Random kernel freezing on process creation / exiting

[greatbridf]

Running sh -c 'echo $PPID' repeatedly in debug mode can trigger this easily.

Possibly a deadlock in process creation / exiting procedure.

[greatbridf]

make run SMP= also freezes, weird...

After applying this chunk, both debug and release mode don't boot now. So this problem might be associated with bootstrap memory allocation...

diff --git a/crates/eonix_hal/src/mm.rs b/crates/eonix_hal/src/mm.rs
index c4b9bb749c49..028c5fff4820 100644
--- a/crates/eonix_hal/src/mm.rs
+++ b/crates/eonix_hal/src/mm.rs
@@ -41,9 +41,9 @@ impl BasicPageAlloc {
     fn alloc_one(&mut self) -> PFN {
         assert_ne!(self.head, self.tail, "No free pages available");
         let mut range = self.ranges[self.head].take().unwrap();
-        range = range.shrink(PAGE_SIZE);
+        let pfn = PFN::from(range.start());
 
-        let pfn = PFN::from(range.end());
+        range = PRange::new(range.start() + PAGE_SIZE, range.end());
 
         if range.len() != 0 {
             self.ranges[self.head] = Some(range);

[greatbridf]

Seems that __kernel_end symbol is placed at the wrong place. We consider the area between __kernel_end and align_ceil(__kernel_end, 0x20000) free and allocate them out. So writes to these memory will overwrite the end of data section and vdso.

Sections:
Idx Name          Size      VMA               LMA               File off  Algn
  0 .bootstrap    00013100  0000000080200000  0000000080200000  00001000  2**12
                  CONTENTS, ALLOC, LOAD, CODE
  1 .text         00092bec  ffffffff80213100  0000000080213100  00014100  2**2
                  CONTENTS, ALLOC, LOAD, READONLY, CODE
  2 .text.syscall_fns 000026f0  ffffffff802a5cec  00000000802a5cec  000a6cec  2**1
                  CONTENTS, ALLOC, LOAD, READONLY, CODE
  3 .rodata       000246d4  ffffffff802a83e0  00000000802a83e0  000a93e0  2**4
                  CONTENTS, ALLOC, LOAD, READONLY, DATA
  4 .rodata.fixups 00000a4c  ffffffff802ccab4  00000000802ccab4  000cdab4  2**0
                  CONTENTS, ALLOC, LOAD, READONLY, DATA
  5 .rodata.syscalls 00000ae0  ffffffff802cd500  00000000802cd500  000ce500  2**3
                  CONTENTS, ALLOC, LOAD, READONLY, DATA
  6 .percpu       00000210  ffffffff802cdfe0  00000000802cdfe0  000cefe0  2**4
                  CONTENTS, ALLOC, LOAD, DATA
  7 .data         00003920  ffffffff802ce1f0  00000000802ce1f0  000cf1f0  2**4
                  CONTENTS, ALLOC, LOAD, DATA
  8 .vdso         00001000  00007f0000000000  00000000802d2000  000d3000  2**12
                  CONTENTS, ALLOC, LOAD, READONLY, CODE
  9 .data.after   00000000  ffffffff802d1b10  ffffffff802d1b10  000d4000  2**0
                  CONTENTS, ALLOC, LOAD, READONLY, DATA
 10 .bss          00002000  ffffffff40000000  00000000802d3000  000d4000  2**4
                  ALLOC

...

ffffffff802d1b10 g       .data.after    0000000000000000 __kernel_end

This patch relieves the bug but doesn't solve it completely. After applying it, the system behaves normally now...

diff --git a/crates/eonix_hal/src/arch/riscv64/mm.rs b/crates/eonix_hal/src/arch/riscv64/mm.rs
index f67646cf5b56..438772876ce6 100644
--- a/crates/eonix_hal/src/arch/riscv64/mm.rs
+++ b/crates/eonix_hal/src/arch/riscv64/mm.rs
@@ -314,17 +322,17 @@ where
         let kernel_end = PAddr::from(__kernel_end as usize - KIMAGE_OFFSET);
         let paddr_after_kimage_aligned = kernel_end.ceil_to(0x200000);
 
-        core::iter::once(PRange::new(kernel_end, paddr_after_kimage_aligned)).chain(
-            self.filter(move |range| range.end() > paddr_after_kimage_aligned)
-                .map(move |range| {
-                    if range.start() < paddr_after_kimage_aligned {
-                        let (_, right) = range.split_at(paddr_after_kimage_aligned);
-                        right
-                    } else {
-                        range
-                    }
-                }),
-        )
+        // core::iter::once(PRange::new(kernel_end, paddr_after_kimage_aligned)).chain(
+        self.filter(move |range| range.end() > paddr_after_kimage_aligned)
+            .map(move |range| {
+                if range.start() < paddr_after_kimage_aligned {
+                    let (_, right) = range.split_at(paddr_after_kimage_aligned);
+                    right
+                } else {
+                    range
+                }
+            })
+        // )
     }
 }

So the linker script should need reworking...