| Message ID | 1553105650-28012-2-git-send-email-john.garry@huawei.com |
|---|---|
| State | Superseded |
| Headers | show |
| Series | Fix system crash for accessing unmapped IO port regions | expand |
Hi John, On Thu, Mar 21, 2019 at 02:14:08AM +0800, John Garry wrote: > Currently when we request an IO port region, the request is made directly > to the top resource, ioport_resource. Let's be explicit here, e.g., Currently request_region() requests an IO port region directly from the top resource, ioport_resource. > There is an issue here, in that drivers may successfully request an IO > port region even if the IO port region has not even been mapped in > (in pci_remap_iospace()). > > This may lead to crashes when the system has no PCI host, or, has a host > but it has failed enumeration, while drivers still attempt to access PCI > IO ports, as below: I don't understand the strategy here. f71882fg is not a driver for a PCI device, so it should work even if there is no PCI host in the system. 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? > 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)$ Please remove the timestamps (because they don't contribute useful information) and indent the example a couple spaces (which is conventional for quoted material). > 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. > Signed-off-by: John Garry <john.garry@huawei.com> > --- > include/linux/ioport.h | 12 +++++++++--- > kernel/resource.c | 28 ++++++++++++++++++++++++++++ > 2 files changed, 37 insertions(+), 3 deletions(-) > > diff --git a/include/linux/ioport.h b/include/linux/ioport.h > index da0ebaec25f0..d7b7e1e08291 100644 > --- a/include/linux/ioport.h > +++ b/include/linux/ioport.h > @@ -217,19 +217,25 @@ static inline bool resource_contains(struct resource *r1, struct resource *r2) > > > /* Convenience shorthand with allocation */ > -#define request_region(start,n,name) __request_region(&ioport_resource, (start), (n), (name), 0) > -#define request_muxed_region(start,n,name) __request_region(&ioport_resource, (start), (n), (name), IORESOURCE_MUXED) > +#define request_region(start,n,name) __request_region_from_children(&ioport_resource, (start), (n), (name), 0) > +#define request_muxed_region(start,n,name) __request_region_from_children(&ioport_resource, (start), (n), (name), IORESOURCE_MUXED) > #define __request_mem_region(start,n,name, excl) __request_region(&iomem_resource, (start), (n), (name), excl) > #define request_mem_region(start,n,name) __request_region(&iomem_resource, (start), (n), (name), 0) > #define request_mem_region_exclusive(start,n,name) \ > __request_region(&iomem_resource, (start), (n), (name), IORESOURCE_EXCLUSIVE) > #define rename_region(region, newname) do { (region)->name = (newname); } while (0) > > -extern struct resource * __request_region(struct resource *, > +extern struct resource *__request_region(struct resource *, > resource_size_t start, > resource_size_t n, > const char *name, int flags); > > +extern struct resource *__request_region_from_children(struct resource *, > + resource_size_t start, > + resource_size_t n, > + const char *name, int flags); > + > + > /* Compatibility cruft */ > #define release_region(start,n) __release_region(&ioport_resource, (start), (n)) > #define release_mem_region(start,n) __release_region(&iomem_resource, (start), (n)) > diff --git a/kernel/resource.c b/kernel/resource.c > index 92190f62ebc5..87ed200eda8b 100644 > --- a/kernel/resource.c > +++ b/kernel/resource.c > @@ -1097,6 +1097,34 @@ resource_size_t resource_alignment(struct resource *res) > > static DECLARE_WAIT_QUEUE_HEAD(muxed_resource_wait); > > +/** > + * __request_region_from_children - create a new busy region from a child > + * @parent: parent resource descriptor > + * @start: resource start address > + * @n: resource region size > + * @name: reserving caller's ID string > + * @flags: IO resource flags > + */ > +struct resource *__request_region_from_children(struct resource *parent, > + resource_size_t start, > + resource_size_t n, > + const char *name, int flags) > +{ > + struct resource *res = __request_region(parent, start, n, name, flags); > + > + if (res && res->parent == parent) { > + /* > + * This is a direct descendent of the parent, which is > + * what we didn't want. > + */ > + __release_region(parent, start, n); > + res = NULL; > + } > + > + return res; > +} > +EXPORT_SYMBOL(__request_region_from_children); > + > /** > * __request_region - create a new busy resource region > * @parent: parent resource descriptor > -- > 2.17.1 >
