[PATCH v3 2/5] efi: Add embedded peripheral firmware support

[PATCH v3 2/5] efi: Add embedded peripheral firmware support

[Luis R. Rodriguez]

On Tue, Apr 24, 2018 at 12:07:01PM -0400, Mimi Zohar wrote:
> On Tue, 2018-04-24 at 17:09 +0200, Hans de Goede wrote:
> > Hi,
> > 
> > On 23-04-18 23:11, Luis R. Rodriguez wrote:
> > > Hans, please see use of READING_FIRMWARE_PREALLOC_BUFFER, we'll need a new ID
> > > and security for this type of request so IMA can reject it if the policy is
> > > configured for it.
> > 
> > Hmm, interesting, actually it seems like the whole existence
> > of READING_FIRMWARE_PREALLOC_BUFFER is a mistake, 

request_firmware_into_buf() was merged without my own review, however,
the ID thing did get review from Mimi:

https://patchwork.kernel.org/patch/9074611/

The ID is not for IMA alone, its for any LSM to decide what to do.
Note Mimi asked for READING_FIRMWARE_DMA if such buffer was in DMA,
otherise READING_FIRMWARE_PREALLOC_BUFFER was suggested.

> > the IMA
> > framework really does not care if we are loading the firmware
> > into memory allocated by the firmware-loader code, or into
> > memory allocated by the device-driver requesting the firmware.

That's up to LSM folks to decide. We have these so far:

#define __kernel_read_file_id(id) \                                             
        id(UNKNOWN, unknown)            \                                       
        id(FIRMWARE, firmware)          \                                       
        id(FIRMWARE_PREALLOC_BUFFER, firmware)  \                               
        id(MODULE, kernel-module)               \                               
        id(KEXEC_IMAGE, kexec-image)            \                               
        id(KEXEC_INITRAMFS, kexec-initramfs)    \                               
        id(POLICY, security-policy)             \                               
        id(X509_CERTIFICATE, x509-certificate)  \                               
        id(MAX_ID, )  

The first type of IDs added was about type of files the kernel
LSMs may want to do different things for.

Mimi why did you want a separate ID for it back before?

I should note now that request_firmware_into_buf() and its
READING_FIRMWARE_PREALLOC_BUFFER was to enable a driver on memory constrained
devices. The files are large (commit says 16 MiB).

I've heard of larger possible files with remoteproc and with Android using
the custom fallback mechanism -- which could mean a proprietary tool
fetching firmware from a random special place on a device.

I could perhaps imagine an LSM which may be aware of such type of
arrangement may want to do its own vetting of some sort, but this
would not be specific to READING_FIRMWARE_PREALLOC_BUFFER, but rather
the custom fallback mechaism.

Whether or not the buffer was preallocated by the driver seems a little
odd for security folks to do something different with it. Security LSM
folks please chime in.

I could see a bit more of a use case for an ID for firmware scraped
from EFI, which Hans' patch will provide. But that *also* should get
good review from other LSM folks.

One of the issues with accepting more IDs loosely is where do we
stop though? If no one really is using READING_FIRMWARE_PREALLOC_BUFFER
I'd say lets remove it. Likewise, for this EFI thing I'd like an idea
if we really are going to have users for it.

If its of any help --

drivers/soc/qcom/mdt_loader.c is the only driver currently using
request_firmware_into_buf() however I'll note qcom_mdt_load() is used in many
other drivers so they are wrappers around request_firmware_into_buf():

drivers/gpu/drm/msm/adreno/a5xx_gpu.c:   * adreno_request_fw() handles this, but qcom_mdt_load() does
drivers/gpu/drm/msm/adreno/a5xx_gpu.c:          ret = qcom_mdt_load(dev, fw, fwname, GPU_PAS_ID,
drivers/gpu/drm/msm/adreno/a5xx_gpu.c:          ret = qcom_mdt_load(dev, fw, newname, GPU_PAS_ID,
drivers/media/platform/qcom/venus/firmware.c:   ret = qcom_mdt_load(dev, mdt, fwname, VENUS_PAS_ID, mem_va, mem_phys,
drivers/remoteproc/qcom_adsp_pil.c:     return qcom_mdt_load(adsp->dev, fw, rproc->firmware, adsp->pas_id,
drivers/remoteproc/qcom_wcnss.c:        return qcom_mdt_load(wcnss->dev, fw, rproc->firmware, WCNSS_PAS_ID,

Are we going to add more IDs for more types of firmware?
What type of *different* decisions could LSMs take if the firmware
was being written to a buffer? Or in this new case that is coming
up, if the file came scraping EFI, would having that information
be useful?

> > As such the current IMA code (from v4.17-rc2) actually does
> > not handle READING_FIRMWARE_PREALLOC_BUFFER at all, 
> 
> Right, it doesn't yet address READING_FIRMWARE_PREALLOC_BUFFER, but
> should.
> 
> Depending on whether the device requesting the firmware has access to
> the DMA memory, before the signature verification, 

It would seem from the original patch review about READING_FIRMWARE_PREALLOC_BUFFER
that this is not a DMA buffer.

The device driver should have access to the buffer pointer with write given
that with request_firmware_into_buf() the driver is giving full write access to
the memory pointer so that the firmware API can stuff the firmware it finds
there.

Firmware signature verification would be up to the device hardware to do upon
load *after* request_firmware_into_buf().

  Luis

> will determine how
> IMA-appraisal addresses READING_FIRMWARE_PREALLOC_BUFFER.
> 
> Mimi
> 
> > here
> > are bits of code from: security/integrity/ima/ima_main.c:
> > 
> > static int read_idmap[READING_MAX_ID] = {
> >          [READING_FIRMWARE] = FIRMWARE_CHECK,
> >          [READING_MODULE] = MODULE_CHECK,
> >          [READING_KEXEC_IMAGE] = KEXEC_KERNEL_CHECK,
> >          [READING_KEXEC_INITRAMFS] = KEXEC_INITRAMFS_CHECK,
> >          [READING_POLICY] = POLICY_CHECK
> > };
> > 
> > int ima_post_read_file(struct file *file, void *buf, loff_t size,
> > 	...
> >          if (!file && read_id == READING_FIRMWARE) {
> >                  if ((ima_appraise & IMA_APPRAISE_FIRMWARE) &&
> >                      (ima_appraise & IMA_APPRAISE_ENFORCE))
> >                          return -EACCES; /* INTEGRITY_UNKNOWN */
> >                  return 0;
> >          }
> > 
> > Which show that the IMA code is not handling
> > READING_FIRMWARE_PREALLOC_BUFFER as it should (I believe it
> > should handle it the same as READING_FIRMWARE).
> > 
> > Now we could fix that, but the only user of
> > READING_FIRMWARE_PREALLOC_BUFFER is the code which originally
> > introduced it:
> > 
> > https://patchwork.kernel.org/patch/9162011/
> > 
> > So I believe it might be better to instead replace it
> > with just READING_FIRMWARE and find another way to tell
> > kernel_read_file() that there is a pre-allocated buffer,
> > perhaps the easiest way there is that  *buf must be
> > NULL when the caller wants kernel_read_file() to
> > vmalloc the mem. This would of course require auditing
> > all callers that the buf which the pass in is initialized
> > to NULL.
> > 
> > Either way adding a third READING_FIRMWARE_FOO to the
> > kernel_read_file_id enum seems like a bad idea, from
> > the IMA pov firmware is firmware.
> > 
> > What this whole exercise has shown me though is that
> > I need to call security_kernel_post_read_file() when
> > loading EFI embedded firmware. I will add a call to
> > security_kernel_post_read_file() for v4 of the patch-set.
> > 
> > > Please Cc Kees in future patches.
> > 
> > Will do.
> > 
> > Regards,
> > 
> > Hans
> > 
> 
> 

-- 
Do not panic
--
To unsubscribe from this list: send the line "unsubscribe linux-security-module" in
the body of a message to majordomo at vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

[PATCH v3 2/5] efi: Add embedded peripheral firmware support

[Mimi Zohar]

