[libcpu/arm64] dealing with mem region out of kernel space (#8847)

* [ofw] dealing with mem region out of kernel space

- Fix parameter checking in _out_of_range() that NULL is excluded for
  fixed mapping
- Split page install with a deferred stage to avoid mapping over
  ARCH_EARLY_MAP_SIZE

Signed-off-by: Shell <[email protected]>

* fixup: restrict vstart for using of RT_NULL

---------

Signed-off-by: Shell <[email protected]>

5 files changed, 73 insertions, 21 deletions

diff --git a/components/mm/mm_aspace.c b/components/mm/mm_aspace.c
index 8370bea8d7..acea6cc080 100644
--- a/components/mm/mm_aspace.c
+++ b/components/mm/mm_aspace.c
@@ -643,11 +643,11 @@ static rt_varea_t _varea_create(void *start, rt_size_t size)
 #define _IS_OVERFLOW(start, length) ((length) > (0ul - (uintptr_t)(start)))
 #define _IS_OVERSIZE(start, length, limit_s, limit_sz) (((length) + (rt_size_t)((char *)(start) - (char *)(limit_start))) > (limit_size))
 
-static inline int _not_in_range(void *start, rt_size_t length,
+static inline int _not_in_range(rt_size_t flags, void *start, rt_size_t length,
                                 void *limit_start, rt_size_t limit_size)
 {
     /* assuming (base + length) will not overflow except (0) */
-    int rc = start != RT_NULL
+    int rc = (flags & MMF_MAP_FIXED || start != RT_NULL)
                ? (_IS_OVERFLOW(start, length) || start < limit_start ||
                   _IS_OVERSIZE(start, length, limit_start, limit_size))
                : length > limit_size;
@@ -684,7 +684,7 @@ int rt_aspace_map(rt_aspace_t aspace, void **addr, rt_size_t length,
         LOG_I("%s(%p, %p, %lx, %lx, %lx, %p, %lx): Invalid input",
             __func__, aspace, addr, length, attr, flags, mem_obj, offset);
     }
-    else if (_not_in_range(*addr, length, aspace->start, aspace->size))
+    else if (_not_in_range(flags, *addr, length, aspace->start, aspace->size))
     {
         err = -RT_EINVAL;
         LOG_I("%s(addr:%p, len:%lx): out of range", __func__, *addr, length);
@@ -716,7 +716,7 @@ int rt_aspace_map_static(rt_aspace_t aspace, rt_varea_t varea, void **addr,
     int err;
 
     if (!aspace || !varea || !addr || !mem_obj || length == 0 ||
-        _not_in_range(*addr, length, aspace->start, aspace->size))
+        _not_in_range(flags, *addr, length, aspace->start, aspace->size))
     {
         err = -RT_EINVAL;
         LOG_W("%s: Invalid input", __func__);
@@ -766,9 +766,9 @@ int _mm_aspace_map_phy(rt_aspace_t aspace, rt_varea_t varea,
         LOG_W("%s: not aligned", __func__);
         err = -RT_EINVAL;
     }
-    else if (_not_in_range(hint->limit_start, hint->limit_range_size, aspace->start,
+    else if (_not_in_range(hint->flags, hint->limit_start, hint->limit_range_size, aspace->start,
                       aspace->size) ||
-        _not_in_range(hint->prefer, hint->map_size, aspace->start,
+        _not_in_range(hint->flags, hint->prefer, hint->map_size, aspace->start,
                       aspace->size))
     {
         LOG_W("%s: not in range", __func__);
@@ -892,7 +892,7 @@ int rt_aspace_unmap(rt_aspace_t aspace, void *addr)
         LOG_I("%s: Invalid input", __func__);
         error = -RT_EINVAL;
     }
-    else if (_not_in_range(addr, 1, aspace->start, aspace->size))
+    else if (_not_in_range(MMF_MAP_FIXED, addr, 1, aspace->start, aspace->size))
     {
         LOG_I("%s: %lx not in range of aspace[%lx:%lx]", __func__, addr,
               aspace->start, (char *)aspace->start + aspace->size);
@@ -1041,7 +1041,7 @@ int rt_aspace_unmap_range(rt_aspace_t aspace, void *addr, size_t length)
         LOG_I("%s: Invalid input", __func__);
         error = -RT_EINVAL;
     }
-    else if (_not_in_range(addr, length, aspace->start, aspace->size))
+    else if (_not_in_range(MMF_MAP_FIXED, addr, length, aspace->start, aspace->size))
     {
         LOG_I("%s: %lx not in range of aspace[%lx:%lx]", __func__, addr,
               aspace->start, (char *)aspace->start + aspace->size);
@@ -1069,7 +1069,7 @@ int rt_aspace_unmap_range(rt_aspace_t aspace, void *addr, size_t length)
 }
 
 void *rt_aspace_mremap_range(rt_aspace_t aspace, void *old_address, size_t old_size,
-                                size_t new_size, int flags, void *new_address)
+                             size_t new_size, int flags, void *new_address)
 {
     void *ret = RT_NULL;
 
