[U-Boot] [PATCH 0/3] sunxi: sun8i-emac: Update DT bindings
The existing sun8i-emac driver in U-Boot uses some preliminary bindings, which matched our own DTs. Now that the Linux kernel got a driver, lets update our probe code to handle those Linux DTs as well. The first patch adds the missing compatible strings for the pinctrl drivers, which is needed for using the sunxi_name_to_gpio() lookup function. Patch 2/3 updates the pinctrl parser used in the sun8i-emac driver, to cope with the new, generic Allwinner pinctrl bindings. The final patch extends the probe routine in the Ethernet driver to deal with both the old and the new bindings. This series allows to copy in the DTs from the latest kernel. Unfortunately right now updating the DTs for the H5 and A64 breaks the build, as the resulting binary (which embeds the DT) gets to large and triggers our new image size check. As the H5 and H3 share most of the DT, we can't just update the H3 DTs either. Hopefully we find some neat trick to work around that.
Cheers, Andre.
Andre Przywara (3): sunxi: gpio: add missing compatible strings net: sun8i-emac: support new pinctrl DT bindings net: sun8i-emac: add support for new EMAC DT binding
drivers/gpio/sunxi_gpio.c | 4 ++ drivers/net/sun8i_emac.c | 107 +++++++++++++++++++++++++++++++++++++--------- 2 files changed, 91 insertions(+), 20 deletions(-)
The sunxi GPIO driver is missing some compatible strings for recent SoCs. While most of the sunxi GPIO code seems to not rely on this (and so works anyway), the sunxi_name_to_gpio() function does and fails at the moment (for instance when resolving the MMC CD pin name). Add the compatible strings for the A64, H5 and V3s, which were missing from the list. This now covers all pinctrl nodes in our own DTs.
Signed-off-by: Andre Przywara andre.przywara@arm.com --- drivers/gpio/sunxi_gpio.c | 4 ++++ 1 file changed, 4 insertions(+)
diff --git a/drivers/gpio/sunxi_gpio.c b/drivers/gpio/sunxi_gpio.c index 3cf01b6e36..90fd7c1596 100644 --- a/drivers/gpio/sunxi_gpio.c +++ b/drivers/gpio/sunxi_gpio.c @@ -354,12 +354,16 @@ static const struct udevice_id sunxi_gpio_ids[] = { ID("allwinner,sun8i-a83t-pinctrl", a_all), ID("allwinner,sun8i-h3-pinctrl", a_all), ID("allwinner,sun8i-r40-pinctrl", a_all), + ID("allwinner,sun8i-v3s-pinctrl", a_all), ID("allwinner,sun9i-a80-pinctrl", a_all), + ID("allwinner,sun50i-a64-pinctrl", a_all), + ID("allwinner,sun50i-h5-pinctrl", a_all), ID("allwinner,sun6i-a31-r-pinctrl", l_2), ID("allwinner,sun8i-a23-r-pinctrl", l_1), ID("allwinner,sun8i-a83t-r-pinctrl", l_1), ID("allwinner,sun8i-h3-r-pinctrl", l_1), ID("allwinner,sun9i-a80-r-pinctrl", l_3), + ID("allwinner,sun50i-a64-r-pinctrl", l_1), { } };
The Linux kernel driver for the Allwinner pin controller gained support for generic properties, which are now also used in the DTs. The sun8i-emac Ethernet driver for new Allwinner MACs reads the pins from the DT, but so far only supported the old binding. Update the parsing routine to cope with both the old and new bindings, so that the newer DTs can be used with U-Boot and its Ethernet driver.
Signed-off-by: Andre Przywara andre.przywara@arm.com --- drivers/net/sun8i_emac.c | 52 ++++++++++++++++++++++++++++++++++++------------ 1 file changed, 39 insertions(+), 13 deletions(-)
diff --git a/drivers/net/sun8i_emac.c b/drivers/net/sun8i_emac.c index 3ccc6b0bb6..df2b857310 100644 --- a/drivers/net/sun8i_emac.c +++ b/drivers/net/sun8i_emac.c @@ -21,6 +21,7 @@ #include <malloc.h> #include <miiphy.h> #include <net.h> +#include <dt-bindings/pinctrl/sun4i-a10.h> #ifdef CONFIG_DM_GPIO #include <asm-generic/gpio.h> #endif @@ -465,30 +466,55 @@ static int parse_phy_pins(struct udevice *dev) }
drive = fdt_getprop_u32_default_node(gd->fdt_blob, offset, 0, - "allwinner,drive", 4); - pull = fdt_getprop_u32_default_node(gd->fdt_blob, offset, 0, - "allwinner,pull", 0); + "drive-strength", ~0); + if (drive != ~0) { + if (drive <= 10) + drive = SUN4I_PINCTRL_10_MA; + else if (drive <= 20) + drive = SUN4I_PINCTRL_20_MA; + else if (drive <= 30) + drive = SUN4I_PINCTRL_30_MA; + else + drive = SUN4I_PINCTRL_40_MA; + } else { + drive = fdt_getprop_u32_default_node(gd->fdt_blob, offset, 0, + "allwinner,drive", 4); + } + + if (fdt_get_property(gd->fdt_blob, offset, "bias-pull-up", NULL)) + pull = SUN4I_PINCTRL_PULL_UP; + else if (fdt_get_property(gd->fdt_blob, offset, "bias-disable", NULL)) + pull = SUN4I_PINCTRL_NO_PULL; + else if (fdt_get_property(gd->fdt_blob, offset, "bias-pull-down", NULL)) + pull = SUN4I_PINCTRL_PULL_DOWN; + else + pull = fdt_getprop_u32_default_node(gd->fdt_blob, offset, 0, + "allwinner,pull", 0); for (i = 0; ; i++) { int pin;
pin_name = fdt_stringlist_get(gd->fdt_blob, offset, "allwinner,pins", i, NULL); - if (!pin_name) - break; - if (pin_name[0] != 'P') - continue; - pin = (pin_name[1] - 'A') << 5; - if (pin >= 26 << 5) + if (!pin_name) { + pin_name = fdt_stringlist_get(gd->fdt_blob, offset, + "pins", i, NULL); + if (!pin_name) + break; + } + + pin = sunxi_name_to_gpio(pin_name); + if (pin < 0) continue; - pin += simple_strtol(&pin_name[2], NULL, 10);
sunxi_gpio_set_cfgpin(pin, SUN8I_GPD8_GMAC); - sunxi_gpio_set_drv(pin, drive); - sunxi_gpio_set_pull(pin, pull); + if (drive != ~0) + sunxi_gpio_set_drv(pin, drive); + if (pull != ~0) + sunxi_gpio_set_pull(pin, pull); }
if (!i) { - printf("WARNING: emac: cannot find allwinner,pins property\n"); + printf("WARNING: emac: cannot find pins property\n"); return -2; }
The Ethernet MAC used in newer Allwinner SoCs (H3, A64, H5) got an upstream Linux driver in v4.15. This one uses a slightly different binding from the original one used by the U-Boot driver. The differences to the old binding are: - The "syscon" address is held in a separate node, referenced via a phandle in the "syscon" property. - The reference to the PHY is held in a property called "phy-handle", not "phy". - The PHY register is at offset 0x30 in the syscon device, not at 0. - The internal PHY is activated when the node, which phy-handle points to, is a child node of an "allwinner,sun8i-h3-mdio-internal" node.
Teach the U-Boot driver how to find its resources in a "new-style" DT, so that we can use a Linux kernel compatible DT for U-Boot as well. This keeps support for the old binding (for now?), to allow a smooth transition.