On Tue, 2018-04-24 at 23:42 +0000, Luis R. Rodriguez wrote:
> On Tue, Apr 24, 2018 at 12:07:01PM -0400, Mimi Zohar wrote:
> > On Tue, 2018-04-24 at 17:09 +0200, Hans de Goede wrote:
> > > Hi,
> > > 
> > > On 23-04-18 23:11, Luis R. Rodriguez wrote:
> > > > Hans, please see use of READING_FIRMWARE_PREALLOC_BUFFER, we'll need a new ID
> > > > and security for this type of request so IMA can reject it if the policy is
> > > > configured for it.
> > > 
> > > Hmm, interesting, actually it seems like the whole existence
> > > of READING_FIRMWARE_PREALLOC_BUFFER is a mistake, 
> 
> request_firmware_into_buf() was merged without my own review, however,
> the ID thing did get review from Mimi:
> 
> https://patchwork.kernel.org/patch/9074611/
> 
> The ID is not for IMA alone, its for any LSM to decide what to do.
> Note Mimi asked for READING_FIRMWARE_DMA if such buffer was in DMA,
> otherise READING_FIRMWARE_PREALLOC_BUFFER was suggested.
> 
> > > the IMA
> > > framework really does not care if we are loading the firmware
> > > into memory allocated by the firmware-loader code, or into
> > > memory allocated by the device-driver requesting the firmware.
> 
> That's up to LSM folks to decide. We have these so far:
> 
> #define __kernel_read_file_id(id) \                                             
>         id(UNKNOWN, unknown)            \                                       
>         id(FIRMWARE, firmware)          \                                       
>         id(FIRMWARE_PREALLOC_BUFFER, firmware)  \                               
>         id(MODULE, kernel-module)               \                               
>         id(KEXEC_IMAGE, kexec-image)            \                               
>         id(KEXEC_INITRAMFS, kexec-initramfs)    \                               
>         id(POLICY, security-policy)             \                               
>         id(X509_CERTIFICATE, x509-certificate)  \                               
>         id(MAX_ID, )  
> 
> The first type of IDs added was about type of files the kernel
> LSMs may want to do different things for.
> 
> Mimi why did you want a separate ID for it back before?

The point of commit a098ecd2fa7d ("firmware: support loading into a
pre-allocated buffer") is to avoid reading the firmware into kernel
memory and then copying it "to it's final resting place". ?My concern
is that if the device driver has access to the buffer, it could access
the buffer prior to the firmware's signature having been verified by
the kernel.

In tightly controlled environments interested in limiting which signed
firmware version is loaded, require's the device driver not having
access to the buffer until after the signature has been verified by
the kernel (eg. IMA-appraisal).

> 
> I should note now that request_firmware_into_buf() and its
> READING_FIRMWARE_PREALLOC_BUFFER was to enable a driver on memory constrained
> devices. The files are large (commit says 16 MiB).
> 
> I've heard of larger possible files with remoteproc and with Android using
> the custom fallback mechanism -- which could mean a proprietary tool
> fetching firmware from a random special place on a device.
> 
> I could perhaps imagine an LSM which may be aware of such type of
> arrangement may want to do its own vetting of some sort, but this
> would not be specific to READING_FIRMWARE_PREALLOC_BUFFER, but rather
> the custom fallback mechaism.
> 
> Whether or not the buffer was preallocated by the driver seems a little
> odd for security folks to do something different with it. Security LSM
> folks please chime in.
> 
> I could see a bit more of a use case for an ID for firmware scraped
> from EFI, which Hans' patch will provide. But that *also* should get
> good review from other LSM folks.
> 
> One of the issues with accepting more IDs loosely is where do we
> stop though? If no one really is using READING_FIRMWARE_PREALLOC_BUFFER
> I'd say lets remove it. Likewise, for this EFI thing I'd like an idea
> if we really are going to have users for it.
> 
> If its of any help --
> 
> drivers/soc/qcom/mdt_loader.c is the only driver currently using
> request_firmware_into_buf() however I'll note qcom_mdt_load() is used in many
> other drivers so they are wrappers around request_firmware_into_buf():
> 
> drivers/gpu/drm/msm/adreno/a5xx_gpu.c:   * adreno_request_fw() handles this, but qcom_mdt_load() does
> drivers/gpu/drm/msm/adreno/a5xx_gpu.c:          ret = qcom_mdt_load(dev, fw, fwname, GPU_PAS_ID,
> drivers/gpu/drm/msm/adreno/a5xx_gpu.c:          ret = qcom_mdt_load(dev, fw, newname, GPU_PAS_ID,
> drivers/media/platform/qcom/venus/firmware.c:   ret = qcom_mdt_load(dev, mdt, fwname, VENUS_PAS_ID, mem_va, mem_phys,
> drivers/remoteproc/qcom_adsp_pil.c:     return qcom_mdt_load(adsp->dev, fw, rproc->firmware, adsp->pas_id,
> drivers/remoteproc/qcom_wcnss.c:        return qcom_mdt_load(wcnss->dev, fw, rproc->firmware, WCNSS_PAS_ID,
> 
> Are we going to add more IDs for more types of firmware?
> What type of *different* decisions could LSMs take if the firmware
> was being written to a buffer? Or in this new case that is coming
> up, if the file came scraping EFI, would having that information
> be useful?
> 
> > > As such the current IMA code (from v4.17-rc2) actually does
> > > not handle READING_FIRMWARE_PREALLOC_BUFFER at all, 
> > 
> > Right, it doesn't yet address READING_FIRMWARE_PREALLOC_BUFFER, but
> > should.
> > 
> > Depending on whether the device requesting the firmware has access to
> > the DMA memory, before the signature verification, 
> 
> It would seem from the original patch review about READING_FIRMWARE_PREALLOC_BUFFER
> that this is not a DMA buffer.

The call sequence:
qcom_mdt_load() ->?qcom_scm_pas_init_image() ->?dma_alloc_coherent()

If dma_alloc_coherent() isn't allocating a DMA buffer, then the
function name is misleading/confusing.

> 
> The device driver should have access to the buffer pointer with write given
> that with request_firmware_into_buf() the driver is giving full write access to
> the memory pointer so that the firmware API can stuff the firmware it finds
> there.
> 
> Firmware signature verification would be up to the device hardware to do upon
> load *after* request_firmware_into_buf().

We're discussing the kernel's signature verification, not the device
hardware's signature verification. ?Can the device driver access the
buffer, before IMA-appraisal has verified the firmware's signature?

Mimi

> 
>   Luis
> 
> > will determine how
> > IMA-appraisal addresses READING_FIRMWARE_PREALLOC_BUFFER.
> > 
> > Mimi
> > 
> > > here
> > > are bits of code from: security/integrity/ima/ima_main.c:
> > > 
> > > static int read_idmap[READING_MAX_ID] = {
> > >          [READING_FIRMWARE] = FIRMWARE_CHECK,
> > >          [READING_MODULE] = MODULE_CHECK,
> > >          [READING_KEXEC_IMAGE] = KEXEC_KERNEL_CHECK,
> > >          [READING_KEXEC_INITRAMFS] = KEXEC_INITRAMFS_CHECK,
> > >          [READING_POLICY] = POLICY_CHECK
> > > };
> > > 
> > > int ima_post_read_file(struct file *file, void *buf, loff_t size,
> > > 	...
> > >          if (!file && read_id == READING_FIRMWARE) {
> > >                  if ((ima_appraise & IMA_APPRAISE_FIRMWARE) &&
> > >                      (ima_appraise & IMA_APPRAISE_ENFORCE))
> > >                          return -EACCES; /* INTEGRITY_UNKNOWN */
> > >                  return 0;
> > >          }
> > > 
> > > Which show that the IMA code is not handling
> > > READING_FIRMWARE_PREALLOC_BUFFER as it should (I believe it
> > > should handle it the same as READING_FIRMWARE).
> > > 
> > > Now we could fix that, but the only user of
> > > READING_FIRMWARE_PREALLOC_BUFFER is the code which originally
> > > introduced it:
> > > 
> > > https://patchwork.kernel.org/patch/9162011/
> > > 
> > > So I believe it might be better to instead replace it
> > > with just READING_FIRMWARE and find another way to tell
> > > kernel_read_file() that there is a pre-allocated buffer,
> > > perhaps the easiest way there is that  *buf must be
> > > NULL when the caller wants kernel_read_file() to
> > > vmalloc the mem. This would of course require auditing
> > > all callers that the buf which the pass in is initialized
> > > to NULL.
> > > 
> > > Either way adding a third READING_FIRMWARE_FOO to the
> > > kernel_read_file_id enum seems like a bad idea, from
> > > the IMA pov firmware is firmware.
> > > 
> > > What this whole exercise has shown me though is that
> > > I need to call security_kernel_post_read_file() when
> > > loading EFI embedded firmware. I will add a call to
> > > security_kernel_post_read_file() for v4 of the patch-set.
> > > 
> > > > Please Cc Kees in future patches.
> > > 
> > > Will do.
> > > 
> > > Regards,
> > > 
> > > Hans
> > > 
> > 
> > 
> 

