A Total Overhaul on the Lunaix's Virtual Memory Model (#26)

[lunaix-os.git] / lunaix-os / kernel / mm / vmap.c

diff --git a/lunaix-os/kernel/mm/vmap.c b/lunaix-os/kernel/mm/vmap.c
index f805adc6b13673c2c6595306c4d29bce5d4f454b..02711fb3f34507094d2e4badb64b3a72782e93f2 100644 (file)
--- a/lunaix-os/kernel/mm/vmap.c
+++ b/lunaix-os/kernel/mm/vmap.c
@@ -2,143 +2,70 @@
 #include <lunaix/mm/valloc.h>
 #include <lunaix/mm/vmm.h>
 #include <lunaix/spike.h>
+#include <lunaix/syslog.h>
 
 #include <sys/mm/mempart.h>
 
 static ptr_t start = VMAP;
+static volatile ptr_t prev_va = 0;
 
-void*
-vmap(ptr_t paddr, size_t size, pt_attr attr, int flags)
+static pte_t*
+__alloc_contig_ptes(pte_t* ptep, size_t base_sz, int n)
 {
-    // next fit
-    assert_msg((paddr & 0xfff) == 0, "vmap: bad alignment");
-    size = ROUNDUP(size, PG_SIZE);
-
-    ptr_t current_addr = start;
-    size_t examed_size = 0, wrapped = 0;
-    x86_page_table* pd = (x86_page_table*)L1_BASE_VADDR;
-
-    while (!wrapped || current_addr < start) {
-        size_t l1inx = L1_INDEX(current_addr);
-        if (!(pd->entry[l1inx])) {
-            // empty 4mb region
-            examed_size += MEM_4M;
-            current_addr = (current_addr & 0xffc00000) + MEM_4M;
-        } else {
-            x86_page_table* ptd = (x86_page_table*)(L2_VADDR(l1inx));
-            size_t i = L2_INDEX(current_addr), j = 0;
-            for (; i < PG_MAX_ENTRIES && examed_size < size; i++, j++) {
-                if (!ptd->entry[i]) {
-                    examed_size += PG_SIZE;
-                } else if (examed_size) {
-                    // found a discontinuity, start from beginning
-                    examed_size = 0;
-                    j++;
-                    break;
-                }
-            }
-            current_addr += j << 12;
-        }
-
-        if (examed_size >= size) {
-            goto done;
+    int _n     = 0;
+    size_t sz  = L0T_SIZE;
+    ptr_t va   = page_addr(ptep_pfn(ptep));
+
+    ptep = mkl0tep(ptep);
+
+    while (_n < n && va < VMAP_END) {
+        pte_t pte = *ptep;
+        if (pte_isnull(pte)) {
+            _n += sz / base_sz;
+        } 
+        else if ((sz / LEVEL_SIZE) < base_sz) {
+            _n = 0;
         }
-
-        if (current_addr >= VMAP_END) {
-            wrapped = 1;
-            examed_size = 0;
-            current_addr = VMAP;
+        else {
+            sz = sz / LEVEL_SIZE;
+            ptep = ptep_step_into(ptep);
+            continue;
         }
-    }
-
-    return NULL;
 
-done:
-    ptr_t alloc_begin = current_addr - examed_size;
-    start = alloc_begin + size;
-
-    if ((flags & VMAP_NOMAP)) {
-        for (size_t i = 0; i < size; i += PG_SIZE) {
-            vmm_set_mapping(VMS_SELF, alloc_begin + i, -1, 0, 0);
+        if (ptep_vfn(ptep) + 1 == LEVEL_SIZE) {
+            ptep = ptep_step_out(++ptep);
+            va += sz;
+            
+            sz = sz * LEVEL_SIZE;
+            continue;
         }
-
-        return (void*)alloc_begin;
-    }
-
-    for (size_t i = 0; i < size; i += PG_SIZE) {
-        vmm_set_mapping(VMS_SELF, alloc_begin + i, paddr + i, attr, 0);
-        pmm_ref_page(paddr + i);
+        
+        va += sz;
+        ptep++;
     }
 
-    return (void*)alloc_begin;
-}
-
-/*
-    This is a kernel memory region that represent a contiguous virtual memory
-   address such that all memory allocation/deallocation can be concentrated
-   into a single big chunk, which will help to mitigate the external
-   fragmentation in the VMAP address domain. It is significant if our
-   allocation granule is single page or in some use cases.
-
-    XXX (vmap_area)
-    A potential performance improvement on pcache? (need more analysis!)
-        -> In exchange of a fixed size buffer pool. (does it worth?)
-*/
-
-struct vmap_area*
-vmap_varea(size_t size, pt_attr attr)
-{
-    ptr_t start = (ptr_t)vmap(0, size, attr ^ PG_PRESENT, VMAP_NOMAP);
-
-    if (!start) {
+    if (va >= VMAP_END) {
         return NULL;
     }
 
-    struct vmap_area* varea = valloc(sizeof(struct vmap_area));
-    *varea =
-      (struct vmap_area){ .start = start, .size = size, .area_attr = attr };
-
-    return varea;
+    va -= base_sz * _n;
+    assert(prev_va < va);
+    
+    prev_va = va;
+    return mkptep_va(ptep_vm_mnt(ptep), va);
 }
 
 ptr_t
-vmap_area_page(struct vmap_area* area, ptr_t paddr, pt_attr attr)
+vmap_ptes_at(pte_t pte, size_t lvl_size, int n)
 {
-    ptr_t current = area->start;
-    size_t bound = current + area->size;
+    pte_t* ptep = mkptep_va(VMS_SELF, start);
+    ptep = __alloc_contig_ptes(ptep, lvl_size, n);
 
-    while (current < bound) {
-        x86_pte_t* pte =
-          (x86_pte_t*)(L2_VADDR(L1_INDEX(current)) | L2_INDEX(current));
-        if (PG_IS_PRESENT(*pte)) {
-            current += PG_SIZE;
-            continue;
-        }
-
-        *pte = NEW_L2_ENTRY(attr | PG_PRESENT, paddr);
-        cpu_flush_page(current);
-        break;
-    }
-
-    return current;
-}
-
-ptr_t
-vmap_area_rmpage(struct vmap_area* area, ptr_t vaddr)
-{
-    ptr_t current = area->start;
-    size_t bound = current + area->size;
-
-    if (current > vaddr || vaddr > bound) {
+    if (!ptep) {
         return 0;
     }
 
-    x86_pte_t* pte =
-      (x86_pte_t*)(L2_VADDR(L1_INDEX(current)) | L2_INDEX(current));
-    ptr_t pa = PG_ENTRY_ADDR(*pte);
-
-    *pte = NEW_L2_ENTRY(0, -1);
-    cpu_flush_page(current);
+    vmm_set_ptes_contig(ptep, pte, lvl_size, n);
 
-    return pa;
+    return page_addr(ptep_pfn(ptep));
 }
\ No newline at end of file