327
262
return parser.parse()
330
# TODO: We can go from 8 byte offset + 4 byte length to a simple lookup,
331
# because the block_offset + length is likely to be repeated. However,
332
# the big win there is to cache across pages, and not just one page
333
# Though if we did cache in a page, we could certainly use a short int.
334
# And this goes from 40 bytes to 30 bytes.
335
# One slightly ugly option would be to cache block offsets in a global.
336
# However, that leads to thread-safety issues, etc.
337
ctypedef struct gc_chk_sha1_record:
338
long long block_offset
339
unsigned int block_length
340
unsigned int record_start
341
unsigned int record_end
345
cdef int _unhexbuf[256]
347
_hexbuf = '0123456789abcdef'
349
cdef _populate_unhexbuf():
351
for i from 0 <= i < 256:
353
for i from 0 <= i < 10: # 0123456789 => map to the raw number
354
_unhexbuf[(i + c'0')] = i
355
for i from 10 <= i < 16: # abcdef => 10, 11, 12, 13, 14, 15, 16
356
_unhexbuf[(i - 10 + c'a')] = i
357
for i from 10 <= i < 16: # ABCDEF => 10, 11, 12, 13, 14, 15, 16
358
_unhexbuf[(i - 10 + c'A')] = i
362
cdef int _unhexlify_sha1(char *as_hex, char *as_bin): # cannot_raise
363
"""Take the hex sha1 in as_hex and make it binary in as_bin
365
Same as binascii.unhexlify, but working on C strings, not Python objects.
372
# binascii does this using isupper() and tolower() and ?: syntax. I'm
373
# guessing a simple lookup array should be faster.
375
for i from 0 <= i < 20:
376
top = _unhexbuf[<unsigned char>(as_hex[j])]
378
bot = _unhexbuf[<unsigned char>(as_hex[j])]
380
if top == -1 or bot == -1:
382
as_bin[i] = <unsigned char>((top << 4) + bot);
386
def _py_unhexlify(as_hex):
387
"""For the test infrastructure, just thunks to _unhexlify_sha1"""
388
if len(as_hex) != 40 or not PyString_CheckExact(as_hex):
389
raise ValueError('not a 40-byte hex digest')
390
as_bin = PyString_FromStringAndSize(NULL, 20)
391
if _unhexlify_sha1(PyString_AS_STRING(as_hex), PyString_AS_STRING(as_bin)):
396
cdef void _hexlify_sha1(char *as_bin, char *as_hex): # cannot_raise
401
for i from 0 <= i < 20:
403
as_hex[j] = _hexbuf[(c>>4)&0xf]
405
as_hex[j] = _hexbuf[(c)&0xf]
409
def _py_hexlify(as_bin):
410
"""For test infrastructure, thunk to _hexlify_sha1"""
411
if len(as_bin) != 20 or not PyString_CheckExact(as_bin):
412
raise ValueError('not a 20-byte binary digest')
413
as_hex = PyString_FromStringAndSize(NULL, 40)
414
_hexlify_sha1(PyString_AS_STRING(as_bin), PyString_AS_STRING(as_hex))
418
cdef int _key_to_sha1(key, char *sha1): # cannot_raise
419
"""Map a key into its sha1 content.
421
:param key: A tuple of style ('sha1:abcd...',)
422
:param sha1: A char buffer of 20 bytes
423
:return: 1 if this could be converted, 0 otherwise
428
if StaticTuple_CheckExact(key) and StaticTuple_GET_SIZE(key) == 1:
429
p_val = <PyObject *>StaticTuple_GET_ITEM(key, 0)
430
elif (PyTuple_CheckExact(key) and PyTuple_GET_SIZE(key) == 1):
431
p_val = PyTuple_GET_ITEM_ptr_object(key, 0)
433
# Not a tuple or a StaticTuple
435
if (PyString_CheckExact_ptr(p_val) and PyString_GET_SIZE_ptr(p_val) == 45):
436
c_val = PyString_AS_STRING_ptr(p_val)
439
if strncmp(c_val, 'sha1:', 5) != 0:
441
if not _unhexlify_sha1(c_val + 5, sha1):
446
def _py_key_to_sha1(key):
447
"""Map a key to a simple sha1 string.
449
This is a testing thunk to the C function.
451
as_bin_sha = PyString_FromStringAndSize(NULL, 20)
452
if _key_to_sha1(key, PyString_AS_STRING(as_bin_sha)):
457
cdef StaticTuple _sha1_to_key(char *sha1):
458
"""Compute a ('sha1:abcd',) key for a given sha1."""
