| Message ID | 20230203113650.78146-1-philmd@linaro.org |
|---|---|
| Headers | show |
| Series | hw: Use QOM alias properties and few QOM/QDev cleanups | expand |
On 03/02/2023 11:36, Philippe Mathieu-Daudé wrote: > These patches are extracted from a QOM/QDev refactor series, > so they are preliminary cleanups noticed while working on it: > > - Use correct type when calling qdev_prop_set_xxx() > - Unify some qdev properties in MIPS models > - Replace intermediate properties by link properties > - Remove DEFINE_PROP_DMAADDR() macro which is used one time > - Use qdev_realize_and_unref() instead of open-coding it > > Philippe Mathieu-Daudé (9): > hw/i386/sgx: Do not open-code qdev_realize_and_unref() > hw/ppc/sam460ex: Correctly set MAL properties > hw/arm/nrf51: QOM-alias 'flash-size' property in SoC object > hw/arm/fsl-imx: QOM-alias 'phy-num' property in SoC object > hw/usb/hcd-ohci: Include missing 'sysbus.h' header > hw/display/sm501: QOM-alias 'dma-offset' property in chipset object > hw/qdev: Remove DEFINE_PROP_DMAADDR() and 'hw/qdev-dma.h' > hw/mips: Declare all length properties as unsigned > hw/mips/itu: Pass SAAR using QOM link property > > hw/arm/fsl-imx25.c | 3 +-- > hw/arm/fsl-imx6.c | 3 +-- > hw/arm/fsl-imx6ul.c | 8 ++++---- > hw/arm/fsl-imx7.c | 12 ++++++------ > hw/arm/microbit.c | 5 ++++- > hw/arm/nrf51_soc.c | 10 +--------- > hw/display/sm501.c | 22 +++++++++++----------- > hw/i386/sgx.c | 5 ++--- > hw/intc/mips_gic.c | 4 ++-- > hw/mips/boston.c | 2 +- > hw/mips/cps.c | 35 ++++++++++++----------------------- > hw/mips/malta.c | 2 +- > hw/misc/mips_cmgcr.c | 2 +- > hw/misc/mips_itu.c | 30 ++++++++++++++++++++---------- > hw/nvram/nrf51_nvm.c | 6 +++++- > hw/ppc/sam460ex.c | 4 ++-- > hw/sh4/r2d.c | 2 +- > hw/usb/hcd-ohci-pci.c | 1 - > hw/usb/hcd-ohci.c | 3 +-- > hw/usb/hcd-ohci.h | 1 + > include/hw/arm/fsl-imx25.h | 1 - > include/hw/arm/fsl-imx6.h | 1 - > include/hw/arm/fsl-imx6ul.h | 2 -- > include/hw/arm/fsl-imx7.h | 1 - > include/hw/arm/nrf51_soc.h | 1 - > include/hw/intc/mips_gic.h | 4 ++-- > include/hw/misc/mips_cmgcr.h | 2 +- > include/hw/misc/mips_itu.h | 9 ++++----- > include/hw/qdev-dma.h | 16 ---------------- > 29 files changed, 84 insertions(+), 113 deletions(-) > delete mode 100644 include/hw/qdev-dma.h I must admit to being slightly nervous about using QOM alias properties in this way, simply because you start creating implicit dependencies between QOM objects. How would this work when trying to build machines from configuration files and/or the monitor? Or are the changes restricted to container devices i.e. those which consist of in-built child devices? ATB, Mark.
