[PATCH 0/2] efi_loader: provide media ID

[AKASHI Takahiro]

On Wed, Sep 28, 2022 at 08:57:43AM +0200, Heinrich Schuchardt wrote:
> 
> 
> On 9/28/22 03:54, Simon Glass wrote:
> > Hi,
> > 
> > On Tue, 27 Sept 2022 at 00:53, Heinrich Schuchardt
> > <heinrich.schuchardt at canonical.com> wrote:
> > > 
> > > 
> > > 
> > > On 9/27/22 03:51, AKASHI Takahiro wrote:
> > > > On Mon, Sep 26, 2022 at 08:06:52AM +0200, Heinrich Schuchardt wrote:
> > > > > 
> > > > > 
> > > > > On 9/16/22 02:58, AKASHI Takahiro wrote:
> > > > > > On Thu, Sep 15, 2022 at 10:02:40PM +0200, Heinrich Schuchardt wrote:
> > > > > > > The medium a device like 'mmc 0' or 'usb 0' points to may change over
> > > > > > > time. Hence device type and number are not sufficient to identify the
> > > > > > > inserted medium. The same is true for the device path generated for
> > > > > > > such a device.
> > > > > > 
> > > > > > Well, it depends on how a device path is generated in U-Boot's UEFI
> > > > > > implementation. I believe that a device path represents an "unique path"
> > > > > > to a given device however this device is enumerated.
> > > > > > In this sense, the current dp_fill()/efi_dp_from_part() is not a right
> > > > > > implementation as it relies on device numbers.
> > > > > > Furthermore, a generated device path here is different from one generated
> > > > > > by EDK2 (even if both software are run on the same board).
> > > > > > 
> > > > > > This is an issue that I used to tackle in
> > > > > > https://lists.denx.de/pipermail/u-boot/2021-November/468216.html
> > > > > > although I have since had no progress.
> > > > > > 
> > > > > > > This is why the EFI_BLOCK_IO_PROTOCOL provides a field
> > > > > > > MediaId.
> > > > > > > 
> > > > > > > Whenever a removable medium is changed or a new block device with a
> > > > > > > previously used device path is created we should provide a different
> > > > > > > MediaID.
> > > > > > > 
> > > > > > > This series adds a field media_id to the block device descriptor and fills
> > > > > > > it after probing. The value of the field is then copied to the
> > > > > > > EFI_BLOCK_IO_PROTOCOL.
> > > > > > 
> > > > > > I'm afraid that your patch doesn't always work as you expect.
> > > > > > When "scsi rescan" or "usb stop; usb start", for instance, is invoked,
> > > > > > all the existing devices and associated blk_desc structures are once freed
> > > > > > and even if nothing is changed, i.e. a device is neither removed nor added,
> > > > > > the exact same structures will be re-created.
> > > > > > With your patch applied, however, a new (and different) "media_id" will be
> > > > > > assigned to an existing device. UEFI User may be notified of "media change".
> > > > > > (To be honest, this is quite unlikely because the current UEFI implementation
> > > > > > doesn't use BLOCK_IO_PROTOCOL internally, say, for file system access.)
> > > > > 
> > > > > This behavior matches what EDK II does if you remove a device and create a
> > > > > new device.
> > > > 
> > > > I don't think that EDK2 has "scsi rescan" or others, which users can invoke
> > > > at any time. Moreover, I believe that EDK2 code (drivers) checks whether a device
> > > > is really changed or not before updating a MediaId.
> > > > 
> > > > > If a device is removed and recreated anything could have happened in between
> > > > > like complete repartitioning. We cannot assume that any cached state is
> > > > > valid anymore even if GUIDs are the same.
> > > > 
> > > > I'm not sure if you fully understand my point.
> > > > My assumption is the case where a device is NOT removed around "scsi rescan"
> > > > (or usb stop/start) and stays online. In this case,
> > > > 1. access to, say, "scsi 0:1", via UEFI BLOCK_IO succeeds
> > > > 2. "scsi rescan"
> > > > 3. access to the same device, "scsi 0:1", via UEFI BLOCK_IO
> > > > currently (3) succeeds, but with your patch, it may potentially fail because
> > > > of media_id altered.
> > > > 
> > > > I admit that it will not happen under the current UEFI implementation because
> > > > non of UEFI applications will survive across command lines and none of information,
> > > > including media_id or handle, can be carried over from (1) to (3).
> > > > But unconditionally incrementing an internally-held media_id, as in your patch,
> > > > is a wrong behavior.
> > > 
> > > The patch issues a new media ID if a new device is probed which only
> > > happens to have the same device number if another device of that number
> > > was removed before.
> > > 
> > > Commands like 'usb scan' don't necessarily issue the same numbers to the
> > > same device as before the command if a new device has been attached in
> > > the meanwhile.
> > > 
> > > Assuming that a new device contains the same medium as an old one
> > > because by chance it has the same device number is definitively unsafe.
> > > 
> > > If a device is probed, we have to assume that it contains a new medium.
> > 
> > Sorry if I repeat myself, but this sort of thing should be handled in
> > the driver model code. Can we get some more progress on integrating
> > the EFI layer better?
> 
> The last mails where about *whether* the media ID should be bumped after a
> block device has been created and not about where we will implement it.

Indeed. I don't care "where" for now, but "how" or "whether".

The most essential issue is that none of U-Boot block device drivers has
ability of detecting media insertion or removal immediately
(due to the lack of interrupt support).
This is even not related to DM or not.

-Takahiro Akashi

> Best regards
> 
> Heinrich