Use MSR to map framebuffer as WC (write-combining) = huge speed diff on real HW

include/arch/x86.h

@@ -0,0 +1,13 @@
+#ifndef X86_H
+#define X86_H
+
+#include <stdbool.h>
+#include <stdint.h>
+
+uint64_t rdmsr(uint32_t msr);
+void cpuid(uint32_t leaf, uint32_t* eax, uint32_t* ebx, uint32_t* ecx, uint32_t* edx);
+void wrmsr(uint32_t msr, uint64_t value);
+bool x86_has_msr();
+void x86_arch_init();
+
+#endif

src/arch/x86/cpuid.c

@@ -0,0 +1,20 @@
+/*
+ * @author  xamidev <xamidev@riseup.net>
+ * @brief   x86 CPU identification
+ * @license GPL-3.0-only
+ */
+
+#include <stdint.h>
+
+/*
+ * cpuid - Wrapper for CPUID instruction
+ * @leaf: Requested leaf
+ * @eax: EAX register value
+ * @ebx: EBX register value
+ * @ecx: ECX register value
+ * @edx: EDX register value
+*/
+void cpuid(uint32_t leaf, uint32_t* eax, uint32_t* ebx, uint32_t* ecx, uint32_t* edx)
+{
+    __asm__ volatile("cpuid" : "=a"(*eax), "=b"(*ebx), "=c"(*ecx), "=d"(*edx) : "a"(leaf));
+}

src/arch/x86/init.c

@@ -0,0 +1,28 @@
+/*
+ * @author  xamidev <xamidev@riseup.net>
+ * @brief   x86 architecture-dependant initialization
+ * @license GPL-3.0-only
+ */
+
+#include <stdint.h>
+#include <arch/x86.h>
+#include <kernel.h>
+
+/*
+ * x86_arch_init - Initialize x86 CPU structures
+ *
+ * This function is responsible for overriding a PAT entry
+ * (to put the framebuffer area in WC mode) only.
+ *
+ * Later, all architecture-dependant init (GDT, IDT, TSS, ...)
+ * should be initialized here, and separate function pointers
+ * should be set up for each arch.
+ */
+void x86_arch_init()
+{
+    uint64_t pat = rdmsr(0x277);
+    pat &= ~(0xFFULL << 8); // Clear PAT1
+    pat |= (0x01ULL << 8); // PAT1 = 0x01 (WC)
+    wrmsr(0x277, pat);
+    DEBUG("Overrode PAT1 entry to set up Write-Combining");
+}

src/arch/x86/msr.c

@@ -0,0 +1,66 @@
+/*
+ * @author  xamidev <xamidev@riseup.net>
+ * @brief   x86 MSR C wrappers
+ * @description
+ * Wrapper functions to access Model Specific Registers
+ *
+ * @license GPL-3.0-only
+ */
+
+#include <stdint.h>
+#include <stdbool.h>
+#include <arch/x86.h>
+
+/*
+ * rdmsr - Read from MSR
+ * @msr: model specific register number 
+ *
+ * Read a 64-bit word from a Model Specific Register.
+ * Wrapper for the "rdmsr" instruction. It originally
+ * outputs to two 32-bit registers (EDX:EAX), so the
+ * function does the job of uniting them as a 64-bit
+ * value for us.
+ *
+ * Return:
+ * <value> - value read from MSR
+ */
+uint64_t rdmsr(uint32_t msr)
+{
+    uint32_t low;
+    uint32_t high;
+
+    __asm__ volatile("rdmsr" : "=a"(low), "=d"(high) : "c"(msr));
+
+    return ((uint64_t)high << 32) | low;
+}
+
+/*
+ * wrmsr - Write to MSR
+ * @msr: model specific register number 
+ *
+ * Write a 64-bit value to a Model Specific Register.
+ */
+void wrmsr(uint32_t msr, uint64_t value)
+{
+    uint32_t low = (uint32_t)(value & 0xFFFFFFFF);
+    uint32_t high = (uint32_t)(value >> 32);
+
+    __asm__ volatile("wrmsr" : : "c"(msr), "a"(low), "d"(high) : "memory");
+}
+
+/*
+ * x86_has_msr - Test for MSR support
+ *
+ * Checks if CPU supports Model Specific Registers
+ * using CPUID.01h:EDX[bit 5].
+ *
+ * Return:
+ * true  - MSR are supported
+ * false - MSR are not supported 
+ */
+bool x86_has_msr()
+{
+    uint32_t eax, ebx, ecx, edx;
+    cpuid(1, &eax, &ebx, &ecx, &edx);
+    return (edx & (1 << 5)) != 0;
+}

src/kmain.c

@@ -4,6 +4,7 @@
  * @license GPL-3.0-only
  */
 
+#include "arch/x86.h"
 #include <stdbool.h>	
 #include <stddef.h>
 #include <limine.h>
@@ -24,6 +25,7 @@
 #include <config.h>
 #include <io/term/flanterm.h>
 #include <io/term/flanterm_backends/fb.h>
+#include <arch/x86.h>
 
 // Limine version used
 __attribute__((used, section(".limine_requests")))
@@ -112,6 +114,8 @@ void kmain()
 	serial_init();
 	timer_init();
 
+	x86_arch_init();
+
 	boot_mem_display();
 	pmm_init(boot_ctx);
 

src/mem/paging.c

@@ -202,9 +202,9 @@ void paging_init(struct boot_context boot_ctx)
     uint64_t fb_size = fb->pitch * fb->height;
     uint64_t fb_pages = (fb_size + PAGE_SIZE-1)/PAGE_SIZE;
 
-    // Map the framebuffer (with cache-disable & write-through)
+    // Map the framebuffer (PWT set, and no PCD means PAT1 [Write-Combining] for this region)
     for (uint64_t i=0; i<fb_pages; i++) {
-        paging_map_page(kernel_pml4, fb_virt+i*PAGE_SIZE, fb_phys+i*PAGE_SIZE, PTE_WRITABLE | PTE_PCD | PTE_PWT);
+        paging_map_page(kernel_pml4, fb_virt+i*PAGE_SIZE, fb_phys+i*PAGE_SIZE, PTE_WRITABLE | PTE_PWT);
         page_count++;
     }
     DEBUG("Mapped %u pages for framebuffer", page_count);