From: Hannes Reinecke <hare@suse.de>
To: sashiko-reviews@lists.linux.dev, Hannes Reinecke <hare@kernel.org>
Cc: linux-nvme@lists.infradead.org
Subject: Re: [PATCH v4 6/9] nvme: allow to pass in key description as dhchap secret
Date: Fri, 20 Nov 2026 15:48:59 +0100 [thread overview]
Message-ID: <c19b8afa-8a14-4ba5-a1f7-2d2a89ec5e9d@suse.de> (raw)
In-Reply-To: <sashiko-outbox-225046@kernel.org>
On 11/20/26 12:31 PM, sashiko-bot@kernel.org wrote:
> Thank you for your contribution! Sashiko AI review found 9 potential
> issues to consider:
> - [High] nvme-fabrics: wrong key revoked in nvmf_free_options() for
> dhchap_ctrl_key
> - [High] nvme-auth: empty/prefix match and self-revocation in
> nvme_auth_extract_key()
> - [High] nvme-fabrics: link failure calling nvme_auth_extract_key()
> without NVME_AUTH
> - [Medium] nvme: unlocked access to dhchap keys and flags in sysfs show
> callbacks
> - [Medium] nvme: broken sysfs visibility for dhchap_key in
> nvme_dev_attrs_are_visible()
> - [Medium] nvme-auth: host_key leak and stale flags on
> nvme_auth_init_ctrl() error
> - [Medium] nvmet: memory leak of keydata buffer in
> nvmet_host_dhchap_ctrl_key_store()
> - [Low] nvme: unsigned count and missing newline in
> dhchap_ctrl_secret_show()
> - [Low] nvme-auth: missing keyring param in nvme_auth_extract_key()
> kernel-doc
> --
>
> commit 2b7a36fa0b1a7939d6a4d7f1cdc10818701d7d40
> Author: Hannes Reinecke <hare@kernel.org>
>
> nvme: allow to pass in key description as dhchap secret
>
> This commit updates nvme_auth_extract_key() to look up pre-populated
> DH-HMAC-CHAP keys in the keyring before falling back to generating an
> ephemeral key from raw secret material, and tracks whether keys were
> auto-generated across the host and target drivers so only generated keys
> are revoked on release.
>
>> diff --git a/drivers/nvme/common/auth.c b/drivers/nvme/common/auth.c
>> index b7092deed9b06..c65bfafe9e3fc 100644
>> --- a/drivers/nvme/common/auth.c
>> +++ b/drivers/nvme/common/auth.c
>> @@ -143,20 +143,32 @@ EXPORT_SYMBOL_GPL(nvme_auth_hmac_hash_len);
>> /**
>> * nvme_auth_extract_key - extract the DH-HMAC-CHAP key
>> *
>> - * @secret: key data
>> - * @secret_len: length of @secret
>> + * @input: key serial or key data
>> + * @input_len: length of @input
>> + * @generated: indicate whether a key has been generated
>> *
>> - * Extracts a dhchap key from @secret.
>> + * Extracts or generates a key from @input.
>> + * @input can either be a key serial or raw key data; in
>> + * the latter case a key is generated from @input and
>> + * @generated is set to 'true'.
>> *
>> * Returns the dhchap key or an error pointer on failure.
>> */
>> -struct key *nvme_auth_extract_key(struct key *keyring, const char *secret,
>> - size_t secret_len)
>> +struct key *nvme_auth_extract_key(struct key *keyring, const char *input,
>> + size_t input_len, bool *generated)
>
> [Severity: Low]
> The kernel-doc comment for nvme_auth_extract_key() omits the @keyring
> parameter, which triggers a W=1 kernel-doc warning. Also, the comment and
> commit message describe @input as a key serial number even though
> nvme_dhchap_psk_lookup() searches by key description, and the prototype in
> include/linux/nvme-auth.h still names the parameters secret and secret_len.