On 25/03/2019 23:32, Bjorn Helgaas wrote: > Hi John, > Hi Bjorn, Thanks for reviewing this. > On Thu, Mar 21, 2019 at 02:14:08AM +0800, John Garry wrote: >> Currently when we request an IO port region, the request is made directly >> to the top resource, ioport_resource. > > Let's be explicit here, e.g., > > Currently request_region() requests an IO port region directly from the > top resource, ioport_resource. ok > >> There is an issue here, in that drivers may successfully request an IO >> port region even if the IO port region has not even been mapped in >> (in pci_remap_iospace()). >> >> This may lead to crashes when the system has no PCI host, or, has a host >> but it has failed enumeration, while drivers still attempt to access PCI >> IO ports, as below: > > I don't understand the strategy here. f71882fg is not a driver for a > PCI device, so it should work even if there is no PCI host in the > system. 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: $more /proc/ioports 0000-0cf7 : PCI Bus 0000:00 0000-001f : dma1 0020-0021 : pic1 0040-0043 : timer0 0050-0053 : timer1 0060-0060 : keyboard 0064-0064 : keyboard 0070-0077 : rtc0 0080-008f : dma page reg 00a0-00a1 : pic2 00c0-00df : dma2 00f0-00ff : fpu So, the idea in the patch is that if PCI Bus 0000:00 does not exist because of no PCI host, then we should fail a request to an IO port region. > > On x86, I think inb/inw/inl from a port where nothing responds > probably just returns ~0, and outb/outw/outl just get dropped. > Shouldn't arm64 do the same, without crashing? That would be ideal and we're doing something similar in patch 2/3. So on ARM64 we have to IO remap the PCI IO resource. If this mapping is not done (due to no PCI host), then any inb/inw/inl calls will crash the system. So in patch 2/3, I am also making the change to the logical PIO inb/inw/inl accessors to discard accesses when no PCI MMIO regions are registered in logical PIO space. This is really a second line of defense (this patch being the first). > >> root@(none)$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)$ > > Please remove the timestamps (because they don't contribute useful > information) and indent the example a couple spaces (which is > conventional for quoted material). ok > >> 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. > >> Signed-off-by: John Garry <john.garry@huawei.com> >> --- Thanks! >> include/linux/ioport.h | 12 +++++++++--- >> kernel/resource.c | 28 ++++++++++++++++++++++++++++ >> 2 files changed, 37 insertions(+), 3 deletions(-) >> Leaving remaing text as a reference. >> diff --git a/include/linux/ioport.h b/include/linux/ioport.h >> index da0ebaec25f0..d7b7e1e08291 100644 >> --- a/include/linux/ioport.h >> +++ b/include/linux/ioport.h >> @@ -217,19 +217,25 @@ static inline bool resource_contains(struct resource *r1, struct resource *r2) >> >> >> /* Convenience shorthand with allocation */ >> -#define request_region(start,n,name) __request_region(&ioport_resource, (start), (n), (name), 0) >> -#define request_muxed_region(start,n,name) __request_region(&ioport_resource, (start), (n), (name), IORESOURCE_MUXED) >> +#define request_region(start,n,name) __request_region_from_children(&ioport_resource, (start), (n), (name), 0) >> +#define request_muxed_region(start,n,name) __request_region_from_children(&ioport_resource, (start), (n), (name), IORESOURCE_MUXED) >> #define __request_mem_region(start,n,name, excl) __request_region(&iomem_resource, (start), (n), (name), excl) >> #define request_mem_region(start,n,name) __request_region(&iomem_resource, (start), (n), (name), 0) >> #define request_mem_region_exclusive(start,n,name) \ >> __request_region(&iomem_resource, (start), (n), (name), IORESOURCE_EXCLUSIVE) >> #define rename_region(region, newname) do { (region)->name = (newname); } while (0) >> >> -extern struct resource * __request_region(struct resource *, >> +extern struct resource *__request_region(struct resource *, >> resource_size_t start, >> resource_size_t n, >> const char *name, int flags); >> >> +extern struct resource *__request_region_from_children(struct resource *, >> + resource_size_t start, >> + resource_size_t n, >> + const char *name, int flags); >> + >> + >> /* Compatibility cruft */ >> #define release_region(start,n) __release_region(&ioport_resource, (start), (n)) >> #define release_mem_region(start,n) __release_region(&iomem_resource, (start), (n)) >> diff --git a/kernel/resource.c b/kernel/resource.c >> index 