462
hexxed = PyString_FromStringAndSize(NULL, 45)
463
c_buf = PyString_AS_STRING(hexxed)
464
memcpy(c_buf, 'sha1:', 5)
465
_hexlify_sha1(sha1, c_buf+5)
466
key = StaticTuple_New(1)
468
StaticTuple_SET_ITEM(key, 0, hexxed)
469
# This is a bit expensive. To parse 120 keys takes 48us, to return them all
470
# can be done in 66.6us (so 18.6us to build them all).
471
# Adding simple hash() here brings it to 76.6us (so computing the hash
472
# value of 120keys is 10us), Intern is 86.9us (another 10us to look and add
473
# them to the intern structure.)
474
# However, since we only intern keys that are in active use, it is probably
475
# a win. Since they would have been read from elsewhere anyway.
476
# We *could* hang the PyObject form off of the gc_chk_sha1_record for ones
477
# that we have deserialized. Something to think about, at least.
478
key = StaticTuple_Intern(key)
482
def _py_sha1_to_key(sha1_bin):
483
"""Test thunk to check the sha1 mapping."""
484
if not PyString_CheckExact(sha1_bin) or PyString_GET_SIZE(sha1_bin) != 20:
485
raise ValueError('sha1_bin must be a str of exactly 20 bytes')
486
return _sha1_to_key(PyString_AS_STRING(sha1_bin))
489
cdef unsigned int _sha1_to_uint(char *sha1): # cannot_raise
490
cdef unsigned int val
491
# Must be in MSB, because that is how the content is sorted
492
val = (((<unsigned int>(sha1[0]) & 0xff) << 24)
493
| ((<unsigned int>(sha1[1]) & 0xff) << 16)
494
| ((<unsigned int>(sha1[2]) & 0xff) << 8)
495
| ((<unsigned int>(sha1[3]) & 0xff) << 0))
499
cdef _format_record(gc_chk_sha1_record *record):
500
# This is inefficient to go from a logical state back to a
501
# string, but it makes things work a bit better internally for now.
502
if record.block_offset >= 0xFFFFFFFF:
503
# %llu is what we really want, but unfortunately it was only added
504
# in python 2.7... :(
505
block_offset_str = str(record.block_offset)
506
value = PyString_FromFormat('%s %u %u %u',
507
PyString_AS_STRING(block_offset_str),
509
record.record_start, record.record_end)
511
value = PyString_FromFormat('%lu %u %u %u',
512
<unsigned long>record.block_offset,
514
record.record_start, record.record_end)
518
cdef class GCCHKSHA1LeafNode:
519
"""Track all the entries for a given leaf node."""
521
cdef gc_chk_sha1_record *records
522
cdef public object last_key
523
cdef gc_chk_sha1_record *last_record
524
cdef public int num_records
525
# This is the number of bits to shift to get to the interesting byte. A
526
# value of 24 means that the very first byte changes across all keys.
527
# Anything else means that there is a common prefix of bits that we can
528
# ignore. 0 means that at least the first 3 bytes are identical, though
529
# that is going to be very rare
530
cdef public unsigned char common_shift
531
# This maps an interesting byte to the first record that matches.
532
# Equivalent to bisect.bisect_left(self.records, sha1), though only taking
533
# into account that one byte.
534
cdef unsigned char offsets[257]
536
def __sizeof__(self):
537
# :( Why doesn't Pyrex let me do a simple sizeof(GCCHKSHA1LeafNode)
538
# like Cython? Explicitly enumerating everything here seems to leave my
539
# size off by 2 (286 bytes vs 288 bytes actual). I'm guessing it is an
540
# alignment/padding issue. Oh well- at least we scale properly with
541
# num_records and are very close to correct, which is what I care
543
# If we ever decide to require cython:
544
# return (sizeof(GCCHKSHA1LeafNode)
545
# + sizeof(gc_chk_sha1_record)*self.num_records)
546
return (sizeof(PyObject) + sizeof(void*) + sizeof(int)
547
+ sizeof(gc_chk_sha1_record*) + sizeof(PyObject *)
548
+ sizeof(gc_chk_sha1_record*) + sizeof(char)
549
+ sizeof(unsigned char)*257
550
+ sizeof(gc_chk_sha1_record)*self.num_records)
552
def __dealloc__(self):
553
if self.records != NULL:
554
PyMem_Free(self.records)
557
def __init__(self, bytes):
558
self._parse_bytes(bytes)
560
self.last_record = NULL
564
if self.num_records > 0:
565
return _sha1_to_key(self.records[0].sha1)
570
if self.num_records > 0:
571
return _sha1_to_key(self.records[self.num_records-1].sha1)
574
cdef StaticTuple _record_to_value_and_refs(self,
575
gc_chk_sha1_record *record):
576
"""Extract the refs and value part of this record."""
577
cdef StaticTuple value_and_refs
578
cdef StaticTuple empty
579
value_and_refs = StaticTuple_New(2)
580
value = _format_record(record)
582
StaticTuple_SET_ITEM(value_and_refs, 0, value)
584
empty = StaticTuple_New(0)
586
StaticTuple_SET_ITEM(value_and_refs, 1, empty)
587
return value_and_refs
589
cdef StaticTuple _record_to_item(self, gc_chk_sha1_record *record):
590
"""Turn a given record back into a fully fledged item.
592
cdef StaticTuple item
594
cdef StaticTuple value_and_refs
596
key = _sha1_to_key(record.sha1)
597
item = StaticTuple_New(2)
599
StaticTuple_SET_ITEM(item, 0, key)
600
value_and_refs = self._record_to_value_and_refs(record)
601
Py_INCREF(value_and_refs)
602
StaticTuple_SET_ITEM(item, 1, value_and_refs)
605
cdef gc_chk_sha1_record* _lookup_record(self, char *sha1) except? NULL:
606
"""Find a gc_chk_sha1_record that matches the sha1 supplied."""
607
cdef int lo, hi, mid, the_cmp
610
# TODO: We can speed up misses by comparing this sha1 to the common
611
# bits, and seeing if the common prefix matches, if not, we don't
612
# need to search for anything because it cannot match
613
# Use the offset array to find the closest fit for this entry
614
# follow that up with bisecting, since multiple keys can be in one
616
# Bisecting dropped us from 7000 comparisons to 582 (4.8/key), using
617
# the offset array dropped us from 23us to 20us and 156 comparisions
619
offset = self._offset_for_sha1(sha1)
620
lo = self.offsets[offset]
621
hi = self.offsets[offset+1]
623
# if hi == 255 that means we potentially ran off the end of the
624
# list, so push it up to num_records
625
# note that if 'lo' == 255, that is ok, because we can start
626
# searching from that part of the list.
627
hi = self.num_records
631
the_cmp = memcmp(self.records[mid].sha1, sha1, 20)
633
return &self.records[mid]
640
def __contains__(self, key):
642
cdef gc_chk_sha1_record *record
643
if _key_to_sha1(key, sha1):
644
# If it isn't a sha1 key, then it won't be in this leaf node
645
record = self._lookup_record(sha1)
648
self.last_record = record
652
def __getitem__(self, key):
654
cdef gc_chk_sha1_record *record
656
if self.last_record != NULL and key is self.last_key:
657
record = self.last_record
658
elif _key_to_sha1(key, sha1):
659
record = self._lookup_record(sha1)
661
raise KeyError('key %r is not present' % (key,))
662
return self._record_to_value_and_refs(record)
665
return self.num_records
670
for i from 0 <= i < self.num_records:
671
PyList_Append(result, _sha1_to_key(self.records[i].sha1))
677
for i from 0 <= i < self.num_records:
678
item = self._record_to_item(&self.records[i])
679
PyList_Append(result, item)
682
cdef int _count_records(self, char *c_content, char *c_end): # cannot_raise
683
"""Count how many records are in this section."""
689
while c_cur != NULL and c_cur < c_end:
690
c_cur = <char *>memchr(c_cur, c'\n', c_end - c_cur);
694
num_records = num_records + 1
697
cdef _parse_bytes(self, bytes):
698
"""Parse the string 'bytes' into content."""
702
cdef Py_ssize_t n_bytes
705
cdef gc_chk_sha1_record *cur_record
707
if not PyString_CheckExact(bytes):
708
raise TypeError('We only support parsing plain 8-bit strings.')
709
# Pass 1, count how many records there will be
710
n_bytes = PyString_GET_SIZE(bytes)
711
c_bytes = PyString_AS_STRING(bytes)
712
c_end = c_bytes + n_bytes
713
if strncmp(c_bytes, 'type=leaf\n', 10):
714
raise ValueError("bytes did not start with 'type=leaf\\n': %r"
717
num_records = self._count_records(c_cur, c_end)
718
# Now allocate the memory for these items, and go to town
719
self.records = <gc_chk_sha1_record*>PyMem_Malloc(num_records *
720
(sizeof(unsigned short) + sizeof(gc_chk_sha1_record)))
721
self.num_records = num_records
722
cur_record = self.records
724
while c_cur != NULL and c_cur < c_end and entry < num_records:
725
c_cur = self._parse_one_entry(c_cur, c_end, cur_record)
726
cur_record = cur_record + 1
728
if (entry != self.num_records
730
or cur_record != self.records + self.num_records):
731
raise ValueError('Something went wrong while parsing.')
732
# Pass 3: build the offset map
733
self._compute_common()
735
cdef char *_parse_one_entry(self, char *c_cur, char *c_end,
736
gc_chk_sha1_record *cur_record) except NULL:
737
"""Read a single sha record from the bytes.
738
:param c_cur: The pointer to the start of bytes
742
if strncmp(c_cur, 'sha1:', 5):
743
raise ValueError('line did not start with sha1: %r'
744
% (safe_string_from_size(c_cur, 10),))
746
c_next = <char *>memchr(c_cur, c'\0', c_end - c_cur)
747
if c_next == NULL or (c_next - c_cur != 40):
748
raise ValueError('Line did not contain 40 hex bytes')
749
if not _unhexlify_sha1(c_cur, cur_record.sha1):
750
raise ValueError('We failed to unhexlify')
752
if c_cur[0] != c'\0':
753
raise ValueError('only 1 null, not 2 as expected')
755
cur_record.block_offset = strtoll(c_cur, &c_next, 10)
756
if c_cur == c_next or c_next[0] != c' ':
757
raise ValueError('Failed to parse block offset')
759
cur_record.block_length = strtoul(c_cur, &c_next, 10)
760
if c_cur == c_next or c_next[0] != c' ':
761
raise ValueError('Failed to parse block length')
763
cur_record.record_start = strtoul(c_cur, &c_next, 10)
764
if c_cur == c_next or c_next[0] != c' ':
765
raise ValueError('Failed to parse block length')
767
cur_record.record_end = strtoul(c_cur, &c_next, 10)
768
if c_cur == c_next or c_next[0] != c'\n':
769
raise ValueError('Failed to parse record end')
773
cdef int _offset_for_sha1(self, char *sha1) except -1:
774
"""Find the first interesting 8-bits of this sha1."""
776
cdef unsigned int as_uint
777
as_uint = _sha1_to_uint(sha1)
778
this_offset = (as_uint >> self.common_shift) & 0xFF
781
def _get_offset_for_sha1(self, sha1):
782
return self._offset_for_sha1(PyString_AS_STRING(sha1))
784
cdef _compute_common(self):
785
cdef unsigned int first
786
cdef unsigned int this
787
cdef unsigned int common_mask
788
cdef unsigned char common_shift
790
cdef int offset, this_offset
792
# The idea with the offset map is that we should be able to quickly
793
# jump to the key that matches a gives sha1. We know that the keys are
794
# in sorted order, and we know that a lot of the prefix is going to be
795
# the same across them.
796
# By XORing the records together, we can determine what bits are set in
798
if self.num_records < 2:
799
# Everything is in common if you have 0 or 1 leaves
800
# So we'll always just shift to the first byte
801
self.common_shift = 24
803
common_mask = 0xFFFFFFFF
804
first = _sha1_to_uint(self.records[0].sha1)
805
for i from 0 < i < self.num_records:
806
this = _sha1_to_uint(self.records[i].sha1)
807
common_mask = (~(first ^ this)) & common_mask
809
while common_mask & 0x80000000 and common_shift > 0:
810
common_mask = common_mask << 1
811
common_shift = common_shift - 1
812
self.common_shift = common_shift
814
max_offset = self.num_records
815
# We cap this loop at 254 records. All the other offsets just get
816
# filled with 0xff as the singleton saying 'too many'.
817
# It means that if we have >255 records we have to bisect the second
818
# half of the list, but this is going to be very rare in practice.
821
for i from 0 <= i < max_offset:
822
this_offset = self._offset_for_sha1(self.records[i].sha1)
823
while offset <= this_offset:
824
self.offsets[offset] = i
827
self.offsets[offset] = max_offset
830
def _get_offsets(self):
833
for i from 0 <= i < 257:
834
PyList_Append(result, self.offsets[i])
838
def _parse_into_chk(bytes, key_length, ref_list_length):
839
"""Parse into a format optimized for chk records."""
840
assert key_length == 1
841
assert ref_list_length == 0
842
return GCCHKSHA1LeafNode(bytes)
845
265
def _flatten_node(node, reference_lists):
846
266
"""Convert a node into the serialized form.