Bug 211117

Do you have the issue if you play without alsa-lib plugins, i.e. with aplay -Dhw:0 or such?  And is this bug specific to the oxygen driver, or it happens on other devices like USB-audio?  I'm wondering which part is problematic with SEV.

Direct play to "hw:D2X,0" works (the copy was just to try to capture the result but the result is the same without it).
-> Meaning it's probably a plugin issue.


And as explained, the exact same configuration on an integrated Realtek ALC 1200A (hda-intel driver) works perfectly fine.
-> Meaning it's probably a driver issue.

So honestly, I'm quite at loss here. :-/

Hm, that's quite puzzling.

Does the behavior change if you call dma_set_mask_and_coherent() with DMA_BIT_MASK(32) explicitly in the oxygen driver?

I'm willing to test but I have no idea how/where to do that (developer but not a kernel driver dev), could you explain?

It's something like:
--- a/sound/pci/oxygen/oxygen_lib.c
+++ b/sound/pci/oxygen/oxygen_lib.c
@@ -650,6 +650,8 @@ int oxygen_pci_probe(struct pci_dev *pci, int index, char *id,
 	pci_set_master(pci);
 	card->private_free = oxygen_card_free;
 
+	dma_set_mask_and_coherent(&pci->dev, DMA_BIT_MASK(32));
+
 	configure_pcie_bridge(pci);
 	oxygen_init(chip);
 	chip->model.init(chip);


I don't expect much by this change, so just to be sure...

Other than that, check the contents of /proc/asound/card*/pcm0p/sub0/hw_params while playing back different cases (with working device, with non-working device, with plugins, etc).  Especially check if it's running in the mmap access mode or not.  e.g. you can run aplay with -M option to enforce the mmap mode for -Dhw:*, too.

Sorry for the delay, here are the test results:

With or without the patch, always the same result (what was working is still working and vice-versa).

Now, regarding the content of /proc/asound/card*/pcm0p/sub0/hw_params

Note: card0 = D2X/virtuoso and card1 = Generic/hda-intel.

The content of the settings for any given test is entirely identical with or without the mem_encrypt=off boot option.

The content of /proc/asound/card0/pcm0p/sub0/hw_params for a test with hw:0,0 is very similar to the content of /proc/asound/card1/pcm0p/sub0/hw_params for the same test with hw:1,0:
- period_size and buffer_size vary but buffer_size/period_size = 4.0 in all cases.
- everything else is identical, beginning with access.

Probably no surprise here, but tested anyway.

Regarding the tests with speaker-test:
- routed_* shows MMAP_INTERLEAVED
- copied_* shows RW_INTERLEAVED (as well as hw:*,0)

Surely enough, the tests with aplay are consistent with the request:
- aplay -D"hw:$card,0" shows RW_INTERLEAVED and works
- aplay -M -D"hw:$card,0" shows MMAP_INTERLEAVED and outputs white noise on D2X

Additional tests with the default system configuration alsa shows MMAP_INTERLEAVED which is once again consistent.

Thanks.  It's clearer now that it's the mmap that causes the problem.

Could you give alsa-info.sh output for further verification?  Run the script with --no-upload option and attach the output to Bugzilla.

And, now let's try other way round; modify HD-audio driver and see whether the issue happens with the mmap mode.  The patch below enforces the 32bit DMA mask and the same buffer allocation type without S/G for HD-audio:

--- a/sound/pci/hda/hda_controller.c
+++ b/sound/pci/hda/hda_controller.c
@@ -720,7 +720,7 @@ int snd_hda_attach_pcm_stream(struct hda_bus *_bus, struct hda_codec *codec,
        int pcm_dev = cpcm->device;
        unsigned int size;
        int s, err;
-       int type = SNDRV_DMA_TYPE_DEV_SG;
+       int type = SNDRV_DMA_TYPE_DEV;
 
        list_for_each_entry(apcm, &chip->pcm_list, list) {
                if (apcm->pcm->device == pcm_dev) {
@@ -759,8 +759,6 @@ int snd_hda_attach_pcm_stream(struct hda_bus *_bus, struct hda_codec *codec,
        size = CONFIG_SND_HDA_PREALLOC_SIZE * 1024;
        if (size > MAX_PREALLOC_SIZE)
                size = MAX_PREALLOC_SIZE;
-       if (chip->uc_buffer)
-               type = SNDRV_DMA_TYPE_DEV_UC_SG;
        snd_pcm_set_managed_buffer_all(pcm, type, chip->card->dev,
                                       size, MAX_PREALLOC_SIZE);
        return 0;
diff --git a/sound/pci/hda/hda_intel.c b/sound/pci/hda/hda_intel.c
index 8d568277088a..bcf3b0cadf5c 100644
--- a/sound/pci/hda/hda_intel.c
+++ b/sound/pci/hda/hda_intel.c
@@ -1967,6 +1967,7 @@ static int azx_first_init(struct azx *chip)
        /* allow 64bit DMA address if supported by H/W */
        if (!(gcap & AZX_GCAP_64OK))
                dma_bits = 32;
+       dma_bits = 32;
        if (!dma_set_mask(&pci->dev, DMA_BIT_MASK(dma_bits))) {
                dma_set_coherent_mask(&pci->dev, DMA_BIT_MASK(dma_bits));
        } else {

