24 -06:00
+++ edited/drivers/serial/8250.c 2004-10-04 12:12:34 -06:00
@@ -1733,7 +1733,7 @@
/*
* If we have a mapbase, then request that as well.
*/
- if (ret == 0 && up->port.flags & UPF_IOREMAP) {
+ if (ret == 0 && up->port.iotype == UPIO_MEM && up->port.mapbase) {
int size = res->end - res->start + 1;
up->port.membase = ioremap(up->port.mapbase, size);
On Mon, Oct 04, 2004 at 12:54:22PM -0600, Alex Williamson wrote:
>
> I'm running into a problem that seems to be caused by this really old
> changeset:
>
> http://linux.bkbits.net:8080/linux-2.5/cset@3d9f67f2BWvXiLsZCFwD-8s_E9AN6A
>
> When I run 'setserial /dev/ttyS1 uart 16450' on an ia64 system w/ MMIO
> UARTs, I get a NAT consumption oops from the kernel. The problem is
> that this code path calls serial8250_release_port() where the membase
> gets cleared. However, the subsequent call to serial8250_request_port()
> doesn't restore membase, causing a read from a bad address. I don't see
> many users of the UPF_IOREMAP flag, so I think the solution is to simply
> make the remap case symmetric to the unmap case. Patch below. Thanks,
Mostly correct reasoning, but the solution is wrong. Consider what
happens if we call request_port where we have set mapbase and pre-
initialised membase for a memory mapped port (eg, PCI card.)
This would cause us to re-ioremap the mapbase, which is wrong. We
must obey the UPF_IOREMAP flag here. Note also that this fix you're
reverting will break 8250 for PPC people...
Could you give further information about the problem you're seeing?
Bear in mind that I know precisely zero about ia64 oopsen so you'll
probably have to explain it to me in detail.
--
Russell King
Linux kernel 2.6 ARM Linux - http://www.arm.linux.org.uk/
maintainer of: 2.6 PCMCIA - http://pcmcia.arm.linux.org.uk/
2.6 Serial core
On Mon, 2004-10-04 at 22:04 +0100, Russell King wrote:
> On Mon, Oct 04, 2004 at 12:54:22PM -0600, Alex Williamson wrote:
> >
> > I'm running into a problem that seems to be caused by this really old
> > changeset:
> >
> > http://linux.bkbits.net:8080/linux-2.5/cset@3d9f67f2BWvXiLsZCFwD-8s_E9AN6A
> >
> > When I run 'setserial /dev/ttyS1 uart 16450' on an ia64 system w/ MMIO
> > UARTs, I get a NAT consumption oops from the kernel. The problem is
> > that this code path calls serial8250_release_port() where the membase
> > gets cleared. However, the subsequent call to serial8250_request_port()
> > doesn't restore membase, causing a read from a bad address. I don't see
> > many users of the UPF_IOREMAP flag, so I think the solution is to simply
> > make the remap case symmetric to the unmap case. Patch below. Thanks,
>
> Mostly correct reasoning, but the solution is wrong. Consider what
> happens if we call request_port where we have set mapbase and pre-
> initialised membase for a memory mapped port (eg, PCI card.)
>
> This would cause us to re-ioremap the mapbase, which is wrong. We
> must obey the UPF_IOREMAP flag here. Note also that this fix you're
> reverting will break 8250 for PPC people...
>
> Could you give further information about the problem you're seeing?
> Bear in mind that I know precisely zero about ia64 oopsen so you'll
> probably have to explain it to me in detail.
Sure. I've see the problem on any MMIO UART on my box:
# cat /proc/tty/driver/serial
serinfo:1.0 driver revision:
0: uart:16550A mmio:0xFF5E0000 irq:49 tx:5327 rx:67 RTS|CTS|DTR|DSR|CD
1: uart:16550A mmio:0xFF5E2000 irq:66 tx:0 rx:0
2: uart:16550A mmio:0xF8031000 irq:64 tx:0 rx:0
3: uart:16550A mmio:0xF8030000 irq:64 tx:0 rx:0
4: uart:16550A mmio:0xF8030010 irq:64 tx:0 rx:0
5: uart:16550A mmio:0xF8030038 irq:64 tx:0 rx:0
The first 2 are dangling off a platform bus, not on PCI. They're
discovered via the 8250_acpi code (or the first one may be found via the
pcdp setup). The last 4 are in PCI space and handled by 8250_pci.
