Message ID | 1498523473-157373-1-git-send-email-shawn.lin@rock-chips.com (mailing list archive) |
---|---|
State | New, archived |
Delegated to: | Bjorn Helgaas |
Headers | show |
[+cc Lorenzo, resending because I fat-fingered the cc line and subject] On Tue, Jun 27, 2017 at 08:31:13AM +0800, Shawn Lin wrote: > If not getting domain number from DT, the domain number will > keep increasing once doing unbind/bind RC drivers. This could > introduce pointless tree view of lspci as shows below: > > -+-[0001:00]---00.0-[01]----00.0 > \-[0000:00]- > > The more test we do, the lengthier it would be. The more serious > issue is that if attaching two hierarchies for two different domains > belonging to two root bridges, so when doing unbind/bind test for one > of them and keep the other, then the domain number would finally > overflow and make the two hierarchies of devices share the some domain > number but actually they shouldn't. So it looks like we need to invent > a new indexing ID mechanism to manage domain number. This patch > introduces idr to achieve our purpose. > > Signed-off-by: Shawn Lin <shawn.lin@rock-chips.com> The "use_dt_domains" logic in of_pci_bus_find_domain_nr() is fairly obtuse. I *think*, now that we have pci_scan_root_bus_bridge() due to Lorenzo's excellent work, the time is ripe for moving the domain number from arch-specific places into struct pci_host_bridge. I suspect that will end up simplifying the CONFIG_PCI_DOMAINS vs CONFIG_PCI_DOMAINS_GENERIC situation, and I wonder whether it might enable some simplification of of_pci_bus_find_domain_nr() as well, which in turn, might make *this* patch simpler. This isn't that big a patch to begin with, so I could apply it as-is and we could do more domain cleanup later. It's just that it's intertwined with the PCI_DOMAINS #ifdefs and maybe there's an opportunity to make this story more readable if those are out of the way. Any thoughts? > --- > > Changes in v5: > - fix Bjorn's comments for v3 as actually I didn't address his comments > for v3 when posted my v4. > v3: https://patchwork.kernel.org/patch/9742003/ > For ACPI case, we never use DT or IDA to get domain numbers and > acpi_pci_bus_find_domain_nr would return the proper value we need > whether _SEG exist or not. For DT or IDA, we hope bridges use one of > them consistently, and this is actually what the pre-existing code > did. So I remove ida_domain now since it complicated the logic of the > code and we could just make use_dt_domains global instead to slightly > achieve our purpose. > > Changes in v4: > - make domain_nr depends on CONFIG_PCI_DOMAINS instead of > CONFIG_PCI_DOMAINS_GENERIC. > > Changes in v3: > - make ida_domain a system-wide property and check it in the code to > combine with use_dt_domains. Also update the comment there. > > Changes in v2: > - add a remove wrapper > - rename use_dt_domains to ida_domain and set this bit > in pci_get_new_domain_nr and test it in the remove wrapper. > > drivers/pci/pci.c | 13 ++++++++++--- > drivers/pci/remove.c | 2 ++ > include/linux/pci.h | 7 +++++-- > 3 files changed, 17 insertions(+), 5 deletions(-) > > diff --git a/drivers/pci/pci.c b/drivers/pci/pci.c > index b58a6b3..9953eaf 100644 > --- a/drivers/pci/pci.c > +++ b/drivers/pci/pci.c > @@ -11,6 +11,7 @@ > #include <linux/kernel.h> > #include <linux/delay.h> > #include <linux/dmi.h> > +#include <linux/idr.h> > #include <linux/init.h> > #include <linux/of.h> > #include <linux/of_pci.h> > @@ -5305,17 +5306,23 @@ static void pci_no_domains(void) > } > > #ifdef CONFIG_PCI_DOMAINS > -static atomic_t __domain_nr = ATOMIC_INIT(-1); > +static DEFINE_IDA(__domain_nr); > > int pci_get_new_domain_nr(void) > { > - return atomic_inc_return(&__domain_nr); > + return ida_simple_get(&__domain_nr, 