Created attachment 294699 [details]
Output of alsa-info.sh

Good news... well, sorta: the sound is now equally horrible on hda-intel with mmap activated.

Note: I had a few error messages on alsa-info.sh execution:
cat: /proc/asound/modules: No such file or directory
cat: /sys/class/sound/hwC1D0/user_pin_configs: No such file or directory
cat: /sys/class/sound/hwC1D0/init_verbs: No such file or directory
cat: /sys/class/sound/hwC1D0/hints: No such file or directory
cat: /sys/class/sound/hwC2D0/user_pin_configs: No such file or directory
cat: /sys/class/sound/hwC2D0/init_verbs: No such file or directory
cat: /sys/class/sound/hwC2D0/hints: No such file or directory

One thing I forgot to mention is that the virtuoso driver is built in-kernel while the hda-intel is built as a module because at first, I didn't intend to build the hda-intel at all since I did not think I would use the embedded chip.

OK, then could you check which change causes the problem on snd-hda-intel: the change in hda_intel.c (setting dma_mask) or in hda-controller.c (setting SNDRV_DMA_TYPE_*)?

I can.

- With only the hda_intel.c (dma_bits = 32) part of the patch, I hear music.

- With only the hda_controller.c (type = SNDRV_DMA_TYPE_DEV) part of the patch, I hear white noise.

OK, thanks.  Then, could you try the change below together with SNDRV_DMA_TYPE_DEV change?  Does it work now?

--- a/sound/core/pcm_native.c
+++ b/sound/core/pcm_native.c
@@ -3713,7 +3713,7 @@ int snd_pcm_lib_default_mmap(struct snd_pcm_substream *substream,
 				area->vm_end - area->vm_start, area->vm_page_prot);
 	}
 #endif /* CONFIG_GENERIC_ALLOCATOR */
-	if (IS_ENABLED(CONFIG_HAS_DMA) && !substream->ops->page &&
+	if (0 && IS_ENABLED(CONFIG_HAS_DMA) && !substream->ops->page &&
 	    (substream->dma_buffer.dev.type == SNDRV_DMA_TYPE_DEV ||
 	     substream->dma_buffer.dev.type == SNDRV_DMA_TYPE_DEV_UC))
 		return dma_mmap_coherent(substream->dma_buffer.dev.dev,

With hda_controller.c AND pcm_native.c patches, the Generic/hda-intel device is working as expected.

pcm_native.c being obviously a common file, I also tested the D2X/virtuoso device and it works too.

So, the test result implies that dma_mmap_coherent() is broken with the memory encryption while the mmap with the page fault still works.

I'll trust you on that. :-)

In any case, I stay available for further tests if needed.

It looks like an issue which is not related directly to audio, right?

Reassigning to Platform_X86 .

Yes, this must be a generic issue.  I'll ask Christoph at first.

This sounds like a mismatch between the encryption bit in the kernel and the encryption bit in userspace. It looks like that should be taken care of by the dma_pgprot() call in dma_mmap_attrs() or in iommu_dma_mmap(). But maybe the force_dma_unencrypted() in arch/x86/mm/mem_encrypt.c needs to understand if the IOMMU is doing the mapping. Since, even if the device doesn't support 48-bit or higher DMA, it will still done encrypted because of the IOMMU. I don't see any dmesg output, is the IOMMU enabled? What happens if you do iommu=pt on the kernel command line?

Alternatively, if you want memory encryption for your bare-metal system, you can see if the BIOS supports TSME (Transparent SME). Then you can remove mem_encrypt=on or add mem_encrypt=off (if enabled by default), and still get memory encryption.

Tested with iommu=pt (and no mem_encrypt option, so "on" by default): still white noise.

Regarding the encryption, I don't really have a _need_ for that. If I use it, it's in part because Virtualbox seems to require it in some use cases (not even entirely sure about that) and mostly because it's activated by default (and I like default settings to work correctly).
I checked the BIOS anyway but did not find any TSME or similar option (thanks for the information, though).

Right now, I'm working with mem_encrypt=off and I can do everything I need for now (no vbox for the moment).
I will continue to follow this bug because I can reproduce the problem easily and do tests if required.