On Mon, Aug 02, 2021 at 10:39:04AM +0800, Kefeng Wang wrote:

On 2021/8/1 23:23, Catalin Marinas wrote:

quoted

On Tue, Jul 20, 2021 at 10:51:03AM +0800, Kefeng Wang wrote:

quoted

There are some fixed locations in the vmalloc area be reserved
in ARM(see iotable_init()) and ARM64(see map_kernel()), but for
pcpu_page_first_chunk(), it calls vm_area_register_early() and
choose VMALLOC_START as the start address of vmap area which
could be conflicted with above address, then could trigger a
BUG_ON in vm_area_add_early().

Let's choose the end of existing address range in vmlist as the
start address instead of VMALLOC_START to avoid the BUG_ON.

Signed-off-by: Kefeng Wang <redacted>
---
  mm/vmalloc.c | 8 +++++---
  1 file changed, 5 insertions(+), 3 deletions(-)

diff --git a/mm/vmalloc.c b/mm/vmalloc.c
index d5cd52805149..a98cf97f032f 100644
--- a/mm/vmalloc.c
+++ b/mm/vmalloc.c

@@ -2238,12 +2238,14 @@ void __init vm_area_add_early(struct vm_struct *vm)
   */
  void __init vm_area_register_early(struct vm_struct *vm, size_t align)
  {
-	static size_t vm_init_off __initdata;
+	unsigned long vm_start = VMALLOC_START;
+	struct vm_struct *tmp;
  	unsigned long addr;
-	addr = ALIGN(VMALLOC_START + vm_init_off, align);
-	vm_init_off = PFN_ALIGN(addr + vm->size) - VMALLOC_START;
+	for (tmp = vmlist; tmp; tmp = tmp->next)
+		vm_start = (unsigned long)tmp->addr + tmp->size;
+	addr = ALIGN(vm_start, align);
  	vm->addr = (void *)addr;
  	vm_area_add_early(vm);

Is there a risk of breaking other architectures? It doesn't look like to
me but I thought I'd ask.

Before this patch, vm_init_off is to record the offset from VMALLOC_START,

but it use VMALLOC_START as start address on the function
vm_area_register_early()

called firstly,  this will cause the BUG_ON.

With this patch, the most important change is that we choose the start
address via

dynamic calculate the 'start' address by traversing the list.

[wkf@localhost linux-next]$ git grep vm_area_register_early
arch/alpha/mm/init.c: vm_area_register_early(&console_remap_vm, PAGE_SIZE);
arch/x86/xen/p2m.c:     vm_area_register_early(&vm, PMD_SIZE *
PMDS_PER_MID_PAGE);
mm/percpu.c:    vm_area_register_early(&vm, PAGE_SIZE);
[wkf@localhost linux-next]$ git grep vm_area_add_early
arch/arm/mm/ioremap.c:  vm_area_add_early(vm);
arch/arm64/mm/mmu.c:    vm_area_add_early(vma);

x86/alpha won't call vm_area_add_early(), only arm64 could call both vm_area_add_early()
and  vm_area_register_early() when this patchset is merged. so it won't break other architectures.

Thanks for checking.

quoted

Also, instead of always picking the end, could we search for a range
that fits?

We only need a space in vmalloc range,  using end or a range in the middle
is not different.

I was thinking of making it more future-proof in case one registers a
vm area towards the end of the range. It's fairly easy to pick a range
in the middle now that you are adding a list traversal.

-- 
Catalin