block: preparations for NVMEM provider

On embedded devices using an eMMC it is common that one or more (hw/sw) partitions on the eMMC are used to store MAC addresses and Wi-Fi calibration EEPROM data.

Typically the NVMEM framework is used to have kernel drivers read and use binary data from EEPROMs, efuses, flash memory (MTD), ...

Using references to NVMEM bits in Device Tree allows the kernel to access and apply e.g. the Ethernet MAC address, which can be a requirement for userland to come up (e.g. for nfsroot).

The goal of this series is to prepare the block subsystem to allow for the implementation of an NVMEM provider similar to other types of non-volatile storage, so the same approach already used for EEPROMs, MTD (raw flashes) and UBI-managed NAND can also be used for devices storing those bits on an eMMC.

Define a device tree schema for block devices and partitions on them, which (similar to how it now works also for UBI volumes) can be matched by one or more properties.

Also add a simple notification API for other subsystems to be notified about additions and removals of block devices, which is going to be used by the block-backed NVMEM provider.

Overall, this enables uniform handling across practially all flash storage types used for this purpose (MTD, UBI, and soon also MMC or and in future maybe also other block devices).
---
Changes since v3 sent on Jun 26th, addressing comments from Jens Axboe:
- improve readability and error-handling in fwnode-matching code
- remove forgotten code from earlier development accessing ddev->parent
- use '#if defined' instead of '#ifdef' in header
- provide inline dummies in case of CONFIG_BLOCK_NOTIFIERS not being set

Changes since v2 sent on May 30th 2024 [1] addressing comments from Hauke Mehrtens (https://patchwork.kernel.org/comment/25892133/)
- Check length of UUID and PARTNAME.
- Remove forgotten fallback to get 'partitions' subnode from parent. It is no longer needed and was a left over from earlier development.
- Split series into 3 parts, one for each affected subsystem. This is the first part covering only the changes needed in the block subsystem. The second part adds the actual nvmem provider to drivers/nvmem/, the third part is going to make use of it for MMC block devices and cover changes in drivers/mmc.

Changes since v1 sent on March 21st 2024 [2]:
- introduce notifications for block device addition and removal for in-kernel users. This allows the nvmem driver to be built as a module and avoids using class_interface and block subsystem internals as suggested in https://patchwork.kernel.org/comment/25771998/ and
https://patchwork.kernel.org/comment/25770441/

This series has previously been submitted as RFC on July 19th 2023[3] and most of the basic idea did not change since. Another round of RFC was submitted on March 5th 2024[4].

[1]: https://patchwork.kernel.org/project/linux-block/list/?series=857192
[2]: https://patchwork.kernel.org/project/linux-block/list/?series=837150&archive=both
[3]: https://patchwork.kernel.org/project/linux-block/list/?series=767565
[4]: https://patchwork.kernel.org/project/linux-block/list/?series=832705

Daniel Golle (4):
dt-bindings: block: add basic bindings for block devices
block: partitions: populate fwnode
block: add support for notifications
block: add new genhd flag GENHD_FL_NVMEM

 .../bindings/block/block-device.yaml          | 22 +++++
 .../devicetree/bindings/block/partition.yaml  | 51 +++++++++++
 .../devicetree/bindings/block/partitions.yaml | 20 +++++
 block/Kconfig                                 |  6 ++
 block/Makefile                                |  1 +
 block/blk-notify.c                            | 87 +++++++++++++++++++
 block/partitions/core.c                       | 70 +++++++++++++++
 include/linux/blkdev.h                        | 13 +++
 8 files changed, 270 insertions(+)
 create mode 100644 Documentation/devicetree/bindings/block/block-device.yaml
 create mode 100644 Documentation/devicetree/bindings/block/partition.yaml
 create mode 100644 Documentation/devicetree/bindings/block/partitions.yaml
 create mode 100644 block/blk-notify.c

2.45.2

Add bindings for block devices which are used to allow referencing nvmem bits on them.