Signed-off-by: Andre Przywara andre.przywara@arm.com --- drivers/net/sun8i_emac.c | 55 ++++++++++++++++++++++++++++++++++++++++++------ 1 file changed, 48 insertions(+), 7 deletions(-)
diff --git a/drivers/net/sun8i_emac.c b/drivers/net/sun8i_emac.c index df2b857310..4ba8959239 100644 --- a/drivers/net/sun8i_emac.c +++ b/drivers/net/sun8i_emac.c @@ -279,7 +279,7 @@ static int sun8i_emac_set_syscon(struct emac_eth_dev *priv) int ret; u32 reg;
- reg = readl(priv->sysctl_reg); + reg = readl(priv->sysctl_reg + 0x30);
if (priv->variant == H3_EMAC) { ret = sun8i_emac_set_syscon_ephy(priv, ®); @@ -310,7 +310,7 @@ static int sun8i_emac_set_syscon(struct emac_eth_dev *priv) return -EINVAL; }
- writel(reg, priv->sysctl_reg); + writel(reg, priv->sysctl_reg + 0x30);
return 0; } @@ -806,17 +806,50 @@ static int sun8i_emac_eth_ofdata_to_platdata(struct udevice *dev) #endif
pdata->iobase = devfdt_get_addr_name(dev, "emac"); + if (pdata->iobase == FDT_ADDR_T_NONE) + pdata->iobase = devfdt_get_addr(dev); + if (pdata->iobase == FDT_ADDR_T_NONE) { + debug("%s: Cannot find MAC base address\n", __func__); + return -EINVAL; + } + priv->sysctl_reg = devfdt_get_addr_name(dev, "syscon"); + if (priv->sysctl_reg == FDT_ADDR_T_NONE) { + const fdt32_t *reg; + + offset = fdtdec_lookup_phandle(gd->fdt_blob, node, "syscon"); + if (offset < 0) { + debug("%s: cannot find syscon node\n", __func__); + return -EINVAL; + } + reg = fdt_getprop(gd->fdt_blob, offset, "reg", NULL); + if (!reg) { + debug("%s: cannot find reg property in syscon node\n", + __func__); + return -EINVAL; + } + priv->sysctl_reg = fdt_translate_address((void *)gd->fdt_blob, + offset, reg); + } else + priv->sysctl_reg -= 0x30; + if (priv->sysctl_reg == FDT_ADDR_T_NONE) { + debug("%s: Cannot find syscon base address\n", __func__); + return -EINVAL; + }
pdata->phy_interface = -1; priv->phyaddr = -1; priv->use_internal_phy = false;
- offset = fdtdec_lookup_phandle(gd->fdt_blob, node, - "phy"); - if (offset > 0) - priv->phyaddr = fdtdec_get_int(gd->fdt_blob, offset, "reg", - -1); + offset = fdtdec_lookup_phandle(gd->fdt_blob, node, "phy"); + if (offset < 0) + offset = fdtdec_lookup_phandle(gd->fdt_blob, node, + "phy-handle"); + if (offset < 0) { + debug("%s: Cannot find PHY address\n", __func__); + return -EINVAL; + } + priv->phyaddr = fdtdec_get_int(gd->fdt_blob, offset, "reg", -1);
phy_mode = fdt_getprop(gd->fdt_blob, node, "phy-mode", NULL);
@@ -841,6 +874,14 @@ static int sun8i_emac_eth_ofdata_to_platdata(struct udevice *dev) if (fdt_getprop(gd->fdt_blob, node, "allwinner,use-internal-phy", NULL)) priv->use_internal_phy = true; + else { + int parent = fdt_parent_offset(gd->fdt_blob, offset); + + if (parent >= 0 && + !fdt_node_check_compatible(gd->fdt_blob, parent, + "allwinner,sun8i-h3-mdio-internal")) + priv->use_internal_phy = true; + } }
priv->interface = pdata->phy_interface;
Hi,
On Mon, Jan 29, 2018 at 01:15:19AM +0000, Andre Przywara wrote:
The existing sun8i-emac driver in U-Boot uses some preliminary bindings, which matched our own DTs. Now that the Linux kernel got a driver, lets update our probe code to handle those Linux DTs as well. The first patch adds the missing compatible strings for the pinctrl drivers, which is needed for using the sunxi_name_to_gpio() lookup function. Patch 2/3 updates the pinctrl parser used in the sun8i-emac driver, to cope with the new, generic Allwinner pinctrl bindings. The final patch extends the probe routine in the Ethernet driver to deal with both the old and the new bindings.
Thanks for posting this
This series allows to copy in the DTs from the latest kernel. Unfortunately right now updating the DTs for the H5 and A64 breaks the build, as the resulting binary (which embeds the DT) gets to large and triggers our new image size check.
Sigh...
As the H5 and H3 share most of the DT, we can't just update the H3 DTs either. Hopefully we find some neat trick to work around that.
Is it just because of the DT size, or because there's more code?
Maxime
Hi,
On 29/01/18 08:51, Maxime Ripard wrote:
Hi,
On Mon, Jan 29, 2018 at 01:15:19AM +0000, Andre Przywara wrote:
The existing sun8i-emac driver in U-Boot uses some preliminary bindings, which matched our own DTs. Now that the Linux kernel got a driver, lets update our probe code to handle those Linux DTs as well. The first patch adds the missing compatible strings for the pinctrl drivers, which is needed for using the sunxi_name_to_gpio() lookup function. Patch 2/3 updates the pinctrl parser used in the sun8i-emac driver, to cope with the new, generic Allwinner pinctrl bindings. The final patch extends the probe routine in the Ethernet driver to deal with both the old and the new bindings.
Thanks for posting this
This series allows to copy in the DTs from the latest kernel. Unfortunately right now updating the DTs for the H5 and A64 breaks the build, as the resulting binary (which embeds the DT) gets to large and triggers our new image size check.
Sigh...
As the H5 and H3 share most of the DT, we can't just update the H3 DTs either. Hopefully we find some neat trick to work around that.
Is it just because of the DT size, or because there's more code?
My impression the code itself is always growing a tiny bit over the weeks, but this time around it's really the DT update. The current A64 .dtbs in U-Boot are around 8KB, mainline is at 13KB. Similar for the H5: going from 9.5KB to 14.5KB.
Since you did a pretty good job already in identifying the code hogs, I couldn't find *easy* mitigations over the weekend. One possible fix is to remove the second .dtb in the Pine64 case, for which I sent a patch Friday night. Another thing that stuck out is the sha256 checksum. It's "default y" if you have FIT. We need FIT for the SPL loader - but we don't do or need the checksum there. Some people do FIT loading for the kernel and initrd in U-Boot proper, I suppose, but I am not sure how many depend on SHA256 checksums in their images. Disabling this buys us another 8KB, which would more than offset the DT grow.
What do you think?
Cheers, Andre.
On Mon, Jan 29, 2018 at 09:44:44AM +0000, Andre Przywara wrote:
Hi,
On 29/01/18 08:51, Maxime Ripard wrote:
Hi,
On Mon, Jan 29, 2018 at 01:15:19AM +0000, Andre Przywara wrote:
The existing sun8i-emac driver in U-Boot uses some preliminary bindings, which matched our own DTs. Now that the Linux kernel got a driver, lets update our probe code to handle those Linux DTs as well. The first patch adds the missing compatible strings for the pinctrl drivers, which is needed for using the sunxi_name_to_gpio() lookup function. Patch 2/3 updates the pinctrl parser used in the sun8i-emac driver, to cope with the new, generic Allwinner pinctrl bindings. The final patch extends the probe routine in the Ethernet driver to deal with both the old and the new bindings.
Thanks for posting this
This series allows to copy in the DTs from the latest kernel. Unfortunately right now updating the DTs for the H5 and A64 breaks the build, as the resulting binary (which embeds the DT) gets to large and triggers our new image size check.
Sigh...
As the H5 and H3 share most of the DT, we can't just update the H3 DTs either. Hopefully we find some neat trick to work around that.
Is it just because of the DT size, or because there's more code?
My impression the code itself is always growing a tiny bit over the weeks, but this time around it's really the DT update. The current A64 .dtbs in U-Boot are around 8KB, mainline is at 13KB. Similar for the H5: going from 9.5KB to 14.5KB.