--
To unsubscribe from this list: send the line "unsubscribe linux-security-module" in
the body of a message to majordomo at vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

[PATCH v3 2/5] efi: Add embedded peripheral firmware support

[Luis R. Rodriguez]

On Wed, Apr 25, 2018 at 01:00:09AM -0400, Mimi Zohar wrote:
> On Tue, 2018-04-24 at 23:42 +0000, Luis R. Rodriguez wrote:
> > On Tue, Apr 24, 2018 at 12:07:01PM -0400, Mimi Zohar wrote:
> > > On Tue, 2018-04-24 at 17:09 +0200, Hans de Goede wrote:
> > > > On 23-04-18 23:11, Luis R. Rodriguez wrote:
> > > > > Hans, please see use of READING_FIRMWARE_PREALLOC_BUFFER, we'll need a new ID
> > > > > and security for this type of request so IMA can reject it if the policy is
> > > > > configured for it.
> > > > 
> > > > Hmm, interesting, actually it seems like the whole existence
> > > > of READING_FIRMWARE_PREALLOC_BUFFER is a mistake, 
> > 
> > request_firmware_into_buf() was merged without my own review, however,
> > the ID thing did get review from Mimi:
> > 
> > https://patchwork.kernel.org/patch/9074611/
> > 
> > The ID is not for IMA alone, its for any LSM to decide what to do.
> > Note Mimi asked for READING_FIRMWARE_DMA if such buffer was in DMA,
> > otherise READING_FIRMWARE_PREALLOC_BUFFER was suggested.
> > 
> > Mimi why did you want a separate ID for it back before?
> 
> The point of commit a098ecd2fa7d ("firmware: support loading into a
> pre-allocated buffer") is to avoid reading the firmware into kernel
> memory and then copying it "to it's final resting place". ?My concern
> is that if the device driver has access to the buffer, it could access
> the buffer prior to the firmware's signature having been verified by
> the kernel.

If request_firmware_into_buf() is used and the firmware was found in
/lib/firmware/ paths then the driver will *not* use the firmware prior
to any LSM doing any firmware signature verification because
kernel_read_file_from_path() and in turn security_kernel_read_file().

The firmware API has a fallback mechanism [0] though, and if that is used then
security_kernel_post_read_file() is used once the firmware is loaded through
the sysfs interface *prior* to handing the firmware data to the driver. As
Hans noted though security_kernel_post_read_file() currently *only* uses
READING_FIRMWARE, so this needs to be fixed. Also note though that LSMs
get a hint of what is going to happen *soon* prior to the fallback
mechanism kicking on as we travere the /lib/firmware/ paths for direct
filesystem loading.

If this is not sufficient to cover LSM appraisals *one* option could be to
have security_kernel_read_file() return a special error of some sort
for READING_FIRMWARE_PREALLOC_BUFFER so that kernel_read_file_from_path()
users could *know* to fatally give up.

Currently the device drivers using request_firmware_into_buf() can end up
getting the buffer with firmware stashed in it without having the kernel do any
firmware signature verification at all through its LSMs. The LSM hooks added to
the firmware loader long ago by Kees via commit 6593d9245bc66 ("firmware_class:
perform new LSM checks") on v3.17 added an LSM for direct filesystem lookups,
but on the fallback mechanism seems to have only added a post LSM hook
security_kernel_fw_from_file().

There is also a custom fallback mechanism [1] which can be used if the path to
the firmware may be out of the /lib/firmware/ paths or perhaps the firmware
requires some very custom fetching of some sort. The only thing this does
though is just *not* issue a uevent when we don't find the firmware and also
sets the timeout to a practically never-ending value. The custom fallback
mechanism is only usable for request_firmware_nowait() though. In retrospect
the custom fallback mechanism is pure crap and these days we've acknowledged
that even in crazy custom firmware fetching cases folks should be able to
accomplish this by relying on uevents and using the firmwared [2] or forking
it, or a different similar proprietary similar solution, which would just
monitor for uevents for firmware and just Do The Right Thing (TM).

Consider some mobile devices which may want to fetch it from some custom
partition which only it can know how to get.

There is a kernel config option which enables the fallback mechanism always,
This is now easily readable as follows:

drivers/base/firmware_loader/fallback_table.c

struct firmware_fallback_config fw_fallback_config = {
	.force_sysfs_fallback = IS_ENABLED(CONFIG_FW_LOADER_USER_HELPER_FALLBACK),
	.loading_timeout = 60,
	.old_timeout = 60,
};

Even if this is used we always do direct fs lookups first.

Android became the primary user of CONFIG_FW_LOADER_USER_HELPER_FALLBACK.

It would be good for us to hear from Android folks if their current use of
request_firmware_into_buf() is designed in practice to *never* use the direct
filesystem firmware loading interface, and always rely instead on the
fallback mechanism.

That would answer help your appraisal question in practice today.

[0] https://www.kernel.org/doc/html/latest/driver-api/firmware/fallback-mechanisms.html
[1] https://www.kernel.org/doc/html/latest/driver-api/firmware/fallback-mechanisms.html#firmware-custom-fallback-mechanism
[2] https://github.com/teg/firmwared

> In tightly controlled environments interested in limiting which signed
> firmware version is loaded, require's the device driver not having
> access to the buffer until after the signature has been verified by
> the kernel (eg. IMA-appraisal).

We may need more work for this for request_firmware_into_buf().

> > I should note now that request_firmware_into_buf() and its
> > READING_FIRMWARE_PREALLOC_BUFFER was to enable a driver on memory constrained
> > devices. The files are large (commit says 16 MiB).
> > 
> > I've heard of larger possible files with remoteproc and with Android using
> > the custom fallback mechanism -- which could mean a proprietary tool
> > fetching firmware from a random special place on a device.
> > 
> > I could perhaps imagine an LSM which may be aware of such type of
> > arrangement may want to do its own vetting of some sort, but this
> > would not be specific to READING_FIRMWARE_PREALLOC_BUFFER, but rather
> > the custom fallback mechaism.
> > 
> > Whether or not the buffer was preallocated by the driver seems a little
> > odd for security folks to do something different with it. Security LSM
> > folks please chime in.
> > 
> > I could see a bit more of a use case for an ID for firmware scraped
> > from EFI, which Hans' patch will provide. But that *also* should get
> > good review from other LSM folks.
> > 
> > One of the issues with accepting more IDs loosely is where do we
> > stop though? If no one really is using READING_FIRMWARE_PREALLOC_BUFFER
> > I'd say lets remove it. Likewise, for this EFI thing I'd like an idea
> > if we really are going to have users for it.
> > 
> > If its of any help --
> > 
> > drivers/soc/qcom/mdt_loader.c is the only driver currently using
> > request_firmware_into_buf() however I'll note qcom_mdt_load() is used in many
> > other drivers so they are wrappers around request_firmware_into_buf():
> > 
> > drivers/gpu/drm/msm/adreno/a5xx_gpu.c:   * adreno_request_fw() handles this, but qcom_mdt_load() does
> > drivers/gpu/drm/msm/adreno/a5xx_gpu.c:          ret = qcom_mdt_load(dev, fw, fwname, GPU_PAS_ID,
> > drivers/gpu/drm/msm/adreno/a5xx_gpu.c:          ret = qcom_mdt_load(dev, fw, newname, GPU_PAS_ID,
> > drivers/media/platform/qcom/venus/firmware.c:   ret = qcom_mdt_load(dev, mdt, fwname, VENUS_PAS_ID, mem_va, mem_phys,
> > drivers/remoteproc/qcom_adsp_pil.c:     return qcom_mdt_load(adsp->dev, fw, rproc->firmware, adsp->pas_id,
> > drivers/remoteproc/qcom_wcnss.c:        return qcom_mdt_load(wcnss->dev, fw, rproc->firmware, WCNSS_PAS_ID,
> > 
> > Are we going to add more IDs for more types of firmware?
> > What type of *different* decisions could LSMs take if the firmware
> > was being written to a buffer? Or in this new case that is coming
> > up, if the file came scraping EFI, would having that information
> > be useful?
> > 
> > > > As such the current IMA code (from v4.17-rc2) actually does
> > > > not handle READING_FIRMWARE_PREALLOC_BUFFER at all, 
> > > 
> > > Right, it doesn't yet address READING_FIRMWARE_PREALLOC_BUFFER, but
> > > should.
> > > 
> > > Depending on whether the device requesting the firmware has access to
> > > the DMA memory, before the signature verification, 
> > 
> > It would seem from the original patch review about READING_FIRMWARE_PREALLOC_BUFFER
> > that this is not a DMA buffer.
> 
> The call sequence:
> qcom_mdt_load() ->?qcom_scm_pas_init_image() ->?dma_alloc_coherent()
> 
> If dma_alloc_coherent() isn't allocating a DMA buffer, then the
> function name is misleading/confusing.

