| Message ID | 20240503144920.61075-1-rf@opensource.cirrus.com |
|---|---|
| State | New |
| Headers | show |
| Series | ALSA: hda/cs_dsp_ctl: Actually remove ALSA controls | expand |
On Fri, 03 May 2024 16:49:20 +0200, Richard Fitzgerald wrote: > > hda_cs_dsp_control_remove() must remove the ALSA control when > deleting all the infrastructure for handling the control. > > Without this it is possible for ALSA controls to be left in > the Soundcard after the amp driver module has been unloaded. > So the get/set callbacks point to code that no longer exists. > > Signed-off-by: Richard Fitzgerald <rf@opensource.cirrus.com> > Fixes: 3233b978af23 ("ALSA: hda: hda_cs_dsp_ctl: Add Library to support CS_DSP ALSA controls") > --- > Note: it would be better to use the control private_free to do the > cleanup, and that is my plan long-term. But that is a larger change > to the code. > > I like to keep bugfix patches as simple as possible so they are > low-risk and easy to cherry-pick into older kernels. So this patch > fixes the bug. Sometime I will send a patch for future kernel > versions that reworks the cleanup to use private_free. I also like to keep as simple as possible :) One slight concern is whether cs_dsp kctls can be deleted at the snd_card removal (disconnect) before this function gets called. That is, snd_card_free() of the main card may delete all associated kctls, and may this function be called afterwards? If yes, this change would lead to a UAF. The structure is so complex and I can't follow immediately, unfortunately... thanks, Takashi > --- > sound/pci/hda/hda_cs_dsp_ctl.c | 4 ++++ > 1 file changed, 4 insertions(+) > > diff --git a/sound/pci/hda/hda_cs_dsp_ctl.c b/sound/pci/hda/hda_cs_dsp_ctl.c > index 463ca06036bf..a42653d3473d 100644 > --- a/sound/pci/hda/hda_cs_dsp_ctl.c > +++ b/sound/pci/hda/hda_cs_dsp_ctl.c > @@ -203,6 +203,10 @@ void hda_cs_dsp_control_remove(struct cs_dsp_coeff_ctl *cs_ctl) > { > struct hda_cs_dsp_coeff_ctl *ctl = cs_ctl->priv; > > + /* Only public firmware controls will have an associated kcontrol */ > + if (ctl && ctl->kctl) > + snd_ctl_remove(ctl->card, ctl->kctl); > + > kfree(ctl); > } > EXPORT_SYMBOL_NS_GPL(hda_cs_dsp_control_remove, SND_HDA_CS_DSP_CONTROLS); > -- > 2.39.2 >
On 03/05/2024 16:17, Takashi Iwai wrote: > On Fri, 03 May 2024 16:49:20 +0200, > Richard Fitzgerald wrote: >> >> hda_cs_dsp_control_remove() must remove the ALSA control when >> deleting all the infrastructure for handling the control. >> >> Without this it is possible for ALSA controls to be left in >> the Soundcard after the amp driver module has been unloaded. >> So the get/set callbacks point to code that no longer exists. >> >> Signed-off-by: Richard Fitzgerald <rf@opensource.cirrus.com> >> Fixes: 3233b978af23 ("ALSA: hda: hda_cs_dsp_ctl: Add Library to support CS_DSP ALSA controls") >> --- >> Note: it would be better to use the control private_free to do the >> cleanup, and that is my plan long-term. But that is a larger change >> to the code. >> >> I like to keep bugfix patches as simple as possible so they are >> low-risk and easy to cherry-pick into older kernels. So this patch >> fixes the bug. Sometime I will send a patch for future kernel >> versions that reworks the cleanup to use private_free. > > I also like to keep as simple as possible :) > > One slight concern is whether cs_dsp kctls can be deleted at the > snd_card removal (disconnect) before this function gets called. > That is, snd_card_free() of the main card may delete all associated > kctls, and may this function be called afterwards? > If yes, this change would lead to a UAF. > That's a good question. This is is safe for the cs35l56 driver because if the soundcard (or HDA codec driver) is removed, the HDA codec will destroy the component binding in its HDA_FIXUP_ACT_FREE. This will cause an unbind() call to the amp driver, which will (indirectly) call this function to remove all the controls. So they will have been removed before the soundcard is cleaned up. But it turns out that the cs35l41 driver doesn't clean up the cs_dsp instance in its unbind() call so the controls _won't_ be cleaned up and a double-free is possible. The firmware handling in the cs35l41 driver is strange and confusing so I'm not sure whether this is a bug or something necessary. > The structure is so complex and I can't follow immediately, > unfortunately... > Yes, I know. When this hda_cs_dsp_ctl code was first submitted to the kernel you suggested using private_free instead of this manual cleanup but for some reason that wasn't implemented. > > thanks, > > Takashi > >> --- >> sound/pci/hda/hda_cs_dsp_ctl.c | 4 ++++ >> 1 file changed, 4 insertions(+) >> >> diff --git a/sound/pci/hda/hda_cs_dsp_ctl.c b/sound/pci/hda/hda_cs_dsp_ctl.c >> index 463ca06036bf..a42653d3473d 100644 >> --- a/sound/pci/hda/hda_cs_dsp_ctl.c >> +++ b/sound/pci/hda/hda_cs_dsp_ctl.c >> @@ -203,6 +203,10 @@ void hda_cs_dsp_control_remove(struct cs_dsp_coeff_ctl *cs_ctl) >> { >> struct hda_cs_dsp_coeff_ctl *ctl = cs_ctl->priv; >> >> + /* Only public firmware controls will have an associated kcontrol */ >> + if (ctl && ctl->kctl) >> + snd_ctl_remove(ctl->card, ctl->kctl); >> + >> kfree(ctl); >> } >> EXPORT_SYMBOL_NS_GPL(hda_cs_dsp_control_remove, SND_HDA_CS_DSP_CONTROLS); >> -- >> 2.39.2 >>