Since you did a pretty good job already in identifying the code hogs, I couldn't find *easy* mitigations over the weekend. One possible fix is to remove the second .dtb in the Pine64 case, for which I sent a patch Friday night.
Since the DT is fed to the C preprocessor, we could also put some #ifdef 0 around the nodes that are never used by U-Boot (like the clocks, timer, psci, dma, GIC, RTC, RSB, etc.)
This should give us some room.
Another thing that stuck out is the sha256 checksum. It's "default y" if you have FIT. We need FIT for the SPL loader - but we don't do or need the checksum there. Some people do FIT loading for the kernel and initrd in U-Boot proper, I suppose, but I am not sure how many depend on SHA256 checksums in their images.
I think there was someone (Tom?) that said that it was useful in some circumstances?
Maxime
On Mon, Jan 29, 2018 at 5:58 PM, Maxime Ripard maxime.ripard@free-electrons.com wrote:
On Mon, Jan 29, 2018 at 09:44:44AM +0000, Andre Przywara wrote:
Hi,
On 29/01/18 08:51, Maxime Ripard wrote:
Hi,
On Mon, Jan 29, 2018 at 01:15:19AM +0000, Andre Przywara wrote:
The existing sun8i-emac driver in U-Boot uses some preliminary bindings, which matched our own DTs. Now that the Linux kernel got a driver, lets update our probe code to handle those Linux DTs as well. The first patch adds the missing compatible strings for the pinctrl drivers, which is needed for using the sunxi_name_to_gpio() lookup function. Patch 2/3 updates the pinctrl parser used in the sun8i-emac driver, to cope with the new, generic Allwinner pinctrl bindings. The final patch extends the probe routine in the Ethernet driver to deal with both the old and the new bindings.
Thanks for posting this
This series allows to copy in the DTs from the latest kernel. Unfortunately right now updating the DTs for the H5 and A64 breaks the build, as the resulting binary (which embeds the DT) gets to large and triggers our new image size check.
Sigh...
As the H5 and H3 share most of the DT, we can't just update the H3 DTs either. Hopefully we find some neat trick to work around that.
Is it just because of the DT size, or because there's more code?
My impression the code itself is always growing a tiny bit over the weeks, but this time around it's really the DT update. The current A64 .dtbs in U-Boot are around 8KB, mainline is at 13KB. Similar for the H5: going from 9.5KB to 14.5KB.
Since you did a pretty good job already in identifying the code hogs, I couldn't find *easy* mitigations over the weekend. One possible fix is to remove the second .dtb in the Pine64 case, for which I sent a patch Friday night.
Since the DT is fed to the C preprocessor, we could also put some #ifdef 0 around the nodes that are never used by U-Boot (like the clocks, timer, psci, dma, GIC, RTC, RSB, etc.)
This should give us some room.
Another possibility is trimming the device tree of unneeded blocks, possibly automatically by a tool. I think this was raised before?
ChenYu
Another thing that stuck out is the sha256 checksum. It's "default y" if you have FIT. We need FIT for the SPL loader - but we don't do or need the checksum there. Some people do FIT loading for the kernel and initrd in U-Boot proper, I suppose, but I am not sure how many depend on SHA256 checksums in their images.
I think there was someone (Tom?) that said that it was useful in some circumstances?
Maxime
-- Maxime Ripard, Free Electrons Embedded Linux and Kernel engineering http://free-electrons.com
-- You received this message because you are subscribed to the Google Groups "linux-sunxi" group. To unsubscribe from this group and stop receiving emails from it, send an email to linux-sunxi+unsubscribe@googlegroups.com. For more options, visit https://groups.google.com/d/optout.
On Mon, Jan 29, 2018 at 06:02:45PM +0800, Chen-Yu Tsai wrote:
On Mon, Jan 29, 2018 at 5:58 PM, Maxime Ripard maxime.ripard@free-electrons.com wrote:
On Mon, Jan 29, 2018 at 09:44:44AM +0000, Andre Przywara wrote:
Hi,
On 29/01/18 08:51, Maxime Ripard wrote:
Hi,
On Mon, Jan 29, 2018 at 01:15:19AM +0000, Andre Przywara wrote:
The existing sun8i-emac driver in U-Boot uses some preliminary bindings, which matched our own DTs. Now that the Linux kernel got a driver, lets update our probe code to handle those Linux DTs as well. The first patch adds the missing compatible strings for the pinctrl drivers, which is needed for using the sunxi_name_to_gpio() lookup function. Patch 2/3 updates the pinctrl parser used in the sun8i-emac driver, to cope with the new, generic Allwinner pinctrl bindings. The final patch extends the probe routine in the Ethernet driver to deal with both the old and the new bindings.
Thanks for posting this
This series allows to copy in the DTs from the latest kernel. Unfortunately right now updating the DTs for the H5 and A64 breaks the build, as the resulting binary (which embeds the DT) gets to large and triggers our new image size check.
Sigh...
As the H5 and H3 share most of the DT, we can't just update the H3 DTs either. Hopefully we find some neat trick to work around that.
Is it just because of the DT size, or because there's more code?
My impression the code itself is always growing a tiny bit over the weeks, but this time around it's really the DT update. The current A64 .dtbs in U-Boot are around 8KB, mainline is at 13KB. Similar for the H5: going from 9.5KB to 14.5KB.
Since you did a pretty good job already in identifying the code hogs, I couldn't find *easy* mitigations over the weekend. One possible fix is to remove the second .dtb in the Pine64 case, for which I sent a patch Friday night.
Since the DT is fed to the C preprocessor, we could also put some #ifdef 0 around the nodes that are never used by U-Boot (like the clocks, timer, psci, dma, GIC, RTC, RSB, etc.)
This should give us some room.
Another possibility is trimming the device tree of unneeded blocks, possibly automatically by a tool. I think this was raised before?
Yes, but it turns out that it's actually hard, since you have no idea with the overlays if the block is still going to be unneeded at runtime.
Maxime
Hi,
On 29/01/18 09:58, Maxime Ripard wrote:
On Mon, Jan 29, 2018 at 09:44:44AM +0000, Andre Przywara wrote:
Hi,
On 29/01/18 08:51, Maxime Ripard wrote:
Hi,
On Mon, Jan 29, 2018 at 01:15:19AM +0000, Andre Przywara wrote:
The existing sun8i-emac driver in U-Boot uses some preliminary bindings, which matched our own DTs. Now that the Linux kernel got a driver, lets update our probe code to handle those Linux DTs as well. The first patch adds the missing compatible strings for the pinctrl drivers, which is needed for using the sunxi_name_to_gpio() lookup function. Patch 2/3 updates the pinctrl parser used in the sun8i-emac driver, to cope with the new, generic Allwinner pinctrl bindings. The final patch extends the probe routine in the Ethernet driver to deal with both the old and the new bindings.
Thanks for posting this
This series allows to copy in the DTs from the latest kernel. Unfortunately right now updating the DTs for the H5 and A64 breaks the build, as the resulting binary (which embeds the DT) gets to large and triggers our new image size check.
Sigh...
As the H5 and H3 share most of the DT, we can't just update the H3 DTs either. Hopefully we find some neat trick to work around that.
Is it just because of the DT size, or because there's more code?
My impression the code itself is always growing a tiny bit over the weeks, but this time around it's really the DT update. The current A64 .dtbs in U-Boot are around 8KB, mainline is at 13KB. Similar for the H5: going from 9.5KB to 14.5KB.
Since you did a pretty good job already in identifying the code hogs, I couldn't find *easy* mitigations over the weekend. One possible fix is to remove the second .dtb in the Pine64 case, for which I sent a patch Friday night.
Since the DT is fed to the C preprocessor, we could also put some #ifdef 0 around the nodes that are never used by U-Boot (like the clocks, timer, psci, dma, GIC, RTC, RSB, etc.)
Well yes, U-Boot itself actually only requires a *tiny* .dtb (I think /aliases, /chosen, the reg of USB and Ethernet). But to be honest I don't want to go there. First it would be a constant churn to keep this up-to-date, but more importantly for proper UEFI boot we just reuse U-Boot's .dtb to pass it on to the kernel. That is actually the purpose of this whole exercise. That already works today (at least for A64), but would benefit from some updates.