0, sizeof(u64), GFP_KERNEL); > +} > + > +static int use_dt_domains = -1; > +void pci_put_domain_nr(struct pci_bus *bus) > +{ > + if (acpi_disabled && use_dt_domains != 1) > + ida_simple_remove(&__domain_nr, bus->domain_nr); > } > > #ifdef CONFIG_PCI_DOMAINS_GENERIC > static int of_pci_bus_find_domain_nr(struct device *parent) > { > - static int use_dt_domains = -1; > int domain = -1; > > if (parent) > diff --git a/drivers/pci/remove.c b/drivers/pci/remove.c > index 73a03d3..01dd1b4 100644 > --- a/drivers/pci/remove.c > +++ b/drivers/pci/remove.c > @@ -157,6 +157,8 @@ void pci_remove_root_bus(struct pci_bus *bus) > list_for_each_entry_safe(child, tmp, > &bus->devices, bus_list) > pci_remove_bus_device(child); > + > + pci_put_domain_nr(bus); > pci_remove_bus(bus); > host_bridge->bus = NULL; > > diff --git a/include/linux/pci.h b/include/linux/pci.h > index 18cc70a..d6be9596 100644 > --- a/include/linux/pci.h > +++ b/include/linux/pci.h > @@ -523,7 +523,7 @@ struct pci_bus { > unsigned char primary; /* number of primary bridge */ > unsigned char max_bus_speed; /* enum pci_bus_speed */ > unsigned char cur_bus_speed; /* enum pci_bus_speed */ > -#ifdef CONFIG_PCI_DOMAINS_GENERIC > +#ifdef CONFIG_PCI_DOMAINS > int domain_nr; > #endif > > @@ -1466,11 +1466,14 @@ static inline int pci_enable_ptm(struct pci_dev *dev, u8 *granularity) > #ifdef CONFIG_PCI_DOMAINS > extern int pci_domains_supported; > int pci_get_new_domain_nr(void); > +void pci_put_domain_nr(struct pci_bus *bus); > #else > enum { pci_domains_supported = 0 }; > static inline int pci_domain_nr(struct pci_bus *bus) { return 0; } > static inline int pci_proc_domain(struct pci_bus *bus) { return 0; } > -static inline int pci_get_new_domain_nr(void) { return -ENOSYS; } > +static inline int pci_get_new_domain_nr(void) > +{ return -ENOSYS; } > +static inline void pci_put_domain_nr(struct pci_bus *bus) { } > #endif /* CONFIG_PCI_DOMAINS */ > > /* > -- > 1.9.1 > >
Hi Bjorn, On 2017/8/12 5:17, Bjorn Helgaas wrote: > [+cc Lorenzo, resending because I fat-fingered the cc line and subject] > > On Tue, Jun 27, 2017 at 08:31:13AM +0800, Shawn Lin wrote: >> If not getting domain number from DT, the domain number will >> keep increasing once doing unbind/bind RC drivers. This could >> introduce pointless tree view of lspci as shows below: >> >> -+-[0001:00]---00.0-[01]----00.0 >> \-[0000:00]- >> >> The more test we do, the lengthier it would be. The more serious >> issue is that if attaching two hierarchies for two different domains >> belonging to two root bridges, so when doing unbind/bind test for one >> of them and keep the other, then the domain number would finally >> overflow and make the two hierarchies of devices share the some domain >> number but actually they shouldn't. So it looks like we need to invent >> a new indexing ID mechanism to manage domain number. This patch >> introduces idr to achieve our purpose. >> >> Signed-off-by: Shawn Lin <shawn.lin@rock-chips.com> > > The "use_dt_domains" logic in of_pci_bus_find_domain_nr() is fairly > obtuse. I *think*, now that we have pci_scan_root_bus_bridge() due to > Lorenzo's excellent work, the time is ripe for moving the domain > number from arch-specific places into struct pci_host_bridge. > > I suspect that will end up simplifying the CONFIG_PCI_DOMAINS vs > CONFIG_PCI_DOMAINS_GENERIC situation, and I wonder whether it might > enable some simplification of of_pci_bus_find_domain_nr() as well, > which in turn, might make *this* patch simpler. > > This isn't that big a patch to begin with, so I could apply it as-is > and we could do more domain cleanup later. It's just that it's > intertwined with the PCI_DOMAINS #ifdefs and maybe there's an > opportunity to make this story more readable if those are out of the > way. Any thoughts? That sounds