On Fri, 03 May 2024 17:31:15 +0200, Richard Fitzgerald wrote: > > On 03/05/2024 16:17, Takashi Iwai wrote: > > On Fri, 03 May 2024 16:49:20 +0200, > > Richard Fitzgerald wrote: > >> > >> hda_cs_dsp_control_remove() must remove the ALSA control when > >> deleting all the infrastructure for handling the control. > >> > >> Without this it is possible for ALSA controls to be left in > >> the Soundcard after the amp driver module has been unloaded. > >> So the get/set callbacks point to code that no longer exists. > >> > >> Signed-off-by: Richard Fitzgerald <rf@opensource.cirrus.com> > >> Fixes: 3233b978af23 ("ALSA: hda: hda_cs_dsp_ctl: Add Library to support CS_DSP ALSA controls") > >> --- > >> Note: it would be better to use the control private_free to do the > >> cleanup, and that is my plan long-term. But that is a larger change > >> to the code. > >> > >> I like to keep bugfix patches as simple as possible so they are > >> low-risk and easy to cherry-pick into older kernels. So this patch > >> fixes the bug. Sometime I will send a patch for future kernel > >> versions that reworks the cleanup to use private_free. > > > > I also like to keep as simple as possible :) > > > > One slight concern is whether cs_dsp kctls can be deleted at the > > snd_card removal (disconnect) before this function gets called. > > That is, snd_card_free() of the main card may delete all associated > > kctls, and may this function be called afterwards? > > If yes, this change would lead to a UAF. > > > > That's a good question. This is is safe for the cs35l56 driver because > if the soundcard (or HDA codec driver) is removed, the HDA codec will > destroy the component binding in its HDA_FIXUP_ACT_FREE. This will cause > an unbind() call to the amp driver, which will (indirectly) call this > function to remove all the controls. So they will have been removed > before the soundcard is cleaned up. > > But it turns out that the cs35l41 driver doesn't clean up the cs_dsp > instance in its unbind() call so the controls _won't_ be cleaned up > and a double-free is possible. The firmware handling in the cs35l41 > driver is strange and confusing so I'm not sure whether this is a bug > or something necessary. OK, then setting kctl->private_free additionally like below could work around it, I suppose. Takashi --- a/sound/pci/hda/hda_cs_dsp_ctl.c +++ b/sound/pci/hda/hda_cs_dsp_ctl.c @@ -97,6 +97,14 @@ static unsigned int wmfw_convert_flags(unsigned int in) return out; } +static void hda_cs_dsp_coeff_free(struct snd_kcontrol *kctl) +{ + struct hda_cs_dsp_coeff_ctl *ctl = snd_kcontrol_chip(kctl); + + if (ctl) + ctl->kctl = NULL; +} + static void hda_cs_dsp_add_kcontrol(struct hda_cs_dsp_coeff_ctl *ctl, const char *name) { struct cs_dsp_coeff_ctl *cs_ctl = ctl->cs_ctl; @@ -130,6 +138,7 @@ static void hda_cs_dsp_add_kcontrol(struct hda_cs_dsp_coeff_ctl *ctl, const char } dev_dbg(cs_ctl->dsp->dev, "Added KControl: %s\n", kcontrol.name); + kctl->private_free = hda_cs_dsp_coeff_free; ctl->kctl = kctl; }
diff --git a/sound/pci/hda/hda_cs_dsp_ctl.c b/sound/pci/hda/hda_cs_dsp_ctl.c index 463ca06036bf..a42653d3473d 100644 --- a/sound/pci/hda/hda_cs_dsp_ctl.c +++ b/sound/pci/hda/hda_cs_dsp_ctl.c @@ -203,6 +203,10 @@ void hda_cs_dsp_control_remove(struct cs_dsp_coeff_ctl *cs_ctl) { struct hda_cs_dsp_coeff_ctl *ctl = cs_ctl->priv; + /* Only public firmware controls will have an associated kcontrol */ + if (ctl && ctl->kctl) + snd_ctl_remove(ctl->card, ctl->kctl); + kfree(ctl); } EXPORT_SYMBOL_NS_GPL(hda_cs_dsp_control_remove, SND_HDA_CS_DSP_CONTROLS);
hda_cs_dsp_control_remove() must remove the ALSA control when deleting all the infrastructure for handling the control. Without this it is possible for ALSA controls to be left in the Soundcard after the amp driver module has been unloaded. So the get/set callbacks point to code that no longer exists. Signed-off-by: Richard Fitzgerald <rf@opensource.cirrus.com> Fixes: 3233b978af23 ("ALSA: hda: hda_cs_dsp_ctl: Add Library to support CS_DSP ALSA controls") --- Note: it would be better to use the control private_free to do the cleanup, and that is my plan long-term. But that is a larger change to the code. I like to keep bugfix patches as simple as possible so they are low-risk and easy to cherry-pick into older kernels. So this patch fixes the bug. Sometime I will send a patch for future kernel versions that reworks the cleanup to use private_free. --- sound/pci/hda/hda_cs_dsp_ctl.c | 4 ++++ 1 file changed, 4 insertions(+)