So I would refrain from tinkering with U-Boot's .dtbs.
This should give us some room.
Another thing that stuck out is the sha256 checksum. It's "default y" if you have FIT. We need FIT for the SPL loader - but we don't do or need the checksum there. Some people do FIT loading for the kernel and initrd in U-Boot proper, I suppose, but I am not sure how many depend on SHA256 checksums in their images.
I think there was someone (Tom?) that said that it was useful in some circumstances?
Yes, I clearly see that it is *useful*, but I wonder how many people would actually miss it today? We would bring it back once we dumped ENV_IS_IN_MMC, so it's only temporarily. I think we can just disable it in some defconfigs, to avoid collateral damage to other boards. If people have a special need, they can always disable the MMC env and enable stuff at their likings, it's just the standard "make .._defconfig; make" process that needs to be fixed with some band-aids for now.
Cheers, Andre.
Hi,
On Mon, Jan 29, 2018 at 10:38:25AM +0000, Andre Przywara wrote:
On 29/01/18 09:58, Maxime Ripard wrote:
On Mon, Jan 29, 2018 at 09:44:44AM +0000, Andre Przywara wrote:
On 29/01/18 08:51, Maxime Ripard wrote:
On Mon, Jan 29, 2018 at 01:15:19AM +0000, Andre Przywara wrote:
The existing sun8i-emac driver in U-Boot uses some preliminary bindings, which matched our own DTs. Now that the Linux kernel got a driver, lets update our probe code to handle those Linux DTs as well. The first patch adds the missing compatible strings for the pinctrl drivers, which is needed for using the sunxi_name_to_gpio() lookup function. Patch 2/3 updates the pinctrl parser used in the sun8i-emac driver, to cope with the new, generic Allwinner pinctrl bindings. The final patch extends the probe routine in the Ethernet driver to deal with both the old and the new bindings.
Thanks for posting this
This series allows to copy in the DTs from the latest kernel. Unfortunately right now updating the DTs for the H5 and A64 breaks the build, as the resulting binary (which embeds the DT) gets to large and triggers our new image size check.
Sigh...
As the H5 and H3 share most of the DT, we can't just update the H3 DTs either. Hopefully we find some neat trick to work around that.
Is it just because of the DT size, or because there's more code?
My impression the code itself is always growing a tiny bit over the weeks, but this time around it's really the DT update. The current A64 .dtbs in U-Boot are around 8KB, mainline is at 13KB. Similar for the H5: going from 9.5KB to 14.5KB.
Since you did a pretty good job already in identifying the code hogs, I couldn't find *easy* mitigations over the weekend. One possible fix is to remove the second .dtb in the Pine64 case, for which I sent a patch Friday night.
Since the DT is fed to the C preprocessor, we could also put some #ifdef 0 around the nodes that are never used by U-Boot (like the clocks, timer, psci, dma, GIC, RTC, RSB, etc.)
Well yes, U-Boot itself actually only requires a *tiny* .dtb (I think /aliases, /chosen, the reg of USB and Ethernet). But to be honest I don't want to go there. First it would be a constant churn to keep this up-to-date,
I'm not too worried about the churn, it would be there only for the time until we fully migrate to the FAT environment, so one-two release now. And we're not syncing the DT very often these days (now that we have support for the EMAC and USB that is all U-Boot cares about).
but more importantly for proper UEFI boot we just reuse U-Boot's .dtb to pass it on to the kernel. That is actually the purpose of this whole exercise. That already works today (at least for A64), but would benefit from some updates.
So I would refrain from tinkering with U-Boot's .dtbs.
That sucks :/
This should give us some room.
Another thing that stuck out is the sha256 checksum. It's "default y" if you have FIT. We need FIT for the SPL loader - but we don't do or need the checksum there. Some people do FIT loading for the kernel and initrd in U-Boot proper, I suppose, but I am not sure how many depend on SHA256 checksums in their images.
I think there was someone (Tom?) that said that it was useful in some circumstances?
Yes, I clearly see that it is *useful*, but I wonder how many people would actually miss it today? We would bring it back once we dumped ENV_IS_IN_MMC, so it's only temporarily.
His words were stronger actually, and he said that we want to keep it.
I think we can just disable it in some defconfigs, to avoid collateral damage to other boards. If people have a special need, they can always disable the MMC env and enable stuff at their likings, it's just the standard "make .._defconfig; make" process that needs to be fixed with some band-aids for now.
I really don't want to go down the "let's fix each defconfig when we find out it broke", this is very likely to be broken with no-one noticing.
Is this issue happening when you sync the whole DT, and would it break if you just convert the EMAC binding?
Otherwise, we might try to revive the DTC garbage collection of unused nodes patches. This would prevent us from using the overlays on such a DT, but that doesn't like like an unfair tradeoff.
Maxime
Hi Maxime,
On Wed, Jan 31, 2018 at 7:21 PM, Maxime Ripard maxime.ripard@free-electrons.com wrote:
Hi,
On Mon, Jan 29, 2018 at 10:38:25AM +0000, Andre Przywara wrote:
On 29/01/18 09:58, Maxime Ripard wrote:
On Mon, Jan 29, 2018 at 09:44:44AM +0000, Andre Przywara wrote:
On 29/01/18 08:51, Maxime Ripard wrote:
On Mon, Jan 29, 2018 at 01:15:19AM +0000, Andre Przywara wrote:
The existing sun8i-emac driver in U-Boot uses some preliminary bindings, which matched our own DTs. Now that the Linux kernel got a driver, lets update our probe code to handle those Linux DTs as well. The first patch adds the missing compatible strings for the pinctrl drivers, which is needed for using the sunxi_name_to_gpio() lookup function. Patch 2/3 updates the pinctrl parser used in the sun8i-emac driver, to cope with the new, generic Allwinner pinctrl bindings. The final patch extends the probe routine in the Ethernet driver to deal with both the old and the new bindings.
Thanks for posting this
This series allows to copy in the DTs from the latest kernel. Unfortunately right now updating the DTs for the H5 and A64 breaks the build, as the resulting binary (which embeds the DT) gets to large and triggers our new image size check.
Sigh...
As the H5 and H3 share most of the DT, we can't just update the H3 DTs either. Hopefully we find some neat trick to work around that.
Is it just because of the DT size, or because there's more code?
My impression the code itself is always growing a tiny bit over the weeks, but this time around it's really the DT update. The current A64 .dtbs in U-Boot are around 8KB, mainline is at 13KB. Similar for the H5: going from 9.5KB to 14.5KB.
Since you did a pretty good job already in identifying the code hogs, I couldn't find *easy* mitigations over the weekend. One possible fix is to remove the second .dtb in the Pine64 case, for which I sent a patch Friday night.
Since the DT is fed to the C preprocessor, we could also put some #ifdef 0 around the nodes that are never used by U-Boot (like the clocks, timer, psci, dma, GIC, RTC, RSB, etc.)
Well yes, U-Boot itself actually only requires a *tiny* .dtb (I think /aliases, /chosen, the reg of USB and Ethernet). But to be honest I don't want to go there. First it would be a constant churn to keep this up-to-date,
I'm not too worried about the churn, it would be there only for the time until we fully migrate to the FAT environment, so one-two release now. And we're not syncing the DT very often these days (now that we have support for the EMAC and USB that is all U-Boot cares about).
but more importantly for proper UEFI boot we just reuse U-Boot's .dtb to pass it on to the kernel. That is actually the purpose of this whole exercise. That already works today (at least for A64), but would benefit from some updates.
So I would refrain from tinkering with U-Boot's .dtbs.
That sucks :/
This should give us some room.
Another thing that stuck out is the sha256 checksum. It's "default y" if you have FIT. We need FIT for the SPL loader - but we don't do or need the checksum there. Some people do FIT loading for the kernel and initrd in U-Boot proper, I suppose, but I am not sure how many depend on SHA256 checksums in their images.
