Message-ID: <2343a281-eb6e-99f9-4600-96931d00d215@huawei.com>
Date:   Wed, 27 Mar 2019 11:24:33 +0000
From:   John Garry <john.garry@...wei.com>
To:     Bjorn Helgaas <helgaas@...nel.org>
CC:     <rafael@...nel.org>, <arnd@...db.de>, <lorenzo.pieralisi@....com>,
        <bp@...e.de>, <linux@...ck-us.net>, <linux-kernel@...r.kernel.org>,
        <linux-pci@...r.kernel.org>, <wangkefeng.wang@...wei.com>,
        <linuxarm@...wei.com>, <andy.shevchenko@...il.com>,
        Catalin Marinas <catalin.marinas@....com>,
        Will Deacon <will.deacon@....com>,
        <linux-arm-kernel@...ts.infradead.org>
Subject: Re: [RFC PATCH v2 1/3] resource: Request IO port regions from
 children of ioport_resource

On 26/03/2019 22:48, Bjorn Helgaas wrote:
> [+cc Catalin, Will, linux-arm-kernel]
>
>> From my checking, the f71882fg hwmon is accessed via the super-io interface
>> on the PCH on x86. The super-io interface is at fixed addresses, those being
>> 0x2e and 0x4e.
>>
>> Please see the following:
>>
>> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/drivers/hwmon/f71805f.c?h=v5.1-rc2#n1621
>>
>> and
>>
>> https://www.intel.com/content/dam/www/public/us/en/documents/datasheets/8-series-chipset-pch-datasheet.pdf
>> (Table 9.2).
>>
>> On x86 systems, these PCH IO ports will be mapped on a PCI bus, like:
>>
>> $more /proc/ioports
>> 0000-0cf7 : PCI Bus 0000:00
>>   0000-001f : dma1
>>   0020-0021 : pic1
>>   0040-0043 : timer0
>>   0050-0053 : timer1
>>   0060-0060 : keyboard
>>   0064-0064 : keyboard
>>   0070-0077 : rtc0
>>   0080-008f : dma page reg
>>   00a0-00a1 : pic2
>>   00c0-00df : dma2
>>   00f0-00ff : fpu
>>
>> So, the idea in the patch is that if PCI Bus 0000:00 does not exist because
>> of no PCI host, then we should fail a request to an IO port region.
>

Hi Bjorn,

> I'm not convinced about this last sentence.
>
> It's true that on most modern systems, including that Intel PCH, the
> Super I/O controller is attached via an LPC bridge on a PCI bus.
>
> But I don't think it's an actual requirement that PCI be involved.
> There certainly once were systems, e.g., PC/104, that had ISA devices
> but no PCI.  Maybe Super I/O attached via ISA is obsolete enough that
> we don't care any more, but I really don't know.

OK, fine. So if this is true, then this patch falls apart.

However I don't know for sure either. I would still like to think that 
these legacy ISA system should still insert a bus resource under 
ioport_resource, from which devices on that bus should request resources.

>
>>> On x86, I think inb/inw/inl from a port where nothing responds
>>> probably just returns ~0, and outb/outw/outl just get dropped.
>>> Shouldn't arm64 do the same, without crashing?
>>
>> That would be ideal and we're doing something similar in patch 2/3.
>>
>> So on ARM64 we have to IO remap the PCI IO resource. If this mapping is not
>> done (due to no PCI host), then any inb/inw/inl calls will crash the system.
>
> My take is that ARM64 is responsible for implementing inb/inw/inl in
> such a way that they don't crash.  I don't think it's practical to
> update all the old ISA drivers or even the core code to work around
> that.

As I mentioned below, I was actually also fixing up inb/inw/inl et al 
for arm64 such that they don't crash the system in this case. This was 
in patch 2/3.

So on arm64 - which defines PCI_IOBASE - we need to IO remap the PCI IO 
space resource. If this is not done and we access PCI IO space, then we 
crash.

However with the introduction of logical PIO space in commit 
031e3601869c, we can test this mapping by ensuring that we have a 
logical PIO region registered. If there is none, then we can discard the 
access.

However this would only be for when INDIRECT_PIO is defined. Maybe I can 
make it work for when INDIRECT_PIO is not defined, or even drop 
!INDIRECT_PIO support.

A final note on hwmon f71882fg: even with the change in 2/3, this driver 
still accesses IO ports 0x2e and 0x4e, which would not be a PCH fixed IO 
port on !x86 systems, so far from ideal.

I saw that in commit 746cdfbf01c0 ("hwmon: Avoid building drivers for 
powerpc that read/write ISA addresses"), PPC would not build these 
drivers, as, like arm, it has no native ISA.