Using setserial to poke the uart type on a devices produces something
like this:
# setserial /dev/ttyS2 uart 16450
setserial[1540]: NaT consumption 17179869216 [1]
Modules linked in:
Pid: 1540, CPU 1, comm: setserial
psr : 0000101008026018 ifs : 8000000000000002 ip : [<a0000001003035a0>]
Not
tainted
ip is at __ia64_readb+0x0/0x20
unat: 0000000000000000 pfs : 0000000000000389 rsc : 0000000000000003
rnat: e0000001004ac458 bsps: e0000001004ac668 pr : 0a40000000169969
ldrs: 0000000000000000 ccv : 0000000000000202 fpsr: 0009804c0270033f
csd : 0000000000000000 ssd : 0000000000000000
b0 : a0000001003ab670 b6 : a000000100002d70 b7 : a0000001003035a0
f6 : 1003e6db6db6db6db6db7 f7 : 000000000000000000000
f8 : 1003e000000000000ef6a f9 : 1003e0000000000068be6
f10 : 1003e0000000051eb851f f11 : 1003e0000000000080000
r1 : a000000100ad35a0 r2 : 0000000000000000 r3 : a000000100a7da6b
r8 : a000000100a7da6a r9 : 0000000000000006 r10 : a000000100662cd0
r11 : 0000000000000002 r12 : e00000003bf27d70 r13 : e00000003bf20000
r14 : 0000000000000002 r15 : a000000100a7db55 r16 : 0000000000000002
r17 : a0000001008d6d60 r18 : 0000000000008f46 r19 : 0000000000000000
r20 : a0000001006bff58 r21 : a0000001003035a0 r22 : 0000000000000000
r23 : 0000000000000005 r24 : a0000001008eabb0 r25 : a000000100a7da58
r26 : a0000001008eaac0 r27 : a0000001008eaac0 r28 : e00000003bf27d88
r29 : e00000000310d028 r30 : 0000000000000001 r31 : e00000003bf27d84
Call Trace:
[<a0000001000168e0>] show_stack+0x80/0xa0
sp=e00000003bf278c0 bsp=e00000003bf21310
[<a000000100017140>] show_regs+0x7e0/0x800
sp=e00000003bf27a90 bsp=e00000003bf212b0
[<a00000010003a8b0>] die+0x150/0x1c0
sp=e00000003bf27aa0 bsp=e00000003bf21270
[<a00000010003a960>] die_if_kernel+0x40/0x60
sp=e00000003bf27aa0 bsp=e00000003bf21240
[<a00000010003b570>] ia64_fault+0x150/0xa40
sp=e00000003bf27aa0 bsp=e00000003bf211f0
[<a00000010000e7e0>] ia64_leave_kernel+0x0/0x260
sp=e00000003bf27ba0 bsp=e00000003bf211f0
[<a0000001003035a0>] __ia64_readb+0x0/0x20
sp=e00000003bf27d70 bsp=e00000003bf211e0
[<a0000001003ab670>] serial_in+0x210/0x220
sp=e00000003bf27d70 bsp=e00000003bf211a8
[<a0000001003aea60>] serial8250_startup+0xc0/0x740
sp=e00000003bf27d70 bsp=e00000003bf21170
[<a0000001003a3900>] uart_startup+0x240/0x440
sp=e00000003bf27d70 bsp=e00000003bf21120
[<a0000001003a55d0>] uart_set_info+0x3f0/0xb40
sp=e00000003bf27d90 bsp=e00000003bf21038
[<a0000001003a70b0>] uart_ioctl+0x2f0/0x3a0
sp=e00000003bf27e20 bsp=e00000003bf20fe8
[<a000000100370500>] tty_ioctl+0x780/0xa20
sp=e00000003bf27e20 bsp=e00000003bf20f90
[<a00000010014c430>] sys_ioctl+0x270/0x720
sp=e00000003bf27e20 bsp=e00000003bf20f00
[<a00000010000e660>] ia64_ret_from_syscall+0x0/0x20
sp=e00000003bf27e30 bsp=e00000003bf20f00
I instrumented serial_in/serial_out to see what what happening (this
time for ttyS1):
# setserial /dev/ttyS1 uart 16450
serial_out() -> writeb(0x1, 0xc0000000ff5e2002)
serial_out() -> writeb(0x7, 0xc0000000ff5e2002)
serial_out() -> writeb(0x0, 0xc0000000ff5e2002)
serial_in() -> readb(0xc0000000ff5e2005)
serial_in() -> readb(0xc0000000ff5e2000)
serial_in() -> readb(0xc0000000ff5e2002)
serial_in() -> readb(0xc0000000ff5e2006)
serial_in() -> readb(0xc0000000ff5e2005)
serial_out() -> writeb(0x3, 0xc0000000ff5e2003)
serial_out() -> writeb(0x8, 0xc0000000ff5e2004)
serial_out() -> writeb(0x5, 0xc0000000ff5e2001)
