Re: [PATCH 1/6] sysfs: Introduce a mechanism to hide static attribute_groups

[Dan Williams]

Pierre-Louis Bossart wrote:
> 
> 
> On 1/30/24 19:46, Greg Kroah-Hartman wrote:
> > From: Dan Williams <dan.j.williams@xxxxxxxxx>
> > 
> > Add a mechanism for named attribute_groups to hide their directory at
> > sysfs_update_group() time, or otherwise skip emitting the group
> > directory when the group is first registered. It piggybacks on
> > is_visible() in a similar manner as SYSFS_PREALLOC, i.e. special flags
> > in the upper bits of the returned mode. To use it, specify a symbol
> > prefix to DEFINE_SYSFS_GROUP_VISIBLE(), and then pass that same prefix
> > to SYSFS_GROUP_VISIBLE() when assigning the @is_visible() callback:
> > 
> > 	DEFINE_SYSFS_GROUP_VISIBLE($prefix)
> > 
> > 	struct attribute_group $prefix_group = {
> > 		.name = $name,
> > 		.is_visible = SYSFS_GROUP_VISIBLE($prefix),
> > 	};
> > 
> > SYSFS_GROUP_VISIBLE() expects a definition of $prefix_group_visible()
> > and $prefix_attr_visible(), where $prefix_group_visible() just returns
> > true / false and $prefix_attr_visible() behaves as normal.
> > 
> > The motivation for this capability is to centralize PCI device
> > authentication in the PCI core with a named sysfs group while keeping
> > that group hidden for devices and platforms that do not meet the
> > requirements. In a PCI topology, most devices will not support
> > authentication, a small subset will support just PCI CMA (Component
> > Measurement and Authentication), a smaller subset will support PCI CMA +
> > PCIe IDE (Link Integrity and Encryption), and only next generation
> > server hosts will start to include a platform TSM (TEE Security
> > Manager).
> > 
> > Without this capability the alternatives are:
> > 
> > * Check if all attributes are invisible and if so, hide the directory.
> >   Beyond trouble getting this to work [1], this is an ABI change for
> >   scenarios if userspace happens to depend on group visibility absent any
> >   attributes. I.e. this new capability avoids regression since it does
> >   not retroactively apply to existing cases.
> > 
> > * Publish an empty /sys/bus/pci/devices/$pdev/tsm/ directory for all PCI
> >   devices (i.e. for the case when TSM platform support is present, but
> >   device support is absent). Unfortunate that this will be a vestigial
> >   empty directory in the vast majority of cases.
> > 
> > * Reintroduce usage of runtime calls to sysfs_{create,remove}_group()
> >   in the PCI core. Bjorn has already indicated that he does not want to
> >   see any growth of pci_sysfs_init() [2].
> > 
> > * Drop the named group and simulate a directory by prefixing all
> >   TSM-related attributes with "tsm_". Unfortunate to not use the naming
> >   capability of a sysfs group as intended.
> > 
> > In comparison, there is a small potential for regression if for some
> > reason an @is_visible() callback had dependencies on how many times it
> > was called. Additionally, it is no longer an error to update a group
> > that does not have its directory already present, and it is no longer a
> > WARN() to remove a group that was never visible.
> > 
> > Link: https://lore.kernel.org/all/2024012321-envious-procedure-4a58@gregkh/ [1]
> > Link: https://lore.kernel.org/linux-pci/20231019200110.GA1410324@bhelgaas/ [2]
> > Signed-off-by: Dan Williams <dan.j.williams@xxxxxxxxx>
> > Signed-off-by: Greg Kroah-Hartman <gregkh@xxxxxxxxxxxxxxxxxxx>
> 
> This patch seems to introduce a regression on our Lunar Lake test
> devices, where we can't boot to an ssh shell. No issues on older devices
> [1]. Bard Liao and I reproduced the same results on different boards.
> 
> We'll need to find someone with direct device access to provide more
> information on the problem, remote testing without ssh is a
> self-negating proposition.
> 
> Is there a dependency on other patches? Our tests are still based on
> 6.7.0-rc3 due to other upstream issues we're currently working through.

The only behavior change I can imagine with this patch is that
->is_visble() callbacks get called extra times for named attribute
groups.

..or if an is_visible() callback was inadvertantly already using the
SYSFS_GROUP_INVISIBLE flag in umode_t result.