Message ID | 1352508532-19241-1-git-send-email-stepanm@codeaurora.org (mailing list archive) |
---|---|
State | New, archived |
Headers | show |
On 11/09/2012 06:48 PM, Stepan Moskovchenko wrote: > Use the cell-index property to construct names for platform > devices, falling back on the existing scheme of using the > device register address if cell-index is not specified. > > The cell-index property is a more useful device identifier, > especially in systems containing several numbered instances > of a particular hardware block, since it more easily > illustrates how devices relate to each other. > > Additionally, userspace software may rely on the classic > <name>.<id> naming scheme to access device attributes in > sysfs, without having to know the physical addresses of > that device on every platform the userspace software may > support. Using cell-index for device naming allows the > device addresses to be hidden from userspace and to be > exposed by logical device number without having to rely on > auxdata to perform name overrides. This allows userspace to > make assumptions about which sysfs nodes map to which > logical instance of a specific hardware block. > > Signed-off-by: Stepan Moskovchenko <stepanm@codeaurora.org> > --- > I had also considered using something like the linux,label property to allow > custom names for platform devices without resorting to auxdata, but the > cell-index approach seems more in line with what cell-index was intended for > and with what the pre-DT platform device naming scheme used to be. Please let > me know if you think there is a better way to accomplish this. > > This is just being sent out as an RFC for now. If there are no objections, I > will send this out as an official patch, along with (or combined with) a patch > to fix up the device names in things like clock tables of any affected > platforms. cell-index is basically deprecated. This has been discussed multiple times in the past. You can use auxdata if you really need to have the old name. Rob > > drivers/of/platform.c | 13 ++++++++++++- > 1 files changed, 12 insertions(+), 1 deletions(-) > > diff --git a/drivers/of/platform.c b/drivers/of/platform.c > index 343ad29..472e374 100644 > --- a/drivers/of/platform.c > +++ b/drivers/of/platform.c > @@ -77,8 +77,9 @@ void of_device_make_bus_id(struct device *dev) > static atomic_t bus_no_reg_magic; > struct device_node *node = dev->of_node; > const u32 *reg; > + u32 cell_index; > u64 addr; > - int magic; > + int magic, ret; > > #ifdef CONFIG_PPC_DCR > /* > @@ -101,6 +102,16 @@ void of_device_make_bus_id(struct device *dev) > #endif /* CONFIG_PPC_DCR */ > > /* > + * For devices with a specified cell-index, use the traditional > + * naming scheme of <name>.<id> > + */ > + ret = of_property_read_u32(node, "cell-index", &cell_index); > + if (ret == 0) { > + dev_set_name(dev, "%s.%d", node->name, cell_index); > + return; > + } > + > + /* > * For MMIO, get the physical address > */ > reg = of_get_property(node, "reg", NULL); > -- > The Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, > hosted by The Linux Foundation >
On Sun, Nov 11, 2012 at 2:32 AM, Rob Herring <robherring2@gmail.com> wrote: > On 11/09/2012 06:48 PM, Stepan Moskovchenko wrote: >> Use the cell-index property to construct names for platform >> devices, falling back on the existing scheme of using the >> device register address if cell-index is not specified. >> >> The cell-index property is a more useful device identifier, >> especially in systems containing several numbered instances >> of a particular hardware block, since it more easily >> illustrates how devices relate to each other. >> >> Additionally, userspace software may rely on the classic >> <name>.<id> naming scheme to access device attributes in >> sysfs, without having to know the physical addresses of >> that device on every platform the userspace software may >> support. Using cell-index for device naming allows the >> device addresses to be hidden from userspace and to be >> exposed by logical device number without having to rely on >> auxdata to perform name overrides. This allows userspace to >> make assumptions about which sysfs nodes map to which >> logical instance of a specific hardware block. >> >> Signed-off-by: Stepan Moskovchenko <stepanm@codeaurora.org> >> --- >> I had also considered using something like the linux,label property to allow >> custom names for platform devices without resorting to auxdata, but the >> cell-index approach seems more in line with what cell-index was intended for >> and with what the pre-DT platform device naming scheme used to be. Please let >> me know if you think there is a better way to accomplish this. >> >> This is just being sent out as an RFC for now. If there are no objections, I >> will send this out as an official patch, along with (or combined with) a patch >> to fix up the device names in things like clock tables of any affected >> platforms. > > cell-index is basically deprecated. This has been discussed multiple > times in the past. You can use auxdata if you really need to have the > old name. Actually, I think it would be fine to use an /aliases entry to set the device name. That's the place to put global namespace information. g.