>
>> {
>> struct key *key;
>>
>> - key = nvme_dhchap_psk_create(keyring, secret, secret_len);
>> + /* Check if @input is a key serial number */
>> + key = nvme_dhchap_psk_lookup(keyring, input);
>> if (!IS_ERR(key)) {
>> + *generated = false;
>> + return key;
>> + }
>> + /* Generate a key from @input data */
>> + key = nvme_dhchap_psk_create(keyring, input, input_len);
>> + if (!IS_ERR(key)) {
>> + *generated = true;
>> pr_debug("generated dhchap key %s\n",
>> key->description);
>> }
>
> [Severity: High]
> Can calling nvme_dhchap_psk_lookup() here match the wrong key, read out of
> bounds, or cause the caller to revoke the key it just looked up?
>
> nvme_dhchap_psk_lookup() searches with nvme_dhchap_psk_match():
>
> drivers/nvme/common/keyring.c:nvme_dhchap_psk_match() {
> ...
> match_id = match_data->raw_data;
> match_len = strlen(match_id);
> pr_debug("%s: match '%s' '%s' len %zd\n",
> __func__, match_id, key->description, match_len);
>
> return !memcmp(key->description, match_id, match_len);
> }
>
> Because nvme_dhchap_psk_match() compares match_len bytes with memcmp()
> instead of strcmp(), and nvme_auth_extract_key() does not reject
> input_len == 0:
>
> 1. Writing a single newline to sysfs dhchap_secret_store() or
> dhchap_ctrl_secret_store() passes an empty string (match_len == 0), so
> !memcmp(key->description, "", 0) matches the first dhchap key in the
> keyring with *generated = false.
> 2. For non-empty inputs, memcmp() prefix-matches any key whose description
> starts with input when match_len < strlen(key->description), and reads
> past the end of the key->description allocation when match_len >
> strlen(key->description) (for example, when passing a raw DHHC-1:...
> secret while a key with a shorter description is in the keyring).
> 3. If nvme_dhchap_psk_lookup() returns the controller's or target host's
> currently active auto-generated key (either via an empty write or by
> writing the current key's UUID description to dhchap_secret_store(),
> dhchap_ctrl_secret_store(), or nvmet_auth_set_key()), the store path
> sees that old_key was auto-generated and calls key_revoke(old_key)
> without checking key == old_key, revoking the key it then keeps active.
>
key parsing has been fixed up with the updated version of the previous
patch.
>> diff --git a/drivers/nvme/host/auth.c b/drivers/nvme/host/auth.c
>> index c189096f825ea..470a6e7e43644 100644
>> --- a/drivers/nvme/host/auth.c
>> +++ b/drivers/nvme/host/auth.c
> [ ... ]
>> @@ -1071,6 +1081,7 @@ int nvme_auth_init_ctrl(struct nvme_ctrl *ctrl)
>> key_serial(ctrl->opts->dhchap_key));
>> return -ENOKEY;
>> }
>> + ctrl->host_key_generated = ctrl->opts->dhchap_key_generated;
>> down_read(&ctrl->host_key->sem);
>> ret = key_validate(ctrl->host_key);
>> up_read(&ctrl->host_key->sem);
>> @@ -1097,6 +1108,8 @@ int nvme_auth_init_ctrl(struct nvme_ctrl *ctrl)
>> key_serial(ctrl->opts->dhchap_ctrl_key));
>> return -ENOKEY;
>> }
>> + ctrl->ctrl_key_generated =
>> + ctrl->opts->dhchap_ctrl_key_generated;
>> down_read(&ctrl->ctrl_key->sem);
>> ret = key_validate(ctrl->ctrl_key);
>> up_read(&ctrl->ctrl_key->sem);
>
> [Severity: Medium]
> If acquiring or validating ctrl->ctrl_key fails in nvme_auth_init_ctrl(),
> does this leak ctrl->host_key and leave stale *_key_generated flags set?