good to me that aftering add IDA domain, we could start considering moving domain number from arch-specific places into the bridge code and may be could also finally remove the macro CONFIG_PCI_DOMAIN* both? > >> --- >> >> Changes in v5: >> - fix Bjorn's comments for v3 as actually I didn't address his comments >> for v3 when posted my v4. >> v3: https://patchwork.kernel.org/patch/9742003/ >> For ACPI case, we never use DT or IDA to get domain numbers and >> acpi_pci_bus_find_domain_nr would return the proper value we need >> whether _SEG exist or not. For DT or IDA, we hope bridges use one of >> them consistently, and this is actually what the pre-existing code >> did. So I remove ida_domain now since it complicated the logic of the >> code and we could just make use_dt_domains global instead to slightly >> achieve our purpose. >> >> Changes in v4: >> - make domain_nr depends on CONFIG_PCI_DOMAINS instead of >> CONFIG_PCI_DOMAINS_GENERIC. >> >> Changes in v3: >> - make ida_domain a system-wide property and check it in the code to >> combine with use_dt_domains. Also update the comment there. >> >> Changes in v2: >> - add a remove wrapper >> - rename use_dt_domains to ida_domain and set this bit >> in pci_get_new_domain_nr and test it in the remove wrapper. >> >> drivers/pci/pci.c | 13 ++++++++++--- >> drivers/pci/remove.c | 2 ++ >> include/linux/pci.h | 7 +++++-- >> 3 files changed, 17 insertions(+), 5 deletions(-) >> >> diff --git a/drivers/pci/pci.c b/drivers/pci/pci.c >> index b58a6b3..9953eaf 100644 >> --- a/drivers/pci/pci.c >> +++ b/drivers/pci/pci.c >> @@ -11,6 +11,7 @@ >> #include <linux/kernel.h> >> #include <linux/delay.h> >> #include <linux/dmi.h> >> +#include <linux/idr.h> >> #include <linux/init.h> >> #include <linux/of.h> >> #include <linux/of_pci.h> >> @@ -5305,17 +5306,23 @@ static void pci_no_domains(void) >> } >> >> #ifdef CONFIG_PCI_DOMAINS >> -static atomic_t __domain_nr = ATOMIC_INIT(-1); >> +static DEFINE_IDA(__domain_nr); >> >> int pci_get_new_domain_nr(void) >> { >> - return atomic_inc_return(&__domain_nr); >> + return ida_simple_get(&__domain_nr, 0, sizeof(u64), GFP_KERNEL); >> +} >> + >> +static int use_dt_domains = -1; >> +void pci_put_domain_nr(struct pci_bus *bus) >> +{ >> + if (acpi_disabled && use_dt_domains != 1) >> + ida_simple_remove(&__domain_nr, bus->domain_nr); >> } >> >> #ifdef CONFIG_PCI_DOMAINS_GENERIC >> static int of_pci_bus_find_domain_nr(struct device *parent) >> { >> - static int use_dt_domains = -1; >> int domain = -1; >> >> if (parent) >> diff --git a/drivers/pci/remove.c b/drivers/pci/remove.c >> index 73a03d3..01dd1b4 100644 >> --- a/drivers/pci/remove.c >> +++ b/drivers/pci/remove.c >> @@ -157,6 +157,8 @@ void pci_remove_root_bus(struct pci_bus *bus) >> list_for_each_entry_safe(child, tmp, >> &bus->devices, bus_list) >> pci_remove_bus_device(child); >> + >> + pci_put_domain_nr(bus); >> pci_remove_bus(bus); >> host_bridge->bus = NULL; >> >> diff --git a/include/linux/pci.h b/include/linux/pci.h >> index 18cc70a..d6be9596 100644 >> --- a/include/linux/pci.h >> +++ b/include/linux/pci.h >> @@ -523,7 +523,7 @@ struct pci_bus { >> unsigned char primary; /* number of primary bridge */ >> unsigned char max_bus_speed; /* enum pci_bus_speed */ >> unsigned char cur_bus_speed; /* enum pci_bus_speed */ >> -#ifdef CONFIG_PCI_DOMAINS_GENERIC >> +#ifdef CONFIG_PCI_DOMAINS >> int domain_nr; >> #endif >> >> @@ -1466,11 +1466,14 @@ static inline int pci_enable_ptm(struct pci_dev *dev, u8 *granularity) >> #ifdef CONFIG_PCI_DOMAINS >> extern int pci_domains_supported; >> int pci_get_new_domain_nr(void); >> +void pci_put_domain_nr(struct pci_bus *bus); >> #else >> enum { pci_domains_supported = 0 }; >> static inline int pci_domain_nr(struct pci_bus *bus) { return 0; } >> static inline int pci_proc_domain(struct pci_bus *bus) { return 0; } >> -static inline int pci_get_new_domain_nr(void) { return -ENOSYS; } >> +static inline int pci_get_new_domain_nr(void) >> +{ return -ENOSYS; } >> +static inline void pci_put_domain_nr(struct pci_bus *bus) { } >> #endif /* CONFIG_PCI_DOMAINS */ >> >> /* >> -- >> 1.9.1 >> >> > > >
