Kernel: Handle some Heap edge cases
This commit is contained in:
parent
2614437ba0
commit
821d065eba
|
@ -9,28 +9,40 @@ namespace Kernel::Memory
|
|||
|
||||
PhysicalRange::PhysicalRange(paddr_t start, size_t size)
|
||||
{
|
||||
ASSERT(start + size > (paddr_t)g_kernel_end);
|
||||
// We can't use the memory ovelapping with kernel
|
||||
if (start + size <= (paddr_t)g_kernel_end)
|
||||
return;
|
||||
|
||||
// Align start to page boundary and after the kernel memory
|
||||
m_start = BAN::Math::max(start, (paddr_t)g_kernel_end);
|
||||
if (auto rem = m_start % PAGE_SIZE)
|
||||
m_start += PAGE_SIZE - rem;
|
||||
|
||||
if (size <= m_start - start)
|
||||
return;
|
||||
|
||||
// Align size to page boundary
|
||||
m_size = size - (m_start - start);
|
||||
if (auto rem = m_size % PAGE_SIZE)
|
||||
m_size -= rem;
|
||||
|
||||
// We need atleast 2 pages
|
||||
m_total_pages = m_size / PAGE_SIZE;
|
||||
if (m_total_pages <= 1)
|
||||
return;
|
||||
|
||||
// FIXME: if total pages is just over multiple of (PAGE_SIZE / sizeof(node)) we might make
|
||||
// couple of pages unallocatable
|
||||
m_total_pages = m_size / PAGE_SIZE;
|
||||
m_list_pages = BAN::Math::div_round_up<uint64_t>(m_total_pages * sizeof(node), PAGE_SIZE);
|
||||
m_reservable_pages = m_total_pages - m_list_pages;
|
||||
|
||||
MMU::get().allocate_range(m_start, m_list_pages * PAGE_SIZE, MMU::Flags::Present);
|
||||
MMU::get().allocate_range(m_start, m_list_pages * PAGE_SIZE, MMU::Flags::ReadWrite | MMU::Flags::Present);
|
||||
|
||||
// Initialize page list so that every page points to the next one
|
||||
node* page_list = (node*)m_start;
|
||||
|
||||
ASSERT((paddr_t)&page_list[m_reservable_pages - 1] <= m_start + m_size);
|
||||
|
||||
for (uint64_t i = 0; i < m_reservable_pages; i++)
|
||||
page_list[i] = { page_list + i - 1, page_list + i + 1 };
|
||||
page_list[ 0 ].next = nullptr;
|
||||
|
@ -128,19 +140,22 @@ namespace Kernel::Memory
|
|||
|
||||
if (mmmt->type == 1)
|
||||
{
|
||||
// We can't use the memory ovelapping with kernel
|
||||
if (mmmt->base_addr + mmmt->length > (paddr_t)g_kernel_end)
|
||||
MUST(m_physical_ranges.push_back({ mmmt->base_addr, mmmt->length }));
|
||||
PhysicalRange range(mmmt->base_addr, mmmt->length);
|
||||
if (range.usable_pages() > 0)
|
||||
MUST(m_physical_ranges.push_back(range));
|
||||
}
|
||||
|
||||
i += mmmt->size + sizeof(uint32_t);
|
||||
}
|
||||
|
||||
size_t total = 0;
|
||||
for (auto& range : m_physical_ranges)
|
||||
{
|
||||
size_t bytes = range.usable_pages() * PAGE_SIZE;
|
||||
dprintln("RAM {8H}->{8H}, {} pages ({}.{} MB)", range.usable_start(), range.usable_end(), range.usable_pages(), bytes / (1 << 20), bytes % (1 << 20) * 1000 / (1 << 20));
|
||||
total += bytes;
|
||||
}
|
||||
dprintln("Total RAM {}.{} MB", total / (1 << 20), total % (1 << 20) * 1000 / (1 << 20));
|
||||
}
|
||||
|
||||
}
|
||||
|
|
Loading…
Reference in New Issue