> On Sun, Nov 11, 2012 at 2:32 AM, Rob Herring <robherring2@gmail.com> > wrote: >> On 11/09/2012 06:48 PM, Stepan Moskovchenko wrote: >>> Use the cell-index property to construct names for platform >>> devices, falling back on the existing scheme of using the >>> device register address if cell-index is not specified. >>> >>> The cell-index property is a more useful device identifier, >>> especially in systems containing several numbered instances >>> of a particular hardware block, since it more easily >>> illustrates how devices relate to each other. >>> >>> Additionally, userspace software may rely on the classic >>> <name>.<id> naming scheme to access device attributes in >>> sysfs, without having to know the physical addresses of >>> that device on every platform the userspace software may >>> support. Using cell-index for device naming allows the >>> device addresses to be hidden from userspace and to be >>> exposed by logical device number without having to rely on >>> auxdata to perform name overrides. This allows userspace to >>> make assumptions about which sysfs nodes map to which >>> logical instance of a specific hardware block. >>> >>> Signed-off-by: Stepan Moskovchenko <stepanm@codeaurora.org> >>> --- >>> I had also considered using something like the linux,label property to >>> allow >>> custom names for platform devices without resorting to auxdata, but the >>> cell-index approach seems more in line with what cell-index was >>> intended for >>> and with what the pre-DT platform device naming scheme used to be. >>> Please let >>> me know if you think there is a better way to accomplish this. >>> >>> This is just being sent out as an RFC for now. If there are no >>> objections, I >>> will send this out as an official patch, along with (or combined with) >>> a patch >>> to fix up the device names in things like clock tables of any affected >>> platforms. >> >> cell-index is basically deprecated. This has been discussed multiple >> times in the past. You can use auxdata if you really need to have the >> old name. > > Actually, I think it would be fine to use an /aliases entry to set the > device name. That's the place to put global namespace information. > > g. > Ah, thank you. I would prefer to stay away from auxdata, since it involves placing more platform-specific data into the kernel, and it is my understanding that auxdata is intended as a temporary measure. The /aliases approach looks interesting, and I'll see what I can do with it - hopefully I can have an RFC / patch soon. It looks like we would want an "inverse" alias lookup- that is, we would need to know which alias corresponds to a given node. Is it possible for a node to have multiple aliases? If so, which shall we use to create the device name? Anyway, I will further look into how these aliases work. Steve
On 11/11/2012 5:45 PM, Stepan Moskovchenko wrote: > >> On Sun, Nov 11, 2012 at 2:32 AM, Rob Herring <robherring2@gmail.com> >> wrote: >>> On 11/09/2012 06:48 PM, Stepan Moskovchenko wrote: >>>> Use the cell-index property to construct names for platform >>>> devices, falling back on the existing scheme of using the >>>> device register address if cell-index is not specified. >>>> >>>> The cell-index property is a more useful device identifier, >>>> especially in systems containing several numbered instances >>>> of a particular hardware block, since it more easily >>>> illustrates how devices relate to each other. >>>> >>>> Additionally, userspace software may rely on the classic >>>> <name>.<id> naming scheme to access device attributes in >>>> sysfs, without having to know the physical addresses of >>>> that device on every platform the userspace software may >>>> support. Using cell-index for device naming allows the >>>> device addresses to be hidden from userspace and to be >>>> exposed by logical device number without having to rely on >>>> auxdata to perform name overrides. This allows userspace to >>>> make assumptions about which sysfs nodes map to which >>>> logical instance of a specific hardware block. >>>> >>>> Signed-off-by: Stepan Moskovchenko <stepanm@codeaurora.org> >>>> --- >>>> I had also considered using something like the linux,label property to >>>> allow >>>> custom names for platform devices without resorting to auxdata, but the >>>> cell-index approach seems more in line with what cell-index was >>>> intended for >>>> and with what the pre-DT platform device naming scheme used to be. >>>> Please let >>>> me know if you think there is a better way to accomplish this. >>>> >>>> This is just being sent out as an RFC for now. If there are no >>>> objections, I >>>> will send this out as an official patch, along with (or combined with) >>>> a patch >>>> to fix up the device names in things like clock tables of any affected >>>> platforms. >>> >>> cell-index is basically deprecated. This has been discussed multiple >>> times in the past. You can use auxdata if you really need to have the >>> old name. >> >> Actually, I think it would be fine to use an /aliases entry to set the >> device name. That's the place to put global namespace information. >> >> g. >> > > Ah, thank you. I would prefer to stay away from auxdata, since it involves > placing more platform-specific data into the kernel, and it is my > understanding that auxdata is intended as a temporary measure. The > /aliases approach looks interesting, and I'll see what I can do with it - > hopefully I can have an RFC / patch soon. It looks like we would want an > "inverse" alias lookup- that is, we would need to know which alias > corresponds to a given node. Is it possible for a node to have multiple > aliases? If so, which shall we use to create the device name? Anyway, I > will further look into how these aliases work. > > Steve Hi Grant, Looking through the alias code, I see that the stem and the alias ID are stored and parsed separately. For the current way of using aliases, this makes sense. However, can you please clarify what you meant by using an /aliases entry to set the device name? The first and most straightforward approach would be to use the entire alias name as the device name, making no distinction between the alias stem and ID. However, since it is possible to have multiple aliases to the same device, which of the aliases shall we use to construct the device name? Additionally, this may cause possible problems for legacy software that expects names in the format of <name>.