On 6/2/23 00:29, Mark Cave-Ayland wrote: > On 03/02/2023 11:36, Philippe Mathieu-Daudé wrote: > >> These patches are extracted from a QOM/QDev refactor series, >> so they are preliminary cleanups noticed while working on it: >> >> - Use correct type when calling qdev_prop_set_xxx() >> - Unify some qdev properties in MIPS models >> - Replace intermediate properties by link properties >> - Remove DEFINE_PROP_DMAADDR() macro which is used one time >> - Use qdev_realize_and_unref() instead of open-coding it >> >> Philippe Mathieu-Daudé (9): >> hw/i386/sgx: Do not open-code qdev_realize_and_unref() >> hw/ppc/sam460ex: Correctly set MAL properties >> hw/arm/nrf51: QOM-alias 'flash-size' property in SoC object >> hw/arm/fsl-imx: QOM-alias 'phy-num' property in SoC object >> hw/usb/hcd-ohci: Include missing 'sysbus.h' header >> hw/display/sm501: QOM-alias 'dma-offset' property in chipset object >> hw/qdev: Remove DEFINE_PROP_DMAADDR() and 'hw/qdev-dma.h' >> hw/mips: Declare all length properties as unsigned >> hw/mips/itu: Pass SAAR using QOM link property >> >> hw/arm/fsl-imx25.c | 3 +-- >> hw/arm/fsl-imx6.c | 3 +-- >> hw/arm/fsl-imx6ul.c | 8 ++++---- >> hw/arm/fsl-imx7.c | 12 ++++++------ >> hw/arm/microbit.c | 5 ++++- >> hw/arm/nrf51_soc.c | 10 +--------- >> hw/display/sm501.c | 22 +++++++++++----------- >> hw/i386/sgx.c | 5 ++--- >> hw/intc/mips_gic.c | 4 ++-- >> hw/mips/boston.c | 2 +- >> hw/mips/cps.c | 35 ++++++++++++----------------------- >> hw/mips/malta.c | 2 +- >> hw/misc/mips_cmgcr.c | 2 +- >> hw/misc/mips_itu.c | 30 ++++++++++++++++++++---------- >> hw/nvram/nrf51_nvm.c | 6 +++++- >> hw/ppc/sam460ex.c | 4 ++-- >> hw/sh4/r2d.c | 2 +- >> hw/usb/hcd-ohci-pci.c | 1 - >> hw/usb/hcd-ohci.c | 3 +-- >> hw/usb/hcd-ohci.h | 1 + >> include/hw/arm/fsl-imx25.h | 1 - >> include/hw/arm/fsl-imx6.h | 1 - >> include/hw/arm/fsl-imx6ul.h | 2 -- >> include/hw/arm/fsl-imx7.h | 1 - >> include/hw/arm/nrf51_soc.h | 1 - >> include/hw/intc/mips_gic.h | 4 ++-- >> include/hw/misc/mips_cmgcr.h | 2 +- >> include/hw/misc/mips_itu.h | 9 ++++----- >> include/hw/qdev-dma.h | 16 ---------------- >> 29 files changed, 84 insertions(+), 113 deletions(-) >> delete mode 100644 include/hw/qdev-dma.h > > I must admit to being slightly nervous about using QOM alias properties > in this way, simply because you start creating implicit dependencies > between QOM objects. How would this work when trying to build machines > from configuration files and/or the monitor? Or are the changes > restricted to container devices i.e. those which consist of in-built > child devices? The latter. All parents forward a property to a contained child. The parent forwarding property is replaced by a link into the child, so accessing the parent property transparently access the child one. The dependencies are already explicit. We can not create a parent without its children (the children creation is implicit when we create the parent object). I thought this was the canonical QOM alias properties use. What is the normal use then?