Hah, by *definition* the device *and* processor has immediate access
to data written *immediately* when dma_alloc_coherent() is used. From
Documentation/DMA-API.txt:

-----------------------------------------------------------------------
Part Ia - Using large DMA-coherent buffers                                      
------------------------------------------                                      
                                                                                
::                                                                              
                                                                                
        void *                                                                  
        dma_alloc_coherent(struct device *dev, size_t size,                     
                           dma_addr_t *dma_handle, gfp_t flag)                  
                                                                                
Consistent memory is memory for which a write by either the device or           
the processor can immediately be read by the processor or device                
without having to worry about caching effects.  (You may however need           
to make sure to flush the processor's write buffers before telling              
devices to read that memory.)  
------------------------------------------------------------------------

Is ptr below

	ret = request_firmware_into_buf(&seg_fw, fw_name, dev,  
					ptr, phdr->p_filesz); 

Also part of the DMA buffer allocated earlier via:

        ret = qcom_scm_pas_init_image(pas_id, fw->data, fw->size);              

Android folks?

> > The device driver should have access to the buffer pointer with write given
> > that with request_firmware_into_buf() the driver is giving full write access to
> > the memory pointer so that the firmware API can stuff the firmware it finds
> > there.
> > 
> > Firmware signature verification would be up to the device hardware to do upon
> > load *after* request_firmware_into_buf().
> 
> We're discussing the kernel's signature verification, not the device
> hardware's signature verification. ?Can the device driver access the
> buffer, before IMA-appraisal has verified the firmware's signature?

It will depend on the above question.

  Luis

> 
> Mimi
> 
> > 
> >   Luis
> > 
> > > will determine how
> > > IMA-appraisal addresses READING_FIRMWARE_PREALLOC_BUFFER.
> > > 
> > > Mimi
> > > 
> > > > here
> > > > are bits of code from: security/integrity/ima/ima_main.c:
> > > > 
> > > > static int read_idmap[READING_MAX_ID] = {
> > > >          [READING_FIRMWARE] = FIRMWARE_CHECK,
> > > >          [READING_MODULE] = MODULE_CHECK,
> > > >          [READING_KEXEC_IMAGE] = KEXEC_KERNEL_CHECK,
> > > >          [READING_KEXEC_INITRAMFS] = KEXEC_INITRAMFS_CHECK,
> > > >          [READING_POLICY] = POLICY_CHECK
> > > > };
> > > > 
> > > > int ima_post_read_file(struct file *file, void *buf, loff_t size,
> > > > 	...
> > > >          if (!file && read_id == READING_FIRMWARE) {
> > > >                  if ((ima_appraise & IMA_APPRAISE_FIRMWARE) &&
> > > >                      (ima_appraise & IMA_APPRAISE_ENFORCE))
> > > >                          return -EACCES; /* INTEGRITY_UNKNOWN */
> > > >                  return 0;
> > > >          }
> > > > 
> > > > Which show that the IMA code is not handling
> > > > READING_FIRMWARE_PREALLOC_BUFFER as it should (I believe it
> > > > should handle it the same as READING_FIRMWARE).
> > > > 
> > > > Now we could fix that, but the only user of
> > > > READING_FIRMWARE_PREALLOC_BUFFER is the code which originally
> > > > introduced it:
> > > > 
> > > > https://patchwork.kernel.org/patch/9162011/
> > > > 
> > > > So I believe it might be better to instead replace it
> > > > with just READING_FIRMWARE and find another way to tell
> > > > kernel_read_file() that there is a pre-allocated buffer,
> > > > perhaps the easiest way there is that  *buf must be
> > > > NULL when the caller wants kernel_read_file() to
> > > > vmalloc the mem. This would of course require auditing
> > > > all callers that the buf which the pass in is initialized
> > > > to NULL.
> > > > 
> > > > Either way adding a third READING_FIRMWARE_FOO to the
> > > > kernel_read_file_id enum seems like a bad idea, from
> > > > the IMA pov firmware is firmware.
> > > > 
> > > > What this whole exercise has shown me though is that
> > > > I need to call security_kernel_post_read_file() when
> > > > loading EFI embedded firmware. I will add a call to
> > > > security_kernel_post_read_file() for v4 of the patch-set.
> > > > 
> > > > > Please Cc Kees in future patches.
> > > > 
> > > > Will do.
> > > > 
> > > > Regards,
> > > > 
> > > > Hans
> > > > 
> > > 
> > > 
> > 
> 
> 

-- 
Do not panic
--
To unsubscribe from this list: send the line "unsubscribe linux-security-module" in
the body of a message to majordomo at vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

[PATCH v3 2/5] efi: Add embedded peripheral firmware support

[Luis R. Rodriguez]

Android folks, poke below. otherwise we'll have no option but to seriously
consider Mimi's patch to prevent these calls when IMA appraisal is enforced:

http://lkml.kernel.org/r/1525182503-13849-7-git-send-email-zohar at linux.vnet.ibm.com

Please read below....

On Wed, Apr 25, 2018 at 05:55:57PM +0000, Luis R. Rodriguez wrote:
> On Wed, Apr 25, 2018 at 01:00:09AM -0400, Mimi Zohar wrote:
> > On Tue, 2018-04-24 at 23:42 +0000, Luis R. Rodriguez wrote:
> > > On Tue, Apr 24, 2018 at 12:07:01PM -0400, Mimi Zohar wrote:
> > > > On Tue, 2018-04-24 at 17:09 +0200, Hans de Goede wrote:
> > > If its of any help --
> > > 
> > > drivers/soc/qcom/mdt_loader.c is the only driver currently using
> > > request_firmware_into_buf() however I'll note qcom_mdt_load() is used in many
> > > other drivers so they are wrappers around request_firmware_into_buf():
> > > 
> > > drivers/gpu/drm/msm/adreno/a5xx_gpu.c:   * adreno_request_fw() handles this, but qcom_mdt_load() does
> > > drivers/gpu/drm/msm/adreno/a5xx_gpu.c:          ret = qcom_mdt_load(dev, fw, fwname, GPU_PAS_ID,
> > > drivers/gpu/drm/msm/adreno/a5xx_gpu.c:          ret = qcom_mdt_load(dev, fw, newname, GPU_PAS_ID,
> > > drivers/media/platform/qcom/venus/firmware.c:   ret = qcom_mdt_load(dev, mdt, fwname, VENUS_PAS_ID, mem_va, mem_phys,
> > > drivers/remoteproc/qcom_adsp_pil.c:     return qcom_mdt_load(adsp->dev, fw, rproc->firmware, adsp->pas_id,
> > > drivers/remoteproc/qcom_wcnss.c:        return qcom_mdt_load(wcnss->dev, fw, rproc->firmware, WCNSS_PAS_ID,
> > > 
> > > > > As such the current IMA code (from v4.17-rc2) actually does
> > > > > not handle READING_FIRMWARE_PREALLOC_BUFFER at all, 
> > > > 
> > > > Right, it doesn't yet address READING_FIRMWARE_PREALLOC_BUFFER, but
> > > > should.
> > > > 
> > > > Depending on whether the device requesting the firmware has access to
> > > > the DMA memory, before the signature verification, 
> > > 
> > > It would seem from the original patch review about READING_FIRMWARE_PREALLOC_BUFFER
> > > that this is not a DMA buffer.