<ID>, since '.' is not a valid character for alias names as defined by the DT spec, although strictly speaking this approach would successfully solve the problem of giving devices predictable and controllable names. Another way an /aliases entry could be used to set the device name is to have a <name>.<ID> naming scheme, where the name comes from node->name (as is done in of_device_make_bus_id) and the ID gets queried using of_alias_get_id(). We would need to create a new alias stem for this purpose, and suppose that something like "platform" would work. The name-setting code would then roughly look as follows: + alias_id = of_alias_get_id(node, "platform"); + if (alias_id != -ENODEV) { + dev_set_name(dev, "%s.%d", node->name, alias_id); + return; + } The downside to this approach is that it imposes the restriction that device ID numbers now have to be unique throughout the system, whereas before only the <name>.<ID> combinations had to be unique. This is the result of only the ID number being present in the alias table, with each such ID number having the "platform" stem, and the restriction that node properties shall have unique names. A third possible solution is to use an alias stem prefix for defining the device name. That is, the alias to set the device name would have some prefix (such as "platform-" for example) and the aliases would look something like platform-<name><ID>. The code to assign device names would find the matching alias containing the "platform-" prefix, strip the prefix, and use the resulting name and ID to construct the device name. This approach would make it more obvious as to which of several aliases is used to set the device name, but it imposes additional structure on the stem names and causes any aliases starting with "platform-" to become magical, which bothers me slightly. Do any of these describe what you intended when you suggested using the /aliases node to set device names, or is there another approach that I have missed? Can you please elaborate further? Thank you Steve
On Mon, 12 Nov 2012 18:48:43 -0800, Stepan Moskovchenko <stepanm@codeaurora.org> wrote: > On 11/11/2012 5:45 PM, Stepan Moskovchenko wrote: > > > >> On Sun, Nov 11, 2012 at 2:32 AM, Rob Herring <robherring2@gmail.com> > >> wrote: > >>> On 11/09/2012 06:48 PM, Stepan Moskovchenko wrote: > >>>> Use the cell-index property to construct names for platform > >>>> devices, falling back on the existing scheme of using the > >>>> device register address if cell-index is not specified. > >>>> > >>>> The cell-index property is a more useful device identifier, > >>>> especially in systems containing several numbered instances > >>>> of a particular hardware block, since it more easily > >>>> illustrates how devices relate to each other. > >>>> > >>>> Additionally, userspace software may rely on the classic > >>>> <name>.<id> naming scheme to access device attributes in > >>>> sysfs, without having to know the physical addresses of > >>>> that device on every platform the userspace software may > >>>> support. Using cell-index for device naming allows the > >>>> device addresses to be hidden from userspace and to be > >>>> exposed by logical device number without having to rely on > >>>> auxdata to perform name overrides. This allows userspace to > >>>> make assumptions about which sysfs nodes map to which > >>>> logical instance of a specific hardware block. > >>>> > >>>> Signed-off-by: Stepan Moskovchenko <stepanm@codeaurora.org> > >>>> --- > >>>> I had also considered using something like the linux,label property to > >>>> allow > >>>> custom names for platform devices without resorting to auxdata, but the > >>>> cell-index approach seems more in line with what cell-index was > >>>> intended for > >>>> and with what the pre-DT platform device naming scheme used to be. > >>>> Please let > >>>> me know if you think there is a better way to accomplish this. > >>>> > >>>> This is just being sent out as an RFC for now. If there are no > >>>> objections, I > >>>> will send this out as an official patch, along with (or combined with) > >>>> a patch > >>>> to fix up the device names in things like clock tables of any affected > >>>> platforms. > >>> > >>> cell-index is basically deprecated. This has been discussed multiple > >>> times in the past. You can use auxdata if you really need to have the > >>> old name. > >> > >> Actually, I think it would be fine to use an /aliases entry to set the > >> device name. That's the place to put global namespace information. > >> > >> g. > >> > > > > Ah, thank you. I would prefer to stay away from auxdata, since it involves > > placing more platform-specific data into the kernel, and it is my > > understanding that auxdata is intended as a temporary measure. The > > /aliases approach looks interesting, and I'll see what I can do with it - > > hopefully I can have an RFC / patch soon. It looks like we would want an > > "inverse" alias lookup- that is, we would need to know which alias > > corresponds to a given node. Is it possible for a node to have multiple > > aliases? yes > > If so, which shall we use to create the device name? Anyway, I > > will further look into how these