Skip to contentSkip to conversation

iwpm: crash fix for large connections test

bug report Dan Carpenter 4 participants 6 messages Nov 15, 2022 – Jan 20, 2023 InfiniBand

lore
Share & tools
lore
#0 lore raw

[ This isn't really the correct patch to blame. Sorry! -dan ]

Hello Faisal Latif,

The patch dafb5587178a: "iwpm: crash fix for large connections test" from Feb 26, 2016, leads to the following Smatch static checker warning:

drivers/infiniband/core/iwpm_msg.c:437 iwpm_register_pid_cb() warn: 'nlmsg_request' was already freed. drivers/infiniband/core/iwpm_msg.c:509 iwpm_add_mapping_cb() warn: 'nlmsg_request' was already freed. drivers/infiniband/core/iwpm_msg.c:607 iwpm_add_and_query_mapping_cb() warn: 'nlmsg_request' was already freed. drivers/infiniband/core/iwpm_msg.c:806 iwpm_mapping_error_cb() warn: 'nlmsg_request' was already freed.

drivers/infiniband/core/iwpm_msg.c

    385 int iwpm_register_pid_cb(struct sk_buff *skb, struct netlink_callback *cb)
    386 {
    387         struct iwpm_nlmsg_request *nlmsg_request = NULL;
    388         struct nlattr *nltb[IWPM_NLA_RREG_PID_MAX];
    389         struct iwpm_dev_data *pm_msg;
    390         char *dev_name, *iwpm_name;
    391         u32 msg_seq;
    392         u8 nl_client;
    393         u16 iwpm_version;
    394         const char *msg_type = "Register Pid response";
    395 
    396         if (iwpm_parse_nlmsg(cb, IWPM_NLA_RREG_PID_MAX,
    397                                 resp_reg_policy, nltb, msg_type))
    398                 return -EINVAL;
    399 
    400         msg_seq = nla_get_u32(nltb[IWPM_NLA_RREG_PID_SEQ]);
    401         nlmsg_request = iwpm_find_nlmsg_request(msg_seq);
    402         if (!nlmsg_request) {
    403                 pr_info("%s: Could not find a matching request (seq = %u)\n",
    404                                  __func__, msg_seq);
    405                 return -EINVAL;
    406         }
    407         pm_msg = nlmsg_request->req_buffer;
    408         nl_client = nlmsg_request->nl_client;
    409         dev_name = (char *)nla_data(nltb[IWPM_NLA_RREG_IBDEV_NAME]);
    410         iwpm_name = (char *)nla_data(nltb[IWPM_NLA_RREG_ULIB_NAME]);
    411         iwpm_version = nla_get_u16(nltb[IWPM_NLA_RREG_ULIB_VER]);
    412 
    413         /* check device name, ulib name and version */
    414         if (strcmp(pm_msg->dev_name, dev_name) ||
    415                         strcmp(iwpm_ulib_name, iwpm_name) ||
    416                         iwpm_version < IWPM_UABI_VERSION_MIN) {
    417 
    418                 pr_info("%s: Incorrect info (dev = %s name = %s version = %u)\n",
    419                                 __func__, dev_name, iwpm_name, iwpm_version);
    420                 nlmsg_request->err_code = IWPM_USER_LIB_INFO_ERR;
    421                 goto register_pid_response_exit;
    422         }
    423         iwpm_user_pid = cb->nlh->nlmsg_pid;
    424         iwpm_ulib_version = iwpm_version;
    425         if (iwpm_ulib_version < IWPM_UABI_VERSION)
    426                 pr_warn_once("%s: Down level iwpmd/pid %d.  Continuing...",
    427                         __func__, iwpm_user_pid);
    428         atomic_set(&echo_nlmsg_seq, cb->nlh->nlmsg_seq);
    429         pr_debug("%s: iWarp Port Mapper (pid = %d) is available!\n",
    430                         __func__, iwpm_user_pid);
    431         iwpm_set_registration(nl_client, IWPM_REG_VALID);
    432 register_pid_response_exit:
    433         nlmsg_request->request_done = 1;
    434         /* always for found nlmsg_request */
    435         kref_put(&nlmsg_request->kref, iwpm_free_nlmsg_request);

The iwpm_free_nlmsg_request() function will free "nlmsg_request"... It's not clear what the "/* always for found nlmsg_request */" comment means. Maybe it means that the refcount won't drop to zero so the free function won't be called?