I think there was someone (Tom?) that said that it was useful in some circumstances?
Yes, I clearly see that it is *useful*, but I wonder how many people would actually miss it today? We would bring it back once we dumped ENV_IS_IN_MMC, so it's only temporarily.
His words were stronger actually, and he said that we want to keep it.
I think we can just disable it in some defconfigs, to avoid collateral damage to other boards. If people have a special need, they can always disable the MMC env and enable stuff at their likings, it's just the standard "make .._defconfig; make" process that needs to be fixed with some band-aids for now.
I really don't want to go down the "let's fix each defconfig when we find out it broke", this is very likely to be broken with no-one noticing.
Is this issue happening when you sync the whole DT, and would it break if you just convert the EMAC binding?
Otherwise, we might try to revive the DTC garbage collection of unused nodes patches. This would prevent us from using the overlays on such a DT, but that doesn't like like an unfair tradeoff.
Stupid question:
As I understand it, the boot process is SPL => Full U-Boot => Linux.
Would it therefore be possible to use a cut-down DT for the SPL (just the bits it cares about) then use a full one afterwards?
I'm guessing that the SPL wants to patch the DT we pass to Linux, would we be able to handle that using overlays?
Thanks,
Hi Julian,
On Wed, Jan 31, 2018 at 07:29:13PM +1100, Julian Calaby wrote:
Hi Maxime,
On Wed, Jan 31, 2018 at 7:21 PM, Maxime Ripard maxime.ripard@free-electrons.com wrote:
Hi,
On Mon, Jan 29, 2018 at 10:38:25AM +0000, Andre Przywara wrote:
On 29/01/18 09:58, Maxime Ripard wrote:
On Mon, Jan 29, 2018 at 09:44:44AM +0000, Andre Przywara wrote:
On 29/01/18 08:51, Maxime Ripard wrote:
On Mon, Jan 29, 2018 at 01:15:19AM +0000, Andre Przywara wrote: > The existing sun8i-emac driver in U-Boot uses some preliminary bindings, > which matched our own DTs. Now that the Linux kernel got a driver, lets > update our probe code to handle those Linux DTs as well. > The first patch adds the missing compatible strings for the pinctrl drivers, > which is needed for using the sunxi_name_to_gpio() lookup function. > Patch 2/3 updates the pinctrl parser used in the sun8i-emac driver, to cope > with the new, generic Allwinner pinctrl bindings. > The final patch extends the probe routine in the Ethernet driver to deal > with both the old and the new bindings.
Thanks for posting this
> This series allows to copy in the DTs from the latest kernel. Unfortunately > right now updating the DTs for the H5 and A64 breaks the build, as the > resulting binary (which embeds the DT) gets to large and triggers our new > image size check.
Sigh...
> As the H5 and H3 share most of the DT, we can't just update the H3 > DTs either. Hopefully we find some neat trick to work around that.
Is it just because of the DT size, or because there's more code?
My impression the code itself is always growing a tiny bit over the weeks, but this time around it's really the DT update. The current A64 .dtbs in U-Boot are around 8KB, mainline is at 13KB. Similar for the H5: going from 9.5KB to 14.5KB.
Since you did a pretty good job already in identifying the code hogs, I couldn't find *easy* mitigations over the weekend. One possible fix is to remove the second .dtb in the Pine64 case, for which I sent a patch Friday night.
Since the DT is fed to the C preprocessor, we could also put some #ifdef 0 around the nodes that are never used by U-Boot (like the clocks, timer, psci, dma, GIC, RTC, RSB, etc.)
Well yes, U-Boot itself actually only requires a *tiny* .dtb (I think /aliases, /chosen, the reg of USB and Ethernet). But to be honest I don't want to go there. First it would be a constant churn to keep this up-to-date,
I'm not too worried about the churn, it would be there only for the time until we fully migrate to the FAT environment, so one-two release now. And we're not syncing the DT very often these days (now that we have support for the EMAC and USB that is all U-Boot cares about).
but more importantly for proper UEFI boot we just reuse U-Boot's .dtb to pass it on to the kernel. That is actually the purpose of this whole exercise. That already works today (at least for A64), but would benefit from some updates.
So I would refrain from tinkering with U-Boot's .dtbs.
That sucks :/
This should give us some room.
Another thing that stuck out is the sha256 checksum. It's "default y" if you have FIT. We need FIT for the SPL loader - but we don't do or need the checksum there. Some people do FIT loading for the kernel and initrd in U-Boot proper, I suppose, but I am not sure how many depend on SHA256 checksums in their images.
I think there was someone (Tom?) that said that it was useful in some circumstances?
Yes, I clearly see that it is *useful*, but I wonder how many people would actually miss it today? We would bring it back once we dumped ENV_IS_IN_MMC, so it's only temporarily.
His words were stronger actually, and he said that we want to keep it.
I think we can just disable it in some defconfigs, to avoid collateral damage to other boards. If people have a special need, they can always disable the MMC env and enable stuff at their likings, it's just the standard "make .._defconfig; make" process that needs to be fixed with some band-aids for now.
I really don't want to go down the "let's fix each defconfig when we find out it broke", this is very likely to be broken with no-one noticing.
Is this issue happening when you sync the whole DT, and would it break if you just convert the EMAC binding?
Otherwise, we might try to revive the DTC garbage collection of unused nodes patches. This would prevent us from using the overlays on such a DT, but that doesn't like like an unfair tradeoff.
Stupid question:
It's not really stupid :)
As I understand it, the boot process is SPL => Full U-Boot => Linux.
Would it therefore be possible to use a cut-down DT for the SPL (just the bits it cares about) then use a full one afterwards?
The thing is, we're not using the DT for the SPL, and the DT size we're discussing about is the one in the main U-Boot binary.
I'm guessing that the SPL wants to patch the DT we pass to Linux, would we be able to handle that using overlays?
In a "standard" setup (or at least the one you described), U-Boot will patch, or apply the overlay to, the DT provided by Linux, not its own, so even if we prevent the overlay usage on U-Boot's own, the only downside would be that the UEFI case Andre was describing would not work with overlays anymore.
Actually, we're not building U-boot's DT with overlay support anyway at the moment, so it's not a regression.
Maxime
Hi Maxime,
On Wed, Jan 31, 2018 at 7:36 PM, Maxime Ripard maxime.ripard@free-electrons.com wrote:
Hi Julian,
On Wed, Jan 31, 2018 at 07:29:13PM +1100, Julian Calaby wrote:
Hi Maxime,
On Wed, Jan 31, 2018 at 7:21 PM, Maxime Ripard maxime.ripard@free-electrons.com wrote:
Hi,
On Mon, Jan 29, 2018 at 10:38:25AM +0000, Andre Przywara wrote:
On 29/01/18 09:58, Maxime Ripard wrote:
On Mon, Jan 29, 2018 at 09:44:44AM +0000, Andre Przywara wrote:
On 29/01/18 08:51, Maxime Ripard wrote: > On Mon, Jan 29, 2018 at 01:15:19AM +0000, Andre Przywara wrote: >> The existing sun8i-emac driver in U-Boot uses some preliminary bindings, >> which matched our own DTs. Now that the Linux kernel got a driver, lets >> update our probe code to handle those Linux DTs as well. >> The first patch adds the missing compatible strings for the pinctrl drivers, >> which is needed for using the sunxi_name_to_gpio() lookup function. >> Patch 2/3 updates the pinctrl parser used in the sun8i-emac driver, to cope >> with the new, generic Allwinner pinctrl bindings. >> The final patch extends the probe routine in the Ethernet driver to deal >> with both the old and the new bindings. > > Thanks for posting this > >> This series allows to copy in the DTs from the latest kernel. Unfortunately >> right now updating the DTs for the H5 and A64 breaks the build, as the >> resulting binary (which embeds the DT) gets to large and triggers our new >> image size check. > > Sigh... > >> As the H5 and H3 share most of the DT, we can't just update the H3 >> DTs either. Hopefully we find some neat trick to work around that. > > Is it just because of the DT size, or because there's more code?