aliases work. Well, why exactly do you want to control the names of devices? Is it so that devices match up with what they are, or is it to make things match up with things like clocks and regulators. If it is the latter, then no, don't do this. Use auxdata. When the kernel requires a specific name for a device it is very much a kernel *internal* detail. It does not make sense to encode that into the device tree when it isn't something part of the binding. > > > > Steve > > Hi Grant, > > Looking through the alias code, I see that the stem and the alias ID are > stored and parsed separately. For the current way of using aliases, this > makes sense. However, can you please clarify what you meant by using an > /aliases entry to set the device name? > > The first and most straightforward approach would be to use the entire > alias name as the device name, making no distinction between the alias > stem and ID. However, since it is possible to have multiple aliases to > the same device, which of the aliases shall we use to construct the > device name? Additionally, this may cause possible problems for legacy > software that expects names in the format of <name>.<ID>, since '.' is > not a valid character for alias names as defined by the DT spec, > although strictly speaking this approach would successfully solve the > problem of giving devices predictable and controllable names. > > Another way an /aliases entry could be used to set the device name is to > have a <name>.<ID> naming scheme, where the name comes from node->name > (as is done in of_device_make_bus_id) and the ID gets queried using > of_alias_get_id(). We would need to create a new alias stem for this > purpose, and suppose that something like "platform" would work. The > name-setting code would then roughly look as follows: > > + alias_id = of_alias_get_id(node, "platform"); > + if (alias_id != -ENODEV) { > + dev_set_name(dev, "%s.%d", node->name, alias_id); > + return; > + } > > The downside to this approach is that it imposes the restriction that > device ID numbers now have to be unique throughout the system, whereas > before only the <name>.<ID> combinations had to be unique. This is the > result of only the ID number being present in the alias table, with each > such ID number having the "platform" stem, and the restriction that node > properties shall have unique names. Again, this all looks like trying to manipulate names to keep the kernel happy. Since the kernel has the restrictions on naming, that is where the fixups should be made. Either by devres or by changing the expected name in the clk/regulator tables. > A third possible solution is to use an alias stem prefix for defining > the device name. That is, the alias to set the device name would have > some prefix (such as "platform-" for example) and the aliases would look > something like platform-<name><ID>. The code to assign device names > would find the matching alias containing the "platform-" prefix, strip > the prefix, and use the resulting name and ID to construct the device > name. This approach would make it more obvious as to which of several > aliases is used to set the device name, but it imposes additional > structure on the stem names and causes any aliases starting with > "platform-" to become magical, which bothers me slightly. And specific to the current Linux implementation details. > Do any of these describe what you intended when you suggested using the > /aliases node to set device names, or is there another approach that I > have missed? Can you please elaborate further? Really, all I was thinking about was allowing the device that has an alias "eth0" to be given a name with 'eth0' in it somewhere. Since names like that are a global namespace, /aliases is the place to get them because there is no chance of colision with that approach. g.
On 11/15/2012 8:10 AM, Grant Likely wrote: > On Mon, 12 Nov 2012 18:48:43 -0800, Stepan Moskovchenko <stepanm@codeaurora.org> wrote: >> On 11/11/2012 5:45 PM, Stepan Moskovchenko wrote: >>> >>>> On Sun, Nov 11, 2012 at 2:32 AM, Rob Herring <robherring2@gmail.com> >>>> wrote: >>>>> On 11/09/2012 06:48 PM, Stepan Moskovchenko wrote: >>>>>> Use the cell-index property to construct names for platform >>>>>> devices, falling back on the existing scheme of using the >>>>>> device register address if cell-index is not specified. >>>>>> >>>>>> The cell-index property is a more useful device identifier, >>>>>> especially in systems containing several numbered instances >>>>>> of a particular hardware block, since it more easily >>>>>> illustrates how devices relate to each other. >>>>>> >>>>>> Additionally, userspace software may rely on the classic >>>>>> <name>.<id> naming scheme to access device attributes in >>>>>> sysfs, without having to know the physical addresses of >>>>>> that device on every platform the userspace software may >>>>>> support. Using cell-index for device naming allows the >>>>>> device addresses to be hidden from userspace and to be >>>>>> exposed by logical device number without having to rely on >>>>>> auxdata to perform name overrides. This allows userspace to >>>>>> make assumptions about which sysfs nodes map to which >>>>>> logical instance of a specific hardware block. >>>>>> >>>>>> Signed-off-by: Stepan Moskovchenko <stepanm@codeaurora.org> >>>>>> --- >>>>>> I had also considered using something like the linux,label property to >>>>>> allow >>>>>> custom names for platform devices without resorting to auxdata, but the >>>>>> cell-index approach seems more in line with what cell-index was >>>>>> intended for >>>>>> and with what the pre-DT platform device naming scheme used to be. >>>>>> Please let >>>>>> me know if you think there is a better way to accomplish this. >>>>>> >>>>>> This is just being sent out as an RFC