On Tue, Aug 15, 2017 at 03:01:48PM +0800, Shawn Lin wrote: > Hi Bjorn, > > On 2017/8/12 5:17, Bjorn Helgaas wrote: > >[+cc Lorenzo, resending because I fat-fingered the cc line and subject] > > > >On Tue, Jun 27, 2017 at 08:31:13AM +0800, Shawn Lin wrote: > >>If not getting domain number from DT, the domain number will > >>keep increasing once doing unbind/bind RC drivers. This could > >>introduce pointless tree view of lspci as shows below: > >> > >>-+-[0001:00]---00.0-[01]----00.0 > >> \-[0000:00]- > >> > >>The more test we do, the lengthier it would be. The more serious > >>issue is that if attaching two hierarchies for two different domains > >>belonging to two root bridges, so when doing unbind/bind test for one > >>of them and keep the other, then the domain number would finally > >>overflow and make the two hierarchies of devices share the some domain > >>number but actually they shouldn't. So it looks like we need to invent > >>a new indexing ID mechanism to manage domain number. This patch > >>introduces idr to achieve our purpose. > >> > >>Signed-off-by: Shawn Lin <shawn.lin@rock-chips.com> > > > >The "use_dt_domains" logic in of_pci_bus_find_domain_nr() is fairly > >obtuse. I *think*, now that we have pci_scan_root_bus_bridge() due to > >Lorenzo's excellent work, the time is ripe for moving the domain > >number from arch-specific places into struct pci_host_bridge. > > > >I suspect that will end up simplifying the CONFIG_PCI_DOMAINS vs > >CONFIG_PCI_DOMAINS_GENERIC situation, and I wonder whether it might > >enable some simplification of of_pci_bus_find_domain_nr() as well, > >which in turn, might make *this* patch simpler. > > > >This isn't that big a patch to begin with, so I could apply it as-is > >and we could do more domain cleanup later. It's just that it's > >intertwined with the PCI_DOMAINS #ifdefs and maybe there's an > >opportunity to make this story more readable if those are out of the > >way. Any thoughts? > > That sounds good to me that aftering add IDA domain, we could start > considering moving domain number from arch-specific places into the > bridge code and may be could also finally remove the macro > CONFIG_PCI_DOMAIN* both? I need to see how this can be implemented (another hook in pci_host_bridge ?) but I suspect we can't get away with arch specific bits - or maybe you are referring to having one single place where the domain is _assigned_ using an arch specific hook (in pci_host_bridge) ? I have to have a look into this, certainly this patch should be considered because that atomic counter deserved more thought, yes. Thanks, Lorenzo > >>Changes in v5: > >>- fix Bjorn's comments for v3 as actually I didn't address his comments > >> for v3 when posted my v4. > >> v3: https://patchwork.kernel.org/patch/9742003/ > >> For ACPI case, we never use DT or IDA to get domain numbers and > >> acpi_pci_bus_find_domain_nr would return the proper value we need > >> whether _SEG exist or not. For DT or IDA, we hope bridges use one of > >> them consistently, and this is actually what the pre-existing code > >> did. So I remove ida_domain now since it complicated the logic of the > >> code and we could just make use_dt_domains global instead to slightly > >> achieve our purpose. > >> > >>Changes in v4: > >>- make domain_nr depends on CONFIG_PCI_DOMAINS instead of > >>CONFIG_PCI_DOMAINS_GENERIC. > >> > >>Changes in v3: > >>- make ida_domain a system-wide property and check it in the