pmap tlb shootdowns

core/percpu.c

@@ -3,12 +3,11 @@
 #include <zinnia/cmdline.h>
 #include <zinnia/compiler.h>
 #include <zinnia/percpu.h>
+#include <zinnia/pmap.h>
 #include <zinnia/utils.h>
+#include <zinnia/vma.h>
 #include <arch/percpu.h>
 
-extern uint8_t __ld_percpu_start[];
-extern uint8_t __ld_percpu_end[];
-
 static _Atomic(size_t) num_cpus = 1;
 
 static size_t max_cpus = SIZE_MAX;
@@ -25,20 +24,33 @@ struct percpu percpu_bsp = {
 };
 
 struct percpu* percpu_new() {
-    const size_t size = __ld_percpu_end - __ld_percpu_start;
-    const size_t num_pages = ALIGN_UP(size, arch_mem_page_size()) / arch_mem_page_size();
-
-    // TODO: Map pages past the end of the kernel.
-    // TODO: Document how allocations work.
-
-    phys_t mem;
-    if (mem_phys_alloc(num_pages, 0, &mem))
+    const size_t pgsz = arch_mem_page_size();
+    const size_t stride = percpu_stride();
+    const size_t id = atomic_load(&num_cpus, ATOMIC_RELAXED);
+    if (id >= (UINTPTR_MAX - (uintptr_t)__ld_percpu_start) / stride)
         return nullptr;
 
-    struct percpu* cpu = HHDM_PTR(mem);
+    phys_t mem;
+    if (mem_phys_alloc(stride / pgsz, 0, &mem))
+        return nullptr;
+
+    const uintptr_t base = (uintptr_t)percpu_cpu(id);
+    for (size_t offset = 0; offset < stride; offset += pgsz) {
+        if (pmap_map(&kernel_vma.pmap, base + offset, mem + offset, PTE_READ | PTE_WRITE, CACHE_ANY)) {
+            while (offset) {
+                offset -= pgsz;
+                pmap_unmap(&kernel_vma.pmap, base + offset);
+            }
+            mem_phys_free(mem, stride / pgsz);
+            return nullptr;
+        }
+    }
+
+    struct percpu* cpu = (struct percpu*)base;
     cpu->self = cpu;
-    cpu->id = atomic_fetch_add(&num_cpus, 1, ATOMIC_RELAXED);
+    cpu->id = id;
     cpu->online = false;
+    atomic_store(&num_cpus, id + 1, ATOMIC_RELEASE);
     return cpu;
 }
 
@@ -46,6 +58,10 @@ size_t percpu_max_cpus() {
     return max_cpus;
 }
 
+size_t percpu_count() {
+    return atomic_load(&num_cpus, ATOMIC_RELAXED);
+}
+
 struct percpu* percpu_get() {
     return arch_percpu_get();
 }