    436         barrier();
--> 437         up(&nlmsg_request->sem);
                    ^^^^^^^^^^^^^

Dereference.

    438         return 0;
    439 }

regards,
dan carpenter

#1 lore raw
Quoted 66 lines · Dan Carpenter

[ This isn't really the correct patch to blame. Sorry! -dan ]

Hello Faisal Latif,

The patch dafb5587178a: "iwpm: crash fix for large connections test" from Feb 26, 2016, leads to the following Smatch static checker warning:

drivers/infiniband/core/iwpm_msg.c:437 iwpm_register_pid_cb() warn: 'nlmsg_request' was already freed. drivers/infiniband/core/iwpm_msg.c:509 iwpm_add_mapping_cb() warn: 'nlmsg_request' was already freed. drivers/infiniband/core/iwpm_msg.c:607 iwpm_add_and_query_mapping_cb() warn: 'nlmsg_request' was already freed. drivers/infiniband/core/iwpm_msg.c:806 iwpm_mapping_error_cb() warn: 'nlmsg_request' was already freed.

drivers/infiniband/core/iwpm_msg.c

    385 int iwpm_register_pid_cb(struct sk_buff *skb, struct netlink_callback *cb)
    386 {
    387         struct iwpm_nlmsg_request *nlmsg_request = NULL;
    388         struct nlattr *nltb[IWPM_NLA_RREG_PID_MAX];
    389         struct iwpm_dev_data *pm_msg;
    390         char *dev_name, *iwpm_name;
    391         u32 msg_seq;
    392         u8 nl_client;
    393         u16 iwpm_version;
    394         const char *msg_type = "Register Pid response";
    395 
    396         if (iwpm_parse_nlmsg(cb, IWPM_NLA_RREG_PID_MAX,
    397                                 resp_reg_policy, nltb, msg_type))
    398                 return -EINVAL;
    399 
    400         msg_seq = nla_get_u32(nltb[IWPM_NLA_RREG_PID_SEQ]);
    401         nlmsg_request = iwpm_find_nlmsg_request(msg_seq);
    402         if (!nlmsg_request) {
    403                 pr_info("%s: Could not find a matching request (seq = %u)\n",
    404                                  __func__, msg_seq);
    405                 return -EINVAL;
    406         }
    407         pm_msg = nlmsg_request->req_buffer;
    408         nl_client = nlmsg_request->nl_client;
    409         dev_name = (char *)nla_data(nltb[IWPM_NLA_RREG_IBDEV_NAME]);
    410         iwpm_name = (char *)nla_data(nltb[IWPM_NLA_RREG_ULIB_NAME]);
    411         iwpm_version = nla_get_u16(nltb[IWPM_NLA_RREG_ULIB_VER]);
    412 
    413         /* check device name, ulib name and version */
    414         if (strcmp(pm_msg->dev_name, dev_name) ||
    415                         strcmp(iwpm_ulib_name, iwpm_name) ||
    416                         iwpm_version < IWPM_UABI_VERSION_MIN) {
    417 
    418                 pr_info("%s: Incorrect info (dev = %s name = %s version = %u)\n",
    419                                 __func__, dev_name, iwpm_name, iwpm_version);
    420                 nlmsg_request->err_code = IWPM_USER_LIB_INFO_ERR;
    421                 goto register_pid_response_exit;
    422         }
    423         iwpm_user_pid = cb->nlh->nlmsg_pid;
    424         iwpm_ulib_version = iwpm_version;
    425         if (iwpm_ulib_version < IWPM_UABI_VERSION)
    426                 pr_warn_once("%s: Down level iwpmd/pid %d.  Continuing...",
    427                         __func__, iwpm_user_pid);
    428         atomic_set(&echo_nlmsg_seq, cb->nlh->nlmsg_seq);
    429         pr_debug("%s: iWarp Port Mapper (pid = %d) is available!\n",
    430                         __func__, iwpm_user_pid);
    431         iwpm_set_registration(nl_client, IWPM_REG_VALID);
    432 register_pid_response_exit:
    433         nlmsg_request->request_done = 1;
    434         /* always for found nlmsg_request */
    435         kref_put(&nlmsg_request->kref, iwpm_free_nlmsg_request);
Quoted 4 lines