for now. If there are no >>>>>> objections, I >>>>>> will send this out as an official patch, along with (or combined with) >>>>>> a patch >>>>>> to fix up the device names in things like clock tables of any affected >>>>>> platforms. >>>>> >>>>> cell-index is basically deprecated. This has been discussed multiple >>>>> times in the past. You can use auxdata if you really need to have the >>>>> old name. >>>> >>>> Actually, I think it would be fine to use an /aliases entry to set the >>>> device name. That's the place to put global namespace information. >>>> >>>> g. >>>> >>> >>> Ah, thank you. I would prefer to stay away from auxdata, since it involves >>> placing more platform-specific data into the kernel, and it is my >>> understanding that auxdata is intended as a temporary measure. The >>> /aliases approach looks interesting, and I'll see what I can do with it - >>> hopefully I can have an RFC / patch soon. It looks like we would want an >>> "inverse" alias lookup- that is, we would need to know which alias >>> corresponds to a given node. Is it possible for a node to have multiple >>> aliases? > > yes > >>> If so, which shall we use to create the device name? Anyway, I >>> will further look into how these aliases work. > > Well, why exactly do you want to control the names of devices? Is it so > that devices match up with what they are, or is it to make things match > up with things like clocks and regulators. If it is the latter, then no, > don't do this. Use auxdata. When the kernel requires a specific name for > a device it is very much a kernel *internal* detail. It does not make > sense to encode that into the device tree when it isn't something part > of the binding. > > >>> >>> Steve >> >> Hi Grant, >> >> Looking through the alias code, I see that the stem and the alias ID are >> stored and parsed separately. For the current way of using aliases, this >> makes sense. However, can you please clarify what you meant by using an >> /aliases entry to set the device name? >> >> The first and most straightforward approach would be to use the entire >> alias name as the device name, making no distinction between the alias >> stem and ID. However, since it is possible to have multiple aliases to >> the same device, which of the aliases shall we use to construct the >> device name? Additionally, this may cause possible problems for legacy >> software that expects names in the format of <name>.<ID>, since '.' is >> not a valid character for alias names as defined by the DT spec, >> although strictly speaking this approach would successfully solve the >> problem of giving devices predictable and controllable names. >> >> Another way an /aliases entry could be used to set the device name is to >> have a <name>.<ID> naming scheme, where the name comes from node->name >> (as is done in of_device_make_bus_id) and the ID gets queried using >> of_alias_get_id(). We would need to create a new alias stem for this >> purpose, and suppose that something like "platform" would work. The >> name-setting code would then roughly look as follows: >> >> + alias_id = of_alias_get_id(node, "platform"); >> + if (alias_id != -ENODEV) { >> + dev_set_name(dev, "%s.%d", node->name, alias_id); >> + return; >> + } >> >> The downside to this approach is that it imposes the restriction that >> device ID numbers now have to be unique throughout the system, whereas >> before only the <name>.<ID> combinations had to be unique. This is the >> result of only the ID number being present in the alias table, with each >> such ID number having the "platform" stem, and the restriction that node >> properties shall have unique names. > > Again, this all looks like trying to manipulate names to keep the kernel > happy. Since the kernel has the restrictions on naming, that is where > the fixups should be made. Either by devres or by changing the expected > name in the clk/regulator tables. > >> A third possible solution is to use an alias stem prefix for defining >> the device name. That is, the alias to set the device name would have >> some prefix (such as "platform-" for example) and the aliases would look >> something like platform-<name><ID>. The code to assign device names >> would find the matching alias containing the "platform-" prefix, strip >> the prefix, and use the resulting name and ID to construct the device >> name. This approach would make it more obvious as to which of several >> aliases is used to set the device name, but it imposes additional >> structure on the stem names and causes any aliases starting with >> "platform-" to become magical, which bothers me slightly. > > And specific to the current Linux implementation details. > >> Do any of these describe what you intended when you suggested using the >> /aliases node to set device names, or is there another approach that I >> have missed? Can you please elaborate further? > > Really, all I was thinking about was allowing the device that has an > alias "eth0" to be given a name with 'eth0' in it somewhere. Since names > like that are a global namespace, /aliases is the place to get them > because there is no chance of colision with that approach. > > g. > Hi Grant, I realize that auxdata is the correct thing to use for keeping the kernel happy (for things like clocks and regulator consumers) but this is not the problem I am trying to solve. My goal is to try to keep userspace happy by trying to create common and predictable names for functionally equivalent devices across different hardware platforms. For instance, two similar SoCs may have an SDCC controller which may be logically referred to as "the first SDCC device", though the physical address of this device may be different on the two SoCs. And, due to the <reg>.