gtcasl / qsim Goto Github PK

Setup script update did not update the example files to pick up the images properly. Please fix.

Just put it up here for discussion The qsim->run(cpu_id, n) will always execute on cpu 0 as there is only one qemu instance for all cores. It is a big impact on the spx frontend to fetch instruction and communicate with qsim though.

The busybox makefile uses "qemu-aarch64" to generate the cpio file, which might not be installed in the system.

'qtm' example in qsim/examples/x86/ gives a segmentation fault at a pthread_cond_wait from Qemu code (qemu_cond_wait at qemu/util/qemu-thread-posix.c).

The newest version (v2.6) of QSim irregularly hangs on the middle of "make tests".
I'm using Amazon EC2's c3.2xlarge instance with Ubuntu-14.04 LTS AMI.

When I run "./sudo_setup.sh", the script hangs at the line "make tests"
after stdout "Copying benchmark binary...".

I debugged it a little, and found that it hangs after the last magic callback with "0xc75c0020", which indicates the context switch (?).

Could you let me know what was wrong?

The setup.sh script removes the dependency check for qemu, and thus it is needed to use "git submodule --init --recursive" command to include all possible libraries inside qemu directory.

It is useful in an interactive environment to know when the system shutdown/reboot happens to cleanly close resources like open trace files.

manifold used to use booted() to determine whether qsim completes an application as follow:
if(!qsim_osd->booted(core_id)) {
queue_item.cb_type = Qsim::QueueItem::TERMINATED;
}

The current version does not update running[i] as before, so an app_end callback is needed in spx to exit the loop.

Let INST be the number of instructions passed to osd.run(). Take barnes.tar for example in a 4-core configuration. When INST < ~5000, a maximum of 2 cores are assigned to barnes, while when INST > 5000, all the 4 cores are active.

Pranith and I have a discussion on how manifold manipulates the queueItem for spx component. Pranith suggested me to create a NEW queueItem inside the manifold and leave original one in qsim unchanged. Yet the existing implementation of queueItem uses "struct", which is hard to extend/add new members/functions by the wrappers. I have to duplicate the queueItem and add new modifications on top of it, kind of ugly. I think we need to rethink how to implement queueItem in a more effective way, say virtual base class.

The reg_cb fix might cause some problem in run(cpu_id, inst). Please refer to my comment to #70.

command line:
./tester 1 ../state.1.a64 arm64/memory.tar

qemu output:
bling callback generation for pid 37.
[ 7.900271 ] memory[37]: unhandled level 2 translation fault (11) at 0x00000000, esr 0x90000006
[ 7.914421 ] pgd = ffffffc0198dc000
[ 7.917219 ] [00000000] *pgd=00000000598de003, *pud=00000000598de003, *pmd=0000000000000000
[ 7.925618 ]
[ 7.928301 ] CPU: 0 PID: 37 Comm: memory Not tainted 4.1.15 #2
[ 7.930261 ] Hardware name: linux,dummy-virt (DT)
[ 7.935133 ] task: ffffffc008cbb000 ti: ffffffc0198f0000 task.ti: ffffffc0198f0000
[ 7.939697 ] pc : [<00000000004009f4>] lr : [<00000000004009ec>] pstate: 60000000
[ 7.945314 ] sp : 0000007fc452b690
[ 7.947019 ] x29: 0000007fc452b690 x28: 0000000000000000
[ 7.950862 ] x27: 0000000000000000 x26: 0000000000000000
[ 7.957142 ] x25: 0000000000000000 x24: 00000000004001a0
[ 7.960651 ] x23: 0000000000000000 x22: 00000000004920d8
[ 7.964208 ] x21: 0000000000492020 x20: 0000000000000000
[ 7.968416 ] x19: 00000000004001a0 x18: 0000000000000000
[ 7.974541 ] x17: 000000000055b000 x16: 0000000000560000
[ 7.979257 ] x15: 0000000000560000 x14: 000000000055a758
[ 7.984332 ] x13: ffffff0000000000 x12: 000000000055d5a0
[ 7.988900 ] x11: 0000000000000003 x10: 0101010101010101
[ 7.995098 ] x9 : 0000000000000001 x8 : 000000000000001c
[ 7.997843 ] x7 : 0000007fc452b578 x6 : 000000000055d800
[ 8.003978 ] x5 : 0000000000000003 x4 : 0000000012078620
[ 8.008771 ] x3 : 0000000012078610 x2 : 0000000000000000
[ 8.014221 ] x1 : 0000000012078610 x0 : 00000000aaaaaaaa
[ 8.018137 ]
Segmentation fault
--- program exit, will shutdown shortly... ---
QEMU: Terminated

The patch file generates a rej file for kernel/sched/core.c.rej. Might causes OS problem for idle() test.

--- kernel/sched/core.c
+++ kernel/sched/core.c
@@ -2312,7 +2322,15 @@ context_switch(struct rq *rq, struct task_struct *prev,
 {
 	struct mm_struct *mm, *oldmm;
 
+	/* tell qsim the pid of the next task or that it is idle
+	 */
+	if (IS_ENABLED(CONFIG_X86) || IS_ENABLED(CONFIG_ARM64)) {
+		if (next == rq->idle)
+			do_cpuid(0x1d1e1d1e);
+		else
+			do_cpuid(0xc75c0000 | (u16)task_tgid_nr(next));
+	}
+
 	prepare_task_switch(rq, prev, next);
 
 	mm = next->mm;
-- 

The only blocker is having a cross compiler for ARM64. Clang natively supports cross compilation, investigate that.

Tag qsim messages with [qsim] so that they do not mix with output from the benchmark.

using the tester framework, I observed some of the mem_cbs and reg_cbs are missing. Need further investigation.

Upstream qemu has release version 2.5 in december. This is relevant to qsim since there are quite a few fixes for deterministic execution in qemu. Plan to upgrade to 2.5 version.

People complained that we are downloading to parent directory instead of embedding in current directory.

Please fix.

Setup script pulls qemu and capstone from my personal repos. Update this.

While at it, fix arguments in examples.

It is hard to create static executables for all benchmarks. Try to explore a way to create a state file with necessary libraries to run dynamic executables.

To keep naming consistency, the QSIM_ARM64_ENDING in qsim-arm64-regs.h should be renamed to QSIM_A64_N_REGS according to qsim-x86-regs.h. Same applied to qsim-arm-regs.h.

When running an application in interactive mode, the get_tid function always returns 0. This is due to the use of unmodified kernel in the x86_64 image.

Build currently fails.

QEMU has a record-replay feature which records the non-deterministic events which affect the program execution. To reliably reproduce an execution, we need to use this feature in qsim.