On 06/02/2023 15:27, Philippe Mathieu-Daudé wrote: > On 6/2/23 00:29, Mark Cave-Ayland wrote: >> On 03/02/2023 11:36, Philippe Mathieu-Daudé wrote: >> >>> These patches are extracted from a QOM/QDev refactor series, >>> so they are preliminary cleanups noticed while working on it: >>> >>> - Use correct type when calling qdev_prop_set_xxx() >>> - Unify some qdev properties in MIPS models >>> - Replace intermediate properties by link properties >>> - Remove DEFINE_PROP_DMAADDR() macro which is used one time >>> - Use qdev_realize_and_unref() instead of open-coding it >>> >>> Philippe Mathieu-Daudé (9): >>> hw/i386/sgx: Do not open-code qdev_realize_and_unref() >>> hw/ppc/sam460ex: Correctly set MAL properties >>> hw/arm/nrf51: QOM-alias 'flash-size' property in SoC object >>> hw/arm/fsl-imx: QOM-alias 'phy-num' property in SoC object >>> hw/usb/hcd-ohci: Include missing 'sysbus.h' header >>> hw/display/sm501: QOM-alias 'dma-offset' property in chipset object >>> hw/qdev: Remove DEFINE_PROP_DMAADDR() and 'hw/qdev-dma.h' >>> hw/mips: Declare all length properties as unsigned >>> hw/mips/itu: Pass SAAR using QOM link property >>> >>> hw/arm/fsl-imx25.c | 3 +-- >>> hw/arm/fsl-imx6.c | 3 +-- >>> hw/arm/fsl-imx6ul.c | 8 ++++---- >>> hw/arm/fsl-imx7.c | 12 ++++++------ >>> hw/arm/microbit.c | 5 ++++- >>> hw/arm/nrf51_soc.c | 10 +--------- >>> hw/display/sm501.c | 22 +++++++++++----------- >>> hw/i386/sgx.c | 5 ++--- >>> hw/intc/mips_gic.c | 4 ++-- >>> hw/mips/boston.c | 2 +- >>> hw/mips/cps.c | 35 ++++++++++++----------------------- >>> hw/mips/malta.c | 2 +- >>> hw/misc/mips_cmgcr.c | 2 +- >>> hw/misc/mips_itu.c | 30 ++++++++++++++++++++---------- >>> hw/nvram/nrf51_nvm.c | 6 +++++- >>> hw/ppc/sam460ex.c | 4 ++-- >>> hw/sh4/r2d.c | 2 +- >>> hw/usb/hcd-ohci-pci.c | 1 - >>> hw/usb/hcd-ohci.c | 3 +-- >>> hw/usb/hcd-ohci.h | 1 + >>> include/hw/arm/fsl-imx25.h | 1 - >>> include/hw/arm/fsl-imx6.h | 1 - >>> include/hw/arm/fsl-imx6ul.h | 2 -- >>> include/hw/arm/fsl-imx7.h | 1 - >>> include/hw/arm/nrf51_soc.h | 1 - >>> include/hw/intc/mips_gic.h | 4 ++-- >>> include/hw/misc/mips_cmgcr.h | 2 +- >>> include/hw/misc/mips_itu.h | 9 ++++----- >>> include/hw/qdev-dma.h | 16 ---------------- >>> 29 files changed, 84 insertions(+), 113 deletions(-) >>> delete mode 100644 include/hw/qdev-dma.h >> >> I must admit to being slightly nervous about using QOM alias properties in this >> way, simply because you start creating implicit dependencies between QOM objects. >> How would this work when trying to build machines from configuration files and/or >> the monitor? Or are the changes restricted to container devices i.e. those which >> consist of in-built child devices? > > The latter. All parents forward a property to a contained child. > > The parent forwarding property is replaced by a link into the child, > so accessing the parent property transparently access the child one. > > The dependencies are already explicit. We can not create a parent > without its children (the children creation is implicit when we > create the parent object). > > I thought this was the canonical QOM alias properties use. What is > the normal use then? The problem I've found with this approach in the past is that it fails when you have more than one child device of the same type. For example imagine the scenario where there is a QEMU device that contains 2 child UARTs and each UART has a property to disable hardware handshaking: if you add a property alias to the container device, it can only map to a single child UART. Furthermore if you then try to alias the UART IRQs onto the container device using qdev_pass_gpios(), then that also fails with 2 UARTs because the gpios from each UART have the same property name. You could then think about solving that problem by using object_property_add_alias() directly to specify a different property name for each UART's mapped property on the container device, but then you end up accessing the child UART properties with different names, but only when using that particular parent container device(!). For this reason I've tended to avoid aliases and setup child objects from the container like this: static void container_init(Object *obj) { object_initialize_child(obj, "uart0", &s->uart0, TYPE_UART); object_initialize_child(obj, "uart1", &s->uart1, TYPE_UART); } And then when configuring the board it is possible to obtain the UART references like this: uart0 = UART(object_resolve_path_component(OBJECT(container), "uart0")); irq0 = qdev_connect_gpio_out(DEVICE(uart0), 0, ... ); uart1 = UART(object_resolve_path_component(OBJECT(container), "uart1")); irq1 = qdev_connect_gpio_out(DEVICE(uart1), 0, ... ); This allows all UART configuration to be done in the same way regardless of the parent container device and number of child devices, and without having to think about using different property names depending upon the container device. One place where it could conceivably be useful is where you have a chip modelled as a device and you want to expose the memory regions and IRQs to an interface such as ISA, but often even that doesn't work (think PCI IRQs for example). The only valid use cases I can think of are the /rtc property (which is an alias to the RTC device, regardless of where it exists in the QOM tree) and perhaps in future adding similar array aliases to the root of the machine that can point to things like block devices, network devices, chardevs and audio devices (i.e. anything that has a corresponding QEMU backend). ATB, Mark.