>
> On failure of key_get(ctrl->opts->dhchap_ctrl_key) or
> key_validate(ctrl->ctrl_key), nvme_auth_init_ctrl() returns -ENOKEY or
> -EKEYREVOKED without calling nvme_auth_clear_host_key(ctrl). In
> nvme_init_ctrl(), a failure from nvme_auth_init_ctrl() jumps to
> out_release_instance before device_initialize(&ctrl->ctrl_device), so
> nvme_auth_free() never runs and the ctrl->host_key reference is leaked.
> The error paths after setting ctrl->host_key_generated and
> ctrl->ctrl_key_generated also clear the key pointer without resetting the
> corresponding generated boolean to false.
>
Fixed with the next version.
>> diff --git a/drivers/nvme/host/fabrics.c b/drivers/nvme/host/fabrics.c
>> index 7ed8618257ca9..cd728768a68fc 100644
>> --- a/drivers/nvme/host/fabrics.c
>> +++ b/drivers/nvme/host/fabrics.c
> [ ... ]
>> @@ -1089,7 +1091,8 @@ static int nvmf_parse_options(struct nvmf_ctrl_options *opts,
>> if (host_secret) {
>> pr_debug("lookup host identity '%s'\n", host_secret);
>> key = nvme_auth_extract_key(opts->keyring, host_secret,
>> - strlen(host_secret));
>> + strlen(host_secret),
>> + &opts->dhchap_key_generated);
>> if (IS_ERR(key)) {
>> ret = PTR_ERR(key);
>> goto out;
>> @@ -1104,7 +1107,8 @@ static int nvmf_parse_options(struct nvmf_ctrl_options *opts,
>> }
>> pr_debug("lookup ctrl identity '%s'\n", ctrl_secret);
>> key = nvme_auth_extract_key(opts->keyring, ctrl_secret,
>> - strlen(ctrl_secret));
>> + strlen(ctrl_secret),
>> + &opts->dhchap_ctrl_key_generated);
>> if (IS_ERR(key)) {
>> ret = PTR_ERR(key);
>> goto out;
>
> [Severity: High]
> Does calling nvme_auth_extract_key() unconditionally in nvmf_parse_options()
> break linking when CONFIG_NVME_FABRICS is enabled without CONFIG_NVME_AUTH?
>
> CONFIG_NVME_FABRICS does not select CONFIG_NVME_AUTH, and
> include/linux/nvme-auth.h does not provide a !CONFIG_NVME_AUTH stub for
> nvme_auth_extract_key(), so builds with CONFIG_NVME_HOST_AUTH disabled (or
> CONFIG_NVME_FABRICS=y and CONFIG_NVME_AUTH=m) fail to link.
>
Fixed with the next version.
>> @@ -1327,15 +1331,23 @@ void nvmf_free_options(struct nvmf_ctrl_options *opts)
>> kfree(opts->host_traddr);
>> kfree(opts->host_iface);
>> if (opts->dhchap_key) {
>> - pr_debug("revoke dhchap host key %08x\n",
>> + if (opts->dhchap_key_generated) {
>> + pr_debug("revoke dhchap host key %08x\n",
>> + key_serial(opts->dhchap_key));
>> + key_revoke(opts->dhchap_key);
>> + }
>> + pr_debug("drop dhchap host key %08x\n",
>> key_serial(opts->dhchap_key));
>> - key_revoke(opts->dhchap_key);
>> key_put(opts->dhchap_key);
>> }
>> if (opts->dhchap_ctrl_key) {
>> - pr_debug("revoke dhchap ctrl key %08x\n",
>> + if (opts->dhchap_ctrl_key_generated) {
>> + pr_debug("revoke dhchap ctrl key %08x\n",
>> + key_serial(opts->dhchap_ctrl_key));
>> + key_revoke(opts->dhchap_key);
>> + }
>> + pr_debug("drop dhchap ctrl key %08x\n",
>> key_serial(opts->dhchap_ctrl_key));
>> - key_revoke(opts->dhchap_key);
>> key_put(opts->dhchap_ctrl_key);
>> }
>> kfree(opts);
>
> [Severity: High]
> In the opts->dhchap_ctrl_key block, should key_revoke(opts->dhchap_key) be
> key_revoke(opts->dhchap_ctrl_key)?