>
>> So in patch 2/3, I am also making the change to the logical PIO inb/inw/inl
>> accessors to discard accesses when no PCI MMIO regions are registered in
>> logical PIO space.
>>
>> This is really a second line of defense (this patch being the first).
>>
>>>> root@(none)$root@(none)$ insmod f71882fg.ko
>>>> [  152.215377] Unable to handle kernel paging request at virtual address ffff7dfffee0002e
>>>> [  152.231299] Mem abort info:
>>>> [  152.236898]   ESR = 0x96000046
>>>> [  152.243019]   Exception class = DABT (current EL), IL = 32 bits
>>>> [  152.254905]   SET = 0, FnV = 0
>>>> [  152.261024]   EA = 0, S1PTW = 0
>>>> [  152.267320] Data abort info:
>>>> [  152.273091]   ISV = 0, ISS = 0x00000046
>>>> [  152.280784]   CM = 0, WnR = 1
>>>> [  152.286730] swapper pgtable: 4k pages, 48-bit VAs, pgdp = (____ptrval____)
>>>> [  152.300537] [ffff7dfffee0002e] pgd=000000000141c003, pud=000000000141d003, pmd=0000000000000000
>>>> [  152.318016] Internal error: Oops: 96000046 [#1] PREEMPT SMP
>>>> [  152.329199] Modules linked in: f71882fg(+)
>>>> [  152.337415] CPU: 8 PID: 2732 Comm: insmod Not tainted 5.1.0-rc1-00002-gab1a0e9200b8-dirty #102
>>>> [  152.354712] Hardware name: Huawei Taishan 2280 /D05, BIOS Hisilicon D05 IT21 Nemo 2.0 RC0 04/18/2018
>>>> [  152.373058] pstate: 80000005 (Nzcv daif -PAN -UAO)
>>>> [  152.382675] pc : logic_outb+0x54/0xb8
>>>> [  152.390017] lr : f71882fg_find+0x64/0x390 [f71882fg]
>>>> [  152.399977] sp : ffff000013393aa0
>>>> [  152.406618] x29: ffff000013393aa0 x28: ffff000008b98b10
>>>> [  152.417278] x27: ffff000013393df0 x26: 0000000000000100
>>>> [  152.427938] x25: ffff801f8c872d30 x24: ffff000011420000
>>>> [  152.438598] x23: ffff801fb49d2940 x22: ffff000011291000
>>>> [  152.449257] x21: 000000000000002e x20: 0000000000000087
>>>> [  152.459917] x19: ffff000013393b44 x18: ffffffffffffffff
>>>> [  152.470577] x17: 0000000000000000 x16: 0000000000000000
>>>> [  152.481236] x15: ffff00001127d6c8 x14: ffff801f8cfd691c
>>>> [  152.491896] x13: 0000000000000000 x12: 0000000000000000
>>>> [  152.502555] x11: 0000000000000003 x10: 0000801feace2000
>>>> [  152.513215] x9 : 0000000000000000 x8 : ffff841fa654f280
>>>> [  152.523874] x7 : 0000000000000000 x6 : 0000000000ffc0e3
>>>> [  152.534534] x5 : ffff000011291360 x4 : ffff801fb4949f00
>>>> [  152.545194] x3 : 0000000000ffbffe x2 : 76e767a63713d500
>>>> [  152.555853] x1 : ffff7dfffee0002e x0 : ffff7dfffee00000
>>>> [  152.566514] Process insmod (pid: 2732, stack limit = 0x(____ptrval____))
>>>> [  152.579968] Call trace:
>>>> [  152.584863]  logic_outb+0x54/0xb8
>>>> [  152.591506]  f71882fg_find+0x64/0x390 [f71882fg]
>>>> [  152.600768]  f71882fg_init+0x38/0xc70 [f71882fg]
>>>> [  152.610031]  do_one_initcall+0x5c/0x198
>>>> [  152.617723]  do_init_module+0x54/0x1b0
>>>> [  152.625237]  load_module+0x1dc4/0x2158
>>>> [  152.632752]  __se_sys_init_module+0x14c/0x1e8
>>>> [  152.641490]  __arm64_sys_init_module+0x18/0x20
>>>> [  152.650404]  el0_svc_common+0x5c/0x100
>>>> [  152.657919]  el0_svc_handler+0x2c/0x80
>>>> [  152.665433]  el0_svc+0x8/0xc
>>>> [  152.671202] Code: d2bfdc00 f2cfbfe0 f2ffffe0 8b000021 (39000034)
>>>> [  152.683434] ---[ end trace fd4f35b610829a48 ]---
>>>> Segmentation fault
>>>> root@(none)$
>>>
>>>> Note that the f71882fg driver correctly calls request_muxed_region().
>>>>
>>>> This issue was originally reported in [1].
>>>>
>>>> This patch changes the functionality of request{muxed_}_region() to
>>>> request a region from a direct child descendent of the top
>>>> ioport_resource.
>>>>
>>>> In this, if the IO port region has not been mapped for a particular IO
>>>> region, the PCI IO resource would also not have been inserted, and so a
>>>> suitable child region will not exist. As such,
>>>> request_{muxed_}region() calls will fail.
>>>>
>>>> A side note: there are many drivers in the kernel which fail to even call
>>>> request_{muxed_}region() prior to IO port accesses, and they also need to
>>>> be fixed (to call request_{muxed_}region(), as appropriate) separately.
>>>>
>>>> [1] https://www.spinics.net/lists/linux-pci/msg49821.html
>>>
>>> Please use a https://lore.kernel.org/ URL instead of spinics.net.
>>
>> ok, I hope that I can find this old thread.
>
> The beauty of lore.kernel.org is that the URL contains the Message-ID, so
> it's easy build the URL and it would contain useful information even if
> lore.kernel.org disappeared:
>
> https://lore.kernel.org/linux-pci/56F209A9.4040304@huawei.com
>

ok, great.

Thanks again,
John

> Bjorn
>

Powered by blists - more mailing lists