<name> naming convention, the sysfs entries associated with a particular device will be a dependent on the physical address of the device. If userspace wants to touch the sysfs entries of what can logically be described as "the first SDCC device", then userspace needs to know the physical address of this device on each SoC variant it may be running on, since the path to the sysfs entries for this device will be based on the physical address of the device. By using a device naming scheme that replaces the physical address with a logical device number, the userspace-facing interface for each device (such as sysfs entries) could be kept common across SoC variants even when device physical addresses can move around but devices still have the same logical assignments. I realize that this problem can be solved by using auxdata to set the device name, but in this case the only purpose of the auxdata would be to keep userspace happy, since all the other in-kernel relationships (for things like clocks and regulators) can already work without having to rely on auxdata. So, introducing auxdata just for consistency of userspace-facing names seems silly and, I am trying to come up with a more appropriate solution. What do you think? Steve
On 11/15/2012 8:10 AM, Grant Likely wrote: > On Mon, 12 Nov 2012 18:48:43 -0800, Stepan Moskovchenko <stepanm@codeaurora.org> wrote: >> On 11/11/2012 5:45 PM, Stepan Moskovchenko wrote: >>> >>>> On Sun, Nov 11, 2012 at 2:32 AM, Rob Herring <robherring2@gmail.com> >>>> wrote: >>>>> On 11/09/2012 06:48 PM, Stepan Moskovchenko wrote: >>>>>> Use the cell-index property to construct names for platform >>>>>> devices, falling back on the existing scheme of using the >>>>>> device register address if cell-index is not specified. >>>>>> >>>>>> The cell-index property is a more useful device identifier, >>>>>> especially in systems containing several numbered instances >>>>>> of a particular hardware block, since it more easily >>>>>> illustrates how devices relate to each other. >>>>>> >>>>>> Additionally, userspace software may rely on the classic >>>>>> <name>.<id> naming scheme to access device attributes in >>>>>> sysfs, without having to know the physical addresses of >>>>>> that device on every platform the userspace software may >>>>>> support. Using cell-index for device naming allows the >>>>>> device addresses to be hidden from userspace and to be >>>>>> exposed by logical device number without having to rely on >>>>>> auxdata to perform name overrides. This allows userspace to >>>>>> make assumptions about which sysfs nodes map to which >>>>>> logical instance of a specific hardware block. >>>>>> >>>>>> Signed-off-by: Stepan Moskovchenko <stepanm@codeaurora.org> >>>>>> --- >>>>>> I had also considered using something like the linux,label property to >>>>>> allow >>>>>> custom names for platform devices without resorting to auxdata, but the >>>>>> cell-index approach seems more in line with what cell-index was >>>>>> intended for >>>>>> and with what the pre-DT platform device naming scheme used to be. >>>>>> Please let >>>>>> me know if you think there is a better way to accomplish this. >>>>>> >>>>>> This is just being sent out as an RFC for now. If there are no >>>>>> objections, I >>>>>> will send this out as an official patch, along with (or combined with) >>>>>> a patch >>>>>> to fix up the device names in things like clock tables of any affected >>>>>> platforms. >>>>> >>>>> cell-index is basically deprecated. This has been discussed multiple >>>>> times in the past. You can use auxdata if you really need to have the >>>>> old name. >>>> >>>> Actually, I think it would be fine to use an /aliases entry to set the >>>> device name. That's the place to put global namespace information. >>>> >>>> g. >>>> >>> >>> Ah, thank you. I would prefer to stay away from auxdata, since it involves >>> placing more platform-specific data into the kernel, and it is my >>> understanding that auxdata is intended as a temporary measure. The >>> /aliases approach looks interesting, and I'll see what I can do with it - >>> hopefully I can have an RFC / patch soon. It looks like we would want an >>> "inverse" alias lookup- that is, we would need to know which alias >>> corresponds to a given node. Is it possible for a node to have multiple >>> aliases? > > yes > >>> If so, which shall we use to create the device name? Anyway, I >>> will further look into how these aliases work. > > Well, why exactly do you want to control the names of devices? Is it so > that devices match up with what they are, or is it to make things match > up with things like clocks and regulators. If it is the latter, then no, > don't do this. Use auxdata. When the kernel requires a specific name for > a device it is very much a kernel *internal* detail. It does not make > sense to encode that into the device tree when it isn't something part > of the binding. > > >>> >>> Steve >> >> Hi Grant, >> >> Looking through the alias code, I see that the stem and the alias ID are >> stored and parsed separately. For the current way of using aliases, this >> makes sense. However, can you please clarify what you meant by using an >> /aliases entry to set the device name? >> >> The first and most straightforward approach would be to use the entire >> alias name as the device name, making no distinction between the alias >> stem and ID. However, since it is possible to have multiple aliases to >> the same device, which of the aliases shall we use to construct the >> device name? Additionally, this may cause possible problems for legacy >> software that expects names in the format of <name>.<ID>, since '.' is >> not a valid character for alias names as defined by the DT spec, >> although strictly speaking this approach would successfully solve the >> problem of giving devices predictable and controllable names. >> >> Another way an /aliases entry could be used to set the device name is to >> have a <name>.