My impression the code itself is always growing a tiny bit over the weeks, but this time around it's really the DT update. The current A64 .dtbs in U-Boot are around 8KB, mainline is at 13KB. Similar for the H5: going from 9.5KB to 14.5KB.
Since you did a pretty good job already in identifying the code hogs, I couldn't find *easy* mitigations over the weekend. One possible fix is to remove the second .dtb in the Pine64 case, for which I sent a patch Friday night.
Since the DT is fed to the C preprocessor, we could also put some #ifdef 0 around the nodes that are never used by U-Boot (like the clocks, timer, psci, dma, GIC, RTC, RSB, etc.)
Well yes, U-Boot itself actually only requires a *tiny* .dtb (I think /aliases, /chosen, the reg of USB and Ethernet). But to be honest I don't want to go there. First it would be a constant churn to keep this up-to-date,
I'm not too worried about the churn, it would be there only for the time until we fully migrate to the FAT environment, so one-two release now. And we're not syncing the DT very often these days (now that we have support for the EMAC and USB that is all U-Boot cares about).
but more importantly for proper UEFI boot we just reuse U-Boot's .dtb to pass it on to the kernel. That is actually the purpose of this whole exercise. That already works today (at least for A64), but would benefit from some updates.
So I would refrain from tinkering with U-Boot's .dtbs.
That sucks :/
This should give us some room.
Another thing that stuck out is the sha256 checksum. It's "default y" if you have FIT. We need FIT for the SPL loader - but we don't do or need the checksum there. Some people do FIT loading for the kernel and initrd in U-Boot proper, I suppose, but I am not sure how many depend on SHA256 checksums in their images.
I think there was someone (Tom?) that said that it was useful in some circumstances?
Yes, I clearly see that it is *useful*, but I wonder how many people would actually miss it today? We would bring it back once we dumped ENV_IS_IN_MMC, so it's only temporarily.
His words were stronger actually, and he said that we want to keep it.
I think we can just disable it in some defconfigs, to avoid collateral damage to other boards. If people have a special need, they can always disable the MMC env and enable stuff at their likings, it's just the standard "make .._defconfig; make" process that needs to be fixed with some band-aids for now.
I really don't want to go down the "let's fix each defconfig when we find out it broke", this is very likely to be broken with no-one noticing.
Is this issue happening when you sync the whole DT, and would it break if you just convert the EMAC binding?
Otherwise, we might try to revive the DTC garbage collection of unused nodes patches. This would prevent us from using the overlays on such a DT, but that doesn't like like an unfair tradeoff.
Stupid question:
It's not really stupid :)
Well it is, because I completely missed the UEFI case =)
As I understand it, the boot process is SPL => Full U-Boot => Linux.
Would it therefore be possible to use a cut-down DT for the SPL (just the bits it cares about) then use a full one afterwards?
The thing is, we're not using the DT for the SPL, and the DT size we're discussing about is the one in the main U-Boot binary.
I'm guessing that the SPL wants to patch the DT we pass to Linux, would we be able to handle that using overlays?
In a "standard" setup (or at least the one you described), U-Boot will patch, or apply the overlay to, the DT provided by Linux, not its own, so even if we prevent the overlay usage on U-Boot's own, the only downside would be that the UEFI case Andre was describing would not work with overlays anymore.
So essentially I'm proposing a fix for a problem that doesn't exist.
I'll leave it at this.
Thanks for replying,
Hi,
On 31/01/18 08:21, Maxime Ripard wrote:
Hi,
On Mon, Jan 29, 2018 at 10:38:25AM +0000, Andre Przywara wrote:
On 29/01/18 09:58, Maxime Ripard wrote:
On Mon, Jan 29, 2018 at 09:44:44AM +0000, Andre Przywara wrote:
On 29/01/18 08:51, Maxime Ripard wrote:
On Mon, Jan 29, 2018 at 01:15:19AM +0000, Andre Przywara wrote:
The existing sun8i-emac driver in U-Boot uses some preliminary bindings, which matched our own DTs. Now that the Linux kernel got a driver, lets update our probe code to handle those Linux DTs as well. The first patch adds the missing compatible strings for the pinctrl drivers, which is needed for using the sunxi_name_to_gpio() lookup function. Patch 2/3 updates the pinctrl parser used in the sun8i-emac driver, to cope with the new, generic Allwinner pinctrl bindings. The final patch extends the probe routine in the Ethernet driver to deal with both the old and the new bindings.
Thanks for posting this
This series allows to copy in the DTs from the latest kernel. Unfortunately right now updating the DTs for the H5 and A64 breaks the build, as the resulting binary (which embeds the DT) gets to large and triggers our new image size check.
Sigh...
As the H5 and H3 share most of the DT, we can't just update the H3 DTs either. Hopefully we find some neat trick to work around that.
Is it just because of the DT size, or because there's more code?
My impression the code itself is always growing a tiny bit over the weeks, but this time around it's really the DT update. The current A64 .dtbs in U-Boot are around 8KB, mainline is at 13KB. Similar for the H5: going from 9.5KB to 14.5KB.
Since you did a pretty good job already in identifying the code hogs, I couldn't find *easy* mitigations over the weekend. One possible fix is to remove the second .dtb in the Pine64 case, for which I sent a patch Friday night.
Since the DT is fed to the C preprocessor, we could also put some #ifdef 0 around the nodes that are never used by U-Boot (like the clocks, timer, psci, dma, GIC, RTC, RSB, etc.)
Well yes, U-Boot itself actually only requires a *tiny* .dtb (I think /aliases, /chosen, the reg of USB and Ethernet). But to be honest I don't want to go there. First it would be a constant churn to keep this up-to-date,
I'm not too worried about the churn, it would be there only for the time until we fully migrate to the FAT environment, so one-two release now. And we're not syncing the DT very often these days (now that we have support for the EMAC and USB that is all U-Boot cares about).
but more importantly for proper UEFI boot we just reuse U-Boot's .dtb to pass it on to the kernel. That is actually the purpose of this whole exercise. That already works today (at least for A64), but would benefit from some updates.
So I would refrain from tinkering with U-Boot's .dtbs.
That sucks :/
This should give us some room.
Another thing that stuck out is the sha256 checksum. It's "default y" if you have FIT. We need FIT for the SPL loader - but we don't do or need the checksum there. Some people do FIT loading for the kernel and initrd in U-Boot proper, I suppose, but I am not sure how many depend on SHA256 checksums in their images.
I think there was someone (Tom?) that said that it was useful in some circumstances?
Yes, I clearly see that it is *useful*, but I wonder how many people would actually miss it today? We would bring it back once we dumped ENV_IS_IN_MMC, so it's only temporarily.
His words were stronger actually, and he said that we want to keep it.
I think we can just disable it in some defconfigs, to avoid collateral damage to other boards. If people have a special need, they can always disable the MMC env and enable stuff at their likings, it's just the standard "make .._defconfig; make" process that needs to be fixed with some band-aids for now.
I really don't want to go down the "let's fix each defconfig when we find out it broke", this is very likely to be broken with no-one noticing.
Yeah, that's probably true. Looks we are really running out of silver bullets :-(
Is this issue happening when you sync the whole DT, and would it break if you just convert the EMAC binding?
Otherwise, we might try to revive the DTC garbage collection of unused nodes patches. This would prevent us from using the overlays on such a DT, but that doesn't like like an unfair tradeoff.
Well, what's the point in updating the DT if we then deviate from it again? I would rather suggest to postpone the DT update, until we ditch the MMC environment. The whole reason I wanted the update is to pass it on to Linux, so garbage collection will kill that purpose.
