From 13497e652a9b9495fbefd52396620bfccc6d7a7c Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?John=20H=C3=B6gberg?= Date: Tue, 13 Feb 2018 12:34:08 +0100 Subject: Handle unaligned binaries in erlang:iolist_to_iovec/1 A binary is a binary as long as its size in bits is evenly divisible by 8, regardless of whether it has a bit offset or not. --- erts/emulator/beam/erl_io_queue.c | 31 +++++++++++++++++++++++-------- 1 file changed, 23 insertions(+), 8 deletions(-) (limited to 'erts/emulator/beam') diff --git a/erts/emulator/beam/erl_io_queue.c b/erts/emulator/beam/erl_io_queue.c index 190ba6bbb9..990c0c1cac 100644 --- a/erts/emulator/beam/erl_io_queue.c +++ b/erts/emulator/beam/erl_io_queue.c @@ -801,12 +801,11 @@ static Eterm iol2v_make_sub_bin(iol2v_state_t *state, Eterm bin_term, ERTS_GET_REAL_BIN(bin_term, orig_pb_term, byte_offset, bit_offset, bit_size); - (void)bit_offset; - (void)bit_size; + ASSERT(bit_size == 0); sb->thing_word = HEADER_SUB_BIN; + sb->bitoffs = bit_offset; sb->bitsize = 0; - sb->bitoffs = 0; sb->orig = orig_pb_term; sb->is_writable = 0; @@ -984,7 +983,7 @@ static int iol2v_append_binary(iol2v_state_t *state, Eterm bin_term) { parent_header = binary_val(parent_binary); binary_size = binary_size(bin_term); - if (bit_offset != 0 || bit_size != 0) { + if (bit_size != 0) { return 0; } else if (binary_size == 0) { state->bytereds_spent += 1; @@ -1026,8 +1025,16 @@ static int iol2v_append_binary(iol2v_state_t *state, Eterm bin_term) { * then just copy it into the accumulator. */ iol2v_expand_acc(state, binary_size); - sys_memcpy(&(state->acc)->orig_bytes[state->acc_size], - binary_data, binary_size); + if (ERTS_LIKELY(bit_offset == 0)) { + sys_memcpy(&(state->acc)->orig_bytes[state->acc_size], + binary_data, binary_size); + } else { + ASSERT(binary_size <= ERTS_UWORD_MAX / 8); + + erts_copy_bits(binary_data, bit_offset, 1, + (byte*)&(state->acc)->orig_bytes[state->acc_size], 0, 1, + binary_size * 8); + } state->acc_size += binary_size; } else { @@ -1038,8 +1045,16 @@ static int iol2v_append_binary(iol2v_state_t *state, Eterm bin_term) { iol2v_expand_acc(state, spill); - sys_memcpy(&(state->acc)->orig_bytes[state->acc_size], - binary_data, spill); + if (ERTS_LIKELY(bit_offset == 0)) { + sys_memcpy(&(state->acc)->orig_bytes[state->acc_size], + binary_data, spill); + } else { + ASSERT(binary_size <= ERTS_UWORD_MAX / 8); + + erts_copy_bits(binary_data, bit_offset, 1, + (byte*)&(state->acc)->orig_bytes[state->acc_size], 0, 1, + spill * 8); + } state->acc_size += spill; -- cgit v1.2.3 From c47c8ed7da8c222c5d3b99d0f109fe126a86c632 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?John=20H=C3=B6gberg?= Date: Tue, 13 Feb 2018 16:12:30 +0100 Subject: Handle unaligned binaries in enif_inspect_iovec --- erts/emulator/beam/erl_nif.c | 124 +++++++++++++++++++++++++------------------ 1 file changed, 71 insertions(+), 53 deletions(-) (limited to 'erts/emulator/beam') diff --git a/erts/emulator/beam/erl_nif.c b/erts/emulator/beam/erl_nif.c index f67b67325d..b96db5d054 100644 --- a/erts/emulator/beam/erl_nif.c +++ b/erts/emulator/beam/erl_nif.c @@ -3396,8 +3396,8 @@ typedef struct { Eterm sublist_start; Eterm sublist_end; - UWord offheap_size; - UWord onheap_size; + UWord referenced_size; + UWord copied_size; UWord iovec_len; } iovec_slice_t; @@ -3407,16 +3407,16 @@ static int examine_iovec_term(Eterm list, UWord max_length, iovec_slice_t *resul result->sublist_start = list; result->sublist_length = 0; - result->offheap_size = 0; - result->onheap_size = 0; + result->referenced_size = 0; + result->copied_size = 0; result->iovec_len = 0; lookahead = result->sublist_start; while (is_list(lookahead)) { - Eterm *binary_header, binary; + UWord byte_size; + Eterm binary; Eterm *cell; - UWord size; cell = list_val(lookahead); binary = CAR(cell); @@ -3425,35 +3425,36 @@ static int examine_iovec_term(Eterm list, UWord max_length, iovec_slice_t *resul return 0; } - size = binary_size(binary); - binary_header = binary_val(binary); + byte_size = binary_size(binary); - if (size > 0) { - /* If we're a sub-binary we'll need to check our underlying binary - * to determine whether we're on-heap or not. */ - if (thing_subtag(*binary_header) == SUB_BINARY_SUBTAG) { - ErlSubBin *sb = (ErlSubBin*)binary_header; + if (byte_size > 0) { + int bit_offset, bit_size; + Eterm parent_binary; + UWord byte_offset; - /* Reject bitstrings */ - if((sb->bitoffs + sb->bitsize) > 0) { - return 0; - } + int requires_copying; - ASSERT(size <= binary_size(sb->orig)); - binary_header = binary_val(sb->orig); + ERTS_GET_REAL_BIN(binary, parent_binary, byte_offset, + bit_offset, bit_size); + + (void)byte_offset; + + if (bit_size != 0) { + return 0; } - if (thing_subtag(*binary_header) == HEAP_BINARY_SUBTAG) { - ASSERT(size <= ERL_ONHEAP_BIN_LIMIT); + /* If we're unaligned or an on-heap binary we'll need to copy + * ourselves over to a temporary buffer. */ + requires_copying = (bit_offset != 0) || + thing_subtag(*binary_val(parent_binary)) == HEAP_BINARY_SUBTAG; - result->iovec_len += 1; - result->onheap_size += size; + if (requires_copying) { + result->copied_size += byte_size; } else { - ASSERT(thing_subtag(*binary_header) == REFC_BINARY_SUBTAG); - - result->iovec_len += 1 + size / MAX_SYSIOVEC_IOVLEN; - result->offheap_size += size; + result->referenced_size += byte_size; } + + result->iovec_len += 1 + byte_size / MAX_SYSIOVEC_IOVLEN; } result->sublist_length += 1; @@ -3473,7 +3474,9 @@ static int examine_iovec_term(Eterm list, UWord max_length, iovec_slice_t *resul return 1; } -static void inspect_raw_binary_data(Eterm binary, ErlNifBinary *result) { +static void marshal_iovec_binary(Eterm binary, ErlNifBinary *copy_buffer, + UWord *copy_offset, ErlNifBinary *result) { + Eterm *parent_header; Eterm parent_binary; @@ -3484,6 +3487,8 @@ static void inspect_raw_binary_data(Eterm binary, ErlNifBinary *result) { ERTS_GET_REAL_BIN(binary, parent_binary, byte_offset, bit_offset, bit_size); + ASSERT(bit_size == 0); + parent_header = binary_val(parent_binary); result->size = binary_size(binary); @@ -3510,24 +3515,50 @@ static void inspect_raw_binary_data(Eterm binary, ErlNifBinary *result) { result->data = &((unsigned char*)&hb->data)[byte_offset]; result->ref_bin = NULL; } + + /* If this isn't an *aligned* refc binary, copy its contents to the buffer + * and reference that instead. */ + + if (result->ref_bin == NULL || bit_offset != 0) { + ASSERT(result->size <= (copy_buffer->size - *copy_offset)); + + if (bit_offset == 0) { + sys_memcpy(©_buffer->data[*copy_offset], + result->data, result->size); + } else { + erts_copy_bits(result->data, bit_offset, 1, + (byte*)©_buffer->data[*copy_offset], 0, 1, + result->size * 8); + } + + result->data = ©_buffer->data[*copy_offset]; + result->ref_bin = copy_buffer->ref_bin; + + *copy_offset += result->size; + } } static int fill_iovec_with_slice(ErlNifEnv *env, iovec_slice_t *slice, ErlNifIOVec *iovec) { - UWord onheap_offset, iovec_idx; - ErlNifBinary onheap_data; + UWord copy_offset, iovec_idx; + ErlNifBinary copy_buffer; Eterm sublist_iterator; - /* Set up a common refc binary for all on-heap binaries. */ - if (slice->onheap_size > 0) { - if (!enif_alloc_binary(slice->onheap_size, &onheap_data)) { + /* Set up a common refc binary for all on-heap and unaligned binaries. */ + if (slice->copied_size > 0) { + if (!enif_alloc_binary(slice->copied_size, ©_buffer)) { return 0; } + } else { +#ifdef DEBUG + copy_buffer.data = NULL; + copy_buffer.size = 0; +#endif } sublist_iterator = slice->sublist_start; - onheap_offset = 0; + copy_offset = 0; iovec_idx = 0; while (sublist_iterator != slice->sublist_end) { @@ -3535,20 +3566,7 @@ static int fill_iovec_with_slice(ErlNifEnv *env, Eterm *cell; cell = list_val(sublist_iterator); - inspect_raw_binary_data(CAR(cell), &raw_data); - - /* If this isn't a refc binary, copy its contents to the onheap buffer - * and reference that instead. */ - if (raw_data.size > 0 && raw_data.ref_bin == NULL) { - ASSERT(onheap_offset < onheap_data.size); - ASSERT(slice->onheap_size > 0); - - sys_memcpy(&onheap_data.data[onheap_offset], - raw_data.data, raw_data.size); - - raw_data.data = &onheap_data.data[onheap_offset]; - raw_data.ref_bin = onheap_data.ref_bin; - } + marshal_iovec_binary(CAR(cell), ©_buffer, ©_offset, &raw_data); while (raw_data.size > 0) { UWord chunk_len = MIN(raw_data.size, MAX_SYSIOVEC_IOVLEN); @@ -3579,16 +3597,16 @@ static int fill_iovec_with_slice(ErlNifEnv *env, erts_refc_inc(&refc_binary->intern.refc, 1); } - if (slice->onheap_size > 0) { + if (slice->copied_size > 0) { /* Transfer ownership to the iovec; we've taken references to it in * the above loop. */ - enif_release_binary(&onheap_data); + enif_release_binary(©_buffer); } } else { - if (slice->onheap_size > 0) { + if (slice->copied_size > 0) { /* Attach the binary to our environment and let the GC take care of * it after returning. */ - enif_make_binary(env, &onheap_data); + enif_make_binary(env, ©_buffer); } } @@ -3635,7 +3653,7 @@ static int create_iovec_from_slice(ErlNifEnv *env, iovec->flags = 0; } - iovec->size = slice->offheap_size + slice->onheap_size; + iovec->size = slice->referenced_size + slice->copied_size; iovec->iovcnt = slice->iovec_len; if(!fill_iovec_with_slice(env, slice, iovec)) { -- cgit v1.2.3