On Wed, Nov 03, 2021 at 04:18:38PM -0700, Gaurav Kashyap wrote: > > However, when keys are hardware wrapped, it can be only unwrapped > by Qualcomm Trustzone. Qualcomm Trustzone is software. There is a mode where it passes the key to the actual HWKM hardware as intended, right? > +/** > + * qcom_scm_derive_sw_secret() - Derive SW secret from wrapped encryption key > + * @wrapped_key: the wrapped key used for inline encryption > + * @wrapped_key_size: size of the wrapped key > + * @sw_secret: the secret to be derived > + * @secret_size: size of the secret derived Please make the semantics of the @secret_size parameter clear. Will the output be at least @secret_size, exactly @secret_size, or at most @secret_size? > + * > + * Derive a SW secret to be used for inline encryption using Qualcomm ICE. > + * > + * Generally, for non-wrapped keys, fscrypt can derive a sw secret from the > + * key in the clear in the linux keyring. > + * > + * However, for wrapped keys, the key needs to be unwrapped, which can be done > + * only by the secure EE. So, it makes sense for the secure EE to derive the > + * sw secret and return to the kernel when wrapped keys are used. It's sort of a layering violation to mention fscrypt here, as this is low-level driver code. fscrypt is just an example user. I recommend documenting this in more general terms, and maybe referring to the "Hardware-wrapped keys" section of Documentation/block/inline-encryption.rst (which my patchset adds) as that is intended to explain derive_sw_secret already. > +int qcom_scm_derive_sw_secret(const u8* wrapped_key, u32 wrapped_key_size, > + u8 *sw_secret, u32 secret_size) > +{ > + struct qcom_scm_desc desc = { > + .svc = QCOM_SCM_SVC_ES, > + .cmd = QCOM_SCM_ES_DERIVE_RAW_SECRET, > + .arginfo = QCOM_SCM_ARGS(4, QCOM_SCM_RW, wrapped_key is const. Should it use QCOM_SCM_RO instead of QCOM_SCM_RW? > + keybuf = dma_alloc_coherent(__scm->dev, wrapped_key_size, &key_phys, > + GFP_KERNEL); > + if (!keybuf) > + return -ENOMEM; > + secretbuf = dma_alloc_coherent(__scm->dev, secret_size, &secret_phys, > + GFP_KERNEL); > + if (!secretbuf) > + return -ENOMEM; In the '!secretbuf' case, this leaks 'keybuf'. Also, my understanding is that the use of dma_alloc_coherent() here is a bit unusual. It would be helpful to leave a comment like: /* * Like qcom_scm_ice_set_key(), we use dma_alloc_coherent() to properly * get a physical address, while guaranteeing that we can zeroize the * key material later using memzero_explicit(). */ > + ret = qcom_scm_call(__scm->dev, &desc, NULL); > + memcpy(sw_secret, secretbuf, secret_size); This is copying out the data even if the SCM call failed. > diff --git a/drivers/firmware/qcom_scm.h b/drivers/firmware/qcom_scm.h > index d92156ceb3ac..de5d4f8fd20d 100644 > --- a/drivers/firmware/qcom_scm.h > +++ b/drivers/firmware/qcom_scm.h > @@ -110,6 +110,7 @@ extern int scm_legacy_call(struct device *dev, const struct qcom_scm_desc *desc, > #define QCOM_SCM_SVC_ES 0x10 /* Enterprise Security */ > #define QCOM_SCM_ES_INVALIDATE_ICE_KEY 0x03 > #define QCOM_SCM_ES_CONFIG_SET_ICE_KEY 0x04 > +#define QCOM_SCM_ES_DERIVE_RAW_SECRET 0x07 Can this be renamed to DERIVE_SW_SECRET? If not, then you probably should call the function qcom_scm_derive_raw_secret() instead of qcom_scm_derive_sw_secret(), since the functions in qcom_scm.c are intended to be thin wrappers around the SCM calls. The naming difference can be dealt with at a higher level. - Eric