<ID> naming scheme, where the name comes from node->name >> (as is done in of_device_make_bus_id) and the ID gets queried using >> of_alias_get_id(). We would need to create a new alias stem for this >> purpose, and suppose that something like "platform" would work. The >> name-setting code would then roughly look as follows: >> >> + alias_id = of_alias_get_id(node, "platform"); >> + if (alias_id != -ENODEV) { >> + dev_set_name(dev, "%s.%d", node->name, alias_id); >> + return; >> + } >> >> The downside to this approach is that it imposes the restriction that >> device ID numbers now have to be unique throughout the system, whereas >> before only the <name>.<ID> combinations had to be unique. This is the >> result of only the ID number being present in the alias table, with each >> such ID number having the "platform" stem, and the restriction that node >> properties shall have unique names. > > Again, this all looks like trying to manipulate names to keep the kernel > happy. Since the kernel has the restrictions on naming, that is where > the fixups should be made. Either by devres or by changing the expected > name in the clk/regulator tables. > >> A third possible solution is to use an alias stem prefix for defining >> the device name. That is, the alias to set the device name would have >> some prefix (such as "platform-" for example) and the aliases would look >> something like platform-<name><ID>. The code to assign device names >> would find the matching alias containing the "platform-" prefix, strip >> the prefix, and use the resulting name and ID to construct the device >> name. This approach would make it more obvious as to which of several >> aliases is used to set the device name, but it imposes additional >> structure on the stem names and causes any aliases starting with >> "platform-" to become magical, which bothers me slightly. > > And specific to the current Linux implementation details. > >> Do any of these describe what you intended when you suggested using the >> /aliases node to set device names, or is there another approach that I >> have missed? Can you please elaborate further? > > Really, all I was thinking about was allowing the device that has an > alias "eth0" to be given a name with 'eth0' in it somewhere. Since names > like that are a global namespace, /aliases is the place to get them > because there is no chance of colision with that approach. > > g. > [I seem to have forgotten to add the CCed recipients in my last reply. Adding them now.] Hi Grant, I realize that auxdata is the correct thing to use for keeping the kernel happy (for things like clocks and regulator consumers) but this is not the problem I am trying to solve. My goal is to try to keep userspace happy by trying to create common and predictable names for functionally equivalent devices across different hardware platforms. For instance, two similar SoCs may have an SDCC controller which may be logically referred to as "the first SDCC device", though the physical address of this device may be different on the two SoCs. And, due to the <reg>.<name> naming convention, the sysfs entries associated with a particular device will be a dependent on the physical address of the device. If userspace wants to touch the sysfs entries of what can logically be described as "the first SDCC device", then userspace needs to know the physical address of this device on each SoC variant it may be running on, since the path to the sysfs entries for this device will be based on the physical address of the device. By using a device naming scheme that replaces the physical address with a logical device number, the userspace-facing interface for each device (such as sysfs entries) could be kept common across SoC variants even when device physical addresses can move around but devices still have the same logical assignments. I realize that this problem can be solved by using auxdata to set the device name, but in this case the only purpose of the auxdata would be to keep userspace happy, since all the other in-kernel relationships (for things like clocks and regulators) can already work without having to rely on auxdata. So, introducing auxdata just for consistency of userspace-facing names seems silly and, I am trying to come up with a more appropriate solution. What do you think? Steve
On Fri, 16 Nov 2012 19:29:09 -0800, Stepan Moskovchenko <stepanm@codeaurora.org> wrote: > On 11/15/2012 8:10 AM, Grant Likely wrote: > > On Mon, 12 Nov 2012 18:48:43 -0800, Stepan Moskovchenko <stepanm@codeaurora.org> wrote: > > > > Well, why exactly do you want to control the names of devices? Is it so > > that devices match up with what they are, or is it to make things match > > up with things like clocks and regulators. If it is the latter, then no, > > don't do this. Use auxdata. When the kernel requires a specific name for > > a device it is very much a kernel *internal* detail. It does not make > > sense to encode that into the device tree when it isn't something part > > of the binding. > > Hi Grant, > I realize that auxdata is the correct thing to use for keeping the > kernel happy (for things like clocks and regulator consumers) but this > is not the problem I am trying to solve. My goal is to try to keep > userspace happy by trying to create common and predictable names for > functionally equivalent devices across different hardware platforms. > For instance, two similar SoCs may have an SDCC controller which may be > logically referred to as "the first SDCC device", though the physical > address of this device may be different on the two SoCs. And, due to the > <reg>.