92190f62ebc5..87ed200eda8b 100644 >> --- a/kernel/resource.c >> +++ b/kernel/resource.c >> @@ -1097,6 +1097,34 @@ resource_size_t resource_alignment(struct resource *res) >> >> static DECLARE_WAIT_QUEUE_HEAD(muxed_resource_wait); >> >> +/** >> + * __request_region_from_children - create a new busy region from a child >> + * @parent: parent resource descriptor >> + * @start: resource start address >> + * @n: resource region size >> + * @name: reserving caller's ID string >> + * @flags: IO resource flags >> + */ >> +struct resource *__request_region_from_children(struct resource *parent, >> + resource_size_t start, >> + resource_size_t n, >> + const char *name, int flags) >> +{ >> + struct resource *res = __request_region(parent, start, n, name, flags); >> + >> + if (res && res->parent == parent) { >> + /* >> + * This is a direct descendent of the parent, which is >> + * what we didn't want. >> + */ >> + __release_region(parent, start, n); >> + res = NULL; >> + } >> + >> + return res; >> +} >> +EXPORT_SYMBOL(__request_region_from_children); >> + >> /** >> * __request_region - create a new busy resource region >> * @parent: parent resource descriptor >> -- >> 2.17.1 >> > > . >
[+cc Catalin, Will, linux-arm-kernel] On Tue, Mar 26, 2019 at 04:33:55PM +0000, John Garry wrote: > On 25/03/2019 23:32, Bjorn Helgaas wrote: > > On Thu, Mar 21, 2019 at 02:14:08AM +0800, John Garry wrote: > > > Currently when we request an IO port region, the request is made directly > > > to the top resource, ioport_resource. > > > > Let's be explicit here, e.g., > > > > Currently request_region() requests an IO port region directly from the > > top resource, ioport_resource. > > ok > > > > There is an issue here, in that drivers may successfully request an IO > > > port region even if the IO port region has not even been mapped in > > > (in pci_remap_iospace()). > > > > > > This may lead to crashes when the system has no PCI host, or, has a host > > > but it has failed enumeration, while drivers still attempt to access PCI > > > IO ports, as below: > > > > I don't understand the strategy here. f71882fg is not a driver for a > > PCI device, so it should work even if there is no PCI host in the > > system. > > 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: > > $more /proc/ioports > 0000-0cf7 : PCI Bus 0000:00 > 0000-001f : dma1 > 0020-0021 : pic1 > 0040-0043 : timer0 > 0050-0053 : timer1 > 0060-0060 : keyboard > 0064-0064 : keyboard > 0070-0077 : rtc0 > 0080-008f : dma page reg > 00a0-00a1 : pic2 > 00c0-00df : dma2 > 00f0-00ff : fpu > > So, the idea in the patch is that if PCI Bus 0000:00 does not exist because > of no PCI host, then we should fail a request to an IO port region. I'm not convinced about this last sentence. It's true that on most modern systems, including that Intel PCH, the Super I/O controller is attached via an LPC bridge on a PCI bus. But I don't think it's an actual requirement that PCI be involved. There certainly once were systems, e.g., PC/104, that had ISA devices but no PCI. Maybe Super I/O attached via ISA is obsolete enough that we don't care any more, but I really don't know. > > On x86, I think inb/inw/inl from a port where nothing responds > > probably just returns ~0, and outb/outw/outl just get dropped. > > Shouldn't arm64 do the same, without crashing? > > That would be ideal and we're doing something similar in patch 2/3. > > So on ARM64 we have to IO remap the PCI IO resource. If this mapping is not > done (due to no PCI host), then any inb/inw/inl calls will crash the system. My take is that ARM64 is responsible for implementing inb/inw/inl in such a way that they don't crash. I don't think it's practical to update all the old ISA drivers or even the core code to work around that. > So in patch 2/3, I am also making the change to the logical PIO inb/inw/inl > accessors to discard accesses when no PCI MMIO regions are registered in > logical PIO space. > > This is really a second line of defense (this patch being the first). > > > > root@(none)$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 Bjorn > > > Signed-off-by: John Garry <john.garry@huawei.com> > > > --- > > Thanks! > > > > include/linux/ioport.h | 12 +++++++++--- > > > kernel/resource.c | 28 ++++++++++++++++++++++++++++ > > > 2 files changed, 37 insertions(+), 3 deletions(-) > > > > > Leaving remaing text as a reference. > > > > diff --git a/include/linux/ioport.h b/include/linux/ioport.h > > > index da0ebaec25f0..d7b7e1e08291 100644 > > > --- a/include/linux/ioport.h > > > +++ b/include/linux/ioport.h > > > @@ -217,19 +217,25 @@ static inline bool resource_contains(struct resource *r1, struct resource *r2) > > > > > > > > > /* Convenience shorthand with allocation */ > > > -#define request_region(start,n,name) __request_region(&ioport_resource, (start), (n), (name), 0) > > > -#define request_muxed_region(start,n,name) __request_region(&ioport_resource, (start), (n), (name), IORESOURCE_MUXED) > > > +#define request_region(start,n,name) __request_region_from_children(&ioport_resource, (start), (n), (name), 0) > > > +#define request_muxed_region(start,n,name) __request_region_from_children(&ioport_resource, (start), (n), (name), IORESOURCE_MUXED) > > > #define __request_mem_region(start,n,name, excl) __request_region(&iomem_resource, (start), (n), (name), excl) > > > #define request_mem_region(start,n,name) __request_region(&iomem_resource, (start), (n), (name), 0) > > > #define request_mem_region_exclusive(start,n,name) \ > > > __request_region(&iomem_resource, (start), (n), (name), IORESOURCE_EXCLUSIVE) > > > #define rename_region(region, newname) do { (region)->name = (newname); } while (0) > > > > > > -extern struct resource * __request_region(struct resource *, > > > +extern struct resource *__request_region(struct resource *, > > > resource_size_t start, > > > resource_size_t n, > > > const char *name, int flags); > > > > > > +extern struct resource *__request_region_from_children(struct resource *, > > > + resource_size_t start, > > > + resource_size_t n, > > > + const char *name, int flags); > > > + > > > + > > > /* Compatibility cruft */ > > > #define release_region(start,n) __release_region(&ioport_resource, (start), (n)) > > > #define release_mem_region(start,n) __release_region(&iomem_resource, (start), (n)) > > > diff --git a/kernel/resource.c b/kernel/resource.c > > > index 92190f62ebc5..87ed200eda8b 100644 > > > --- a/kernel/resource.c > > > +++ b/kernel/resource.c > > > @@ -1097,6 +1097,34 @@ resource_size_t resource_alignment(struct resource *res) > > > > > > static DECLARE_WAIT_QUEUE_HEAD(muxed_resource_wait); > > > > > > +/** > > > + * __request_region_from_children - create a new busy region from a child > > > + * @parent: parent resource descriptor > > > + * @start: resource start address > > > + * @n: resource region size > > > + * @name: reserving caller's ID string > > > + * @flags: IO resource flags > > > + */ > > > +struct resource *__request_region_from_children(struct resource *parent, > > > + resource_size_t start, > > > + resource_size_t n, > > > + const char *name, int flags) > > > +{ > > > + struct resource *res = __request_region(parent, start, n, name, flags); > > > + > > > + if (res && res->parent == parent) { > > > + /* > > > + * This is a direct descendent of the parent, which is > > > + * what we didn't want. > > > + */ > > > + __release_region(parent, start, n); > > > + res = NULL; > > > + } > > > + > > > + return res; > > > +} > > > +EXPORT_SYMBOL(__request_region_from_children); > > > + > > > /** > > > * __request_region - create a new busy resource region > > > * @parent: parent resource descriptor > > > -- > > > 2.17.1 > > > > > > > . > > > >
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 >
On Tue, Mar 26, 2019 at 05:48:10PM -0500, Bjorn Helgaas wrote: [...] > 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. > > > > 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. The problem is that those drivers are accessing a resource that does not exist in practice, it is taken for granted on x86 systems (and on IA64) because that was an actual bus (actual or emulated) and was made part of the architecture. The ISA space is not necessarily tied to PCI, at least not always. Side note: these drivers can't be compiled on PPC, it would be good to understand why, I have a hunch it can be related. > > 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 Yes, the bottom line is what Arnd outlined in the thread above. ISA IO port space is not necessarily PCI but it does not exist architecturally on ARM systems. Taking the example of IA64, the ISA space is memory mapped (like any other arch except for x86) but IIUC the virtual mapping for the ISA port space _always_ exists on IA64 so this issue won't happen. Arnd pointed out a solution in the thread above but I need to check if that's feasible. Lorenzo