>
> Passing opts->dhchap_key here accesses it after key_put(opts->dhchap_key)
> has already dropped the reference (risking a use-after-free), revokes a
> pre-populated host key when only the controller key was auto-generated, and
> leaves the auto-generated opts->dhchap_ctrl_key unrevoked.
>
Fixed with the next version.
>> diff --git a/drivers/nvme/host/sysfs.c b/drivers/nvme/host/sysfs.c
>> index 1311f5e29b004..47caed2ce3978 100644
>> --- a/drivers/nvme/host/sysfs.c
>> +++ b/drivers/nvme/host/sysfs.c
>> @@ -828,13 +828,14 @@ static ssize_t dhchap_secret_show(struct device *dev,
>> down_read(&key->sem);
>> if (key_validate(key))
>> count = sysfs_emit(buf, "<invalidated>\n");
>> - else {
>> + else if (ctrl->host_key_generated) {
>> count = key->type->read(key, buf, PAGE_SIZE);
>> if (count > 0) {
>> buf[count] = '\n';
>> count++;
>> }
>> - }
>> + } else
>> + count = sysfs_emit(buf, "%s\n", key->description);
>> up_read(&key->sem);
>> return count;
>> }
> [ ... ]
>> @@ -869,19 +872,25 @@ static ssize_t dhchap_secret_store(struct device *dev,
>> up_read(&key->sem);
>> if (ret) {
>> dev_warn(ctrl->dev, "key %08x invalidated\n", key_serial(key));
>> - dev_dbg(ctrl->dev, "revoke host key %08x\n", key_serial(key));
>> - key_revoke(key);
>> + if (generated) {
>> + dev_dbg(ctrl->dev, "revoke host key %08x\n", key_serial(key));
>> + key_revoke(key);
>> + synchronize_rcu();
>> + }
>> key_put(key);
>> kfree_sensitive(dhchap_secret);
>> return ret;
>> }
>> mutex_lock(&ctrl->dhchap_auth_mutex);
>> old_key = ctrl->host_key;
>> - dev_dbg(ctrl->dev, "revoke host key %08x\n",
>> - key_serial(old_key));
>> - key_revoke(old_key);
>> -
>> + if (ctrl->host_key_generated) {
>> + dev_dbg(ctrl->dev, "revoke host key %08x\n",
>> + key_serial(old_key));
>> + key_revoke(old_key);
>> + synchronize_rcu();
>> + }
>> ctrl->host_key = key;
>> + ctrl->host_key_generated = generated;
>> mutex_unlock(&ctrl->dhchap_auth_mutex);
>> key_put(old_key);
>> kfree_sensitive(dhchap_secret);
>
> [Severity: Medium]
> Can dhchap_secret_show(), dhchap_ctrl_secret_show(), dhchap_key_show(), and
> dhchap_ctrl_key_show() race with concurrent stores and trigger a
> use-after-free or read a mismatched *_key_generated flag?
>
> Unlike nvmet_host_dhchap_key_show() in drivers/nvme/target/configfs.c, the
> host sysfs show callbacks read ctrl->host_key / ctrl->ctrl_key and
> ctrl->host_key_generated / ctrl->ctrl_key_generated without holding
> ctrl->dhchap_auth_mutex or pinning the key with key_get():
>
> CPU0 CPU1
> dhchap_secret_show() dhchap_secret_store()
> key = ctrl->host_key;
> mutex_lock(&ctrl->dhchap_auth_mutex);
> old_key = ctrl->host_key;
> ctrl->host_key = new_key;
> ctrl->host_key_generated = true;
> mutex_unlock(&ctrl->dhchap_auth_mutex);
> key_put(old_key);
> down_read(&key->sem); <-- use-after-free
>
> Even if old_key still has another reference (such as when old_key was a
> pre-populated user key in the keyring), CPU0 can pair old_key with the newly
> written ctrl->host_key_generated == true and dump the pre-populated key's
> raw secret via key->type->read() instead of its description.