serial_in() -> readb(0xc0000000ff5e2005)
serial_in() -> readb(0xc0000000ff5e2000)
serial_in() -> readb(0xc0000000ff5e2002)
serial_in() -> readb(0xc0000000ff5e2006)
serial_out() -> writeb(0x5, 0xc0000000ff5e2001)
serial_out() -> writeb(0x93, 0xc0000000ff5e2003)
serial_out() -> writeb(0xc, 0xc0000000ff5e2000)
serial_out() -> writeb(0x0, 0xc0000000ff5e2001)
serial_out() -> writeb(0x13, 0xc0000000ff5e2003)
serial_out() -> writeb(0x1, 0xc0000000ff5e2002)
serial_out() -> writeb(0x81, 0xc0000000ff5e2002)
serial_out() -> writeb(0x8, 0xc0000000ff5e2004)
serial_out() -> writeb(0xb, 0xc0000000ff5e2004)
serial_out() -> writeb(0x8, 0xc0000000ff5e2004)
serial_out() -> writeb(0x0, 0xc0000000ff5e2001)
serial_out() -> writeb(0x0, 0xc0000000ff5e2004)
serial_in() -> readb(0xc0000000ff5e2003)
serial_out() -> writeb(0x13, 0xc0000000ff5e2003)
serial_out() -> writeb(0x1, 0xc0000000ff5e2002)
serial_out() -> writeb(0x7, 0xc0000000ff5e2002)
serial_out() -> writeb(0x0, 0xc0000000ff5e2002)
serial_in() -> readb(0xc0000000ff5e2000)
serial_in() -> readb(0x5)
As you can see, we completely lost membase between the last 2 reads
and are only dealing with the offset. This is what causes the stack
trace in the readb(). I suspect a PCI MMIO UART would fail just as
badly on other architectures as well. Is PPC somehow dependent on the
UPF_IOREMAP flag, or would it be sufficient to check that membase is
NULL before calling ioremap? I see exactly one instance of a driver
setting UPF_IOREMAP, which is why I took the path I did. Thanks,
Alex
--
Alex Williamson HP Linux & Open Source Lab
How do we proceed with this? There's obviously an iounmap/ioremap
mismatch in the serial8250_release_port/serial8250_request_port path for
memory mapped UARTs. Should we only iounmap UARTs w/ the UPF_IOREMAP
flag? Should we test membase for NULL and ioremap? Do we need to add
the UPF_IOREMAP flag to all MMIO UARTs? The current asymmetric test
makes setserial unusable on any MMIO UARTs w/o the UPF_IOREMAP flag (so
everyone except mpc52xx_uart AFAICT). Does anyone else have MMIO UARTs
(other than mpc52xx) that can confirm this? Just try to toggle the uart
type on a non-console UART. Thanks,
Alex
On Mon, 2004-10-04 at 15:43 -0600, Alex Williamson wrote:
> On Mon, 2004-10-04 at 22:04 +0100, Russell King wrote:
> > On Mon, Oct 04, 2004 at 12:54:22PM -0600, Alex Williamson wrote:
> > >
> > > I'm running into a problem that seems to be caused by this really old
> > > changeset:
> > >
> > > http://linux.bkbits.net:8080/linux-2.5/cset@3d9f67f2BWvXiLsZCFwD-8s_E9AN6A
> > >
> > > When I run 'setserial /dev/ttyS1 uart 16450' on an ia64 system w/ MMIO
> > > UARTs, I get a NAT consumption oops from the kernel. The problem is
> > > that this code path calls serial8250_release_port() where the membase
> > > gets cleared. However, the subsequent call to serial8250_request_port()
> > > doesn't restore membase, causing a read from a bad address. I don't see
> > > many users of the UPF_IOREMAP flag, so I think the solution is to simply
> > > make the remap case symmetric to the unmap case. Patch below. Thanks,
> >
> > Mostly correct reasoning, but the solution is wrong. Consider what
> > happens if we call request_port where we have set mapbase and pre-
> > initialised membase for a memory mapped port (eg, PCI card.)
