Add support for cacheless PIM

pim-os/Cargo.toml

@@ -6,6 +6,9 @@ version = "0.1.0"
 edition = "2021"
 forced-target = "aarch64-unknown-none"
 
+[features]
+cacheless = []
+
 [dependencies]
 aarch64-cpu = "9.4.0"
 half = { version = "2.3.1", default-features = false }

pim-os/src/main.rs

@@ -1,8 +1,8 @@
 #![no_std]
 #![no_main]
 
+use aarch64_cpu::asm::barrier;
 use core::{
-    arch::global_asm,
     fmt::Write,
     panic::PanicInfo,
     sync::atomic::{compiler_fence, Ordering},
@@ -15,13 +15,11 @@ use pim::{
 };
 use pim_isa::BankMode;
 use uart::Uart0;
-
+mod boot;
 mod m5ops;
 mod pim;
 mod uart;
 
-global_asm!(include_str!("start.s"));
-
 #[no_mangle]
 pub extern "C" fn entry() -> ! {
     let mut pim_state = PimState::new(&TEST_KERNEL);
@@ -52,6 +50,7 @@ pub extern "C" fn entry() -> ! {
     // Invalidate and flush array just in case
     compute_array.invalidate_flush();
     dummy_array.invalidate_flush();
+    barrier::dsb(barrier::SY);
 
     pim_state.set_bank_mode(BankMode::PimAllBank);
     compute_array.0[0].execute_instruction_read();
@@ -62,6 +61,7 @@ pub extern "C" fn entry() -> ! {
     pim_state.set_bank_mode(BankMode::SingleBank);
 
     compute_array.invalidate();
+    barrier::dsb(barrier::SY);
 
     writeln!(
         &mut uart,

pim-os/src/pim/array.rs

@@ -1,3 +1,4 @@
+use aarch64_cpu::asm::barrier;
 use core::arch::asm;
 use half::f16;
 
@@ -17,8 +18,13 @@ impl Default for BankArray {
 
 impl BankArray {
     pub fn execute_instruction_read(&self) {
-        self.invalidate_single_bank(0);
+        if !cfg!(cacheless) {
+            self.invalidate_single_bank(0);
+            barrier::dsb(barrier::SY);
+        }
+
         self.read_data();
+        barrier::dsb(barrier::SY);
     }
 
     pub fn read_data(&self) {
@@ -30,9 +36,18 @@ impl BankArray {
     }
 
     pub fn execute_instruction_write(&mut self) {
-        self.preload_zero();
+        if !cfg!(cacheless) {
+            self.preload_zero();
+            barrier::dsb(barrier::SY);
+        }
+
         self.write_data();
-        self.invalidate_flush_single_bank(0);
+
+        if !cfg!(cacheless) {
+            self.invalidate_flush_single_bank(0);
+        }
+
+        barrier::dsb(barrier::SY);
     }
 
     pub fn write_data(&mut self) {
@@ -53,7 +68,6 @@ impl BankArray {
         unsafe {
             // Invalidate first bank
             asm!("dc ivac, {val}", val = in(reg) &self.0[idx]);
-            asm!("dsb sy");
         }
     }
 
@@ -67,7 +81,6 @@ impl BankArray {
         unsafe {
             // Invalidate and flush first bank
             asm!("dc civac, {val}", val = in(reg) &self.0[idx]);
-            asm!("dsb sy");
         }
     }
 
@@ -81,7 +94,6 @@ impl BankArray {
         unsafe {
             // Preload first bank
             asm!("dc zva, {val}", val = in(reg) &self.0[idx]);
-            asm!("dsb sy");
         }
     }
 }

pim-os/src/pim/state.rs

@@ -1,10 +1,11 @@
+use crate::boot;
+
 use super::config::PimWriter;
 use pim_isa::{BankMode, Kernel, PimConfig};
 
 #[derive(Debug)]
 pub struct PimState {
     kernel: Kernel,
-    bank_mode: BankMode,
     writer: PimWriter,
 }
 
@@ -12,18 +13,24 @@ impl PimState {
     pub fn new(kernel: &Kernel) -> Self {
         Self {
             kernel: kernel.clone(),
-            bank_mode: BankMode::SingleBank,
             writer: PimWriter,
         }
     }
 