---
 .../bindings/block/block-device.yaml          | 22 ++++++++
 .../devicetree/bindings/block/partition.yaml  | 51 +++++++++++++++++++
 .../devicetree/bindings/block/partitions.yaml | 20 ++++++++
 3 files changed, 93 insertions(+)
 create mode 100644 Documentation/devicetree/bindings/block/block-device.yaml
 create mode 100644 Documentation/devicetree/bindings/block/partition.yaml
 create mode 100644 Documentation/devicetree/bindings/block/partitions.yaml

new file mode 100644
index 000000000000..c83ea525650b
@@ -0,0 +1,22 @@
+# SPDX-License-Identifier: GPL-2.0-only OR BSD-2-Clause
+%YAML 1.2
+---
+$id: http://devicetree.org/schemas/block/block-device.yaml#
+$schema: http://devicetree.org/meta-schemas/core.yaml#
+
+title: block storage device
+
+description: |
+  This binding is generic and describes a block-oriented storage device.
+
+maintainers:
+  - Daniel Golle <[email protected]>
+
+properties:
+  partitions:
+    $ref: /schemas/block/partitions.yaml
+
+  nvmem-layout:
+    $ref: /schemas/nvmem/layouts/nvmem-layout.yaml#
+
+unevaluatedProperties: false

new file mode 100644
index 000000000000..fd84c3ba8493
@@ -0,0 +1,20 @@
+# SPDX-License-Identifier: GPL-2.0-only OR BSD-2-Clause
+%YAML 1.2
+---
+$id: http://devicetree.org/schemas/block/partitions.yaml#
+$schema: http://devicetree.org/meta-schemas/core.yaml#
+
+title: Partitions on block devices
+
+description: |
+  This binding is generic and describes the content of the partitions container
+  node.
+
+maintainers:
+  - Daniel Golle <[email protected]>
+
+patternProperties:
+  "^block-partition-.+$":
+    $ref: partition.yaml
+
+unevaluatedProperties: false

2.45.2

Let block partitions to be represented by a firmware node and hence allow them to being referenced e.g. for use with blk-nvmem.

---
 block/partitions/core.c | 70 +++++++++++++++++++++++++++++++++++++++++
 1 file changed, 70 insertions(+)

index ab76e64f0f6c..edf582de741f 100644
@@ -10,6 +10,8 @@
 #include <linux/ctype.h>
 #include <linux/vmalloc.h>
 #include <linux/raid/detect.h>
+#include <linux/property.h>
+
 #include "check.h"
 
 static int (*const check_part[])(struct parsed_partitions *) = {

@@ -281,6 +283,72 @@ static ssize_t whole_disk_show(struct device *dev,
 }
 static const DEVICE_ATTR(whole_disk, 0444, whole_disk_show, NULL);
 
+static bool part_meta_match(const char *attr, const char *member, size_t length)
+{
+	/* check if length of attr exceeds specified maximum length */
+	if (strnlen(attr, length) == length)
+		return false;
+
+	/* return true if strings match */
+	return !strncmp(attr, member, length);
+}
+
+static struct fwnode_handle *find_partition_fwnode(struct block_device *bdev)
+{
+	struct fwnode_handle *fw_parts, *fw_part;
+	struct device *ddev = disk_to_dev(bdev->bd_disk);
+	const char *partname, *uuid;
+	u32 partno;
+	bool got_uuid, got_partname, got_partno;
+
+	fw_parts = device_get_named_child_node(ddev, "partitions");
+	if (!fw_parts)
+		return NULL;
+
+	fwnode_for_each_child_node(fw_parts, fw_part) {
+		got_uuid = false;
+		got_partname = false;
+		got_partno = false;
+		/*
+		 * In case 'uuid' is defined in the partitions firmware node require
+		 * partition meta info being present and the specified uuid to match.
+		 */
+		got_uuid = !fwnode_property_read_string(fw_part, "uuid", &uuid);
+		if (got_uuid && (!bdev->bd_meta_info ||
+				 !part_meta_match(uuid, bdev->bd_meta_info->uuid,
+						  PARTITION_META_INFO_UUIDLTH)))
+			continue;
+
+		/*
+		 * In case 'partname' is defined in the partitions firmware node require
+		 * partition meta info being present and the specified volname to match.
+		 */
+		got_partname = !fwnode_property_read_string(fw_part, "partname",
+							    &partname);
+		if (got_partname && (!bdev->bd_meta_info ||
+				     !part_meta_match(partname,
+						      bdev->bd_meta_info->volname,
+						      PARTITION_META_INFO_VOLNAMELTH)))
+			continue;
+
+		/*
+		 * In case 'partno' is defined in the partitions firmware node the
+		 * specified partno needs to match.
+		 */
+		got_partno = !fwnode_property_read_u32(fw_part, "partno", &partno);
+		if (got_partno && bdev_partno(bdev) != partno)
+			continue;
+
+		/* Skip if no matching criteria is present in firmware node */
+		if (!got_uuid && !got_partname && !got_partno)
+			continue;
+
+		return fw_part;
+	}
+
+	return NULL;
+}
+
 /*
  * Must be called either with open_mutex held, before a disk can be opened or
  * after all disk users are gone.

