[RFC 0/2] fixed order for mmc devices/mmc block devices

[RFC 0/2] fixed order for mmc devices/mmc block devices

[stefan-XLVq0VzYD2Y]

From: Stefan Agner <stefan-XLVq0VzYD2Y@public.gmane.org>

Nowadays SoCs usually have multiple SD/MMC host adapters. Often
the first is connected to an internal eMMC, however sometimes it
is the second controller that is connected to the eMMC.

If the second controller is connected to the eMMC, the number
assigned to the eMMC block device varies depending on whether the
external SD-Card is present or not.
The boot device could be specified using partition UUIDs, but
those can change when recreating the partition table. Especially
for embedded devices it would be preferable to have a reliable
assigned block device number.

The first patch aims to give a way to specify the logical SD/MMC
host order. This is achieved by extending the device tree using
aliases:

aliases {
        mmc0 = "/sdhci@78000600";
        mmc1 = "/sdhci@78000200";
};

The second patch then uses the SD/MMC host index as preferred id
for the block device.

In theory, one SD/MMC host could control multiple cards. For this
case, this patch is somewhat ugly. It would only assign the first
host/card combination with a predictable block number. However,
for the case at hand this would be sufficient. 

Stefan Agner (2):
  mmc: read mmc alias from device tree
  mmc: use SD/MMC host ID for block device name ID

 drivers/mmc/card/block.c |  3 ++-
 drivers/mmc/core/host.c  | 24 +++++++++++++++++++-----
 2 files changed, 21 insertions(+), 6 deletions(-)

-- 
1.9.1

[RFC 1/2] mmc: read mmc alias from device tree

[stefan-XLVq0VzYD2Y]

From: Stefan Agner <stefan-XLVq0VzYD2Y@public.gmane.org>

To get the SD/MMC host device ID, read the alias from the device
tree.

This is useful in case a SoC has multipe SD/MMC host controllers while
the second controller should logically be the first device (e.g. if
the second controller is connected to an internal eMMC). Combined
with block device numbering using MMC/SD host device ID, this
results in predictable name assignment of the internal eMMC block
device.

Signed-off-by: Stefan Agner <stefan-XLVq0VzYD2Y@public.gmane.org>
---
 drivers/mmc/core/host.c | 24 +++++++++++++++++++-----
 1 file changed, 19 insertions(+), 5 deletions(-)

diff --git a/drivers/mmc/core/host.c b/drivers/mmc/core/host.c
index 49bc403..29f44cf 100644
--- a/drivers/mmc/core/host.c
+++ b/drivers/mmc/core/host.c
@@ -448,8 +448,8 @@ EXPORT_SYMBOL(mmc_of_parse);
  */
 struct mmc_host *mmc_alloc_host(int extra, struct device *dev)
 {
-	int err;
 	struct mmc_host *host;
+	int of_id = -1, id = -1;
 
 	host = kzalloc(sizeof(struct mmc_host) + extra, GFP_KERNEL);
 	if (!host)
@@ -459,12 +459,26 @@ struct mmc_host *mmc_alloc_host(int extra, struct device *dev)
 	host->rescan_disable = 1;
 	idr_preload(GFP_KERNEL);
 	spin_lock(&mmc_host_lock);
-	err = idr_alloc(&mmc_host_idr, host, 0, 0, GFP_NOWAIT);
-	if (err >= 0)
-		host->index = err;
+
+	if (dev->of_node)
+		of_id = of_alias_get_id(dev->of_node, "mmc");
+
+	if (of_id >= 0) {
+		id = idr_alloc(&mmc_host_idr, host, of_id, of_id + 1,
+				GFP_NOWAIT);
+		if (id < 0)
+			dev_warn(dev, "/aliases ID %d not available\n", of_id);
+	}
+
+	if (id < 0)
+		id = idr_alloc(&mmc_host_idr, host, 0, 0, GFP_NOWAIT);
+
+	if (id >= 0)
+		host->index = id;
+
 	spin_unlock(&mmc_host_lock);
 	idr_preload_end();
-	if (err < 0)
+	if (id < 0)
 		goto free;
 
 	dev_set_name(&host->class_dev, "mmc%d", host->index);
-- 
1.9.1

Re: [RFC 1/2] mmc: read mmc alias from device tree

[Doug Anderson]

