ssumpf / foc Goto Github PK

Currently, serial input gets only fetched by the kernel when the '-serial_esc' option is set. But when using the kernel debugger console as a terminal backend in Genode, the '-serial_esc' option prevents the use of the ESC key in applications and causes problems when ESC-sequences are received. So, an adaptation is needed to fetch serial input also when the '-serial_esc' option is not set.

When building Fiasco.OC for Arndale with GCC 4.9, an overlap occurs at boot time:

New region for list regions:        [ 40090000,  40095fdf] {     5fe0} Sigma0 sigma0
overlaps with:              [ 40002000,  40093fff] {    92000} Kern   -05-17-22-00/arndale/foc_arndale/kernel/fiasco.oc/fiasco

In the Thread::user_single_step() function, the statement

regs()->flags(enable ? user_flags() | EFLAGS_TF : user_flags() & ~EFLAGS_TF);

can clear the interrupt flag, because with the Ex_regs: Change 'user_flags' patch applied, the user_flags() function does not return the previous CPU flags anymore, but the thread state.

While a CPU executes the idle thread, it receives no cross-CPU TLB flush requests. Therefore, the TLB needs to get flushed when leaving the idle thread to compensate for any missed flush requests.

See also http://www.mail-archive.com/[email protected]/msg05881.html

When the Genode exception handler receives an exception, it saves the thread state given to it by the kernel. When replying to the exception, the saved (and possibly modified via Cpu_session::state()) state is given to the kernel again, who applies it to the thread.

There are some situations where the thread state in the kernel changes without modification of the saved exception handler state, for example when Cpu_session::single_step() gets called. In those cases, the exception handler restores an outdated state. For x86 I observed those unnoticed changes with the single-step flag in Trap_state::_flags and on ARM with Trap_state::'pf_address, Trap_state::error_code, Trap_state:::tpidruro and Trap_state::km_lr.

It's not clear yet what exactly causes each of the unnoticed state changes, so a quick fix would be to not apply these parts of the kernel thread state from the exception reply.

Related to genodelabs/genode#665

In trustzone normal side execution must not interfere with secure side. Each execution mode has assigned a type of interrupt, the normal side only uses the IRQ and secure side the FIQ.
By default when not using TrustZone Secure Side, Fiasco.OC try to handle (enable/disable) both interrupts, the IRQ and FIQ. We think that there is a problem, when the kernel running on TrustZone Normal Side try to write the FIQ bit, and interrupts remains disabled (the operation is cancelled).

In the RaspberryPI platform the USB interrupts must be filtered to avoid the overload of 8000 irqs per second.