Quotas appear to be working fine, and I have an explanation for the results I am seeing. The maildirsize value is correct, even though it does not agree with du, which is also correct.

Maildirsize calculates the number of bytes in a message. du calculates the amount of disk space allocated to the file, used and unused. A small message might contain only 1000 bytes, but it might require 4 disk blocks or 2048 bytes to store the file.

Consequently, if there are thousands of small messages, as there are in this case, actuall disk usage for an account can exceed the soft quota limit by a significant amount, due to internal fragmentation. A file system like reiserfs that does tail packing may help mitigate the impact in this situation.

Thanks for everyone's help. It would be useful if this were included in the documentation on how the maildirsize is calculated. Those who want actual usage to be closer to the intended limit should account for a certain percentage of internal fragmentation based on the intended size and the average expected size of a message when determining the quota (or also set a maximum number of messages).

I noticed an odd problem lately of one account going over quota by nearly 20%.