@@ -1077,7 +1077,8 @@ void *rt_aspace_mremap_range(rt_aspace_t aspace, void *old_address, size_t old_s
     {
         LOG_I("%s: Invalid input", __func__);
     }
-    else if (_not_in_range(old_address, old_size, aspace->start, aspace->size))
+    else if (_not_in_range(MMF_MAP_FIXED, old_address, old_size,
+                           aspace->start, aspace->size))
     {
         LOG_I("%s: %lx not in range of aspace[%lx:%lx]", __func__, old_address,
               aspace->start, (char *)aspace->start + aspace->size);
@@ -1282,8 +1283,8 @@ int rt_aspace_load_page(rt_aspace_t aspace, void *addr, rt_size_t npage)
         err = -RT_ENOENT;
     }
     else if ((char *)addr >= end || (rt_size_t)addr & ARCH_PAGE_MASK ||
-             _not_in_range(addr, npage << ARCH_PAGE_SHIFT, varea->start,
-                           varea->size))
+             _not_in_range(MMF_MAP_FIXED, addr, npage << ARCH_PAGE_SHIFT,
+                           varea->start, varea->size))
     {
         LOG_W("%s: Unaligned parameter or out of range", __func__);
         err = -RT_EINVAL;
@@ -1310,7 +1311,8 @@ int rt_varea_map_page(rt_varea_t varea, void *vaddr, void *page)
         LOG_W("%s: page is not in kernel space", __func__);
         err = -RT_ERROR;
     }
-    else if (_not_in_range(vaddr, ARCH_PAGE_SIZE, varea->start, varea->size))
+    else if (_not_in_range(MMF_MAP_FIXED, vaddr, ARCH_PAGE_SIZE,
+                           varea->start, varea->size))
     {
         LOG_W("%s(%p,%lx): not in range of varea(%p,%lx)", __func__,
             vaddr, ARCH_PAGE_SIZE, varea->start, varea->size);
@@ -1349,7 +1351,8 @@ int rt_varea_map_range(rt_varea_t varea, void *vaddr, void *paddr, rt_size_t len
         LOG_W("%s(%p,%p,%p,%lx): invalid input", __func__, varea, vaddr, paddr, length);
         err = -RT_EINVAL;
     }
-    else if (_not_in_range(vaddr, length, varea->start, varea->size))
+    else if (_not_in_range(MMF_MAP_FIXED, vaddr, length,
+                           varea->start, varea->size))
     {
         LOG_W("%s(%p,%lx): not in range of varea(%p,%lx)", __func__,
             vaddr, length, varea->start, varea->size);
@@ -1382,7 +1385,8 @@ int rt_varea_unmap_range(rt_varea_t varea, void *vaddr, rt_size_t length)
         LOG_W("%s(%p,%p,%lx): invalid input", __func__, varea, vaddr, length);
         err = -RT_EINVAL;
     }
-    else if (_not_in_range(vaddr, length, varea->start, varea->size))
+    else if (_not_in_range(MMF_MAP_FIXED, vaddr, length,
+                           varea->start, varea->size))
     {
         LOG_W("%s(%p,%lx): not in range of varea(%p,%lx)", __func__,
             vaddr, length, varea->start, varea->size);
diff --git a/libcpu/aarch64/common/mmu.c b/libcpu/aarch64/common/mmu.c
index 9583b9d207..106705c677 100644
--- a/libcpu/aarch64/common/mmu.c
+++ b/libcpu/aarch64/common/mmu.c
@@ -51,8 +51,9 @@
 #define MMU_TBL_PAGE_4k_LEVEL  3
 #define MMU_TBL_LEVEL_NR       4
 
+/* restrict virtual address on usage of RT_NULL */
 #ifndef KERNEL_VADDR_START
-#define KERNEL_VADDR_START ARCH_TEXT_OFFSET
+#define KERNEL_VADDR_START 0x1000
 #endif
 
 volatile unsigned long MMUTable[512] __attribute__((aligned(4 * 1024)));
diff --git a/libcpu/aarch64/common/mmu.h b/libcpu/aarch64/common/mmu.h
index f780c5eb97..e2599d2d30 100644
--- a/libcpu/aarch64/common/mmu.h
+++ b/libcpu/aarch64/common/mmu.h
@@ -89,6 +89,10 @@ struct mem_desc
 
 #define ARCH_MAP_FAILED ((void *)0x1ffffffffffff)
 