Stefan,

On Tue, Apr 1, 2014 at 1:35 PM,  <stefan-XLVq0VzYD2Y@public.gmane.org> wrote:
> From: Stefan Agner <stefan-XLVq0VzYD2Y@public.gmane.org>
>
> To get the SD/MMC host device ID, read the alias from the device
> tree.
>
> This is useful in case a SoC has multipe SD/MMC host controllers while
> the second controller should logically be the first device (e.g. if
> the second controller is connected to an internal eMMC). Combined
> with block device numbering using MMC/SD host device ID, this
> results in predictable name assignment of the internal eMMC block
> device.
>
> Signed-off-by: Stefan Agner <stefan-XLVq0VzYD2Y@public.gmane.org>

Yes, please.  It is very handy to have mmc devices be enumerated in a
predictable way.

Reviewed-by: Doug Anderson <dianders-F7+t8E8rja9g9hUCZPvPmw@public.gmane.org>
Tested-by: Doug Anderson <dianders-F7+t8E8rja9g9hUCZPvPmw@public.gmane.org>

Re: [RFC 1/2] mmc: read mmc alias from device tree

[Doug Anderson]

Hi,

On Tue, Apr 1, 2014 at 1:35 PM,  <stefan-XLVq0VzYD2Y@public.gmane.org> wrote:
> From: Stefan Agner <stefan-XLVq0VzYD2Y@public.gmane.org>
>
> To get the SD/MMC host device ID, read the alias from the device
> tree.
>
> This is useful in case a SoC has multipe SD/MMC host controllers while
> the second controller should logically be the first device (e.g. if
> the second controller is connected to an internal eMMC). Combined
> with block device numbering using MMC/SD host device ID, this
> results in predictable name assignment of the internal eMMC block
> device.
>
> Signed-off-by: Stefan Agner <stefan-XLVq0VzYD2Y@public.gmane.org>
> ---
>  drivers/mmc/core/host.c | 24 +++++++++++++++++++-----
>  1 file changed, 19 insertions(+), 5 deletions(-)
>
> diff --git a/drivers/mmc/core/host.c b/drivers/mmc/core/host.c
> index 49bc403..29f44cf 100644
> --- a/drivers/mmc/core/host.c
> +++ b/drivers/mmc/core/host.c
> @@ -448,8 +448,8 @@ EXPORT_SYMBOL(mmc_of_parse);
>   */
>  struct mmc_host *mmc_alloc_host(int extra, struct device *dev)
>  {
> -       int err;
>         struct mmc_host *host;
> +       int of_id = -1, id = -1;
>
>         host = kzalloc(sizeof(struct mmc_host) + extra, GFP_KERNEL);
>         if (!host)
> @@ -459,12 +459,26 @@ struct mmc_host *mmc_alloc_host(int extra, struct device *dev)
>         host->rescan_disable = 1;
>         idr_preload(GFP_KERNEL);
>         spin_lock(&mmc_host_lock);
> -       err = idr_alloc(&mmc_host_idr, host, 0, 0, GFP_NOWAIT);
> -       if (err >= 0)
> -               host->index = err;
> +
> +       if (dev->of_node)
> +               of_id = of_alias_get_id(dev->of_node, "mmc");

Not that there's much chance of such an old patch landing, but just in
case anyone finds this Dmitry points out that this causes a
"scheduling while atomic".  A fix is to move the "of_alias_get_id" up
above the spin_lock():

https://chromium-review.googlesource.com/#/c/259916/

-Doug

[RFC 2/2] mmc: use SD/MMC host ID for block device name ID

[stefan]

From: Stefan Agner <stefan@agner.ch>

By using the SD/MMC host device ID as a starting point for block
device numbering, one can reliably predict the first block device
name (at least for the first controller). This is especially useful
for SoCs with multiple SD/MMC host controller, where the controller
with index 0 is connected to a eMMC device.

Usually the first controller gets the first block device name ID,
however this is not guaranteed. Also if the first controller is
aliased as second controller and visa-versa (using device tree
aliases), the block device name ID assignation is not ordered by
the SD/MMC host device ID (since mmc_rescan is called in order of
the memory mapped pheripherial addresses).