On 6/2/23 22:54, Mark Cave-Ayland wrote: > On 06/02/2023 15:27, Philippe Mathieu-Daudé wrote: > >> On 6/2/23 00:29, Mark Cave-Ayland wrote: >>> On 03/02/2023 11:36, Philippe Mathieu-Daudé wrote: >>> >>>> These patches are extracted from a QOM/QDev refactor series, >>>> so they are preliminary cleanups noticed while working on it: >>>> >>>> - Use correct type when calling qdev_prop_set_xxx() >>>> - Unify some qdev properties in MIPS models >>>> - Replace intermediate properties by link properties >>>> - Remove DEFINE_PROP_DMAADDR() macro which is used one time >>>> - Use qdev_realize_and_unref() instead of open-coding it >>> I must admit to being slightly nervous about using QOM alias >>> properties in this way, simply because you start creating implicit >>> dependencies between QOM objects. How would this work when trying to >>> build machines from configuration files and/or the monitor? Or are >>> the changes restricted to container devices i.e. those which consist >>> of in-built child devices? >> >> The latter. All parents forward a property to a contained child. >> >> The parent forwarding property is replaced by a link into the child, >> so accessing the parent property transparently access the child one. >> >> The dependencies are already explicit. We can not create a parent >> without its children (the children creation is implicit when we >> create the parent object). >> >> I thought this was the canonical QOM alias properties use. What is >> the normal use then? > > The problem I've found with this approach in the past is that it fails > when you have more than one child device of the same type. > > For example imagine the scenario where there is a QEMU device that > contains 2 child UARTs and each UART has a property to disable hardware > handshaking: if you add a property alias to the container device, it can > only map to a single child UART. Furthermore if you then try to alias > the UART IRQs onto the container device using qdev_pass_gpios(), then > that also fails with 2 UARTs because the gpios from each UART have the > same property name. I noticed some qdev gpio namespace issues. Thanks for pointing that qdev_pass_gpios() restriction. > You could then think about solving that problem by using > object_property_add_alias() directly to specify a different property > name for each UART's mapped property on the container device, but then > you end up accessing the child UART properties with different names, but > only when using that particular parent container device(!). > > For this reason I've tended to avoid aliases and setup child objects > from the container like this: > > static void container_init(Object *obj) > { > object_initialize_child(obj, "uart0", &s->uart0, TYPE_UART); > object_initialize_child(obj, "uart1", &s->uart1, TYPE_UART); > } Hmm OK, this is the approach used in IMX: @@ -120,8 +120,12 @@ static void fsl_imx6ul_init(Object *obj) * Ethernet */ for (i = 0; i < FSL_IMX6UL_NUM_ETHS; i++) { + g_autofree gchar *propname = g_strdup_printf("fec%d-phy-num", i + 1); snprintf(name, NAME_SIZE, "eth%d", i); object_initialize_child(obj, name, &s->eth[i], TYPE_IMX_ENET); + qdev_prop_set_uint32(DEVICE(&s->eth[i]), "phy-num", i); + object_property_add_alias(obj, propname, + OBJECT(&s->eth[i]), "phy-num"); } But then this is how it is used today: static Property fsl_imx6ul_properties[] = { - DEFINE_PROP_UINT32("fec1-phy-num", FslIMX6ULState, phy_num[0], 0), - DEFINE_PROP_UINT32("fec2-phy-num", FslIMX6ULState, phy_num[1], 1), DEFINE_PROP_END_OF_LIST(), }; What do you mean by "you end up accessing the child UART properties with different names, but only when using that particular parent container device(!)."