In addition to my previous letter I have to mention a couple more newly discovered details.

1. UEFI Shell (ShellPkg) actually has 3 functions dedicated to connecting controllers and essentially doing the same thing:

- ConnectAllEfi
ShellPkg/Library/UefiShellLevel2CommandsLib/Load.c

This one locates all handles and runs connect on them. It is the one we mentioned in the bug.

- LoadPciRomConnectAllDriversToAllControllers
ShellPkg/Library/UefiShellDebug1CommandsLib/LoadPciRom.c

This one is similar to ConnectAllEfi. The only difference is that it additionally checks for Ctrl+C via ShellGetExecutionBreakFlag.

- ConnectControllers
ShellPkg/Library/UefiShellDriver1CommandsLib/Connect.c

This one is more complex, as it supports explicitly connecting specified controllers, however, for connecting all controllers it locates handles with gEfiDevicePathProtocolGuid. I.e. exactly what we ask. I believe that all these functions should behave the same way in correspondence to gBS->ConnectController at the very least, and most likely there should be a library responsible for connection and disconnection.

2. We also checked EFI 1.1 Shell, and confirmed that it consistently has checks for device path presence before running connect on the handle[1][2]. This makes us believe that our proposal is not really a firmware bug, but rather a limitation of the legacy specification.

Considering all these discoveries, we believe our suggested change is legit depending on the minimum supported UEFI version. Therefore we propose submitting a ControllerConnection library with 5 functions:

- ConnectController
- DisconnectController
- GetAllControllers
- ConnectAllControllers
- DisconnectAllControllers

Adopting it throughout the code will let the distribitor to be responsible for choosing what is right for his applications (or firmwares).

Best wishes,
Vitaly

[1] https://sourceforge.net/p/efi-shell/code/64/tree/trunk/Shell/LoadPciRom/LoadPciRom.c#l528
[2] https://sourceforge.net/p/efi-shell/code/64/tree/trunk/Shell/load/load.c#l312


14 янв. 2020 г., в 11:36, vit9696 <vit9696@protonmail.com> написал(а):

Ray,

We are quite reluctant to have patches in EDK II for a large amount of widely adopted firmwares. Patches eventually break and require maintenance cost, and currently we are trying to get rid of them all. We believe that EDK II Shell is supposed to work on real world platforms and not only the ones that theoretically support the specification. It is always hard to adopt changes based on third-party bugs, and we very well understand your concern, yet it is something we have to do to stay beneficial to the end user.

Best wishes,
Vitaly

On Tue, Jan 14, 2020 at 05:53, Ni, Ray <ray.ni@intel.com> wrote:

Vitaly,

I still have concern to modify the EDKII code to workaround a firmware bug.

Can you just change in your local version?

 

Thanks,

Ray

 

From: devel@edk2.groups.io <devel@edk2.groups.io> On Behalf Of Vitaly Cheptsov via Groups.Io
Sent: Tuesday, January 14, 2020 4:47 AM
To: Laszlo Ersek <lersek@redhat.com>; devel@edk2.groups.io; Ni, Ray <ray.ni@intel.com>; Gao, Zhichao <zhichao.gao@intel.com>
Subject: Re: [edk2-devel] [PATCH 1/1] ShellPkg: Do not connect handles without device paths

 
Thanks all for your input,

 

These explanations seem sufficient to us that it is not a good idea to change the behaviour for everyone. Even so, we still need this to be configurable in some way, as having to patch EDK II is impracticable.

 

We believe there are three possible routes to approach this problem:

 

  1. Introduce a separate ControllerConnectionLib library for this function. While it is small, we found several places in our code that need to call it beyond UEFI Shell. This way different implementations could be used depending on the chosen library.
  2. Introduce a ConnectRequiresDevicePath PCD, which will choose the preferred logic.
  3. Introduce a -dp Shell argument for affected commands the way Lazslo suggested.

 

We believe either route or a combination of multiple routes have their own benefits, and would suggest either going with 1+2 or with 3. Any approach is fine for us.

 

We will submit V2 of the patch after hearing the opinions.

 

Best wishes,
Vitaly
 

On Mon, Jan 13, 2020 at 20:55, Laszlo Ersek <lersek@redhat.com> wrote:

On 01/13/20 12:56, Ni, Ray wrote:
> We shouldn't assume that a DriverBindingStart() can only start on a handle with device path installed. DevicePath protocol is just a special protocol.
> It's possible that a bus driver starts on a host controller handle and creates multiple children, each with only a Specific_IO protocol installed.
> Certain device driver can start on the children handle and open the Specific_IO protocol BY_DRIVER.
> I am not sure if certain today's network drivers may work like this. It's allowed per UEFI spec.

I agree.

Under "10.2 EFI Device Path Protocol", the spec says, "If the handle
does not logically map to a physical device, the handle may not
necessarily support the device path protocol."

I think gBS->ConnectController() and
EFI_DRIVER_BINDING_PROTOCOL.Supported() should work on such handles.

If we'd like to work around related issues in drivers, then I'd suggest
new command line options for the "load", "connect", "reconnect" shell
commands (maybe more), for filtering out handles that do not carry
device paths. Such command line options could be added as an extension,
IIUC, such as "-_option". I.e., I believe it's not necessary to start
with UEFI Shell Spec updates.

Thanks
Laszlo