To be very clear I believe Stephen implied this was not DMA buffer. Mimi
asked for READING_FIRMWARE_DMA if it was:

https://patchwork.kernel.org/patch/9074611/

> > The call sequence:
> > qcom_mdt_load() ->?qcom_scm_pas_init_image() ->?dma_alloc_coherent()
> > 
> > If dma_alloc_coherent() isn't allocating a DMA buffer, then the
> > function name is misleading/confusing.
> 
> Hah, by *definition* the device *and* processor has immediate access
> to data written *immediately* when dma_alloc_coherent() is used. From
> Documentation/DMA-API.txt:
> 
> -----------------------------------------------------------------------
> Part Ia - Using large DMA-coherent buffers                                      
> ------------------------------------------                                      
>                                                                                 
> ::                                                                              
>                                                                                 
>         void *                                                                  
>         dma_alloc_coherent(struct device *dev, size_t size,                     
>                            dma_addr_t *dma_handle, gfp_t flag)                  
>                                                                                 
> Consistent memory is memory for which a write by either the device or           
> the processor can immediately be read by the processor or device                
> without having to worry about caching effects.  (You may however need           
> to make sure to flush the processor's write buffers before telling              
> devices to read that memory.)  
> ------------------------------------------------------------------------
> 
> Is ptr below
> 
> 	ret = request_firmware_into_buf(&seg_fw, fw_name, dev,  
> 					ptr, phdr->p_filesz); 
> 
> Also part of the DMA buffer allocated earlier via:
> 
>         ret = qcom_scm_pas_init_image(pas_id, fw->data, fw->size);              
> 
> Android folks?

Android folks?

> > > The device driver should have access to the buffer pointer with write given
> > > that with request_firmware_into_buf() the driver is giving full write access to
> > > the memory pointer so that the firmware API can stuff the firmware it finds
> > > there.
> > > 
> > > Firmware signature verification would be up to the device hardware to do upon
> > > load *after* request_firmware_into_buf().
> > 
> > We're discussing the kernel's signature verification, not the device
> > hardware's signature verification. ?Can the device driver access the
> > buffer, before IMA-appraisal has verified the firmware's signature?
> 
> It will depend on the above question.

  Luis
--
To unsubscribe from this list: send the line "unsubscribe linux-security-module" in
the body of a message to majordomo at vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

[PATCH v3 2/5] efi: Add embedded peripheral firmware support

[Martijn Coenen]

On Thu, May 3, 2018 at 5:21 PM, Luis R. Rodriguez <mcgrof@kernel.org> wrote:
> Android folks, poke below. otherwise we'll have no option but to seriously
> consider Mimi's patch to prevent these calls when IMA appraisal is enforced:

Sorry, figuring out who's the right person to answer this, will get
back to you ASAP.

Martijn

>
> http://lkml.kernel.org/r/1525182503-13849-7-git-send-email-zohar at linux.vnet.ibm.com
>
> Please read below....
>
> On Wed, Apr 25, 2018 at 05:55:57PM +0000, Luis R. Rodriguez wrote:
>> On Wed, Apr 25, 2018 at 01:00:09AM -0400, Mimi Zohar wrote:
>> > On Tue, 2018-04-24 at 23:42 +0000, Luis R. Rodriguez wrote:
>> > > On Tue, Apr 24, 2018 at 12:07:01PM -0400, Mimi Zohar wrote:
>> > > > On Tue, 2018-04-24 at 17:09 +0200, Hans de Goede wrote:
>> > > If its of any help --
>> > >
>> > > drivers/soc/qcom/mdt_loader.c is the only driver currently using
>> > > request_firmware_into_buf() however I'll note qcom_mdt_load() is used in many
>> > > other drivers so they are wrappers around request_firmware_into_buf():
>> > >
>> > > drivers/gpu/drm/msm/adreno/a5xx_gpu.c:   * adreno_request_fw() handles this, but qcom_mdt_load() does
>> > > drivers/gpu/drm/msm/adreno/a5xx_gpu.c:          ret = qcom_mdt_load(dev, fw, fwname, GPU_PAS_ID,
>> > > drivers/gpu/drm/msm/adreno/a5xx_gpu.c:          ret = qcom_mdt_load(dev, fw, newname, GPU_PAS_ID,
>> > > drivers/media/platform/qcom/venus/firmware.c:   ret = qcom_mdt_load(dev, mdt, fwname, VENUS_PAS_ID, mem_va, mem_phys,
>> > > drivers/remoteproc/qcom_adsp_pil.c:     return qcom_mdt_load(adsp->dev, fw, rproc->firmware, adsp->pas_id,
>> > > drivers/remoteproc/qcom_wcnss.c:        return qcom_mdt_load(wcnss->dev, fw, rproc->firmware, WCNSS_PAS_ID,
>> > >
>> > > > > As such the current IMA code (from v4.17-rc2) actually does
>> > > > > not handle READING_FIRMWARE_PREALLOC_BUFFER at all,
>> > > >
>> > > > Right, it doesn't yet address READING_FIRMWARE_PREALLOC_BUFFER, but
>> > > > should.
>> > > >
>> > > > Depending on whether the device requesting the firmware has access to
>> > > > the DMA memory, before the signature verification,
>> > >
>> > > It would seem from the original patch review about READING_FIRMWARE_PREALLOC_BUFFER
>> > > that this is not a DMA buffer.
>
> To be very clear I believe Stephen implied this was not DMA buffer. Mimi
> asked for READING_FIRMWARE_DMA if it was:
>
> https://patchwork.kernel.org/patch/9074611/
>
>> > The call sequence:
>> > qcom_mdt_load() -> qcom_scm_pas_init_image() -> dma_alloc_coherent()
>> >
>> > If dma_alloc_coherent() isn't allocating a DMA buffer, then the
>> > function name is misleading/confusing.
>>
>> Hah, by *definition* the device *and* processor has immediate access
>> to data written *immediately* when dma_alloc_coherent() is used. From
>> Documentation/DMA-API.txt:
>>
>> -----------------------------------------------------------------------
>> Part Ia - Using large DMA-coherent buffers
>> ------------------------------------------
>>
>> ::
>>
>>         void *
>>         dma_alloc_coherent(struct device *dev, size_t size,
>>                            dma_addr_t *dma_handle, gfp_t flag)
>>
>> Consistent memory is memory for which a write by either the device or
>> the processor can immediately be read by the processor or device
>> without having to worry about caching effects.  (You may however need
>> to make sure to flush the processor's write buffers before telling
>> devices to read that memory.)
>> ------------------------------------------------------------------------
>>
>> Is ptr below
>>
>>       ret = request_firmware_into_buf(&seg_fw, fw_name, dev,
>>                                       ptr, phdr->p_filesz);
>>
>> Also part of the DMA buffer allocated earlier via:
>>
>>         ret = qcom_scm_pas_init_image(pas_id, fw->data, fw->size);
>>
>> Android folks?
>
> Android folks?
>
>> > > The device driver should have access to the buffer pointer with write given
>> > > that with request_firmware_into_buf() the driver is giving full write access to
>> > > the memory pointer so that the firmware API can stuff the firmware it finds
>> > > there.
>> > >
>> > > Firmware signature verification would be up to the device hardware to do upon
>> > > load *after* request_firmware_into_buf().
>> >
>> > We're discussing the kernel's signature verification, not the device
>> > hardware's signature verification.  Can the device driver access the
>> > buffer, before IMA-appraisal has verified the firmware's signature?
>>
>> It will depend on the above question.
>
>   Luis
--
To unsubscribe from this list: send the line "unsubscribe linux-security-module" in
the body of a message to majordomo at vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

[PATCH v3 2/5] efi: Add embedded peripheral firmware support

[Martijn Coenen]

On Wed, Apr 25, 2018 at 10:55 AM, Luis R. Rodriguez <mcgrof@kernel.org> wrote:
> Android became the primary user of CONFIG_FW_LOADER_USER_HELPER_FALLBACK.
>
> It would be good for us to hear from Android folks if their current use of
> request_firmware_into_buf() is designed in practice to *never* use the direct
> filesystem firmware loading interface, and always rely instead on the
> fallback mechanism.