The iwpm_free_nlmsg_request() function will free "nlmsg_request"... It's not clear what the "/* always for found nlmsg_request */" comment means. Maybe it means that the refcount won't drop to zero so the free function won't be called?

I think so. The nlmsg_request reference counter is elevated when it is found in iwpm_find_nlmsg_request(). So I assume that it will be at least 2 before call to kref_put(). Most likely, nlmsg_request->sem prevents from parallel threads to decrease that reference counter.

BTW, not IWPM expert.

Thanks

#2 lore raw
Quoted 86 lines

Subject: Re: [bug report] iwpm: crash fix for large connections test

Quoted 9 lines · Dan Carpenter

[ This isn't really the correct patch to blame. Sorry! -dan ]

Hello Faisal Latif,

The patch dafb5587178a: "iwpm: crash fix for large connections test" from Feb 26, 2016, leads to the following Smatch static checker warning:

drivers/infiniband/core/iwpm_msg.c:437 iwpm_register_pid_cb() warn:

'nlmsg_request' was already freed.

Quoted 1 line

drivers/infiniband/core/iwpm_msg.c:509 iwpm_add_mapping_cb() warn:

'nlmsg_request' was already freed.

Quoted 1 line

drivers/infiniband/core/iwpm_msg.c:607

iwpm_add_and_query_mapping_cb() warn: 'nlmsg_request' was already freed.

Quoted 1 line

drivers/infiniband/core/iwpm_msg.c:806 iwpm_mapping_error_cb() warn:

'nlmsg_request' was already freed.

Quoted 3 lines

drivers/infiniband/core/iwpm_msg.c
385 int iwpm_register_pid_cb(struct sk_buff *skb, struct netlink_callback

*cb)

Quoted 10 lines
    386 {
    387         struct iwpm_nlmsg_request *nlmsg_request = NULL;
    388         struct nlattr *nltb[IWPM_NLA_RREG_PID_MAX];
    389         struct iwpm_dev_data *pm_msg;
    390         char *dev_name, *iwpm_name;
    391         u32 msg_seq;
    392         u8 nl_client;
    393         u16 iwpm_version;
    394         const char *msg_type = "Register Pid response";
    395
Quoted 8 lines
    396         if (iwpm_parse_nlmsg(cb, IWPM_NLA_RREG_PID_MAX,
    397                                 resp_reg_policy, nltb, msg_type))
    398                 return -EINVAL;
    399
    400         msg_seq = nla_get_u32(nltb[IWPM_NLA_RREG_PID_SEQ]);
    401         nlmsg_request = iwpm_find_nlmsg_request(msg_seq);
    402         if (!nlmsg_request) {
    403                 pr_info("%s: Could not find a matching request (seq =

%u)\n",

Quoted 6 lines
    404                                  __func__, msg_seq);
    405                 return -EINVAL;
    406         }
    407         pm_msg = nlmsg_request->req_buffer;
    408         nl_client = nlmsg_request->nl_client;
    409         dev_name = (char

*)nla_data(nltb[IWPM_NLA_RREG_IBDEV_NAME]);

Quoted 1 line

410 iwpm_name = (char

*)nla_data(nltb[IWPM_NLA_RREG_ULIB_NAME]);

Quoted 1 line

411 iwpm_version =

nla_get_u16(nltb[IWPM_NLA_RREG_ULIB_VER]);