> >
> > This would cause us to re-ioremap the mapbase, which is wrong. We
> > must obey the UPF_IOREMAP flag here. Note also that this fix you're
> > reverting will break 8250 for PPC people...
> >
> > Could you give further information about the problem you're seeing?
> > Bear in mind that I know precisely zero about ia64 oopsen so you'll
> > probably have to explain it to me in detail.
>
> Sure. I've see the problem on any MMIO UART on my box:
>
> # cat /proc/tty/driver/serial
> serinfo:1.0 driver revision:
> 0: uart:16550A mmio:0xFF5E0000 irq:49 tx:5327 rx:67 RTS|CTS|DTR|DSR|CD
> 1: uart:16550A mmio:0xFF5E2000 irq:66 tx:0 rx:0
> 2: uart:16550A mmio:0xF8031000 irq:64 tx:0 rx:0
> 3: uart:16550A mmio:0xF8030000 irq:64 tx:0 rx:0
> 4: uart:16550A mmio:0xF8030010 irq:64 tx:0 rx:0
> 5: uart:16550A mmio:0xF8030038 irq:64 tx:0 rx:0
>
> The first 2 are dangling off a platform bus, not on PCI. They're
> discovered via the 8250_acpi code (or the first one may be found via the
> pcdp setup). The last 4 are in PCI space and handled by 8250_pci.
>
> Using setserial to poke the uart type on a devices produces something
> like this:
>
> # setserial /dev/ttyS2 uart 16450
> setserial[1540]: NaT consumption 17179869216 [1]
> Modules linked in:
>
> Pid: 1540, CPU 1, comm: setserial
> psr : 0000101008026018 ifs : 8000000000000002 ip : [<a0000001003035a0>]
> Not
> tainted
> ip is at __ia64_readb+0x0/0x20
> unat: 0000000000000000 pfs : 0000000000000389 rsc : 0000000000000003
> rnat: e0000001004ac458 bsps: e0000001004ac668 pr : 0a40000000169969
> ldrs: 0000000000000000 ccv : 0000000000000202 fpsr: 0009804c0270033f
> csd : 0000000000000000 ssd : 0000000000000000
> b0 : a0000001003ab670 b6 : a000000100002d70 b7 : a0000001003035a0
> f6 : 1003e6db6db6db6db6db7 f7 : 000000000000000000000
> f8 : 1003e000000000000ef6a f9 : 1003e0000000000068be6
> f10 : 1003e0000000051eb851f f11 : 1003e0000000000080000
> r1 : a000000100ad35a0 r2 : 0000000000000000 r3 : a000000100a7da6b
> r8 : a000000100a7da6a r9 : 0000000000000006 r10 : a000000100662cd0
> r11 : 0000000000000002 r12 : e00000003bf27d70 r13 : e00000003bf20000
> r14 : 0000000000000002 r15 : a000000100a7db55 r16 : 0000000000000002
> r17 : a0000001008d6d60 r18 : 0000000000008f46 r19 : 0000000000000000
> r20 : a0000001006bff58 r21 : a0000001003035a0 r22 : 0000000000000000
> r23 : 0000000000000005 r24 : a0000001008eabb0 r25 : a000000100a7da58
> r26 : a0000001008eaac0 r27 : a0000001008eaac0 r28 : e00000003bf27d88
> r29 : e00000000310d028 r30 : 0000000000000001 r31 : e00000003bf27d84
>
> Call Trace:
> [<a0000001000168e0>] show_stack+0x80/0xa0
> sp=e00000003bf278c0 bsp=e00000003bf21310
> [<a000000100017140>] show_regs+0x7e0/0x800
> sp=e00000003bf27a90 bsp=e00000003bf212b0
> [<a00000010003a8b0>] die+0x150/0x1c0