     // TODO return token and return to singlebank when dropped
     pub fn set_bank_mode(&mut self, bank_mode: BankMode) {
-        self.bank_mode = bank_mode;
+        if cfg!(cacheless) {
+            match bank_mode {
+                BankMode::SingleBank => unsafe { boot::set_page_table_cache() },
+                BankMode::AllBank => (),
+                BankMode::PimAllBank => unsafe { boot::set_page_table_non_cache() },
+            }
+        }
+
         self.writer.write(
             serde_json_core::to_string::<PimConfig, 1024>(&PimConfig {
                 kernel: self.kernel.clone(),
-                bank_mode: self.bank_mode,
+                bank_mode,
             })
             .unwrap()
             .as_str(),

pim-os/src/start.s

@@ -1,4 +1,4 @@
-.globl _start
+
 .extern LD_STACK_PTR
 
 // Put a 64-bit value with little endianness.
@@ -36,39 +36,55 @@ BLOCK_1GB (ADDR << 29), 0, 0x74C
 .set ADDR, ADDR+2
 .endr
 
+.align 12
+ttb0_base_non_cache:
+.set ADDR, 0x000
+.rept 0x02
+BLOCK_1GB (ADDR << 29), 0, 0x740
+.set ADDR, ADDR+2
+.endr
+.rept 0x3E
+BLOCK_1GB (ADDR << 29), 0, 0x740
+.set ADDR, ADDR+2
+.endr
+
+.globl _start
 _start:
     ldr     x30, =LD_STACK_PTR
     mov     sp, x30
 
     // Initialize translation table control registers
-    LDR X1, =0x13520 // 64GB space 4KB granularity Inner-shareable. Normal Inner and Outer Cacheable.
-    MSR TCR_EL3, X1
+    ldr x1, =0x13520 // 64GB space 4KB granularity Inner-shareable. Normal Inner and Outer Cacheable.
+    msr tcr_el3, x1
 
-    LDR X1, =0xFF440400
-    MSR MAIR_EL3, X1 // ATTR0 Device-nGnRnE ATTR1 Device. ATTR2 Normal Non-Cacheable. ATTR3 Normal Cacheable.
+    ldr x1, =0xFF440400
+    msr mair_el3, x1 // ATTR0 Device-nGnRnE ATTR1 Device. ATTR2 Normal Non-Cacheable. ATTR3 Normal Cacheable.
 
-    ADR X0, ttb0_base
-    MSR TTBR0_EL3, X0
+    bl set_page_table_cache
 
     // Enable MMU and caches
-    // It is implemented in the CPUECTLR register.
-    MRS X0, S3_1_C15_C2_1
-    ORR X0, X0, #(0x1 << 6) // The SMP bit.
-    MSR S3_1_C15_C2_1, X0
-
-    // Enable caches and the MMU.
-    MRS X0, SCTLR_EL3
-    ORR X0, X0, #(0x1 << 2) // The C bit (data cache).
-    ORR X0, X0, #(0x1 << 12) // The I bit (instruction cache).
-    ORR X0, X0, #0x1 // The M bit (MMU).
-    MSR SCTLR_EL3, X0
-    DSB SY
-    ISB
+    mrs x0, sctlr_el3
+    orr x0, x0, #(0x1 << 2) // The C bit (data cache).
+    orr x0, x0, #(0x1 << 12) // The I bit (instruction cache).
+    orr x0, x0, #0x1 // The M bit (MMU).
+    msr sctlr_el3, x0
+    dsb sy
+    isb
 
     bl entry
 
-.equ PSCI_SYSTEM_OFF, 0x84000008
-.globl system_off
-system_off:
-    ldr     x0, =PSCI_SYSTEM_OFF
-    hvc     #0
+.globl set_page_table_cache
+set_page_table_cache:
+    adr x0, ttb0_base
+    msr ttbr0_el3, x0
+    tlbi alle3
+    isb
+    ret
+
+.globl set_page_table_non_cache
+set_page_table_non_cache:
+    adr x0, ttb0_base_non_cache
+    msr ttbr0_el3, x0
+    tlbi alle3
+    isb
+    ret