It's hard to answer this question for Android in general. As far as I
can tell the reasons we use CONFIG_FW_LOADER_USER_HELPER(_FALLBACK)
are:
1) We have multiple different paths on our devices where firmware can
be located, and the direct loader only supports one custom path
2) Most of those paths are not mounted by the time the corresponding
drivers are loaded, because pretty much all Android kernels today are
built without module support, and therefore drivers are loaded well
before the firmware partition is mounted
3) I think we use _FALLBACK because doing this with uevents is just
the easiest thing to do; our init code has a firmware helper that
deals with this and searches the paths that we care about

2) will change at some point, because Android is moving towards a
model where device-specific peripheral drivers will be loaded as
modules, and since those modules would likely come from the same
partition as the firmware, it's possible that the direct load would
succeed (depending on whether the custom path is configured there or
not). But I don't think we can rely on the direct loader even in those
cases, unless we could configure it with multiple custom paths.

I have no reason to believe request_firmware_into_buf() is special in
this regard; drivers that depend on it may have their corresponding
firmware in different locations, so just depending on the direct
loader would not be good enough.

>
> Is ptr below
>
>         ret = request_firmware_into_buf(&seg_fw, fw_name, dev,
>                                         ptr, phdr->p_filesz);
>
> Also part of the DMA buffer allocated earlier via:
>
>         ret = qcom_scm_pas_init_image(pas_id, fw->data, fw->size);
>
> Android folks?

I think the Qualcomm folks owning this (Andy, David, Bjorn, already
cc'd here) are better suited to answer that question.

Thanks,
Martijn
--
To unsubscribe from this list: send the line "unsubscribe linux-security-module" in
the body of a message to majordomo at vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

[PATCH v3 2/5] efi: Add embedded peripheral firmware support

[Luis R. Rodriguez]

On Fri, May 04, 2018 at 12:44:37PM -0700, Martijn Coenen wrote:
> On Wed, Apr 25, 2018 at 10:55 AM, Luis R. Rodriguez <mcgrof@kernel.org> wrote:
> > Android became the primary user of CONFIG_FW_LOADER_USER_HELPER_FALLBACK.
> >
> > It would be good for us to hear from Android folks if their current use of
> > request_firmware_into_buf() is designed in practice to *never* use the direct
> > filesystem firmware loading interface, and always rely instead on the
> > fallback mechanism.
> 
> It's hard to answer this question for Android in general. As far as I
> can tell the reasons we use CONFIG_FW_LOADER_USER_HELPER(_FALLBACK)
> are:
> 1) We have multiple different paths on our devices where firmware can
> be located, and the direct loader only supports one custom path
> 2) Most of those paths are not mounted by the time the corresponding
> drivers are loaded, because pretty much all Android kernels today are
> built without module support, and therefore drivers are loaded well
> before the firmware partition is mounted
> 3) I think we use _FALLBACK because doing this with uevents is just
> the easiest thing to do; our init code has a firmware helper that
> deals with this and searches the paths that we care about
> 
> 2) will change at some point, because Android is moving towards a
> model where device-specific peripheral drivers will be loaded as
> modules, and since those modules would likely come from the same
> partition as the firmware, it's possible that the direct load would
> succeed (depending on whether the custom path is configured there or
> not). But I don't think we can rely on the direct loader even in those
> cases, unless we could configure it with multiple custom paths.
> 
> I have no reason to believe request_firmware_into_buf() is special in
> this regard; drivers that depend on it may have their corresponding
> firmware in different locations, so just depending on the direct
> loader would not be good enough.

Thanks! This is very useful! This provides yet-another justification and use
case to document for the fallback mechanism. I'll go and extend it.

> >
> > Is ptr below
> >
> >         ret = request_firmware_into_buf(&seg_fw, fw_name, dev,
> >                                         ptr, phdr->p_filesz);
> >
> > Also part of the DMA buffer allocated earlier via:
> >
> >         ret = qcom_scm_pas_init_image(pas_id, fw->data, fw->size);
> >
> > Android folks?
> 
> I think the Qualcomm folks owning this (Andy, David, Bjorn, already
> cc'd here) are better suited to answer that question.

Andy, David, Bjorn?

  Luis
--
To unsubscribe from this list: send the line "unsubscribe linux-security-module" in
the body of a message to majordomo at vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

[PATCH v3 2/5] efi: Add embedded peripheral firmware support

[Luis R. Rodriguez]

On Tue, May 08, 2018 at 03:38:05PM +0000, Luis R. Rodriguez wrote:
> On Fri, May 04, 2018 at 12:44:37PM -0700, Martijn Coenen wrote:
> > On Wed, Apr 25, 2018 at 10:55 AM, Luis R. Rodriguez <mcgrof@kernel.org> wrote:
> > > Android became the primary user of CONFIG_FW_LOADER_USER_HELPER_FALLBACK.
> > >
> > > It would be good for us to hear from Android folks if their current use of
> > > request_firmware_into_buf() is designed in practice to *never* use the direct
> > > filesystem firmware loading interface, and always rely instead on the
> > > fallback mechanism.
> > 
> > It's hard to answer this question for Android in general. As far as I
> > can tell the reasons we use CONFIG_FW_LOADER_USER_HELPER(_FALLBACK)
> > are:
> > 1) We have multiple different paths on our devices where firmware can
> > be located, and the direct loader only supports one custom path

FWIW I'd love to consider patches to address this, if this is something
you may find a need for in the future to *avoid* the fallback, however
would like a clean solution.

> > 2) Most of those paths are not mounted by the time the corresponding
> > drivers are loaded, because pretty much all Android kernels today are
> > built without module support, and therefore drivers are loaded well
> > before the firmware partition is mounted

I've given this some more thought and you can address this with initramfs,
this is how other Linux distributions are addressing this. One way to
address this automatically is to scrape the drivers built-in or needed early on
boot in initamfs and if the driver has a MODULE_FIRMWARE() its respective
firmware is added to initramfs as well.

If you *don't* use initramfs, then yes you can obviously run into issues
where your firmware may not be accessible if the driver is somehow loaded
early.

> > 3) I think we use _FALLBACK because doing this with uevents is just
> > the easiest thing to do; our init code has a firmware helper that
> > deals with this and searches the paths that we care about
> > 
> > 2) will change at some point, because Android is moving towards a
> > model where device-specific peripheral drivers will be loaded as
> > modules, and since those modules would likely come from the same
> > partition as the firmware, it's possible that the direct load would
> > succeed (depending on whether the custom path is configured there or
> > not). But I don't think we can rely on the direct loader even in those
> > cases, unless we could configure it with multiple custom paths.

Using initramfs will help, but because of the custom path needs -- you're
right, we don't have anything for that yet, its also a bit unclear if
something nice and clean can be drawn up for it. So perhaps dealing with
the fallback mechanism is the way to go for this for sure, since we already
have support for it.

Just keep in mind that the fallback mechanism costs you about ~13436 bytes.

So, if someone comes up with a clean interface for custom paths I'd love
to consider it to avoid those 13436 bytes.

  Luis
--
To unsubscribe from this list: send the line "unsubscribe linux-security-module" in
the body of a message to majordomo at vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

[PATCH v3 2/5] efi: Add embedded peripheral firmware support

[Bjorn Andersson]

On Tue 08 May 09:10 PDT 2018, Luis R. Rodriguez wrote:

> On Tue, May 08, 2018 at 03:38:05PM +0000, Luis R. Rodriguez wrote:
> > On Fri, May 04, 2018 at 12:44:37PM -0700, Martijn Coenen wrote:
> > > On Wed, Apr 25, 2018 at 10:55 AM, Luis R. Rodriguez <mcgrof@kernel.org> wrote:
[..]
> > > 2) Most of those paths are not mounted by the time the corresponding
> > > drivers are loaded, because pretty much all Android kernels today are
> > > built without module support, and therefore drivers are loaded well
> > > before the firmware partition is mounted
> 
> I've given this some more thought and you can address this with initramfs,
> this is how other Linux distributions are addressing this. One way to
> address this automatically is to scrape the drivers built-in or needed early on
> boot in initamfs and if the driver has a MODULE_FIRMWARE() its respective
> firmware is added to initramfs as well.
> 

That could be done, but it would not change the fact that the
/sys/class/firmware is ABI and you may not break it.


