| Age | Commit message (Collapse) | Author | Files | Lines |
|---|
|
The memory manager maintains the first free page as the page after the
`_end` synthetic emitted by the linker. This value is stored in a
translation unit local variable. This value is only ever written to
from `init_early_alloc` which is static and only ever invoked from
`pk_vm_init`. Furthermore, the value that `first_free_page` is ever set
to is computed as a rounding of the _address_ of `_end`. Because the
address of the symbol cannot change during execution of a normal
program, this is effectively a constant, making the computed value a
"constant" which can be re-materialized. Now, with the knowledge that
the value is effectively a constant that can be re-materialized and the
fact that the value is ever written to at a single position, we can
simply re-materialize the value if it was ever changed in
`free_page_addr`. This will allow the 8-byte value to be truncated to
1-byte.
Now, we can inline `__early_pgalloc_align`, and because the combination
of `__early_alloc` and `__early_pgalloc_align` is small, we can inline
that again at the two sites locally. This changes the
`__augment_page_freelist` to re-materialize the constant when needed for
the allocation.
The re-materialization however uses a pc-relative addressing, which now
computes a different value than expected - the address has become a VA
rather than a PA. This results in the address computed by
`free_page_addr` (which is the result of the `__early_pgalloc_align`) to
be a virtual address after the relocation, which then propagates through
`__early_alloc` to the value in `__augment_page_freelist`, which is then
consumed by `__page_alloc`, which will treat the now VA as a PA and
perform an additional translation to a VA.
Mark the value as `volatile` to indicate that the value must be read at
all points to thwart the size optimization of the compiler resulting in
a mis-compilation resulting in the eventual invalid memory access during
the `memset` that follows the allocation.
Thanks to @nzmichaelh for the help in tracking this down!
|
|
Eliminate excess reference count when associating fd with file.
Fixes #258
|
|
`SYS_getcwd` is different from `getcwd` in that the return value is < 0
on failure otherwise it is the length of the string. The proxy kernel
was treating 0 as success and all other values as error. As a result,
we would never return a valid value for `getcwd`.
The following program now executes properly with the Proxy Kernel:
```c
#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>
#include <linux/limits.h>
int main(int argc, char **argv) {
unsigned char buffer[PATH_MAX + 1] = {0};
if (getcwd(buffer, PATH_MAX))
printf("cwd: %s\n", buffer);
return EXIT_SUCCESS;
}
```
|
|
|
|
|
|
Resolves #216
|
|
The LLVM IAS currently does not support the older spelling for the CSR.
Update the references to the modern name.
|
|
This replaces use of the old `sbadaddr` CSR name with the current
`stval` name. The old spelling is not supported by the LLVM IAS,
however, the modern spelling is supported by both LLVM and binutils.
|
|
Using recent compilers we get the following error message:
../pk/pk.c: In function 'run_loaded_program.constprop':
../pk/pk.c:177:3: error: both arguments to '__builtin___clear_cache'
must be pointers
177 | __clear_cache(0, 0);
| ^~~~~~~~~~~~~~~~~~~
Let's use the existing function __riscv_flush_icache(),
give it a header with a prototype and use it to
emits the FENCE.I instruction directly.
See #239
Suggested-by: Andrew Waterman <andrew@sifive.com>
Signed-off-by: Christoph Muellner <cmuellner@linux.com>
|
|
|
|
This assumes that stval is populated with the opcode on illegal
instruction exceptions. But since we're only using the opcode for
error reporting, it's OK if this assumption is violated.
|
|
|
|
|
|
|
|
|
|
|
|
Previously, the pk would always run from virtual address MEM_START.
Instead, remap it into the negative virtual addresses, allowing user
processes to expand beyond MEM_START.
|
|
|
|
Return the old brk if mmap fails, rather than just asserting out
|
|
Estimate available memory and return -1 from mmap if not enough
is available, rather than assert-failing.
|
|
This will improve flexibility going forward.
|
|
Enforced with sstatus.SUM.
|
|
|
|
|
|
|
|
|
|
The system call should return the stat struct in kernel format, not
libc format.
|
|
The linux kernel pushes argc as an int, not an uintptr_t.
(The offset to the next element is still sizeof(uintptr_t).)
|
|
This attempt at rounding leads to wrong results, for example:
314689951 cycles
314690101 instructions
0.90 CPI
With my change results in:
314689951 cycles
314690101 instructions
0.99 CPI
I think this was supposed to be part of rounding behaviour but it doesn't work if only the final digit does it and there is no carry...
Instead I changed it to truncate after the second digit
|
|
77a5df569451571d608650a34183d53df99790ec)
|
|
|
|
|
|
This prevents the counters from easily overflowing for a 32-bit pk.
|
|
|
|
For systems where physical memory is mapped to a start address
different from 0x80000000, allow the default to be overridden
using the --with-mem-start argument to the configure script.
Signed-off-by: Gabriel Somlo <gsomlo@gmail.com>
|
|
Closes #161
|
|
|
|
|
|
|
|
For eventual PMP purposes
|
|
|
|
from m-mode. (#117)
|
|
|
|
Otherwise, the emulation code will refuse to emulate FP instructions.
|
|
This makes it more compatible with recent glibc.
|
|
|
|
We now automatically detect everything that the platform interface used
to be used for, so it's now obsolete!
|
|
This checks to see if a hart can't boot Linux by looking for a
compatible "mmu-type" field. If the hart can't boot Linux, then bbl
masks it off.
|
|
|
|
|