On 28/03/2019 17:46, Lorenzo Pieralisi wrote: > On Tue, Mar 26, 2019 at 05:48:10PM -0500, Bjorn Helgaas wrote: > > [...] > Hi Lorenzo, >> 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. >> >>>> 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. > > The problem is that those drivers are accessing a resource that does not > exist in practice, it is taken for granted on x86 systems (and on IA64) > because that was an actual bus (actual or emulated) and was made part of > the architecture. The ISA space is not necessarily tied to PCI, > at least not always. > > Side note: these drivers can't be compiled on PPC, it would be > good to understand why, I have a hunch it can be related. I mentioned this earlier: I saw that in commits like 746cdfbf01c0 ("hwmon: Avoid building drivers forpowerpc that read/write ISA addresses"), PPC would not build these drivers, as, like arm, it has no native ISA. However I still don't think just avoiding compiling these drivers for certain archs solves the problem. [...] >>>>> [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 > > Yes, the bottom line is what Arnd outlined in the thread above. > > ISA IO port space is not necessarily PCI but it does not exist > architecturally on ARM systems. > > Taking the example of IA64, the ISA space is memory mapped (like any > other arch except for x86) but IIUC the virtual mapping for the ISA > port space _always_ exists on IA64 so this issue won't happen. > > Arnd pointed out a solution in the thread above but I need to check > if that's feasible. I doubt that it can work now. Since we when introduced the concept of logical PIO space, this IO space became sparely populated by 2 regions - MMIO and indirect IO - so we cannot grow it as we map in regions. I also don't think it works for when we IO unmap regions. Thanks, John > > Lorenzo > > _______________________________________________ > linux-arm-kernel mailing list > linux-arm-kernel@lists.infradead.org > http://lists.infradead.org/mailman/listinfo/linux-arm-kernel > > . >
On Fri, Mar 29, 2019 at 10:42:17AM +0000, John Garry wrote: > On 28/03/2019 17:46, Lorenzo Pieralisi wrote: > >On Tue, Mar 26, 2019 at 05:48:10PM -0500, Bjorn Helgaas wrote: > > > >[...] > > > > Hi Lorenzo, > > > >>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. > >> > >>>>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. > > > >The problem is that those drivers are accessing a resource that does not > >exist in practice, it is taken for granted on x86 systems (and on IA64) > >because that was an actual bus (actual or emulated) and was made part of > >the architecture. The ISA space is not necessarily tied to PCI, > >at least not always. > > > >Side note: these drivers can't be compiled on PPC, it would be > >good to understand why, I have a hunch it can be related. > > I mentioned this earlier: > > I saw that in commits like 746cdfbf01c0 ("hwmon: Avoid building drivers > forpowerpc that read/write ISA addresses"), PPC would not build these > drivers, as, like arm, it has no native ISA. > > However I still don't think just avoiding compiling these drivers for > certain archs solves the problem. No it does not but I would like to understand how relevant is fixing those drivers (that should not use ISA IO space without first claiming their resources, for the records) given that PPC did not even try and apparently that's not a problem. > > [...] > > >>>>>[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 > > > >Yes, the bottom line is what Arnd outlined in the thread above. > > > >ISA IO port space is not necessarily PCI but it does not exist > >architecturally on ARM systems. > > > >Taking the example of IA64, the ISA space is memory mapped (like any > >other arch except for x86) but IIUC the virtual mapping for the ISA > >port space _always_ exists on IA64 so this issue won't happen. > > > >Arnd pointed out a solution in the thread above but I need to check > >if that's feasible. > > I doubt that it can work now. > > Since we when introduced the concept of logical PIO space, this IO space > became sparely populated by 2 regions - MMIO and indirect IO - so we cannot > grow it as we map in regions. I also don't think it works for when we IO > unmap regions. I do not have the full picture but I suspect that, apart from x86/IA64, this is a common issue across architectures, I am trying to untangle how ARM 32-bit deals with this (if it does). Lorenzo
