Handle invalid UTF-8 byte sequences gracefully by replacing them with 0xFFFD1425-fix-graceful-surrogate-handling

CouchDB's Erlang JSON parser allows storing of invalid UTF-8 byte sequences.
The Query Server inside CouchDB fails upon encountering these byte sequences.
The view process fails for the current batch of document updates. The result
is that the view is invariably broken. Only removing the document in question
solves this otherwise, but finding that is hard as the `log()` inside the
Query Server dies with the invalid byte sequence because our protocol is
synchronous and map results an `log()` messages generated therein are
submitted together.

This patch replaces invalid bytes with the the surrogate chacracter 0xFFFD.

Closes COUCHDB-1425.

Patch by Sam Rijs <recv@awesan.de> and Paul Davis.

Eventually, this should be fixed at the HTTP level, so that no documents
with invalid byte sequences can be written to CouchDB. The jiffy encoder
we'll get with BigCouch will do that for us. This is a fix for the releases
until then.

2 files changed, 17 insertions, 13 deletions

diff --git a/THANKS.in b/THANKS.in
index 4ebf3f015..db0ac07d1 100644
--- a/THANKS.in
+++ b/THANKS.in
@@ -94,6 +94,7 @@ suggesting improvements or submitting changes. Some of these people are:
  * Fedor Indutny <fedor@indutny.com>
  * Tim Blair
  * Tady Walsh <hello@tady.me>
+ * Sam Rijs <recv@awesam.de>
 # Authors from commit 6c976bd and onwards are auto-inserted. If you are merging
 # a commit from a non-committer, you should not add an entry to this file. When
 # `bootstrap` is run, the actual THANKS file will be generated.
diff --git a/src/couchdb/priv/couch_js/utf8.c b/src/couchdb/priv/couch_js/utf8.c
index d60642671..2d23cc204 100644
--- a/src/couchdb/priv/couch_js/utf8.c
+++ b/src/couchdb/priv/couch_js/utf8.c
@@ -66,24 +66,31 @@ enc_charbuf(const jschar* src, size_t srclen, char* dst, size_t* dstlenp)
         c = *src++;
         srclen--;
 
-        if((c >= 0xDC00) && (c <= 0xDFFF)) goto bad_surrogate;
-        
-        if(c < 0xD800 || c > 0xDBFF)
+        if(c <= 0xD7FF || c >= 0xE000)
         {
-            v = c;
+            v = (uint32) c;
         }
-        else
+        else if(c >= 0xD800 && c <= 0xDBFF)
         {
             if(srclen < 1) goto buffer_too_small;
             c2 = *src++;
             srclen--;
-            if ((c2 < 0xDC00) || (c2 > 0xDFFF))
+            if(c2 >= 0xDC00 && c2 <= 0xDFFF)
+            {
+                v = (uint32) (((c - 0xD800) << 10) + (c2 - 0xDC00) + 0x10000);
+            }
+            else
             {
-                c = c2;
-                goto bad_surrogate;
+                // Invalid second half of surrogate pair
+                v = (uint32) 0xFFFD;
             }
-            v = ((c - 0xD800) << 10) + (c2 - 0xDC00) + 0x10000;
         }
+        else
+        {
+            // Invalid first half surrogate pair
+            v = (uint32) 0xFFFD;
+        }
+
         if(v < 0x0080)
         {
             /* no encoding necessary - performance hack */
@@ -109,10 +116,6 @@ enc_charbuf(const jschar* src, size_t srclen, char* dst, size_t* dstlenp)
     *dstlenp = (origDstlen - dstlen);
     return JS_TRUE;
 
-bad_surrogate:
-    *dstlenp = (origDstlen - dstlen);
-    return JS_FALSE;
-
 buffer_too_small:
     *dstlenp = (origDstlen - dstlen);
     return JS_FALSE;