code to > >>combine with use_dt_domains. Also update the comment there. > >> > >>Changes in v2: > >>- add a remove wrapper > >>- rename use_dt_domains to ida_domain and set this bit > >>in pci_get_new_domain_nr and test it in the remove wrapper. > >> > >> drivers/pci/pci.c | 13 ++++++++++--- > >> drivers/pci/remove.c | 2 ++ > >> include/linux/pci.h | 7 +++++-- > >> 3 files changed, 17 insertions(+), 5 deletions(-) > >> > >>diff --git a/drivers/pci/pci.c b/drivers/pci/pci.c > >>index b58a6b3..9953eaf 100644 > >>--- a/drivers/pci/pci.c > >>+++ b/drivers/pci/pci.c > >>@@ -11,6 +11,7 @@ > >> #include <linux/kernel.h> > >> #include <linux/delay.h> > >> #include <linux/dmi.h> > >>+#include <linux/idr.h> > >> #include <linux/init.h> > >> #include <linux/of.h> > >> #include <linux/of_pci.h> > >>@@ -5305,17 +5306,23 @@ static void pci_no_domains(void) > >> } > >> #ifdef CONFIG_PCI_DOMAINS > >>-static atomic_t __domain_nr = ATOMIC_INIT(-1); > >>+static DEFINE_IDA(__domain_nr); > >> int pci_get_new_domain_nr(void) > >> { > >>- return atomic_inc_return(&__domain_nr); > >>+ return ida_simple_get(&__domain_nr, 0, sizeof(u64), GFP_KERNEL); > >>+} > >>+ > >>+static int use_dt_domains = -1; > >>+void pci_put_domain_nr(struct pci_bus *bus) > >>+{ > >>+ if (acpi_disabled && use_dt_domains != 1) > >>+ ida_simple_remove(&__domain_nr, bus->domain_nr); > >> } > >> #ifdef CONFIG_PCI_DOMAINS_GENERIC > >> static int of_pci_bus_find_domain_nr(struct device *parent) > >> { > >>- static int use_dt_domains = -1; > >> int domain = -1; > >> if (parent) > >>diff --git a/drivers/pci/remove.c b/drivers/pci/remove.c > >>index 73a03d3..01dd1b4 100644 > >>--- a/drivers/pci/remove.c > >>+++ b/drivers/pci/remove.c > >>@@ -157,6 +157,8 @@ void pci_remove_root_bus(struct pci_bus *bus) > >> list_for_each_entry_safe(child, tmp, > >> &bus->devices, bus_list) > >> pci_remove_bus_device(child); > >>+ > >>+ pci_put_domain_nr(bus); > >> pci_remove_bus(bus); > >> host_bridge->bus = NULL; > >>diff --git a/include/linux/pci.h b/include/linux/pci.h > >>index 18cc70a..d6be9596 100644 > >>--- a/include/linux/pci.h > >>+++ b/include/linux/pci.h > >>@@ -523,7 +523,7 @@ struct pci_bus { > >> unsigned char primary; /* number of primary bridge */ > >> unsigned char max_bus_speed; /* enum pci_bus_speed */ > >> unsigned char cur_bus_speed; /* enum pci_bus_speed */ > >>-#ifdef CONFIG_PCI_DOMAINS_GENERIC > >>+#ifdef CONFIG_PCI_DOMAINS > >> int domain_nr; > >> #endif > >>@@ -1466,11 +1466,14 @@ static inline int pci_enable_ptm(struct pci_dev *dev, u8 *granularity) > >> #ifdef CONFIG_PCI_DOMAINS > >> extern int pci_domains_supported; > >> int pci_get_new_domain_nr(void); > >>+void pci_put_domain_nr(struct pci_bus *bus); > >> #else > >> enum { pci_domains_supported = 0 }; > >> static inline int pci_domain_nr(struct pci_bus *bus) { return 0; } > >> static inline int pci_proc_domain(struct pci_bus *bus) { return 0; } > >>-static inline int pci_get_new_domain_nr(void) { return -ENOSYS; } > >>+static inline int pci_get_new_domain_nr(void) > >>+{ return -ENOSYS; } > >>+static inline void pci_put_domain_nr(struct pci_bus *bus) { } > >> #endif /* CONFIG_PCI_DOMAINS */ > >> /* > >>-- > >>1.9.1 > >> > >> > > > > > > >
On Tue, Aug 15, 2017 at 12:43:16PM +0100, Lorenzo Pieralisi wrote: > On Tue, Aug 15, 2017 at 03:01:48PM +0800, Shawn Lin wrote: > > Hi Bjorn, > > > > On 2017/8/12 5:17, Bjorn Helgaas wrote: > > >[+cc Lorenzo, resending because I fat-fingered the cc line and subject] > > > > > >On Tue, Jun 27, 2017 at 08:31:13AM +0800, Shawn Lin wrote: > > >>If not getting domain number from DT, the domain number will > > >>keep increasing once doing unbind/bind RC drivers. This could > > >>introduce pointless tree view of lspci as shows below: > > >> > > >>-+-[0001:00]---00.0-[01]----00.0 > > >> \-[0000:00]- > > >> > > >>The more test we do, the lengthier it would be. The more serious > > >>issue is that if attaching two hierarchies for two different domains > > >>belonging to two root bridges, so when doing unbind/bind test for one > > >>of them and keep the other, then the domain number would finally > > >>overflow and make the two hierarchies of devices share the some domain > > >>number but actually they shouldn't. So it looks like we need to invent > > >>a new indexing ID mechanism to manage domain number. This patch > > >>introduces idr to achieve our purpose. > > >> > > >>Signed-off-by: Shawn Lin <shawn.lin@rock-chips.com> > > > > > >The "use_dt_domains" logic in of_pci_bus_find_domain_nr() is fairly > > >obtuse. I *think*, now that we have pci_scan_root_bus_bridge() due to > > >Lorenzo's excellent work, the time is ripe for moving the domain > > >number from arch-specific places into struct pci_host_bridge. > > > > > >I suspect that will end up simplifying the CONFIG_PCI_DOMAINS vs > > >CONFIG_PCI_DOMAINS_GENERIC situation, and I wonder whether it might > > >enable some simplification of of_pci_bus_find_domain_nr() as well, > > >which in turn, might make *this* patch simpler. > > > > > >This isn't that big a patch to begin with, so I could apply it as-is > > >and we could do more domain cleanup later. It's just that it's > > >intertwined with the PCI_DOMAINS #ifdefs and maybe there's an > > >opportunity to make this story more readable if those are out of the > > >way. Any thoughts? > > > > That sounds good to me that aftering add IDA domain, we could start > > considering moving domain number from arch-specific places into the > > bridge code and may be could also finally remove the macro > > CONFIG_PCI_DOMAIN* both? > > I need to see how this can be implemented (another hook in > pci_host_bridge ?) but I suspect we can't get away with arch > specific bits - or maybe you are referring to having one single > place where the domain is _assigned_ using an arch specific hook > (in pci_host_bridge) ? I have to have a look into this, certainly > this patch should be considered because that atomic counter deserved > more thought, yes. What I was hoping (and I haven't thought this all through) was that we could: - add "domain" to struct pci_host_bridge - have callers of pci_scan_root_bus_bridge() assign bridge->domain alongside their existing bridge->busnr, bridge->ops, etc. assignments. This would pull a little of the messiness of pci_bus_find_domain_nr() into the bridge drivers, but they would know a priori whether to use ACPI or DT, so we wouldn't need quite as much guesswork. - replace the pci_bus_find_domain_nr() call in pci_register_host_bridge() with "bus->domain_nr = bridge->domain" - replace the arch-specific pci_domain_nr() implementations with a generic one - add IDA alloc to the DT domain number alloc path Bjorn
On Tue, Aug 15, 2017 at 07:23:30AM -0500, Bjorn Helgaas wrote: > On Tue, Aug 15, 2017 at 12:43:16PM +0100, Lorenzo Pieralisi wrote: > > On Tue, Aug 15, 2017 at 03:01:48PM +0800, Shawn Lin wrote: > > > Hi Bjorn, > > > > > > On 2017/8/12 5:17, Bjorn Helgaas wrote: > > > >[+cc Lorenzo, resending because I fat-fingered the cc line and subject] > > > > > > > >On Tue, Jun 27, 2017 at 08:31:13AM +0800, Shawn Lin wrote: > > > >>If not getting domain number from DT, the domain number will > > > >>keep increasing once doing unbind/bind RC drivers. This could > > > >>introduce pointless tree view of lspci as shows below: > > > >> > > > >>-+-[0001:00]---00.0-[01]----00.0 > > > >> \-[0000:00]- > > > >> > > > >>The more test we do, the lengthier it would be. The more serious > > > >>issue is that if attaching two hierarchies for two different domains > > > >>belonging to two root bridges, so when doing