<name> naming convention, the sysfs entries associated with a > particular device will be a dependent on the physical address of the device. > > If userspace wants to touch the sysfs entries of what can logically be > described as "the first SDCC device", then userspace needs to know the > physical address of this device on each SoC variant it may be running > on, since the path to the sysfs entries for this device will be based on > the physical address of the device. By using a device naming scheme that > replaces the physical address with a logical device number, the > userspace-facing interface for each device (such as sysfs entries) could > be kept common across SoC variants even when device physical addresses > can move around but devices still have the same logical assignments. Okay, so the thinking is that if the generated name of a device can be manipulated, say as 'serial0', then userspace can easily find the device. Correct? If so, then be careful. Userspace is not supposed to ever rely on a particular path to a device. Instead, userspace is supposed to wait for a uevent to announce a device's existence, and then use the data in the uevent attribute. We /could/ use a device tree alias to manipulate the name of the device, but as several people have pointed out there can be more than one alias to a node. Which one do we use? I know I suggested using aliases a couple of weeks ago, but I now think it is a bad idea after mulling it over a bit. What if instead we added something like OF_ALIASES to the uevent attribute? Then userspace would have access to all the aliases for a device. Heck, even a shell script can parse the uevent attribute. There is also precedence for exporting OF data using a uevent. This is from the versatile device tree support: # cat /sys/devices/amba.0/uevent OF_NAME=amba OF_FULLNAME=/amba OF_COMPATIBLE_0=arm,amba-bus OF_COMPATIBLE_N=1 MODALIAS=of:NambaT<NULL>Carm,amba-bus > I realize that this problem can be solved by using auxdata to set the > device name, but in this case the only purpose of the auxdata would be > to keep userspace happy, since all the other in-kernel relationships > (for things like clocks and regulators) can already work without having > to rely on auxdata. So, introducing auxdata just for consistency of > userspace-facing names seems silly and, I am trying to come up with a > more appropriate solution. Yeah, auxdata is completely inappropriate for this. > > What do you think? > > Steve > > -- > The Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, > hosted by The Linux Foundation
On 11/20/2012 8:19 AM, Grant Likely wrote: > What if instead we added something like OF_ALIASES to the uevent > attribute? Then userspace would have access to all the aliases for a > device. Heck, even a shell script can parse the uevent attribute. There > is also precedence for exporting OF data using a uevent. This is from > the versatile device tree support: > > # cat /sys/devices/amba.0/uevent > OF_NAME=amba > OF_FULLNAME=/amba > OF_COMPATIBLE_0=arm,amba-bus > OF_COMPATIBLE_N=1 > MODALIAS=of:NambaT<NULL>Carm,amba-bus This sounds like a good idea. I have just sent out a new patch to do this. I followed the OF_COMPATIBLE uevent convention for listing the aliases, but I can follow another convention if someone objects. Steve
diff --git a/drivers/of/platform.c b/drivers/of/platform.c index 343ad29..472e374 100644 --- a/drivers/of/platform.c +++ b/drivers/of/platform.c @@ -77,8 +77,9 @@ void of_device_make_bus_id(struct device *dev) static atomic_t bus_no_reg_magic; struct device_node *node = dev->of_node; const u32 *reg; + u32 cell_index; u64 addr; - int magic; + int magic, ret; #ifdef CONFIG_PPC_DCR /* @@ -101,6 +102,16 @@ void of_device_make_bus_id(struct device *dev) #endif /* CONFIG_PPC_DCR */ /* + * For devices with a specified cell-index, use the traditional + * naming scheme of <name>.<id> + */ + ret = of_property_read_u32(node, "cell-index", &cell_index); + if (ret == 0) { + dev_set_name(dev, "%s.%d", node->name, cell_index); + return; + } + + /* * For MMIO, get the physical address */ reg = of_get_property(node, "reg", NULL);
Use the cell-index property to construct names for platform devices, falling back on the existing scheme of using the device register address if cell-index is not specified. The cell-index property is a more useful device identifier, especially in systems containing several numbered instances of a particular hardware block, since it more easily illustrates how devices relate to each other. Additionally, userspace software may rely on the classic <name>.<id> naming scheme to access device attributes in sysfs, without having to know the physical addresses of that device on every platform the userspace software may support. Using cell-index for device naming allows the device addresses to be hidden from userspace and to be exposed by logical device number without having to rely on auxdata to perform name overrides. This allows userspace to make assumptions about which sysfs nodes map to which logical instance of a specific hardware block. Signed-off-by: Stepan Moskovchenko <stepanm@codeaurora.org> --- I had also considered using something like the linux,label property to allow custom names for platform devices without resorting to auxdata, but the cell-index approach seems more in line with what cell-index was intended for and with what the pre-DT platform device naming scheme used to be. Please let me know if you think there is a better way to accomplish this. This is just being sent out as an RFC for now. If there are no objections, I will send this out as an official patch, along with (or combined with) a patch to fix up the device names in things like clock tables of any affected platforms. drivers/of/platform.c | 13 ++++++++++++- 1 files changed, 12 insertions(+), 1 deletions(-) -- The Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, hosted by The Linux Foundation