So what about this plan: - We need to do something *now* about the breakage of the 64-bit sunxi *_defconfigs, which do not build with origin/master (for me). I tried GCC 7.1.0, 7.3.0 and 6.4.0, also the actual merge point of the sunxi branch in addition to origin/master. The fix I sent (remove the Pine64 "non-plus" .dtb from the FIT image) was supposed do the trick, but in contrast to my test on Friday doesn't work (anymore?). bananapi_m64_defconfig for instance fails also (with a single .dtb). Can someone confirm this? Maybe it's worth to see what made U-Boot proper grow in the last few days/weeks? - Regardless of this we take the patches from this series, so U-Boot *can* deal with both DTs. - We wait with any DT updates until we remove the MMC environment support, so that we have enough space to give U-Boot the proper .dts.
What is the current plan for the MMC environment? Remove it (or not enable it by default) in 2018.05?
Having the support for both DT bindings in the EMAC driver *now* would allow people to build their own firmware, *configuring* out the MMC environment already and not suffer from any limits anymore. It would just be the _defconfig build which keeps compatibility, for users who care.
Cheers, Andre.
Hi Andre,
On Wed, Jan 31, 2018 at 01:40:23PM +0000, Andre Przywara wrote:
Is this issue happening when you sync the whole DT, and would it break if you just convert the EMAC binding?
Otherwise, we might try to revive the DTC garbage collection of unused nodes patches. This would prevent us from using the overlays on such a DT, but that doesn't like like an unfair tradeoff.
Well, what's the point in updating the DT if we then deviate from it again? I would rather suggest to postpone the DT update, until we ditch the MMC environment. The whole reason I wanted the update is to pass it on to Linux, so garbage collection will kill that purpose.
Not really, the garbage collection I had in mind is: https://patchwork.kernel.org/patch/3840371/
ie, removing only nodes marked as such that are not referenced anywhere in your tree (typically pinctrl nodes). It should totally be usable from Linux, since the usable part of the DT will remain untouched.
As I was saying, the only noticeable "glitch" would be that you don't have nodes in the base DT anymore that used to be there, and overlays won't be able to apply properly if they depended on those nodes.
However, that's not really bad, because: - U-Boot doesn't build its DT with the overlay support anyway - Linux in itself is pretty dumb when it comes to applying overlays at the moment, so there's not a lot of things you can do. The only user in tree (as far as I'm aware) is the FPGA manager, but it doesn't really apply to our SoCs.
So what about this plan:
- We need to do something *now* about the breakage of the 64-bit sunxi
*_defconfigs, which do not build with origin/master (for me). I tried GCC 7.1.0, 7.3.0 and 6.4.0, also the actual merge point of the sunxi branch in addition to origin/master. The fix I sent (remove the Pine64 "non-plus" .dtb from the FIT image) was supposed do the trick, but in contrast to my test on Friday doesn't work (anymore?). bananapi_m64_defconfig for instance fails also (with a single .dtb).
Yes, we should definitely merge that patch.
Can someone confirm this? Maybe it's worth to see what made U-Boot proper grow in the last few days/weeks?
I guess I'm (at least partially) to blame with the introduction of the FAT environment...
How much size do we need?
I've looked at the gains we could have with the dtc patch mentionned above if we flag all the PIO nodes, and its roughly around 500-800 bytes for an A64 DTS. Would it be enough?
- Regardless of this we take the patches from this series, so U-Boot
*can* deal with both DTs.
We should merge those patches, but I don't get why we should maintain the compatibility with the old bindings. Yes, I know what you are going to say, but the deal that was made about those bindings when they were introduced was that we would simply update those bindings when Linux would have decided.
And not maintain any kind of compatibility. Apparently, that deal has changed without consulting or notifying anyone. Great.
And since most of that discussion has been around using the U-Boot DT from Linux, I'm not even sure what the pro would be to keep the old binding around.
- We wait with any DT updates until we remove the MMC environment
support, so that we have enough space to give U-Boot the proper .dts.
What is the current plan for the MMC environment? Remove it (or not enable it by default) in 2018.05?
Not enabling it by default in 2018.05, and removing the migration code later on seems like a good idea yes.
Having the support for both DT bindings in the EMAC driver *now* would allow people to build their own firmware, *configuring* out the MMC environment already and not suffer from any limits anymore.
I understand your frustration, and honestly, it frustrates me even more to see that the hunt of config option and the discussion we had a month ago are already rendered useless, and we need to do it again.
It would just be the _defconfig build which keeps compatibility, for users who care.
We do agree on that :) Maxime
Hi,
On 01/02/18 08:26, Maxime Ripard wrote:
Hi Andre,
On Wed, Jan 31, 2018 at 01:40:23PM +0000, Andre Przywara wrote:
Is this issue happening when you sync the whole DT, and would it break if you just convert the EMAC binding?
Otherwise, we might try to revive the DTC garbage collection of unused nodes patches. This would prevent us from using the overlays on such a DT, but that doesn't like like an unfair tradeoff.
Well, what's the point in updating the DT if we then deviate from it again? I would rather suggest to postpone the DT update, until we ditch the MMC environment. The whole reason I wanted the update is to pass it on to Linux, so garbage collection will kill that purpose.
Not really, the garbage collection I had in mind is: https://patchwork.kernel.org/patch/3840371/
Ah, OK, sorry. Did miss that.
ie, removing only nodes marked as such that are not referenced anywhere in your tree (typically pinctrl nodes). It should totally be usable from Linux, since the usable part of the DT will remain untouched.
As I was saying, the only noticeable "glitch" would be that you don't have nodes in the base DT anymore that used to be there, and overlays won't be able to apply properly if they depended on those nodes.
However, that's not really bad, because:
- U-Boot doesn't build its DT with the overlay support anyway
- Linux in itself is pretty dumb when it comes to applying overlays at the moment, so there's not a lot of things you can do. The only user in tree (as far as I'm aware) is the FPGA manager, but it doesn't really apply to our SoCs.
Yeah, that makes sense and would be fine for me. However I am not sure it will gain us enough to be significant. My impression is that the DT growth is mostly due to the AXP803 nodes. A lot of them are unreferenced as well, would that be covered by that method also?
So what about this plan:
- We need to do something *now* about the breakage of the 64-bit sunxi
*_defconfigs, which do not build with origin/master (for me). I tried GCC 7.1.0, 7.3.0 and 6.4.0, also the actual merge point of the sunxi branch in addition to origin/master. The fix I sent (remove the Pine64 "non-plus" .dtb from the FIT image) was supposed do the trick, but in contrast to my test on Friday doesn't work (anymore?). bananapi_m64_defconfig for instance fails also (with a single .dtb).
Yes, we should definitely merge that patch.
Can someone confirm this? Maybe it's worth to see what made U-Boot proper grow in the last few days/weeks?
I guess I'm (at least partially) to blame with the introduction of the FAT environment...
How much size do we need?
I saw it being off by roughly 3KB, unfortunately. But that is my particular setup here, I think that differs with compiler versions and what not. Plus I just realised that ATF is of course also part of the .itb, and more or less by accident I had a slightly bigger version pointed to by my bl31.bin link (being 5KB bigger than usual). So sorry for the noise, with the standard 32K ATF image sunxi64 boards with one .dtb build fine. Well, 936 bytes below the limit ;-). It's just the Pine64 double .dtb that needs fixing immediately.
I've looked at the gains we could have with the dtc patch mentionned above if we flag all the PIO nodes, and its roughly around 500-800 bytes for an A64 DTS. Would it be enough?
So not for the issue I reported yesterday, but that seems to be moot now. That's why I was originally suggesting SHA256, which gives us 8KB, so some room for the inevitable growth here and there, even after the merge window. We might keep that in mind for emergency cases ;-)
- Regardless of this we take the patches from this series, so U-Boot
*can* deal with both DTs.
We should merge those patches, but I don't get why we should maintain the compatibility with the old bindings. Yes, I know what you are going to say, but the deal that was made about those bindings when they were introduced was that we would simply update those bindings when Linux would have decided.
And not maintain any kind of compatibility. Apparently, that deal has changed without consulting or notifying anyone. Great.
