As tonfa advised checking for basestring instead of str because unicode is 
also valid in this case.

diff --git a/mercurial/hgweb/protocol.py b/mercurial/hgweb/protocol.py
--- a/mercurial/hgweb/protocol.py
+++ b/mercurial/hgweb/protocol.py
@@ -179,8 +179,8 @@
             raise ErrorResponse(HTTP_OK, inst)
         except (OSError, IOError), inst:
             error = getattr(inst, 'strerror', 'Unknown error')
-            if not isinstance(error, str):
-                error = 'Error: %s' % str(error)
+            if not isinstance(error, basestring):
+                error = 'Error: %s' % inst.__class__.__name__
             if inst.errno == errno.ENOENT:
                 code = HTTP_NOT_FOUND
             else:

I would do this, as an error message of "Error: None" would not be very 
useful. A way to actually test this would be nice also.

diff --git a/mercurial/hgweb/protocol.py b/mercurial/hgweb/protocol.py
--- a/mercurial/hgweb/protocol.py
+++ b/mercurial/hgweb/protocol.py
@@ -180,7 +180,7 @@
         except (OSError, IOError), inst:
             error = getattr(inst, 'strerror', 'Unknown error')
             if not isinstance(error, str):
-                error = 'Error: %s' % str(error)
+                error = 'Error: %s' % inst.__class__.__name__
             if inst.errno == errno.ENOENT:
                 code = HTTP_NOT_FOUND
             else:

Changing to testing

See http://hg.intevation.org/mercurial/crew/rev/5dc09507b90e
(hgweb: fix attribute error in error response (issue2060))

adding a str call would work.

However in the case below it is happening because error = None. I think the
intention in this situation was to have text = 'Unknown error', especially
as returning 'None' as an error will confuse people.

As a fix I would suggest the following (although this hasn't been properly
tested it should work):

diff -r ef3668450cd0 mercurial/hgweb/protocol.py
--- a/mercurial/hgweb/protocol.py	Tue Mar 30 13:09:25 2010 -0500
+++ b/mercurial/hgweb/protocol.py	Thu Apr 01 15:30:11 2010 +0100
@@ -179,6 +179,8 @@
             raise ErrorResponse(HTTP_OK, inst)
         except (OSError, IOError), inst:
             error = getattr(inst, 'strerror', 'Unknown error')
+            if type(error) != str:
+                error = 'Unknown error: %s' % str(error)
             if inst.errno == errno.ENOENT:
                 code = HTTP_NOT_FOUND
             else:

We probably need to add a str() call, but it's weird that strerror isn't a 
string...

text = str(error).replace(repo.root + os.path.sep, '')

This bug should happen in 1.4 too.

This is resulting in a Error 500, and it occurred several times over the
last week. Each time, I had to restart the server.

I am running an almost-up-to-date stable changeset.

[Wed Feb 24 11:30:32 2010] [error] [client 10.0.8.53] mod_wsgi (pid=20511):
Exception occurred processing WSGI script '/var/hg/hgwebdir.wsgi'.
[Wed Feb 24 11:30:32 2010] [error] [client 10.0.8.53] Traceback (most recent
call last):
[Wed Feb 24 11:30:32 2010] [error] [client 10.0.8.53]   File
"/usr/local/lib/python2.4/site-packages/mercurial/hgweb/hgwebdir_mod.py",
line 103, in __call__
[Wed Feb 24 11:30:32 2010] [error] [client 10.0.8.53]     return
self.run_wsgi(req)
[Wed Feb 24 11:30:32 2010] [error] [client 10.0.8.53]   File
"/usr/local/lib/python2.4/site-packages/mercurial/hgweb/hgwebdir_mod.py",
line 160, in run_wsgi
[Wed Feb 24 11:30:32 2010] [error] [client 10.0.8.53]     return
hgweb(repo).run_wsgi(req)
[Wed Feb 24 11:30:32 2010] [error] [client 10.0.8.53]   File
"/usr/local/lib/python2.4/site-packages/mercurial/hgweb/hgweb_mod.py", line
123, in run_wsgi
[Wed Feb 24 11:30:32 2010] [error] [client 10.0.8.53]     return
method(self.repo, req)
[Wed Feb 24 11:30:32 2010] [error] [client 10.0.8.53]   File
"/usr/local/lib/python2.4/site-packages/mercurial/hgweb/protocol.py", line
196, in unbundle
[Wed Feb 24 11:30:32 2010] [error] [client 10.0.8.53]     text =
error.replace(repo.root + os.path.sep, '')
[Wed Feb 24 11:30:32 2010] [error] [client 10.0.8.53] AttributeError:
'NoneType' object has no attribute 'replace'

I have a buildbot hook sending lot of twisted warnings about Signals in the
logs also, might be related. I don't think so though.