And it doesn't change the fact that the ramdisk would have to be over
100mb to facilitate this.

> If you *don't* use initramfs, then yes you can obviously run into issues
> where your firmware may not be accessible if the driver is somehow loaded
> early.
> 

This is still a problem that lacks a solution.

> > > 3) I think we use _FALLBACK because doing this with uevents is just
> > > the easiest thing to do; our init code has a firmware helper that
> > > deals with this and searches the paths that we care about
> > > 
> > > 2) will change at some point, because Android is moving towards a
> > > model where device-specific peripheral drivers will be loaded as
> > > modules, and since those modules would likely come from the same
> > > partition as the firmware, it's possible that the direct load would
> > > succeed (depending on whether the custom path is configured there or
> > > not). But I don't think we can rely on the direct loader even in those
> > > cases, unless we could configure it with multiple custom paths.
> 
> Using initramfs will help, but because of the custom path needs -- you're
> right, we don't have anything for that yet, its also a bit unclear if
> something nice and clean can be drawn up for it. So perhaps dealing with
> the fallback mechanism is the way to go for this for sure, since we already
> have support for it.
> 
> Just keep in mind that the fallback mechanism costs you about ~13436 bytes.
> 

Remember that putting the firmware in the ramdisk would cost about
10000x (yes, ten thousand times) more ram.

> So, if someone comes up with a clean interface for custom paths I'd love
> to consider it to avoid those 13436 bytes.
> 

Combined with a way of synchronizing this with the availability of the
firmware, this would be a nice thing!

Regards,
Bjorn
--
To unsubscribe from this list: send the line "unsubscribe linux-security-module" in
the body of a message to majordomo at vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

[PATCH v3 2/5] efi: Add embedded peripheral firmware support

[Luis R. Rodriguez]

On Thu, Jun 07, 2018 at 09:49:50AM -0700, Bjorn Andersson wrote:
> On Tue 08 May 09:10 PDT 2018, Luis R. Rodriguez wrote:
> 
> > On Tue, May 08, 2018 at 03:38:05PM +0000, Luis R. Rodriguez wrote:
> > > On Fri, May 04, 2018 at 12:44:37PM -0700, Martijn Coenen wrote:
> > > > On Wed, Apr 25, 2018 at 10:55 AM, Luis R. Rodriguez <mcgrof@kernel.org> wrote:
> [..]
> > > > 2) Most of those paths are not mounted by the time the corresponding
> > > > drivers are loaded, because pretty much all Android kernels today are
> > > > built without module support, and therefore drivers are loaded well
> > > > before the firmware partition is mounted
> > 
> > I've given this some more thought and you can address this with initramfs,
> > this is how other Linux distributions are addressing this. One way to
> > address this automatically is to scrape the drivers built-in or needed early on
> > boot in initamfs and if the driver has a MODULE_FIRMWARE() its respective
> > firmware is added to initramfs as well.
> > 
> 
> That could be done, but it would not change the fact that the
> /sys/class/firmware is ABI and you may not break it.

Right, this is now well documented and also the latest changes to the firmware
API have made the sysfs fallback loader an option through a sysctl knob. The
code should be much easier to follow and test now.

> And it doesn't change the fact that the ramdisk would have to be over
> 100mb to facilitate this.

Indeed, this is now acknowledged in the latest Kconfig for the firmware loader.

> > If you *don't* use initramfs, then yes you can obviously run into issues
> > where your firmware may not be accessible if the driver is somehow loaded
> > early.
> > 
> 
> This is still a problem that lacks a solution.

The firmwared solution capturing uevents and using the sysfs fallback
mechanism should resolve this. Its also now properly documented on the
firmware loader Kconfig.

> > > > 3) I think we use _FALLBACK because doing this with uevents is just
> > > > the easiest thing to do; our init code has a firmware helper that
> > > > deals with this and searches the paths that we care about
> > > > 
> > > > 2) will change at some point, because Android is moving towards a
> > > > model where device-specific peripheral drivers will be loaded as
> > > > modules, and since those modules would likely come from the same
> > > > partition as the firmware, it's possible that the direct load would
> > > > succeed (depending on whether the custom path is configured there or
> > > > not). But I don't think we can rely on the direct loader even in those
> > > > cases, unless we could configure it with multiple custom paths.
> > 
> > Using initramfs will help, but because of the custom path needs -- you're
> > right, we don't have anything for that yet, its also a bit unclear if
> > something nice and clean can be drawn up for it. So perhaps dealing with
> > the fallback mechanism is the way to go for this for sure, since we already
> > have support for it.
> > 
> > Just keep in mind that the fallback mechanism costs you about ~13436 bytes.
> > 
> 
> Remember that putting the firmware in the ramdisk would cost about
> 10000x (yes, ten thousand times) more ram.

Indeed, this is now documented on the Kconfig too.

> > So, if someone comes up with a clean interface for custom paths I'd love
> > to consider it to avoid those 13436 bytes.
> > 
> 
> Combined with a way of synchronizing this with the availability of the
> firmware, this would be a nice thing!

The firmwared solution seems to be the way to go for now.

  Luis
--
To unsubscribe from this list: send the line "unsubscribe linux-security-module" in
the body of a message to majordomo at vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

[PATCH v3 2/5] efi: Add embedded peripheral firmware support

[Luis R. Rodriguez]

On Tue, May 08, 2018 at 03:38:05PM +0000, Luis R. Rodriguez wrote:
> On Fri, May 04, 2018 at 12:44:37PM -0700, Martijn Coenen wrote:
> > On Wed, Apr 25, 2018 at 10:55 AM, Luis R. Rodriguez <mcgrof@kernel.org> wrote:
> > > Is ptr below
> > >
> > >         ret = request_firmware_into_buf(&seg_fw, fw_name, dev,
> > >                                         ptr, phdr->p_filesz);
> > >
> > > Also part of the DMA buffer allocated earlier via:
> > >
> > >         ret = qcom_scm_pas_init_image(pas_id, fw->data, fw->size);
> > >
> > > Android folks?
> > 
> > I think the Qualcomm folks owning this (Andy, David, Bjorn, already
> > cc'd here) are better suited to answer that question.
> 
> Andy, David, Bjorn?

Andy, David, Bjorn?

Note: as-is we have no option but to assume this is DMA memory for now.
We cannot keep IMA's guarantees with the current prealloc firmware API
buffer, so I've suggested:

a) The prealloc buffer API be expanded to enable the caller to descrbe it
b) Have the qcom driver say this is DMA
c) IMA would reject it to ensure it stays true to what it needs to gaurantee

d) Future platforms which want to use IMA but want to trust DMA buffers
   would need to devise a way to describe IMA can trust some of these
   calls.

I'll leave it up to you guys (Andy, David, Bjorn) to come up with the code for
d) once and if you guys want to use IMA later.  But since what is pressing here
is to stay to true to IMA, with a-c IMA would reject such calls for now.

  Luis
--
To unsubscribe from this list: send the line "unsubscribe linux-security-module" in
the body of a message to majordomo at vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Do Qualcomm drivers use DMA buffers for request_firmware_into_buf()?

[Luis R. Rodriguez]

On Fri, Jun 01, 2018 at 09:23:46PM +0200, Luis R. Rodriguez wrote:
> On Tue, May 08, 2018 at 03:38:05PM +0000, Luis R. Rodriguez wrote:
> > On Fri, May 04, 2018 at 12:44:37PM -0700, Martijn Coenen wrote:
> > > 
> > > I think the Qualcomm folks owning this (Andy, David, Bjorn, already
> > > cc'd here) are better suited to answer that question.
> > 
> > Andy, David, Bjorn?
> 
> Andy, David, Bjorn?

A month now with no answer...