@@ -355,6 +423,8 @@ static struct block_device *add_partition(struct gendisk *disk, int partno,
 			goto out_put;
 	}
 
+	device_set_node(pdev, find_partition_fwnode(bdev));
+
 	/* delay uevent until 'holders' subdir is created */
 	dev_set_uevent_suppress(pdev, 1);
 	err = device_add(pdev);

2.45.2

Hi Christoph,

thank you for reviewing.

Should I use 80 chars as limit everywhere?

At least up to my understanding there isn't any other to know a partitions UUID or volume name.

If there is another way to access this information I'd happily make use of it, but I couldn't find any.

In my opinion that makes things easier. The coding style allows to exceed it for individual lines where it improves readability, which is a bit of an odd case.

That is true, but except for the early dev_t parsing we actually never use these partition uuids in the kernel at all. Also in all the normal file system references either use the file system uuid, or the device provided one for devices that support them. Most of that is handled entirely in userspace, though - the kernel largely operates just on the dev_t.

Add notifier block to notify other subsystems about the addition or removal of block devices.

---
 block/Kconfig          |  6 +++
 block/Makefile         |  1 +
 block/blk-notify.c     | 87 ++++++++++++++++++++++++++++++++++++++++++
 include/linux/blkdev.h | 11 ++++++
 4 files changed, 105 insertions(+)
 create mode 100644 block/blk-notify.c

index 5b623b876d3b..67cd4f92378a 100644
@@ -209,6 +209,12 @@ config BLK_INLINE_ENCRYPTION_FALLBACK
 	  by falling back to the kernel crypto API when inline
 	  encryption hardware is not present.
 
+config BLOCK_NOTIFIERS
+	bool "Enable support for notifications in block layer"
+	help
+	  Enable this option to provide notifiers for other subsystems
+	  upon addition or removal of block devices.
+
 source "block/partitions/Kconfig"
 
 config BLK_MQ_PCI

index ddfd21c1a9ff..a131fa7d6b26 100644
@@ -38,3 +38,4 @@ obj-$(CONFIG_BLK_INLINE_ENCRYPTION)	+= blk-crypto.o blk-crypto-profile.o \
 					   blk-crypto-sysfs.o
 obj-$(CONFIG_BLK_INLINE_ENCRYPTION_FALLBACK)	+= blk-crypto-fallback.o
 obj-$(CONFIG_BLOCK_HOLDER_DEPRECATED)	+= holder.o
+obj-$(CONFIG_BLOCK_NOTIFIERS) += blk-notify.o

index b2f1362c4681..c6e9fb16f76e 100644
@@ -1687,4 +1687,15 @@ static inline bool bdev_can_atomic_write(struct block_device *bdev)
 
