On 6/10/20 3:57 AM, Vladimir Sementsov-Ogievskiy wrote:
08.06.2020 21:26, Eric Blake wrote:
Ever since commit 36683283 (v2.8), the server code asserts that error
strings sent to the client are well-formed per the protocol by not
exceeding the maximum string length of 4096. At the time the server
first started sending error messages, the assertion could not be
triggered, because messages were completely under our control.
However, over the years, we have added latent scenarios where a client
could trigger the server to attempt an error message that would
include the client's information if it passed other checks first:
- requesting NBD_OPT_INFO/GO on an export name that is not present
(commit 0cfae925 in v2.12 echoes the name)
- requesting NBD_OPT_LIST/SET_META_CONTEXT on an export name that is
not present (commit e7b1948d in v2.12 echoes the name)
At the time, those were still safe because we flagged names larger
than 256 bytes with a different message; but that changed in commit
93676c88 (v4.2) when we raised the name limit to 4096 to match the NBD
string limit. (That commit also failed to change the magic number
4096 in nbd_negotiate_send_rep_err to the just-introduced named
constant.) So with that commit, long client names appended to server
text can now trigger the assertion, and thus be used as a denial of
service attack against a server. As a mitigating factor, if the
server requires TLS, the client cannot trigger the problematic paths
unless it first supplies TLS credentials, and such trusted clients are
less likely to try to intentionally crash the server.
Reported-by: Xueqiang Wei <[email protected]>
CC: [email protected]
Fixes: https://bugzilla.redhat.com/1843684 CVE-2020-10761
Fixes: 93676c88d7
Signed-off-by: Eric Blake <[email protected]>
---
nbd/server.c | 28 +++++++++++++++++++++++++---
tests/qemu-iotests/143 | 4 ++++
tests/qemu-iotests/143.out | 2 ++
3 files changed, 31 insertions(+), 3 deletions(-)
diff --git a/nbd/server.c b/nbd/server.c
index 02b1ed080145..ec130303586d 100644
--- a/nbd/server.c
+++ b/nbd/server.c
@@ -217,7 +217,7 @@ nbd_negotiate_send_rep_verr(NBDClient *client,
uint32_t type,
msg = g_strdup_vprintf(fmt, va);
len = strlen(msg);
- assert(len < 4096);
+ assert(len < NBD_MAX_STRING_SIZE);
trace_nbd_negotiate_send_rep_err(msg);
ret = nbd_negotiate_send_rep_len(client, type, len, errp);
if (ret < 0) {
@@ -231,6 +231,27 @@ nbd_negotiate_send_rep_verr(NBDClient *client,
uint32_t type,
return 0;
}
+/*
+ * Truncate a potentially-long user-supplied string into something
+ * more suitable for an error reply.
+ */
+static const char *
+nbd_truncate_name(const char *name)
+{
+#define SANE_LENGTH 80
+ static char buf[SANE_LENGTH + 3 + 1]; /* Trailing '...', NUL */
s/NUL/NULL/
NULL is the pointer (typically 4 or 8 bytes); NUL is the character
(exactly one byte in all multi-byte-encodings like UTF-8, or
sizeof(wchar_t) when using wide characters).
Hmm. It may break if we use it in parallel in two coroutines or
threads.. Not sure, is it possible now, neither of course will it be
possible in future.
After some testing (including adding some temporary sleep() into the
code), it looks like 'qemu-nbd -e 2' is currently serialized (we don't
start responding to a second client until we are done negotiating with
the first); on that grounds, we are not risking that information leaks
from one client to another. But you are correct that it is not obvious,
and that if we do have a situation where two threads can try to allow an
NBD connection, then this static buffer could leak information from one
client to another. So I'll need to post a v2.
I'd avoid creating functions returning instead use g_strdup_printf(), like
char *tmp = g_strdup_printf("%.80s...", name);
( OR, if you want explicit constant: g_strdup_printf("%.*s...",
SANE_LENGTH, name) )
... report error ...
g_free(tmp)
Using g_strdup_printf also is safer as we don't need to care about buf
size.
malloc'ing the buffer is not too bad; error messages are not the hot
path. I'll change it along those lines for v2.
@@ -996,7 +1017,8 @@ static int nbd_negotiate_meta_queries(NBDClient
*client,
meta->exp = nbd_export_find(export_name);
if (meta->exp == NULL) {
return nbd_opt_drop(client, NBD_REP_ERR_UNKNOWN, errp,
- "export '%s' not present", export_name);
+ "export '%s' not present",
+ nbd_truncate_name(export_name));
}
Hmm, maybe instead of assertion, shrink message in
nbd_negotiate_send_rep_verr() too?
This will save us from forgotten (or future) uses of the function.
Truncating in nbd_negotiate_send_rep_verr is arbitrary; it does not have
the context of what makes sense to truncate. With an artificially short
length, and a client request for "longname_from_the_client", the
difference would be between:
export "longnam..." not present
export "longname_from_the_cl...
As a user, I prefer the first form, but truncating in
nbd_negotiate_send_rep_verr is the second form.
Shrinking name is better, as it provides better message on result. But
generally shrink
all two long messages in nbd_negotiate_send_rep_verr() (maybe, together
with error_report())
seems a good thing for me.
What would error_report() accomplish? Logging on the server that we are
truncating a message sent back to the client doesn't help the client.
This is not a case of protocol violation, and a server sending an error
message to the client is not out of the ordinary.
--
Eric Blake, Principal Software Engineer
Red Hat, Inc. +1-919-301-3226
Virtualization: qemu.org | libvirt.org