Perhaps someone who has this hardware can find out empirically for us, as
follows (mm folks is this right?):

page = virt_to_page(address);
if (!page)
   fail closed...
if (page_zone(page) == ZONE_DMA || page_zone(page) == ZONE_DMA32)
	this is a DMA buffer
else
	not DMA!

Note that when request_firmware_into_buf() was being reviewed Mimi had asked back
in 2016 [0] that if a DMA buffer was going to be used READING_FIRMWARE_DMA should be
used otherwise READING_FIRMWARE_PREALLOC_BUFFER was fine.

If it is a DMA buffer *now*, why / how did this change?

[0] https://patchwork.kernel.org/patch/9074611/

  Luis
--
To unsubscribe from this list: send the line "unsubscribe linux-security-module" in
the body of a message to majordomo at vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Do Qualcomm drivers use DMA buffers for request_firmware_into_buf()?

[Bjorn Andersson]

On Wed 06 Jun 13:32 PDT 2018, Luis R. Rodriguez wrote:

> On Fri, Jun 01, 2018 at 09:23:46PM +0200, Luis R. Rodriguez wrote:
> > On Tue, May 08, 2018 at 03:38:05PM +0000, Luis R. Rodriguez wrote:
> > > On Fri, May 04, 2018 at 12:44:37PM -0700, Martijn Coenen wrote:
> > > > 
> > > > I think the Qualcomm folks owning this (Andy, David, Bjorn, already
> > > > cc'd here) are better suited to answer that question.
> > > 
> > > Andy, David, Bjorn?
> > 
> > Andy, David, Bjorn?
> 
> A month now with no answer...
> 

The patch at the top of this thread doesn't interest me and you didn't
bother sending your question To me.

As a matter of fact I'm confused to what the actual question is.

> Perhaps someone who has this hardware can find out empirically for us, as
> follows (mm folks is this right?):
> 
> page = virt_to_page(address);
> if (!page)
>    fail closed...
> if (page_zone(page) == ZONE_DMA || page_zone(page) == ZONE_DMA32)
> 	this is a DMA buffer
> else
> 	not DMA!
> 

Where do you want to put this?

> Note that when request_firmware_into_buf() was being reviewed Mimi had asked back
> in 2016 [0] that if a DMA buffer was going to be used READING_FIRMWARE_DMA should be
> used otherwise READING_FIRMWARE_PREALLOC_BUFFER was fine.
> 
> If it is a DMA buffer *now*, why / how did this change?
> 

>From what I can see [0] says is to use READING_FIRMWARE_PREALLOC_BUFFER
regardless of where the memory comes from.

Regards,
Bjorn

> [0] https://patchwork.kernel.org/patch/9074611/
> 
>   Luis
--
To unsubscribe from this list: send the line "unsubscribe linux-security-module" in
the body of a message to majordomo at vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Do Qualcomm drivers use DMA buffers for request_firmware_into_buf()?

[Vlastimil Babka]

On 06/07/2018 06:33 PM, Greg Kroah-Hartman wrote:
> On Thu, Jun 07, 2018 at 06:23:01PM +0200, Ard Biesheuvel wrote:
>> On 7 June 2018 at 18:18, Bjorn Andersson <bjorn.andersson@linaro.org> wrote:
>>> On Wed 06 Jun 13:32 PDT 2018, Luis R. Rodriguez wrote:
>>>
>>>> On Fri, Jun 01, 2018 at 09:23:46PM +0200, Luis R. Rodriguez wrote:
>>>>> On Tue, May 08, 2018 at 03:38:05PM +0000, Luis R. Rodriguez wrote:
>>>>>> On Fri, May 04, 2018 at 12:44:37PM -0700, Martijn Coenen wrote:
>>>>>>>
>>>>>>> I think the Qualcomm folks owning this (Andy, David, Bjorn, already
>>>>>>> cc'd here) are better suited to answer that question.
>>>>>>
>>>>>> Andy, David, Bjorn?
>>>>>
>>>>> Andy, David, Bjorn?
>>>>
>>>> A month now with no answer...
>>>>
>>>
>>> The patch at the top of this thread doesn't interest me and you didn't
>>> bother sending your question To me.
>>>
>>> As a matter of fact I'm confused to what the actual question is.
>>>
>>
>> The actual question is whether it is really required that the firmware
>> is loaded by the kernel into a buffer that is already mapped for DMA
>> at that point, and thus accessible by the device.
>>
>> To me, it is not entirely clear what the nature is of the firmware
>> that we are talking about, since it seems to be getting passed to the
>> secure world as well?
>>
>> In any case, the preferred model in terms of validation/sig checking is
>>
>> 1) allocate a CPU accessible buffer
>>
>> 2) request the firmware into it (which may include a sig check under the hood)
>>
>> 3) map the buffer for DMA to the device so it can load the firmware.
>>
>> 4) kick off the DMA transfer.
>>
>> The use of dma_alloc_coherent() for this purpose seems unnecessary,
>> given that the DMA transfer is not bidirectional. Would it be possible
>> to replace it with something like the above sequence?
> 
> Why not just use kmalloc, it will always return a DMAable buffer.

DMAble in what sense? For devices that can't handle physical addresses
above 16M you need to pass __GFP_DMA to get those, from ZONE_DMA.
Otherwise it can return anything from lowmem. That's for x86_64, some
other arches have different DMA zone.

> Is the problem that vmalloc() might not?

vmalloc() could only be used as an alternative if you used kvmalloc(),
otherwise kmalloc() won't give you anything from vmalloc

> We need to drop the whole DMA zone crud, it confuses everyone who sees
> it and was primarily for really really old systems.

Yeah that would be nice.

> greg k-h
> 

--
To unsubscribe from this list: send the line "unsubscribe linux-security-module" in
the body of a message to majordomo at vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

[PATCH v3 2/5] efi: Add embedded peripheral firmware support

[Bjorn Andersson]

On Tue 24 Apr 22:00 PDT 2018, Mimi Zohar wrote:
> On Tue, 2018-04-24 at 23:42 +0000, Luis R. Rodriguez wrote:
> > On Tue, Apr 24, 2018 at 12:07:01PM -0400, Mimi Zohar wrote:
> > > On Tue, 2018-04-24 at 17:09 +0200, Hans de Goede wrote:
[..]
> > > > As such the current IMA code (from v4.17-rc2) actually does
> > > > not handle READING_FIRMWARE_PREALLOC_BUFFER at all, 
> > > 
> > > Right, it doesn't yet address READING_FIRMWARE_PREALLOC_BUFFER, but
> > > should.
> > > 
> > > Depending on whether the device requesting the firmware has access to
> > > the DMA memory, before the signature verification, 
> > 
> > It would seem from the original patch review about READING_FIRMWARE_PREALLOC_BUFFER
> > that this is not a DMA buffer.
> 
> The call sequence:
> qcom_mdt_load() ->?qcom_scm_pas_init_image() ->?dma_alloc_coherent()
> 

qcom_mdt_load() is passed a struct firmware object, which "data" is
passed into qcom_scm_pas_init_image(), which uses dma_alloc_coherent()
to allocate scratch memory to perform a call into secure world. So the
dma_alloc_coherent() here has nothing to do with firmware loading.

qcom_mdt_load() will then use request_firmware_into_buf() to load other
files into the passed void *mem_region, which either comes from a
memremap() or dma_alloc_coherent() call.

> If dma_alloc_coherent() isn't allocating a DMA buffer, then the
> function name is misleading/confusing.
> 

It does allocate memory.

> > 
> > The device driver should have access to the buffer pointer with write given
> > that with request_firmware_into_buf() the driver is giving full write access to
> > the memory pointer so that the firmware API can stuff the firmware it finds
> > there.
> > 
> > Firmware signature verification would be up to the device hardware to do upon
> > load *after* request_firmware_into_buf().
> 
> We're discussing the kernel's signature verification, not the device
> hardware's signature verification. ?Can the device driver access the
> buffer, before IMA-appraisal has verified the firmware's signature?
> 

I would expect that it's possible to read the content of the buffer from
a secondary execution context before the request_firmware_into_buf() has
verified the content of the buffer, but I would be considered a
seriously broken driver.

Regards,
Bjorn
--
To unsubscribe from this list: send the line "unsubscribe linux-security-module" in
the body of a message to majordomo at vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html