On 07/01/22 14:33, Richard W.M. Jones wrote:
On Fri, Jul 01, 2022 at 02:31:32PM +0200, Laszlo Ersek wrote:
On 07/01/22 12:25, Richard W.M. Jones wrote:
On Fri, Jul 01, 2022 at 12:08:51PM +0200, Laszlo Ersek wrote:
On 07/01/22 11:56, Laszlo Ersek wrote:
This code runs (in virt-v2v, for example) on the following call path:
main [v2v/v2v.ml] create_standard_options [common/mltools/tools_utils.ml] Getopt.parse parse_key_selector [common/mltools/tools_utils.ml] List.push_back
(* at this point, cmdline_options.ks * has been populated *)
convert [convert/convert.ml] (* skipped in virt-v2v, because it * enables networking anyway: *) key_store_requires_network [common/mltools/tools_utils.ml]
g#set_network g#launch
inspect_decrypt [common/mltools/tools_utils.ml] c_inspect_decrypt [common/mltools/tools_utils.ml] guestfs_int_mllib_inspect_decrypt [common/mltools/tools_utils-c.c]
loop over cmdline_options.ks: key_store_import_key [common/options/keys.c]
inspect_do_decrypt [common/options/decrypt.c] decrypt_mountables [common/options/decrypt.c] MACRO CHECK HERE guestfs_clevis_luks_unlock free_key_store [common/options/keys.c]
This is why I said "it's too late to report this error in decrypt_mountables()"!
In comparison, here's how the C tools work (such as virt-cat):
main [cat/cat.c] OPTION_key [common/options/options.h] key_store_add_from_selector [common/options/keys.c] key_store_import_key [common/options/keys.c]
/* "ks" has been populated */
key_store_requires_network [common/options/keys.c] guestfs_set_network guestfs_launch inspect_mount [common/options/options.h] inspect_mount_handle [common/options/inspect.c] inspect_do_decrypt [common/options/decrypt.c] decrypt_mountables [common/options/decrypt.c] MACRO CHECK HERE guestfs_clevis_luks_unlock free_key_store [common/options/keys.c]
Note the following important facts:
(1) The macro check is reached, in both kinds of tools, way after the options have been parsed, and quite a bit after the appliance has been launched.
(2) The OCaml and C *option parsers* share *nothing* related to keys. Consider:
- OCaml does not call key_store_add_from_selector() at all,
- while OCaml does call key_store_import_key(), it does so *outside* of option parsing. Namely, "parse_key_selector" records the command line options in an OCaml-only data structure (cmdline_options.ks). That is then translated to an *ephemeral* C-language keystore much later, with key_store_import_key() -- only in guestfs_int_mllib_inspect_decrypt()! That is, after the appliance is running.
- Correspondingly: in the C tools, the C-language keystore is freed *after* inspection completes (that is, free_key_store() is called *after* inspect_mount() returns), but in OCaml, the ephemeral keystore is released (correctly) *inside* guestfs_int_mllib_inspect_decrypt() -- that is, internally to the outer function "inspect_decrypt".
If we want to catch the lack of GUESTFS_HAVE_CLEVIS_LUKS_UNLOCK during option parsing, then I need to implement a brand new C function, and wrap it for OCaml too (OCaml cannot check the C-language macro GUESTFS_HAVE_CLEVIS_LUKS_UNLOCK).
For this: because we already have "key_store_requires_network" (same name, but totally separate impl in C and OCaml), I'd have to rename that function (both impls) to "key_store_contains_clevis", and use it for two purposes:
- right after option parsing, if the key store contains clevis, but the feature test macro is missing, bail.
- just before launching the appliance, if the key store contains clevis, enable networking.
There could be an easier solution for this BTW. Namely:
- For the C language tools, extend key_store_add_from_selector() with the GUESTFS_HAVE_CLEVIS_LUKS_UNLOCK macro check. (Review the call tree to see what that's right.) This would be a small modification to
OK, I see, this sounds like a better / earlier check (don't even add a clevis key if we know it's going t fail later).
- For the OCaml language tools, extend "parse_key_selector" with the macro check. For that, I'll add a very small helper function.
And same here.
Both of these changes only affect patch "options, mltools/tools_utils: parse "--key ID:clevis" options".
Regarding the present patch, I'll keep the macro check, but replace the error message with abort() -- as it should never be reached.
Yup.
I'll post v3, hopefully soon.
OK, this seems better, although TBH the late check in virt-v2v wasn't really that bad.
I'm totally happy to merge this version (v2) as well!
It's just that you wrote
To be clear on why this is correct, it's because this code is only used during option parsing in command line tools, where we want to exit when an error happens.
and letting that "slide" un-answered would have been dishonest from me, knowing that the error wouldn't be found during option parsing.
That is: I don't insist on v3 at all. I made the proposal for v3 in order to satisfy the (perceived) requirement from you.
If you are OK with the v2 behavior however, so am I :) Please let me know which way you prefer.
TBH I'm fine with v2, but if you've done significant work on v3, then we could do that. You decide!
OK, let's go with v2 then -- I wanted to be sure of your opinion of v3's necessity before starting it, so I've not written one line of code for it yet :) I'll merge all four series hopefully today. Thanks, Laszlo