On 29/03/2019 12:22, Lorenzo Pieralisi wrote: >>> > >Side note: these drivers can't be compiled on PPC, it would be >>> > >good to understand why, I have a hunch it can be related. >> > >> > I mentioned this earlier: >> > >> > I saw that in commits like 746cdfbf01c0 ("hwmon: Avoid building drivers >> > forpowerpc that read/write ISA addresses"), PPC would not build these >> > drivers, as, like arm, it has no native ISA. >> > >> > However I still don't think just avoiding compiling these drivers for >> > certain archs solves the problem. > No it does not but I would like to understand how relevant is fixing > those drivers (that should not use ISA IO space without first claiming > their resources, for the records) given that PPC did not even try and > apparently that's not a problem. > Hi Lorenzo, Those drivers should still be fixed up separately. The tricky part in this series is making the resource claim fail if there is no IO space mapped/accessible at the addresses requested. However I would still like to fix up the low level IO port accessors to discard accesses when no IO space is mapped. Thanks, John >> > >> > [...] >> > >>>>>>> > >>>>>[1] https://www.spinics.net/lists/linux-pci
diff --git a/include/linux/ioport.h b/include/linux/ioport.h index da0ebaec25f0..d7b7e1e08291 100644 --- a/include/linux/ioport.h +++ b/include/linux/ioport.h @@ -217,19 +217,25 @@ static inline bool resource_contains(struct resource *r1, struct resource *r2) /* Convenience shorthand with allocation */ -#define request_region(start,n,name) __request_region(&ioport_resource, (start), (n), (name), 0) -#define request_muxed_region(start,n,name) __request_region(&ioport_resource, (start), (n), (name), IORESOURCE_MUXED) +#define request_region(start,n,name) __request_region_from_children(&ioport_resource, (start), (n), (name), 0) +#define request_muxed_region(start,n,name) __request_region_from_children(&ioport_resource, (start), (n), (name), IORESOURCE_MUXED) #define __request_mem_region(start,n,name, excl) __request_region(&iomem_resource, (start), (n), (name), excl) #define request_mem_region(start,n,name) __request_region(&iomem_resource, (start), (n), (name), 0) #define request_mem_region_exclusive(start,n,name) \ __request_region(&iomem_resource, (start), (n), (name), IORESOURCE_EXCLUSIVE) #define rename_region(region, newname) do { (region)->name = (newname); } while (0) -extern struct resource * __request_region(struct resource *, +extern struct resource *__request_region(struct resource *, resource_size_t start, resource_size_t n, const char *name, int flags); +extern struct resource *__request_region_from_children(struct resource *, + resource_size_t start, + resource_size_t n, + const char *name, int flags); + + /* Compatibility cruft */ #define release_region(start,n) __release_region(&ioport_resource, (start), (n)) #define release_mem_region(start,n) __release_region(&iomem_resource, (start), (n)) diff --git a/kernel/resource.c b/kernel/resource.c index 92190f62ebc5..87ed200eda8b 100644 --- a/kernel/resource.c +++ b/kernel/resource.c @@ -1097,6 +1097,34 @@ resource_size_t resource_alignment(struct resource *res) static DECLARE_WAIT_QUEUE_HEAD(muxed_resource_wait); +/** + * __request_region_from_children - create a new busy region from a child + * @parent: parent resource descriptor + * @start: resource start address + * @n: resource region size + * @name: reserving caller's ID string + * @flags: IO resource flags + */ +struct resource *__request_region_from_children(struct resource *parent, + resource_size_t start, + resource_size_t n, + const char *name, int flags) +{ + struct resource *res = __request_region(parent, start, n, name, flags); + + if (res && res->parent == parent) { + /* + * This is a direct descendent of the parent, which is + * what we didn't want. + */ + __release_region(parent, start, n); + res = NULL; + } + + return res; +} +EXPORT_SYMBOL(__request_region_from_children); + /** * __request_region - create a new busy resource region * @parent: parent resource descriptor
Currently when we request an IO port region, the request is made directly to the top resource, ioport_resource. There is an issue here, in that drivers may successfully request an IO port region even if the IO port region has not even been mapped in (in pci_remap_iospace()). This may lead to crashes when the system has no PCI host, or, has a host but it has failed enumeration, while drivers still attempt to access PCI IO ports, as below: 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 Signed-off-by: John Garry <john.garry@huawei.com> --- include/linux/ioport.h | 12 +++++++++--- kernel/resource.c | 28 ++++++++++++++++++++++++++++ 2 files changed, 37 insertions(+), 3 deletions(-) -- 2.17.1