Quoted 7 lines
    412
    413         /* check device name, ulib name and version */
    414         if (strcmp(pm_msg->dev_name, dev_name) ||
    415                         strcmp(iwpm_ulib_name, iwpm_name) ||
    416                         iwpm_version < IWPM_UABI_VERSION_MIN) {
    417
    418                 pr_info("%s: Incorrect info (dev = %s name = %s version =

%u)\n",

Quoted 8 lines
    419                                 __func__, dev_name, iwpm_name, iwpm_version);
    420                 nlmsg_request->err_code = IWPM_USER_LIB_INFO_ERR;
    421                 goto register_pid_response_exit;
    422         }
    423         iwpm_user_pid = cb->nlh->nlmsg_pid;
    424         iwpm_ulib_version = iwpm_version;
    425         if (iwpm_ulib_version < IWPM_UABI_VERSION)
    426                 pr_warn_once("%s: Down level iwpmd/pid %d.

Continuing...",

Quoted 10 lines
    427                         __func__, iwpm_user_pid);
    428         atomic_set(&echo_nlmsg_seq, cb->nlh->nlmsg_seq);
    429         pr_debug("%s: iWarp Port Mapper (pid = %d) is available!\n",
    430                         __func__, iwpm_user_pid);
    431         iwpm_set_registration(nl_client, IWPM_REG_VALID);
    432 register_pid_response_exit:
    433         nlmsg_request->request_done = 1;
    434         /* always for found nlmsg_request */
    435         kref_put(&nlmsg_request->kref, iwpm_free_nlmsg_request);
Quoted 4 lines

The iwpm_free_nlmsg_request() function will free "nlmsg_request"... It's not clear what the "/* always for found nlmsg_request */" comment means. Maybe it means that the refcount won't drop to zero so the free function won't be called?

Quoted 5 lines

I think so. The nlmsg_request reference counter is elevated when it is found in iwpm_find_nlmsg_request(). So I assume that it will be at least 2 before call to kref_put(). Most likely, nlmsg_request->sem prevents from parallel threads to decrease that reference counter.

I agree with Leon. The ref count should be 2 here.
However, I don't see why the kref_put() can't be moved after the up(&nlmsg_request->sem) to get rid of the warning.

Regards,

Mustafa

#3 lore raw
Quoted 10 lines · Mustafa
Quoted 10 lines
Quoted 9 lines
    432 register_pid_response_exit:
    433         nlmsg_request->request_done = 1;
    434         /* always for found nlmsg_request */
    435         kref_put(&nlmsg_request->kref, iwpm_free_nlmsg_request);

The iwpm_free_nlmsg_request() function will free "nlmsg_request"... It's not clear what the "/* always for found nlmsg_request */" comment means. Maybe it means that the refcount won't drop to zero so the free function won't be called?

Quoted 9 lines
Quoted 5 lines

I think so. The nlmsg_request reference counter is elevated when it is found in iwpm_find_nlmsg_request(). So I assume that it will be at least 2 before call to kref_put(). Most likely, nlmsg_request->sem prevents from parallel threads to decrease that reference counter.

I agree with Leon. The ref count should be 2 here.
However, I don't see why the kref_put() can't be moved after the up(&nlmsg_request->sem) to get rid of the warning.

Let's not expend too much time trying to silence this warning. One way to silence the warning would be to do:

	kref_put(&nlmsg_request->kref, NULL);

I'm conficted about this approach, but no good can come from calling iwpm_free_nlmsg_request() on this path.

A better way to silence the warning would be to do:

drivers/infiniband/core/iwpm_util.c +1 -1
index 358a2db38d23..4f819e6c1b09 100644
@@ -357,7 +357,7 @@ struct iwpm_nlmsg_request *iwpm_find_nlmsg_request(__u32 echo_seq)
 			    inprocess_list) {
 		if (nlmsg_request->nlmsg_seq == echo_seq) {
 			found_request = nlmsg_request;
-			kref_get(&nlmsg_request->kref);
+			kref_get(&found_request->kref);
 			break;
 		}
 	}

But the best way would be to make Smatch parse iwpm_find_nlmsg_request() correctly as-is.

regards,
dan carpenter

#4 lore raw
In reply to Ismail, Mustafa

So the background here is that Smatch sees this:

	kref_put(&nlmsg_request->kref, iwpm_free_nlmsg_request);

and correctly says "if we call iwpm_free_nlmsg_request() then dereferencing nlmsg_request is a use after free". However, the code is holding two references at this point so it will never call iwpm_free_nlmsg_request().