>
Fixed with the next version.
>> @@ -906,13 +915,14 @@ static ssize_t dhchap_ctrl_secret_show(struct device *dev,
>> down_read(&key->sem);
>> if (key_validate(key))
>> count = sysfs_emit(buf, "<invalidated>");
>> - else {
>> + else if (ctrl->ctrl_key_generated) {
>> count = key->type->read(key, buf, PAGE_SIZE);
>> if (count > 0) {
>> buf[count] = '\n';
>> count++;
>> }
>> - }
>> + } else
>> + count = sysfs_emit(buf, "%s\n", key->description);
>> up_read(&key->sem);
>> return count;
>> }
>
> [Severity: Low]
> In dhchap_ctrl_secret_show(), count is declared as unsigned size_t (unlike
> ssize_t in dhchap_secret_show()), so a negative error return from
> key->type->read() would wrap to a large positive value, pass count > 0, and
> write out of bounds at buf[count] = '\n'. Also,
> sysfs_emit(buf, "<invalidated>") is missing the trailing '\n' present in the
> other show callbacks.
>
Fixed with the next version.
> [ ... ]
>
>> @@ -1066,8 +1088,12 @@ static umode_t nvme_dev_attrs_are_visible(struct kobject *kobj,
>> #ifdef CONFIG_NVME_HOST_AUTH
>> if (a == &dev_attr_dhchap_secret.attr && !ctrl->opts)
>> return 0;
>> + if (a == &dev_attr_dhchap_key.attr && ctrl->host_key_generated)
>> + return 0;
>> if (a == &dev_attr_dhchap_ctrl_secret.attr && !ctrl->opts)
>> return 0;
>> + if (a == &dev_attr_dhchap_ctrl_key.attr && ctrl->ctrl_key_generated)
>> + return 0;
>> #endif
>>
>> return a->mode;
>
> [Severity: Medium]
> Should the visibility checks for dev_attr_dhchap_key.attr and
> dev_attr_dhchap_ctrl_key.attr also check !ctrl->opts, and how is attribute
> visibility updated when keys change at runtime?
>
> Without a !ctrl->opts check, dhchap_key and dhchap_ctrl_key are exposed in
> sysfs on non-Fabrics controllers (such as PCIe controllers, where
> ctrl->opts is NULL and the *_key_generated fields are false) even though
> dhchap_secret and dhchap_ctrl_secret are hidden. In addition, because
> nvme_dev_attrs_are_visible() is only evaluated when the sysfs group is
> created and dhchap_secret_store() / dhchap_ctrl_secret_store() do not call
> sysfs_update_group(), switching a controller at runtime between an
> auto-generated secret and a pre-populated key description leaves dhchap_key
> and dhchap_ctrl_key stuck in their initial visibility state.
>
Fixed with the new version.
>> diff --git a/drivers/nvme/target/configfs.c b/drivers/nvme/target/configfs.c
>> index 880cf5151588a..319947aa0a40f 100644
>> --- a/drivers/nvme/target/configfs.c
>> +++ b/drivers/nvme/target/configfs.c
> [ ... ]
>> @@ -2305,13 +2306,14 @@ static ssize_t nvmet_host_dhchap_ctrl_key_show(struct config_item *item,
>> down_read(&key->sem);
>> if (key_validate(key))
>> ret = sprintf(page, "<invalidated>\n");
>> - else {
>> + else if (host->dhchap_ctrl_key_generated) {
>> ret = key->type->read(key, page, PAGE_SIZE);
>> if (ret > 0) {
>> page[ret] = '\n';
>> ret++;
>> }
>> - }
>> + } else
>> + ret = sprintf(page, "%s\n", key->description);
>> up_read(&key->sem);
>> key_put(key);
>> }
>
> [Severity: Medium]
> Does nvmet_host_dhchap_ctrl_key_store() just below this function leak
> keydata on every write?