 #define DEFINE_IO_COMP_BATCH(name)	struct io_comp_batch name = { }
 
+
+#define BLK_DEVICE_ADD		1
+#define BLK_DEVICE_REMOVE	2
+#if defined(CONFIG_BLOCK_NOTIFIERS)
+void blk_register_notify(struct notifier_block *nb);
+void blk_unregister_notify(struct notifier_block *nb);
+#else
+static inline void blk_register_notify(struct notifier_block *nb) { };
+static inline void blk_unregister_notify(struct notifier_block *nb) { };
+#endif
+
 #endif /* _LINUX_BLKDEV_H */

2.45.2

Notification for what? I really hate the concept of random modular code being able to hook into device discovery / partition scanning. And not actually having a user for it is a complete no-go.

Adding a dedicated notification interface (instead of using block internals in the nvmem driver) has been requested in a previous review by Bart Van Assche:

https://patchwork.kernel.org/comment/25771998/

Quote from that previous review comment:

---

So that's what I did consequently. Using the notification interface the NVMEM driver can live in drivers/nvmem/ and doesn't need to be using block internals.

The user will be the nvmem provider, you can see the code in earlier versions of the patchset where I had included it:

https://patchwork.kernel.org/project/linux-block/patch/96554d6b4d9fa72f936c2c476eb0b023cdd60a64.1717031992.git.daniel@makrotopia.org/

Being another subsystem I thought it'd be better to deal with the block related things first, and once that has been sorted out I will move on to add the NVMEM driver and make the necessary changes for using it on eMMC.

Thank you for taking a look at this. Let me know if anything else is not clear or needs to be changed.

Cheers

Daniel

It is rather hard to review an interface without the users.

I still dislike the idea of notifications from bdev discovery / partition scanning into the users of them. We have one such users in the MD legacy autodetect code, but that has largely been considered at bad idea and distros tend to use mdadm based assembly from initramfs instead. Which IMHO feels like the right thing for nvmem as well, just have an nvmem provider that opens a file for a user provided path and use kernel_read on it. This can covered block devices, character devices and even regular files. It will require initramfs support, but that is pretty much used everywhere for storage discovery and setup anyway.

Hi Christoph,

I understand that, please take a look at previous iterations of this very series or use the link to the patch which I have sent before (see above).

The problem there is that then we cannot use Device Tree to device the NVMEM layouts, and reference NVMEM bits to the dirvers which need them. Hence also the definition of the NVMEM layout would have to happen in userspace, inside an initramfs. I know that having an initramfs is common for classic desktop or server distributions, but the same is not true on more simple embedded devices such as router/firewall or WiFi access point appliances running OpenWrt.

Carrying all that board-specific knowledge in the form of scripts identifying the board is not exactly nice, nor is creating an individual initramfs for each of the 1000+ boards supported by OpenWrt, for example. Extracting the layout information from /sys/firmware/devicetree in userspace just to then somehow use libblkid to identify the block device and throw that information back at the kernel which requested it e.g. using a firmware hotplug call is an option, but imho an unnecessary complication. And obviously it would still prevent things like nfsroot (which requires the MAC address and potentially a WiFi calibration data to be setup) from working out of the box.

Doing the detour via userspace only for devices with an eMMC would be different from how it is done for any other type of backing storage such as simple I2C EEPROMs, (SPI-)flashes or UBI volumes.
Having a unified approach for all of them would make things much more convenient, as typically the actual layouts are even the same and can be reused accross devices of the same vendor. GL-iNet or ASUS router devices are a good example for that: The more high-end ones come with an eMMC and use the same NVMEM layout inside a GPT partition than mid-field devices on SPI-NAND or UBI, and older and/or lower-end devices on an SPI-NOR flash MTD partition.

To better understand the situation, maybe look at a few examples for which we are currently already using this patchset downstream at OpenWrt:

Inside a GPT partition on an eMMC:
https://git.openwrt.org/?p=openwrt/openwrt.git;a=blob;f=target/linux/mediatek/dts/mt7981b-unielec-u7981-01-emmc.dts;h=abd4d4e59d74cc0d24e8c9e10e16c0859844e003;hb=HEAD#l39

On raw SPI-NAND (already possible with vanilla Linux):
https://git.openwrt.org/?p=openwrt/openwrt.git;a=blob;f=target/linux/mediatek/dts/mt7981b-unielec-u7981-01-nand.dts;h=230a612a34f7d370eb09e92284950d9949bf10cd;hb=HEAD#l45

Inside a UBI volume (also already possible with vanilla Linux):
https://git.openwrt.org/?p=openwrt/openwrt.git;a=blob;f=target/linux/mediatek/dts/mt7986a-asus-tuf-ax4200.dts;h=e40602fa215e1a677b534c85b767e824af041518;hb=HEAD#l255

Inside the boot hwpartition of the eMMC:
https://git.openwrt.org/?p=openwrt/openwrt.git;a=blob;f=target/linux/mediatek/dts/mt7981b-glinet-gl-mt2500.dts;h=15818a90fca02955879d1bcca55ba2153e390621;hb=HEAD#l156

Inside a GPT partition on an eMMC:
https://git.openwrt.org/?p=openwrt/openwrt.git;a=blob;f=target/linux/mediatek/dts/mt7986a-glinet-gl-mt6000.dts;h=fd0e1a69157ed0f27c32376aabab0fcc85ce23b9;hb=HEAD#l317

https://git.openwrt.org/?p=openwrt/openwrt.git;a=blob;f=target/linux/mediatek/dts/mt7988a-smartrg-mt-stuart.dtsi;h=2b468f9bb311141d083e51ca5edaa7dce253c91c;hb=HEAD#l504

Those are just the examples I have been working on myself, so it was easy to come up with them. There are also ASUS devices with Qualcomm SoC using the same layout as the MediaTek-based devices inside a UBI volume.

Once a unified way to describe the loaction of the NVMEM bits is also present in upstream Linux, all those downstream device trees could be submitted for inclusion in Linux, and one could install a general-purpose OS like Debian **which wouldn't need to know anything about the details of where to read MAC addresses or calibration data from**, all hardware-specific knowledge would reside in DT.

The failure of defining this in a nice way results in very ugly situations such as distributions carrying the (board-specific!) calibration data for very few but very common single-board computers like the RaspberryPi in their rootfs or even in initramfs. Each distribution with a different level of hacks to give them individual MAC addresses and to load the correct file... And while this doesn't look so bad for systems which anyway come only with removable microSD storage, it **does* get ugly when it comes to systems which do store that information on their eMMC (typically "appliances" rather than SBCs meant for tinkerers).

Please take all that into consiration and also note that obviously, on systems not making use of any of that, you may simple not enable CONFIG_BLOCK_NOTIFIERS -- other than in previous iterations of the patchset this is all completely optional now.

Cheers

Daniel

Maybe it needs to become more common so that we don't need crazy kernel workarounds for something that can be trivially done with a few lines of userspace code?

Add new flag to destinguish block devices which may act as an NVMEM provider.

---
 include/linux/blkdev.h | 2 ++
 1 file changed, 2 insertions(+)

index c6e9fb16f76e..c6c6153dbede 100644
@@ -81,11 +81,13 @@ struct partition_meta_info {
  * ``GENHD_FL_NO_PART``: partition support is disabled.  The kernel will not
  * scan for partitions from add_disk, and users can't add partitions manually.
  *
+ * ``GENHD_FL_NVMEM``: the block device should be considered as NVMEM provider.
  */
 enum {
 	GENHD_FL_REMOVABLE			= 1 << 0,
 	GENHD_FL_HIDDEN				= 1 << 1,
 	GENHD_FL_NO_PART			= 1 << 2,
+	GENHD_FL_NVMEM				= 1 << 3,
 };
 
 enum {

2.45.2

2-4 look fine to me now, but I don't know anything about the device bindings so someone qualified should review that before it gets queued up.

Jens Axboe