Smatch already checks to see if we are holding two references, but it doesn't parse this code correctly. Smatch could be fixed, but there are other places with similar warnings that are more difficult to fix.

What we could do is create a kref_no_release() function that just calls WARN(). This would silence the warning and, I think, this would make the code more readable.

What do other people think?

regards,
dan carpenter

include/linux/kref.h +5 -0
index d32e21a2538c..f40089f61aa6 100644
@@ -45,6 +45,11 @@ static inline void kref_get(struct kref *kref)
 	refcount_inc(&kref->refcount);
 }
 
+static inline void kref_no_release(struct kref *kref)
+{
+	WARN(1, "Unexpected kref release");
+}
+
 /**
  * kref_put - decrement refcount for object.
  * @kref: object.
drivers/infiniband/core/iwpm_msg.c +4 -4
index 3c9a9869212b..1acb96bfaf9c 100644
@@ -432,7 +432,7 @@ int iwpm_register_pid_cb(struct sk_buff *skb, struct netlink_callback *cb)
 register_pid_response_exit:
 	nlmsg_request->request_done = 1;
 	/* always for found nlmsg_request */
-	kref_put(&nlmsg_request->kref, iwpm_free_nlmsg_request);
+	kref_put(&nlmsg_request->kref, kref_no_release);
 	barrier();
 	up(&nlmsg_request->sem);
 	return 0;
@@ -504,7 +504,7 @@ int iwpm_add_mapping_cb(struct sk_buff *skb, struct netlink_callback *cb)
 add_mapping_response_exit:
 	nlmsg_request->request_done = 1;
 	/* always for found request */
-	kref_put(&nlmsg_request->kref, iwpm_free_nlmsg_request);
+	kref_put(&nlmsg_request->kref, kref_no_release);
 	barrier();
 	up(&nlmsg_request->sem);
 	return 0;
@@ -602,7 +602,7 @@ int iwpm_add_and_query_mapping_cb(struct sk_buff *skb,
 query_mapping_response_exit:
 	nlmsg_request->request_done = 1;
 	/* always for found request */
-	kref_put(&nlmsg_request->kref, iwpm_free_nlmsg_request);
+	kref_put(&nlmsg_request->kref, kref_no_release);
 	barrier();
 	up(&nlmsg_request->sem);
 	return 0;
@@ -801,7 +801,7 @@ int iwpm_mapping_error_cb(struct sk_buff *skb, struct netlink_callback *cb)
 	nlmsg_request->err_code = err_code;
 	nlmsg_request->request_done = 1;
 	/* always for found request */
-	kref_put(&nlmsg_request->kref, iwpm_free_nlmsg_request);
+	kref_put(&nlmsg_request->kref, kref_no_release);
 	barrier();
 	up(&nlmsg_request->sem);
 	return 0;
#5 lore raw
Quoted 13 lines · Dan Carpenter

So the background here is that Smatch sees this:

	kref_put(&nlmsg_request->kref, iwpm_free_nlmsg_request);

and correctly says "if we call iwpm_free_nlmsg_request() then dereferencing nlmsg_request is a use after free". However, the code is holding two references at this point so it will never call iwpm_free_nlmsg_request().

Smatch already checks to see if we are holding two references, but it doesn't parse this code correctly. Smatch could be fixed, but there are other places with similar warnings that are more difficult to fix.

Quoted 5 lines

What we could do is create a kref_no_release() function that just calls WARN(). This would silence the warning and, I think, this would make the code more readable.

What do other people think?

Sure, that looks semi-decent if it helps out with the automated tools.

thanks

greg k-h