>
> drivers/nvme/target/configfs.c:nvmet_host_dhchap_ctrl_key_store() {
> ...
> keydata = kstrndup(page, len, GFP_KERNEL);
> if (!keydata)
> return -ENOMEM;
>
> ret = nvmet_auth_set_key(host, keydata, true);
> ...
> return ret < 0 ? ret : count;
> }
>
> Unlike nvmet_host_dhchap_key_store(), which calls kfree_sensitive(keydata)
> after nvmet_auth_set_key(), nvmet_host_dhchap_ctrl_key_store() returns
> without freeing keydata, leaking the buffer and leaving key material in slab
> memory.
>
Already fixed with the update to a previous patch.
Cheers,
Hannes
--
Dr. Hannes Reinecke Kernel Storage Architect
hare@suse.de +49 911 74053 688
SUSE Software Solutions GmbH, Frankenstr. 146, 90461 Nürnberg
HRB 36809 (AG Nürnberg), GF: I. Totev, A. McDonald, W. Knoblich
next prev parent reply other threads:[~2026-11-20 14:49 UTC|newest]
Thread overview: 28+ messages / expand[flat|nested] mbox.gz Atom feed top
2026-11-20 11:13 [PATCH v4 0/9] nvme-auth: switch to use the kernel keyring Hannes Reinecke
2026-11-20 11:14 ` [PATCH v4 1/9] nvme-auth: modify nvme_auth_transform_key() to return status Hannes Reinecke
2026-11-20 11:19 ` sashiko-bot
2026-11-20 11:42 ` Hannes Reinecke
2026-11-20 11:14 ` [PATCH v4 2/9] nvme-keyring: add 'dhchap' key type Hannes Reinecke
2026-11-20 11:28 ` sashiko-bot
2026-11-20 12:28 ` Hannes Reinecke
2026-11-20 11:14 ` [PATCH v4 3/9] nvme-auth: switch to use 'struct key' Hannes Reinecke
2026-11-20 11:31 ` sashiko-bot
2026-11-20 13:08 ` Hannes Reinecke
2026-11-20 11:14 ` [PATCH v4 4/9] nvme: parse dhchap keys during option parsing Hannes Reinecke
2026-11-20 11:32 ` sashiko-bot
2026-11-20 13:55 ` Hannes Reinecke
2026-11-20 11:14 ` [PATCH v4 5/9] nvmet-auth: parse dhchap key from configfs attribute Hannes Reinecke
2026-11-20 11:28 ` sashiko-bot
2026-11-20 14:15 ` Hannes Reinecke
2026-11-20 11:14 ` [PATCH v4 6/9] nvme: allow to pass in key description as dhchap secret Hannes Reinecke
2026-11-20 11:31 ` sashiko-bot
2026-11-20 14:48 ` Hannes Reinecke [this message]
2026-11-20 11:14 ` [PATCH v4 7/9] nvme-auth: wait for authentication to finish when changing keys Hannes Reinecke
2026-11-20 11:26 ` sashiko-bot
2026-11-20 14:52 ` Hannes Reinecke
2026-11-20 11:14 ` [PATCH v4 8/9] nvme-fabrics: allow to pass in keyring by name Hannes Reinecke
2026-11-20 11:37 ` sashiko-bot
2026-11-20 15:02 ` Hannes Reinecke
2026-11-20 11:14 ` [PATCH v4 9/9] nvmet: add configfs attribute 'dhchap_keyring' Hannes Reinecke
2026-11-20 11:39 ` sashiko-bot
2026-11-20 15:27 ` Hannes Reinecke
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=c19b8afa-8a14-4ba5-a1f7-2d2a89ec5e9d@suse.de \
--to=hare@suse.de \
--cc=hare@kernel.org \
--cc=linux-nvme@lists.infradead.org \
--cc=sashiko-reviews@lists.linux.dev \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox