When threads and the GIL are enabled, then the GC mutex is not needed. The
gc_mutex field is never used in this case because of:
#if MICROPY_PY_THREAD && !MICROPY_PY_THREAD_GIL
#define GC_ENTER() mp_thread_mutex_lock(&MP_STATE_MEM(gc_mutex), 1)
#define GC_EXIT() mp_thread_mutex_unlock(&MP_STATE_MEM(gc_mutex))
#else
#define GC_ENTER()
#define GC_EXIT()
#endif
So, we can completely remove gc_mutex everywhere when MICROPY_PY_THREAD
&& !MICROPY_PY_THREAD_GIL.
The older "bool has_finaliser" gets recast as GC_ALLOC_FLAG_HAS_FINALISER=1
so this is a backwards compatible change to the signature. Since bool gets
implicitly converted to 1 this patch doesn't include conversion of all
calls.
Otherwise there is the possibility that n_free starts out non-zero from the
previous iteration, which may have found a few (but not enough) free blocks
at the end of the heap. If this is the case, and if the very first blocks
that are scanned the second time around (starting at
gc_last_free_atb_index) are found to give enough memory (including the
blocks at the end of the heap from the previous iteration that left n_free
non-zero) then memory will be allocated starting before the location that
gc_last_free_atb_index points to, most likely leading to corruption.
This serious bug did not manifest itself in the past because a gc_collect
always resets gc_last_free_atb_index to point to the start of the GC heap,
and the first block there is almost always allocated to a long-lived
object (eg entries from sys.path, or mounted filesystem objects), which
means that n_free would be reset at the start of the search loop.
But with threading enabled with the GIL disabled it is possible to trigger
the bug via the following sequence of events:
1. Thread A runs gc_alloc, fails to find enough memory, and has a non-zero
n_free at the end of the search.
2. Thread A calls gc_collect and frees a bunch of blocks on the GC heap.
3. Just after gc_collect finishes in thread A, thread B takes gc_mutex and
does an allocation, moving gc_last_free_atb_index to point to the
interior of the heap, to a place where there is most likely a run of
available blocks.
4. Thread A regains gc_mutex and does its second search for free memory,
starting with a non-zero n_free. Since it's likely that the first block
it searches is available it will allocate memory which overlaps with the
memory before gc_last_free_atb_index.
DEBUG_printf and MICROPY_DEBUG_PRINTER is now used instead of normal
printf, and a fault is fixed in mp_obj_class_lookup with debugging enabled;
see issue #3999. Debugging can now be enabled on all ports including when
nan-boxing is used.
This patch adds the gc_sweep_all() function which does a garbage collection
without tracing any root pointers, so frees all the memory, and most
importantly runs any remaining finalisers.
This helps primarily for soft reset: it will close any open files, any open
sockets, and help to get the system back to a clean state upon soft reset.
Without this, if GC threshold is hit and there is not enough memory left to
satisfy the request, gc_collect() will run a second time and the search for
memory will happen again and will fail again.
Thanks to @adritium for pointing out this issue, see #3786.
This patch moves the start of the root pointer section in mp_state_ctx_t
so that it skips entries that are not pointers and don't need scanning.
Previously, the start of the root pointer section was at the very beginning
of the mp_state_ctx_t struct (which is the beginning of mp_state_thread_t).
This was the original assembler version of the NLR code was hard-coded to
have the nlr_top pointer at the start of this state structure. But now
that the NLR code is partially written in C there is no longer this
restriction on the location of nlr_top (and a comment to this effect has
been removed in this patch).
So now the root pointer section starts part way through the
mp_state_thread_t structure, after the entries which are not root pointers.
This patch also moves the non-pointer entries for MICROPY_ENABLE_SCHEDULER
outside the root pointer section.
Moving non-pointer entries out of the root pointer section helps to make
the GC more precise and should help to prevent some cases of collectable
garbage being kept.
This patch also has a measurable improvement in performance of the
pystone.py benchmark: on unix x86-64 and stm32 there was an improvement of
roughly 0.6% (tested with both gcc 7.3 and gcc 8.1).
This macro is written out explicitly in the two locations that it is used
and then the code is optimised, opening possibilities for further
optimisations and reducing code size:
unix: -48
minimal CROSS=1: -32
stm32: -32
This patch introduces the MICROPY_ENABLE_PYSTACK option (disabled by
default) which enables a "Python stack" that allows to allocate and free
memory in a scoped, or Last-In-First-Out (LIFO) way, similar to alloca().
A new memory allocation API is introduced along with this Py-stack. It
includes both "local" and "nonlocal" LIFO allocation. Local allocation is
intended to be equivalent to using alloca(), whereby the same function must
free the memory. Nonlocal allocation is where another function may free
the memory, so long as it's still LIFO.
Follow-up patches will convert all uses of alloca() and VLA to the new
scoped allocation API. The old behaviour (using alloca()) will still be
available, but when MICROPY_ENABLE_PYSTACK is enabled then alloca() is no
longer required or used.
The benefits of enabling this option are (or will be once subsequent
patches are made to convert alloca()/VLA):
- Toolchains without alloca() can use this feature to obtain correct and
efficient scoped memory allocation (compared to using the heap instead
of alloca(), which is slower).
- Even if alloca() is available, enabling the Py-stack gives slightly more
efficient use of stack space when calling nested Python functions, due to
the way that compilers implement alloca().
- Enabling the Py-stack with the stackless mode allows for even more
efficient stack usage, as well as retaining high performance (because the
heap is no longer used to build and destroy stackless code states).
- With Py-stack and stackless enabled, Python-calling-Python is no longer
recursive in the C mp_execute_bytecode function.
The micropython.pystack_use() function is included to measure usage of the
Python stack.
Accessing them will crash immediately instead still working for some time,
until overwritten by some other data, leading to much less deterministic
crashes.
These checks are assumed to be true in all cases where gc_realloc is
called with a valid pointer, so no need to waste code space and time
checking them in a non-debug build.
Header files that are considered internal to the py core and should not
normally be included directly are:
py/nlr.h - internal nlr configuration and declarations
py/bc0.h - contains bytecode macro definitions
py/runtime0.h - contains basic runtime enums
Instead, the top-level header files to include are one of:
py/obj.h - includes runtime0.h and defines everything to use the
mp_obj_t type
py/runtime.h - includes mpstate.h and hence nlr.h, obj.h, runtime0.h,
and defines everything to use the general runtime support functions
Additional, specific headers (eg py/objlist.h) can be included if needed.
If a finaliser raises an exception then it must not propagate through the
GC sweep function. This patch protects against such a thing by running
finaliser code via the mp_call_function_1_protected call.
This patch also adds scheduler lock/unlock calls around the finaliser
execution to further protect against any possible reentrancy issues: the
memory manager is already locked when doing a collection, but we also don't
want to allow any scheduled code to run, KeyboardInterrupts to interupt the
code, nor threads to switch.
There can be stray pointers in memory blocks that are not properly zero'd
after allocation. This patch adds a new config option to always zero all
allocated memory (via gc_alloc and gc_realloc) and hence help to eliminate
stray pointers.
See issue #2195.
Currently, MicroPython runs GC when it could not allocate a block of memory,
which happens when heap is exhausted. However, that policy can't work well
with "inifinity" heaps, e.g. backed by a virtual memory - there will be a
lot of swap thrashing long before VM will be exhausted. Instead, in such
cases "allocation threshold" policy is used: a GC is run after some number of
allocations have been made. Details vary, for example, number or total amount
of allocations can be used, threshold may be self-adjusting based on GC
outcome, etc.
This change implements a simple variant of such policy for MicroPython. Amount
of allocated memory so far is used for threshold, to make it useful to typical
finite-size, and small, heaps as used with MicroPython ports. And such GC policy
is indeed useful for such types of heaps too, as it allows to better control
fragmentation. For example, if a threshold is set to half size of heap, then
for an application which usually makes big number of small allocations, that
will (try to) keep half of heap memory in a nice defragmented state for an
occasional large allocation.
For an application which doesn't exhibit such behavior, there won't be any
visible effects, except for GC running more frequently, which however may
affect performance. To address this, the GC threshold is configurable, and
by default is off so far. It's configured with gc.threshold(amount_in_bytes)
call (can be queries without an argument).
Previously, if there was chain of allocated blocks ending with the last
block of heap, it wasn't included in number of 1/2-block or max block
size stats.
GC_EXIT() can cause a pending thread (waiting on the mutex) to be
scheduled right away. This other thread may trigger a garbage
collection. If the pointer to the newly-allocated block (allocated by
the original thread) is not computed before the switch (so it's just left
as a block number) then the block will be wrongly reclaimed.
This patch makes sure the pointer is computed before allowing any thread
switch to occur.
By using a single, global mutex, all memory-related functions (alloc,
free, realloc, collect, etc) are made thread safe. This means that only
one thread can be in such a function at any one time.
Address printed was truncated anyway and in general confusing to outsider.
A line which dumps it is still left in the source, commented, for peculiar
cases when it may be needed (e.g. when running under debugger).
These are typical consumers of large chunks of memory, so it's useful to
see at least their number (how much memory isn't clearly shown, as the data
for these objects is allocated elsewhere).
Previously, mark operation weren't logged at all, while it's quite useful
to see cascade of marks in case of over-marking (and in other cases too).
Previously, sweep was logged for each block of object in memory, but that
doesn't make much sense and just lead to longer output, harder to parse
by a human. Instead, log sweep only once per object. This is similar to
other memory manager operations, e.g. an object is allocated, then freed.
Or object is allocated, then marked, otherwise swept (one log entry per
operation, with the same memory address in each case).
Ideally we'd use %zu for size_t args, but that's unlikely to be supported
by all runtimes, and we would then need to implement it in mp_printf.
So simplest and most portable option is to use %u and cast the argument
to uint(=unsigned int).
Note: reason for the change is that UINT_FMT can be %llu (size suitable
for mp_uint_t) which is wider than size_t and prints incorrect results.
size_t is the correct type to use to count things related to the size of
the address space. Using size_t (instead of mp_uint_t) is important for
the efficiency of ports that configure mp_uint_t to larger than the
machine word size.
This allows the mp_obj_t type to be configured to something other than a
pointer-sized primitive type.
This patch also includes additional changes to allow the code to compile
when sizeof(mp_uint_t) != sizeof(void*), such as using size_t instead of
mp_uint_t, and various casts.
Currently, the only place that clears the bit is in gc_collect.
So if a block with a finalizer is allocated, and subsequently
freed, and then the block is reallocated with no finalizer then
the bit remains set.
This could also be fixed by having gc_alloc clear the bit, but
I'm pretty sure that free is called way less than alloc, so doing
it in free is more efficient.
Previous to this patch all interned strings lived in their own malloc'd
chunk. On average this wastes N/2 bytes per interned string, where N is
the number-of-bytes for a quanta of the memory allocator (16 bytes on 32
bit archs).
With this patch interned strings are concatenated into the same malloc'd
chunk when possible. Such chunks are enlarged inplace when possible,
and shrunk to fit when a new chunk is needed.
RAM savings with this patch are highly varied, but should always show an
improvement (unless only 3 or 4 strings are interned). New version
typically uses about 70% of previous memory for the qstr data, and can
lead to savings of around 10% of total memory footprint of a running
script.
Costs about 120 bytes code size on Thumb2 archs (depends on how many
calls to gc_realloc are made).