? I tend to see QOM modelling as matching hardware design, so if a container is used, there is a similar relationship / hierarchy in the hardware, then accessing the children via a particular parent container path sounds the correct way. QOM indexed child must have the same meaning in the hardware layout. > And then when configuring the board it is possible to obtain the UART > references like this: > > uart0 = UART(object_resolve_path_component(OBJECT(container), > "uart0")); > irq0 = qdev_connect_gpio_out(DEVICE(uart0), 0, ... ); > > uart1 = UART(object_resolve_path_component(OBJECT(container), > "uart1")); > irq1 = qdev_connect_gpio_out(DEVICE(uart1), 0, ... ); > > This allows all UART configuration to be done in the same way regardless > of the parent container device and number of child devices, and without > having to think about using different property names depending upon the > container device. OK I think this is the same explanation as what I just wrote earlier. > One place where it could conceivably be useful is where you have a chip > modelled as a device and you want to expose the memory regions and IRQs > to an interface such as ISA, but often even that doesn't work (think PCI > IRQs for example). IRQ wiring is an unsolved problem in my TODO, in particular when a bus is involved... > The only valid use cases I can think of are the /rtc property (which is > an alias to the RTC device, regardless of where it exists in the QOM > tree) and perhaps in future adding similar array aliases to the root of > the machine that can point to things like block devices, network > devices, chardevs and audio devices (i.e. anything that has a > corresponding QEMU backend). Hmm I see, but this seems a very restrictive use of QOM link property concept IMHO. For me a QOM link allows to share pointer to any QOM object (with the QOM type checked). Am I abusing the concept? BTW DEFINE_PROP_xxx() macros are a QDev concept. In particular DEFINE_PROP_LINK(). The 'realize' step is also a QDev concept... Markus suggested I watch Paolo's QOM talk to use the standard terminology from the expert. I suppose this is "QOM exegesis and apocalypse" from 2014. Thanks for the brainstorming and clarifications! Phil.
On 06/02/2023 23:04, Philippe Mathieu-Daudé wrote: > On 6/2/23 22:54, Mark Cave-Ayland wrote: >> On 06/02/2023 15:27, Philippe Mathieu-Daudé wrote: >> >>> On 6/2/23 00:29, Mark Cave-Ayland wrote: >>>> On 03/02/2023 11:36, Philippe Mathieu-Daudé wrote: >>>> >>>>> These patches are extracted from a QOM/QDev refactor series, >>>>> so they are preliminary cleanups noticed while working on it: >>>>> >>>>> - Use correct type when calling qdev_prop_set_xxx() >>>>> - Unify some qdev properties in MIPS models >>>>> - Replace intermediate properties by link properties >>>>> - Remove DEFINE_PROP_DMAADDR() macro which is used one time >>>>> - Use qdev_realize_and_unref() instead of open-coding it > > >>>> I must admit to being slightly nervous about using QOM alias properties in this >>>> way, simply because you start creating implicit dependencies between QOM objects. >>>> How would this work when trying to build machines from configuration files and/or >>>> the monitor? Or are the changes restricted to container devices i.e. those which >>>> consist of in-built child devices? >>> >>> The latter. All parents forward a property to a contained child. >>> >>> The parent forwarding property is replaced by a link into the child, >>> so accessing the parent property transparently access the child one. >>> >>> The dependencies are already explicit. We can not create a parent >>> without its children (the children creation is implicit when we >>> create the parent object). >>> >>> I thought this was the canonical QOM alias properties use. What is >>> the normal use then? >> >> The problem I've found with this approach in the past is that it fails when you >> have more than one child device of the same type. >> >> For example imagine the scenario where there is a QEMU device that contains 2 child >> UARTs and each UART has a property to disable hardware handshaking: if you add a >> property alias to the container device, it can only map to a single child UART. >> Furthermore if you then try to alias the UART IRQs onto the container device using >> qdev_pass_gpios(), then that also fails with 2 UARTs because the gpios from each >> UART have the same property name. > > I noticed some qdev gpio namespace issues. Thanks for pointing that > qdev_pass_gpios() restriction. > >> You could then think about solving that problem by using >> object_property_add_alias() directly to specify a different property name for each >> UART's mapped property on the container device, but then you end up accessing the >> child UART properties with different names, but only when using that particular >> parent container device(!). >> >> For this reason I've tended to avoid aliases and setup child objects from the >> container like this: >> >> static void container_init(Object *obj) >> { >> object_initialize_child(obj, "uart0", &s->uart0, TYPE_UART); >> object_initialize_child(obj, "uart1", &s->uart1, TYPE_UART); >> } > > Hmm OK, this is the approach used in IMX: > > @@ -120,8 +120,12 @@ static void fsl_imx6ul_init(Object *obj) > * Ethernet > */ > for (i = 0; i < FSL_IMX6UL_NUM_ETHS; i++) { > + g_autofree gchar *propname = g_strdup_printf("fec%d-phy-num", i + 1); > snprintf(name, NAME_SIZE, "eth%d", i); > object_initialize_child(obj, name, &s->eth[i], TYPE_IMX_ENET); > + qdev_prop_set_uint32(DEVICE(&s->eth[i]), "phy-num", i); > + object_property_add_alias(obj, propname, > + OBJECT(&s->eth[i]), "phy-num"); > } > > But then this is how it is used today: > > static Property fsl_imx6ul_properties[] = { > - DEFINE_PROP_UINT32("fec1-phy-num", FslIMX6ULState, phy_num[0], 0), > - DEFINE_PROP_UINT32("fec2-phy-num", FslIMX6ULState, phy_num[1], 1), > DEFINE_PROP_END_OF_LIST(), > }; > > What do you mean by "you end up accessing the child UART properties with > different names, but only when using that particular parent container > device(!)."? I tend to see QOM modelling as matching hardware design, > so if a container is used, there is a similar relationship / hierarchy > in the hardware, then accessing the children via a particular parent > container path sounds the correct way. QOM indexed child must have the > same meaning in the hardware layout. > >> And then when configuring the board it is possible to obtain the UART references >> like this: >> >> uart0 = UART(object_resolve_path_component(OBJECT(container), "uart0")); >> irq0 = qdev_connect_gpio_out(DEVICE(uart0), 0, ... ); >> >> uart1 = UART(object_resolve_path_component(OBJECT(container), "uart1")); >> irq1 = qdev_connect_gpio_out(DEVICE(uart1), 0, ... ); >> >> This allows all UART configuration to be done in the same way regardless of the >> parent container device and number of child devices, and without having to think >> about using different property names depending upon the container device. > > OK I think this is the same explanation as what I just wrote earlier. Yes indeed, I think we're on the same page here now. >> One place where it could conceivably be useful is where you have a chip modelled as >> a device and you want to expose the memory regions and IRQs to an interface such as >> ISA, but often even that doesn't work (think PCI IRQs for example). > > IRQ wiring is an unsolved problem in my TODO, in particular when a bus > is involved... > >> The only valid use cases I can think of are the /rtc property (which is an alias to >> the RTC device, regardless of where it exists in the QOM tree) and perhaps in >> future adding similar array aliases to the root of the machine that can point to >> things like block devices, network devices, chardevs and audio devices (i.e. >> anything that has a corresponding QEMU backend). > > Hmm I see, but this seems a very restrictive use of QOM link property > concept IMHO. For me a QOM link allows to share pointer to any QOM > object (with the QOM type checked). Am I abusing the concept? > > BTW DEFINE_PROP_xxx() macros are a QDev concept. In particular > DEFINE_PROP_LINK(). The 'realize' step is also a QDev concept... I thought we were talking alias properties here? I'm sure I replied separately to the RFC thread containing qdev_set_prop_link() - nothing wrong with link properties at all, but exposing them via qdev was only done recently by Markus (to avoid around passing opaques IIRC) and I'm not sure how that would be represented/parsed on the command line, if indeed that is still considered to be the way forward. > Markus suggested I watch Paolo's QOM talk to use the standard > terminology from the expert. I suppose this is "QOM exegesis and > apocalypse" from 2014. > > Thanks for the brainstorming and clarifications! No problem, hopefully by starting to document some of these issues it will help decided the future direction as to how this should all work. ATB, Mark.