> sp=e00000003bf27aa0 bsp=e00000003bf21270
> [<a00000010003a960>] die_if_kernel+0x40/0x60
> sp=e00000003bf27aa0 bsp=e00000003bf21240
> [<a00000010003b570>] ia64_fault+0x150/0xa40
> sp=e00000003bf27aa0 bsp=e00000003bf211f0
> [<a00000010000e7e0>] ia64_leave_kernel+0x0/0x260
> sp=e00000003bf27ba0 bsp=e00000003bf211f0
> [<a0000001003035a0>] __ia64_readb+0x0/0x20
> sp=e00000003bf27d70 bsp=e00000003bf211e0
> [<a0000001003ab670>] serial_in+0x210/0x220
> sp=e00000003bf27d70 bsp=e00000003bf211a8
> [<a0000001003aea60>] serial8250_startup+0xc0/0x740
> sp=e00000003bf27d70 bsp=e00000003bf21170
> [<a0000001003a3900>] uart_startup+0x240/0x440
> sp=e00000003bf27d70 bsp=e00000003bf21120
> [<a0000001003a55d0>] uart_set_info+0x3f0/0xb40
> sp=e00000003bf27d90 bsp=e00000003bf21038
> [<a0000001003a70b0>] uart_ioctl+0x2f0/0x3a0
> sp=e00000003bf27e20 bsp=e00000003bf20fe8
> [<a000000100370500>] tty_ioctl+0x780/0xa20
> sp=e00000003bf27e20 bsp=e00000003bf20f90
> [<a00000010014c430>] sys_ioctl+0x270/0x720
> sp=e00000003bf27e20 bsp=e00000003bf20f00
> [<a00000010000e660>] ia64_ret_from_syscall+0x0/0x20
> sp=e00000003bf27e30 bsp=e00000003bf20f00
>
>
> I instrumented serial_in/serial_out to see what what happening (this
> time for ttyS1):
>
> # setserial /dev/ttyS1 uart 16450
> serial_out() -> writeb(0x1, 0xc0000000ff5e2002)
> serial_out() -> writeb(0x7, 0xc0000000ff5e2002)
> serial_out() -> writeb(0x0, 0xc0000000ff5e2002)
> serial_in() -> readb(0xc0000000ff5e2005)
> serial_in() -> readb(0xc0000000ff5e2000)
> serial_in() -> readb(0xc0000000ff5e2002)
> serial_in() -> readb(0xc0000000ff5e2006)
> serial_in() -> readb(0xc0000000ff5e2005)
> serial_out() -> writeb(0x3, 0xc0000000ff5e2003)
> serial_out() -> writeb(0x8, 0xc0000000ff5e2004)
> serial_out() -> writeb(0x5, 0xc0000000ff5e2001)
> serial_in() -> readb(0xc0000000ff5e2005)
> serial_in() -> readb(0xc0000000ff5e2000)
> serial_in() -> readb(0xc0000000ff5e2002)
> serial_in() -> readb(0xc0000000ff5e2006)
> serial_out() -> writeb(0x5, 0xc0000000ff5e2001)
> serial_out() -> writeb(0x93, 0xc0000000ff5e2003)
> serial_out() -> writeb(0xc, 0xc0000000ff5e2000)
> serial_out() -> writeb(0x0, 0xc0000000ff5e2001)
> serial_out() -> writeb(0x13, 0xc0000000ff5e2003)
> serial_out() -> writeb(0x1, 0xc0000000ff5e2002)
> serial_out() -> writeb(0x81, 0xc0000000ff5e2002)
> serial_out() -> writeb(0x8, 0xc0000000ff5e2004)
> serial_out() -> writeb(0xb, 0xc0000000ff5e2004)
> serial_out() -> writeb(0x8, 0xc0000000ff5e2004)
> serial_out() -> writeb(0x0, 0xc0000000ff5e2001)
> serial_out() -> writeb(0x0, 0xc0000000ff5e2004)
> serial_in() -> readb(0xc0000000ff5e2003)
> serial_out() -> writeb(0x13, 0xc0000000ff5e2003)
> serial_out() -> writeb(0x1, 0xc0000000ff5e2002)
> serial_out() -> writeb(0x7, 0xc0000000ff5e2002)