unbind/bind test for one > > > >>of them and keep the other, then the domain number would finally > > > >>overflow and make the two hierarchies of devices share the some domain > > > >>number but actually they shouldn't. So it looks like we need to invent > > > >>a new indexing ID mechanism to manage domain number. This patch > > > >>introduces idr to achieve our purpose. > > > >> > > > >>Signed-off-by: Shawn Lin <shawn.lin@rock-chips.com> > > > > > > > >The "use_dt_domains" logic in of_pci_bus_find_domain_nr() is fairly > > > >obtuse. I *think*, now that we have pci_scan_root_bus_bridge() due to > > > >Lorenzo's excellent work, the time is ripe for moving the domain > > > >number from arch-specific places into struct pci_host_bridge. > > > > > > > >I suspect that will end up simplifying the CONFIG_PCI_DOMAINS vs > > > >CONFIG_PCI_DOMAINS_GENERIC situation, and I wonder whether it might > > > >enable some simplification of of_pci_bus_find_domain_nr() as well, > > > >which in turn, might make *this* patch simpler. > > > > > > > >This isn't that big a patch to begin with, so I could apply it as-is > > > >and we could do more domain cleanup later. It's just that it's > > > >intertwined with the PCI_DOMAINS #ifdefs and maybe there's an > > > >opportunity to make this story more readable if those are out of the > > > >way. Any thoughts? > > > > > > That sounds good to me that aftering add IDA domain, we could start > > > considering moving domain number from arch-specific places into the > > > bridge code and may be could also finally remove the macro > > > CONFIG_PCI_DOMAIN* both? > > > > I need to see how this can be implemented (another hook in > > pci_host_bridge ?) but I suspect we can't get away with arch > > specific bits - or maybe you are referring to having one single > > place where the domain is _assigned_ using an arch specific hook > > (in pci_host_bridge) ? I have to have a look into this, certainly > > this patch should be considered because that atomic counter deserved > > more thought, yes. > > What I was hoping (and I haven't thought this all through) was that we > could: > > - add "domain" to struct pci_host_bridge > > - have callers of pci_scan_root_bus_bridge() assign bridge->domain > alongside their existing bridge->busnr, bridge->ops, etc. > assignments. This would pull a little of the messiness of > pci_bus_find_domain_nr() into the bridge drivers, but they would > know a priori whether to use ACPI or DT, so we wouldn't need quite > as much guesswork. > > - replace the pci_bus_find_domain_nr() call in > pci_register_host_bridge() with "bus->domain_nr = bridge->domain" > > - replace the arch-specific pci_domain_nr() implementations with a > generic one > > - add IDA alloc to the DT domain number alloc path Yes, if we accept that arch code has to play a role in setting the domain number I think that's doable but I have to have a look into ACPI for this to work since this means that I have to convert x86/ia64 (and powerpc, not sure about this) to the new bus scanning API. For the DT host bridges and arches I have already converted that should be relatively easy (well, another big series), I have to have a proper look into it. Yes, overall it makes perfect sense. Thanks, Lorenzo
diff --git a/drivers/pci/pci.c b/drivers/pci/pci.c index b58a6b3..9953eaf 100644 --- a/drivers/pci/pci.c +++ b/drivers/pci/pci.c @@ -11,6 +11,7 @@ #include <linux/kernel.h> #include <linux/delay.h> #include <linux/dmi.h> +#include <linux/idr.h> #include <linux/init.h> #include <linux/of.h> #include <linux/of_pci.h> @@ -5305,17 +5306,23 @@ static void pci_no_domains(void) } #ifdef CONFIG_PCI_DOMAINS -static atomic_t __domain_nr = ATOMIC_INIT(-1); +static DEFINE_IDA(__domain_nr); int pci_get_new_domain_nr(void) { - return atomic_inc_return(&__domain_nr); + return ida_simple_get(&__domain_nr, 0, sizeof(u64), GFP_KERNEL); +} + +static int use_dt_domains = -1; +void pci_put_domain_nr(struct pci_bus *bus) +{ + if (acpi_disabled && use_dt_domains != 1) + ida_simple_remove(&__domain_nr, bus->domain_nr); } #ifdef CONFIG_PCI_DOMAINS_GENERIC static int of_pci_bus_find_domain_nr(struct device *parent) { - static int use_dt_domains = -1; int domain = -1; if (parent) diff --git a/drivers/pci/remove.c b/drivers/pci/remove.c index 73a03d3..01dd1b4 100644 --- a/drivers/pci/remove.c +++ b/drivers/pci/remove.c @@ -157,6 +157,8 @@ void pci_remove_root_bus(struct pci_bus *bus) list_for_each_entry_safe(child, tmp, &bus->devices, bus_list) pci_remove_bus_device(child); + + pci_put_domain_nr(bus); pci_remove_bus(bus); host_bridge->bus = NULL; diff --git a/include/linux/pci.h b/include/linux/pci.h index 18cc70a..d6be9596 100644 --- a/include/linux/pci.h +++ b/include/linux/pci.h @@ -523,7 +523,7 @@ struct pci_bus { unsigned char primary; /* number of primary bridge */ unsigned char max_bus_speed; /* enum pci_bus_speed */ unsigned char cur_bus_speed; /* enum pci_bus_speed */ -#ifdef CONFIG_PCI_DOMAINS_GENERIC +#ifdef CONFIG_PCI_DOMAINS int domain_nr; #endif @@ -1466,11 +1466,14 @@ static inline int pci_enable_ptm(struct pci_dev *dev, u8 *granularity) #ifdef CONFIG_PCI_DOMAINS extern int pci_domains_supported; int pci_get_new_domain_nr(void); +void pci_put_domain_nr(struct pci_bus *bus); #else enum { pci_domains_supported = 0 }; static inline int pci_domain_nr(struct pci_bus *bus) { return 0; } static inline int pci_proc_domain(struct pci_bus *bus) { return 0; } -static inline int pci_get_new_domain_nr(void) { return -ENOSYS; } +static inline int pci_get_new_domain_nr(void) +{ return -ENOSYS; } +static inline void pci_put_domain_nr(struct pci_bus *bus) { } #endif /* CONFIG_PCI_DOMAINS */ /*
If not getting domain number from DT, the domain number will keep increasing once doing unbind/bind RC drivers. This could introduce pointless tree view of lspci as shows below: -+-[0001:00]---00.0-[01]----00.0 \-[0000:00]- The more test we do, the lengthier it would be. The more serious issue is that if attaching two hierarchies for two different domains belonging to two root bridges, so when doing unbind/bind test for one of them and keep the other, then the domain number would finally overflow and make the two hierarchies of devices share the some domain number but actually they shouldn't. So it looks like we need to invent a new indexing ID mechanism to manage domain number. This patch introduces idr to achieve our purpose. Signed-off-by: Shawn Lin <shawn.lin@rock-chips.com> --- Changes in v5: - fix Bjorn's comments for v3 as actually I didn't address his comments for v3 when posted my v4. v3: https://patchwork.kernel.org/patch/9742003/ For ACPI case, we never use DT or IDA to get domain numbers and acpi_pci_bus_find_domain_nr would return the proper value we need whether _SEG exist or not. For DT or IDA, we hope bridges use one of them consistently, and this is actually what the pre-existing code did. So I remove ida_domain now since it complicated the logic of the code and we could just make use_dt_domains global instead to slightly achieve our purpose. Changes in v4: - make domain_nr depends on CONFIG_PCI_DOMAINS instead of CONFIG_PCI_DOMAINS_GENERIC. Changes in v3: - make ida_domain a system-wide property and check it in the code to combine with use_dt_domains. Also update the comment there. Changes in v2: - add a remove wrapper - rename use_dt_domains to ida_domain and set this bit in pci_get_new_domain_nr and test it in the remove wrapper. drivers/pci/pci.c | 13 ++++++++++--- drivers/pci/remove.c | 2 ++ include/linux/pci.h | 7 +++++-- 3 files changed, 17 insertions(+), 5 deletions(-)