Signed-off-by: Stefan Agner <stefan@agner.ch>
---
 drivers/mmc/card/block.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/drivers/mmc/card/block.c b/drivers/mmc/card/block.c
index 7b5424f..03626ed 100644
--- a/drivers/mmc/card/block.c
+++ b/drivers/mmc/card/block.c
@@ -2045,7 +2045,8 @@ static struct mmc_blk_data *mmc_blk_alloc_req(struct mmc_card *card,
 	 * index anymore so we keep track of a name index.
 	 */
 	if (!subname) {
-		md->name_idx = find_first_zero_bit(name_use, max_devices);
+		md->name_idx = find_next_zero_bit(name_use, max_devices,
+				card->host->index);
 		__set_bit(md->name_idx, name_use);
 	} else
 		md->name_idx = ((struct mmc_blk_data *)
-- 
1.9.1

Re: [RFC 2/2] mmc: use SD/MMC host ID for block device name ID

[Stephen Warren]

On 04/01/2014 02:35 PM, stefan@agner.ch wrote:
> From: Stefan Agner <stefan@agner.ch>
> 
> By using the SD/MMC host device ID as a starting point for block
> device numbering, one can reliably predict the first block device
> name (at least for the first controller).

That's not true. There's no guarantee that a device name/ID gets
released as soon as the SD card is removed; something might still have
it mounted for example.

The correct solution here is to use filesystem or partition UUIDs to
identify the device/partition, not to attempt to assign static device IDs.

Re: [RFC 2/2] mmc: use SD/MMC host ID for block device name ID

[Stefan Agner]

Am 2014-04-01 22:47, Stephen Warren wrote:
> That's not true. There's no guarantee that a device name/ID gets
> released as soon as the SD card is removed; something might still have
> it mounted for example.
Yes. Also when booted there are other solutions to get static block
device name (e.g. /dev/disk/by-path), but the problem at hand aims for a
solution for the root file system.

Also, eMMC don't get removed, and those internal eMMC devices are the
reason of this thread... (see cover letter)

 
> The correct solution here is to use filesystem or partition UUIDs to
> identify the device/partition, not to attempt to assign static device IDs.
Yes, I'm aware of that solution. However, when recreating the partition
table, those UUIDs do change, which in turn needs a change of the kernel
arguments. Of course, one can use scripts to work around this, but its
easier to just boot from the block device at a given location (say eMMC,
first partition). Then, using UUID is also not as fast as using a device
name directly, since all block devices get scanned. This is not optimal
when trying to optimize boot speed, but might be negligible. I actually
never measured that. 

There are the /dev/disk/by-path/ which aim at a similar goal, but those
names are generated by udev. I'm open to suggestions solve this issue in
a more fashionable way...

Re: [RFC 2/2] mmc: use SD/MMC host ID for block device name ID

[Doug Anderson]

Stefan,

On Tue, Apr 1, 2014 at 1:35 PM,  <stefan-XLVq0VzYD2Y@public.gmane.org> wrote:
> From: Stefan Agner <stefan-XLVq0VzYD2Y@public.gmane.org>
>
> By using the SD/MMC host device ID as a starting point for block
> device numbering, one can reliably predict the first block device
> name (at least for the first controller). This is especially useful
> for SoCs with multiple SD/MMC host controller, where the controller
> with index 0 is connected to a eMMC device.
>
> Usually the first controller gets the first block device name ID,
> however this is not guaranteed. Also if the first controller is
> aliased as second controller and visa-versa (using device tree
> aliases), the block device name ID assignation is not ordered by
> the SD/MMC host device ID (since mmc_rescan is called in order of
> the memory mapped pheripherial addresses).
>
> Signed-off-by: Stefan Agner <stefan-XLVq0VzYD2Y@public.gmane.org>
> ---
>  drivers/mmc/card/block.c | 3 ++-
>  1 file changed, 2 insertions(+), 1 deletion(-)

As Stephen points out, "mmcblk" IDs for SD cards are not actually
guaranteed.  Also UUID is a better API for things to use.

That being said, your patch helps me a lot when doing development.  As
you said, I'm guaranteed that if eMMC is no-removable and enumerated
at boot that it will be a predictable ID and I can use it in my
scripts.  Also this helps keep me from having to do lots of extra
thinking to figure out whether I happened to have an SD card plugged
in at boot or I didn't.

I'd love to see this land.

Reviewed-by: Doug Anderson <dianders-F7+t8E8rja9g9hUCZPvPmw@public.gmane.org>
Tested-by: Doug Anderson <dianders-F7+t8E8rja9g9hUCZPvPmw@public.gmane.org>