> serial_out() -> writeb(0x0, 0xc0000000ff5e2002)
> serial_in() -> readb(0xc0000000ff5e2000)
> serial_in() -> readb(0x5)
>
> As you can see, we completely lost membase between the last 2 reads
> and are only dealing with the offset. This is what causes the stack
> trace in the readb(). I suspect a PCI MMIO UART would fail just as
> badly on other architectures as well. Is PPC somehow dependent on the
> UPF_IOREMAP flag, or would it be sufficient to check that membase is
> NULL before calling ioremap? I see exactly one instance of a driver
> setting UPF_IOREMAP, which is why I took the path I did. Thanks,
>
> Alex
>
--
Alex Williamson HP Linux & Open Source Lab
The iounmap/ioremap path in serial8250_release/request_port is
terribly unbalanced. The UPF_IOREMAP flag is used to determine if a
port gets ioremap'd, but plays no part in whether it gets iounmap'd.
It's easy to see how an MMIO serial port can be passed through
uart_set_info and end up with an unmapped membase. The results is a
non-functional UART or worse. I've tried to generate some discussion on
the proper fix for this, but I haven't succeeded. I propose the patch
below as a safe compromise. An MMIO uart w/ a mapbase, but no membase
doesn't seem viable to me. Thanks,
Alex
--
Signed-off-by: Alex Williamson <[email protected]>
===== drivers/serial/8250.c 1.77 vs edited =====
--- 1.77/drivers/serial/8250.c 2004-10-25 06:05:26 -06:00
+++ edited/drivers/serial/8250.c 2004-10-28 13:18:01 -06:00
@@ -1858,7 +1858,11 @@
/*
* If we have a mapbase, then request that as well.
*/
- if (ret == 0 && up->port.flags & UPF_IOREMAP) {
+ if (ret == 0 && ((up->port.flags & UPF_IOREMAP) ||
+ (up->port.iotype == UPIO_MEM &&
+ up->port.mapbase &&
+ !up->port.membase))) {
+
int size = res->end - res->start + 1;
up->port.membase = ioremap(up->port.mapbase, size);
On Thu, Oct 28, 2004 at 01:24:55PM -0600, Alex Williamson wrote:
> The iounmap/ioremap path in serial8250_release/request_port is
> terribly unbalanced. The UPF_IOREMAP flag is used to determine if a
> port gets ioremap'd, but plays no part in whether it gets iounmap'd.
> It's easy to see how an MMIO serial port can be passed through
> uart_set_info and end up with an unmapped membase. The results is a
> non-functional UART or worse. I've tried to generate some discussion on
> the proper fix for this, but I haven't succeeded. I propose the patch
> below as a safe compromise. An MMIO uart w/ a mapbase, but no membase
> doesn't seem viable to me. Thanks,
I'm going to be doing a major update to serial today, and these patches
are made rather unnecessary by that. Thanks for pointing out the problem
anyway.
--
Russell King
Linux kernel 2.6 ARM Linux - http://www.arm.linux.org.uk/
maintainer of: 2.6 PCMCIA - http://pcmcia.arm.linux.org.uk/
2.6 Serial core