And since most of that discussion has been around using the U-Boot DT from Linux, I'm not even sure what the pro would be to keep the old binding around.
You could have saved your breath ;-) I am totally fine with removing support for the old binding. And (in contrast to Linux!) I would like to consider U-Boot the source of the .dtb, so we can do changes. And indeed the old bindings were never agreed upon.
The reason I kept them in is simply to not break bisectability. The moment you remove the support, you break the board's boot. This is easily fixed with a follow up patch to update the DT, but unless you want to squash code and DT changes into one patch, we should keep them, at least for some commits. This is especially true if we don't know yet when the DT will be updated.
Solutions I see: 1) We keep support for the old bindings in. Once we merged the DT update (which could be months away, possibly, due to the size issue), we add a patch to remove support for the old bindings. 2) We queue those patches, plus update just the nodes affected (EMAC + respective pinctrl nodes) to the Linux ones *now*. This should not increase the size of the DT. Then we can immediately add a patch to remove support for the old binding. So we have this double support in for just two commits or so.
I guess there is not much to say against 2), I am happy to send patches for that.
- We wait with any DT updates until we remove the MMC environment
support, so that we have enough space to give U-Boot the proper .dts.
What is the current plan for the MMC environment? Remove it (or not enable it by default) in 2018.05?
Not enabling it by default in 2018.05, and removing the migration code later on seems like a good idea yes.
Having the support for both DT bindings in the EMAC driver *now* would allow people to build their own firmware, *configuring* out the MMC environment already and not suffer from any limits anymore.
I understand your frustration, and honestly, it frustrates me even more to see that the hunt of config option and the discussion we had a month ago are already rendered useless, and we need to do it again.
I know what you mean!
So quick summary: - I update this series to add one patch to update just the EMAC DT node, plus one patch to remove support for the old binding. - DT sync with Linux is postponed for now until we can afford the size. - We queue the Pine64 non-plus .dtb patch, to fix the build for pine64_plus_defconfig. - We keep an eye on the build size for the next weeks, to spot regressions, especially before the 2018.03 release. - We start chilling the champagne for the day MMC env support gets disabled.
Thanks a lot for wading through this mess ;-)
Cheers, Andre.
It would just be the _defconfig build which keeps compatibility, for users who care.
We do agree on that :) Maxime
Hi,
On Thu, Feb 01, 2018 at 09:53:22AM +0000, Andre Przywara wrote:
ie, removing only nodes marked as such that are not referenced anywhere in your tree (typically pinctrl nodes). It should totally be usable from Linux, since the usable part of the DT will remain untouched.
As I was saying, the only noticeable "glitch" would be that you don't have nodes in the base DT anymore that used to be there, and overlays won't be able to apply properly if they depended on those nodes.
However, that's not really bad, because:
- U-Boot doesn't build its DT with the overlay support anyway
- Linux in itself is pretty dumb when it comes to applying overlays at the moment, so there's not a lot of things you can do. The only user in tree (as far as I'm aware) is the FPGA manager, but it doesn't really apply to our SoCs.
Yeah, that makes sense and would be fine for me. However I am not sure it will gain us enough to be significant. My impression is that the DT growth is mostly due to the AXP803 nodes. A lot of them are unreferenced as well, would that be covered by that method also?
Only the nodes marked as such can be removed if they are unreferenced, and unfortunately, while the pinctrl nodes are harmless to remove, the AXP nodes have a side effect. Indeed, if the regulator node isn't in the DT anymore, the kernel will not disable it if it's unused. So I'm not sure it would be a good idea :/
So what about this plan:
- We need to do something *now* about the breakage of the 64-bit sunxi
*_defconfigs, which do not build with origin/master (for me). I tried GCC 7.1.0, 7.3.0 and 6.4.0, also the actual merge point of the sunxi branch in addition to origin/master. The fix I sent (remove the Pine64 "non-plus" .dtb from the FIT image) was supposed do the trick, but in contrast to my test on Friday doesn't work (anymore?). bananapi_m64_defconfig for instance fails also (with a single .dtb).
Yes, we should definitely merge that patch.
Can someone confirm this? Maybe it's worth to see what made U-Boot proper grow in the last few days/weeks?
I guess I'm (at least partially) to blame with the introduction of the FAT environment...
How much size do we need?
I saw it being off by roughly 3KB, unfortunately. But that is my particular setup here, I think that differs with compiler versions and what not. Plus I just realised that ATF is of course also part of the .itb, and more or less by accident I had a slightly bigger version pointed to by my bl31.bin link (being 5KB bigger than usual). So sorry for the noise, with the standard 32K ATF image sunxi64 boards with one .dtb build fine. Well, 936 bytes below the limit ;-).
Ok, good. I guess one of the related discussion would be to add the bl31 binary to our travis-ci build, so that we actually catch those kind of issues.
It's just the Pine64 double .dtb that needs fixing immediately.
Ack. Consider it merged.
I've looked at the gains we could have with the dtc patch mentionned above if we flag all the PIO nodes, and its roughly around 500-800 bytes for an A64 DTS. Would it be enough?
So not for the issue I reported yesterday, but that seems to be moot now. That's why I was originally suggesting SHA256, which gives us 8KB, so some room for the inevitable growth here and there, even after the merge window. We might keep that in mind for emergency cases ;-)
Yay, two bullets left :)
- Regardless of this we take the patches from this series, so U-Boot
*can* deal with both DTs.
We should merge those patches, but I don't get why we should maintain the compatibility with the old bindings. Yes, I know what you are going to say, but the deal that was made about those bindings when they were introduced was that we would simply update those bindings when Linux would have decided.
And not maintain any kind of compatibility. Apparently, that deal has changed without consulting or notifying anyone. Great.
And since most of that discussion has been around using the U-Boot DT from Linux, I'm not even sure what the pro would be to keep the old binding around.
You could have saved your breath ;-) I am totally fine with removing support for the old binding. And (in contrast to Linux!) I would like to consider U-Boot the source of the .dtb, so we can do changes. And indeed the old bindings were never agreed upon.
The reason I kept them in is simply to not break bisectability. The moment you remove the support, you break the board's boot. This is easily fixed with a follow up patch to update the DT, but unless you want to squash code and DT changes into one patch, we should keep them, at least for some commits. This is especially true if we don't know yet when the DT will be updated.
Solutions I see:
- We keep support for the old bindings in. Once we merged the DT update
(which could be months away, possibly, due to the size issue), we add a patch to remove support for the old bindings. 2) We queue those patches, plus update just the nodes affected (EMAC + respective pinctrl nodes) to the Linux ones *now*. This should not increase the size of the DT. Then we can immediately add a patch to remove support for the old binding. So we have this double support in for just two commits or so.
I guess there is not much to say against 2), I am happy to send patches for that.
2 definitely works for me
- We wait with any DT updates until we remove the MMC environment
support, so that we have enough space to give U-Boot the proper .dts.
What is the current plan for the MMC environment? Remove it (or not enable it by default) in 2018.05?
Not enabling it by default in 2018.05, and removing the migration code later on seems like a good idea yes.
Having the support for both DT bindings in the EMAC driver *now* would allow people to build their own firmware, *configuring* out the MMC environment already and not suffer from any limits anymore.
I understand your frustration, and honestly, it frustrates me even more to see that the hunt of config option and the discussion we had a month ago are already rendered useless, and we need to do it again.
I know what you mean!
So quick summary:
- I update this series to add one patch to update just the EMAC DT node,
plus one patch to remove support for the old binding.
- DT sync with Linux is postponed for now until we can afford the size.
- We queue the Pine64 non-plus .dtb patch, to fix the build for
pine64_plus_defconfig.
- We keep an eye on the build size for the next weeks, to spot
regressions, especially before the 2018.03 release.
Yep
- We start chilling the champagne for the day MMC env support gets disabled.
It's already in a cool place.
Maxime
participants (5)
-
Andre Przywara -
Chen-Yu Tsai -
Julian Calaby -
Maxime Ripard -
Maxime Ripard