+#define ARCH_EARLY_MAP_SIZE         (0x40000000)
+/* this is big enough for even 16TB first-time mapping */
+#define ARCH_PAGE_INIT_THRESHOLD    (0x10000000)
+
 #ifndef __ASSEMBLY__
 
 struct rt_aspace;
diff --git a/libcpu/aarch64/common/setup.c b/libcpu/aarch64/common/setup.c
index 5afcebfe8f..27c647bc6e 100644
--- a/libcpu/aarch64/common/setup.c
+++ b/libcpu/aarch64/common/setup.c
@@ -284,6 +284,15 @@ void rt_hw_common_setup(void)
         .end = (rt_size_t)kernel_end,
     }, RT_TRUE);
 
+#ifndef RT_USING_SMART
+    rt_fdt_commit_memregion_early(&(rt_region_t)
+    {
+        .name = "null",
+        .start = (rt_size_t)RT_NULL,
+        .end = (rt_size_t)RT_NULL + ARCH_PAGE_SIZE,
+    }, RT_TRUE);
+#endif /* !RT_USING_SMART */
+
     if (rt_fdt_prefetch(fdt_ptr))
     {
         /* Platform cannot be initialized */
@@ -332,7 +341,10 @@ void rt_hw_common_setup(void)
     if (!rt_fdt_commit_memregion_request(&mem_region, &mem_region_nr, RT_FALSE))
     {
         rt_ubase_t best_offset = ~0UL;
-        rt_region_t *usable_mem_region = mem_region, *page_region = RT_NULL, init_page_region = { 0 };
+        rt_region_t *usable_mem_region = mem_region, *page_region = RT_NULL;
+        rt_region_t init_page_region = { 0 };
+        rt_region_t defer_hi = { 0 };
+        rt_err_t error;
 
         LOG_I("Usable memory:");
 
@@ -369,6 +381,16 @@ void rt_hw_common_setup(void)
 
         RT_ASSERT(page_region != RT_NULL);
 
+        /* don't map more than ARCH_EARLY_MAP_SIZE */
+        if (page_region->end - page_region->start > ARCH_PAGE_INIT_THRESHOLD)
+        {
+            defer_hi.name = page_region->name;
+            defer_hi.end = page_region->end;
+            defer_hi.start = RT_ALIGN_DOWN(page_region->start + ARCH_PAGE_INIT_THRESHOLD,
+                                           ARCH_SECTION_SIZE);
+            page_region->end = defer_hi.start;
+        }
+
         init_page_region.start = page_region->start - PV_OFFSET;
         init_page_region.end = page_region->end - PV_OFFSET;
 
@@ -389,13 +411,34 @@ void rt_hw_common_setup(void)
 
         mem_region = usable_mem_region;
 
+        if (defer_hi.start)
+        {
+            /* to virt address */
+            init_page_region.start = defer_hi.start - PV_OFFSET;
+            init_page_region.end = defer_hi.end - PV_OFFSET;
+            error = rt_page_install(init_page_region);
+
+            if (error)
+            {
+                LOG_W("Deferred page installation FAILED:");
+                LOG_W("  %-*.s [%p, %p]", RT_NAME_MAX,
+                    defer_hi.name, defer_hi.start, defer_hi.end);
+            }
+            else
+            {
+                LOG_I("Deferred page installation SUCCEED:");
+                LOG_I("  %-*.s [%p, %p]", RT_NAME_MAX,
+                      defer_hi.name, defer_hi.start, defer_hi.end);
+            }
+        }
+
         for (int i = 0; i < mem_region_nr; ++i, ++mem_region)
         {
             if (mem_region != page_region && mem_region->name)
             {
-                mem_region->start -= PV_OFFSET;
-                mem_region->end -= PV_OFFSET;
-                rt_page_install(*mem_region);
+                init_page_region.start = mem_region->start - PV_OFFSET;
+                init_page_region.end = mem_region->end - PV_OFFSET;
+                rt_page_install(init_page_region);
             }
         }
     }
diff --git a/libcpu/aarch64/cortex-a/entry_point.S b/libcpu/aarch64/cortex-a/entry_point.S
index b93978ea6e..361c9d4f67 100644
--- a/libcpu/aarch64/cortex-a/entry_point.S
+++ b/libcpu/aarch64/cortex-a/entry_point.S
@@ -278,7 +278,7 @@ init_mmu_early:
     get_phy x1, .early_tbl1_page
 
     get_pvoff x2 x3
-    ldr     x2, =0x40000000         /* Map 1G memory for kernel space */
+    ldr     x2, =ARCH_EARLY_MAP_SIZE    /* Map 1G memory for kernel space */
     bl      rt_hw_mem_setup_early
 
     b       enable_mmu_early