Eu encontrei o motivo. Problema foi causado pela configuração padrão do módulo do Apache mod_reqtimeout:
<IfModule reqtimeout_module>
# mod_reqtimeout limits the time waiting on the client to prevent an
# attacker from causing a denial of service by opening many connections
# but not sending requests. This file tries to give a sensible default
# configuration, but it may be necessary to tune the timeout values to
# the actual situation. Note that it is also possible to configure
# mod_reqtimeout per virtual host.
# Wait max 20 seconds for the first byte of the request line+headers
# From then, require a minimum data rate of 500 bytes/s, but don't
# wait longer than 40 seconds in total.
# Note: Lower timeouts may make sense on non-ssl virtual hosts but can
# cause problem with ssl enabled virtual hosts: This timeout includes
# the time a browser may need to fetch the CRL for the certificate. If
# the CRL server is not reachable, it may take more than 10 seconds
# until the browser gives up.
RequestReadTimeout header=20-40,minrate=500
# Wait max 10 seconds for the first byte of the request body (if any)
# From then, require a minimum data rate of 500 bytes/s
RequestReadTimeout body=10,minrate=500
</IfModule>
Eu acho que os clientes BlackBerry foram mais afetados porque o envio do corpo da solicitação via infra-estrutura RIM BIS leva mais tempo.
Defina o valor para 100 segundos e monitore se os clientes ainda são afetados.