On 9/30/20 11:52 AM, Chautru, Nicolas wrote:
> Hi Tom,
>
>> From: Tom Rix <t...@redhat.com>
>> On 9/28/20 5:29 PM, Nicolas Chautru wrote:
>>> Adding LDPC decode and encode processing operations
>>>
>>> Signed-off-by: Nicolas Chautru <nicolas.chau...@intel.com>
>>> Acked-by: Liu Tianjiao <tianjiao....@intel.com>
>>> Acked-by: Dave Burley <dave.bur...@accelercomm.com>
>>> ---
>>> doc/guides/bbdevs/features/acc100.ini | 8 +-
>>> drivers/baseband/acc100/rte_acc100_pmd.c | 1625
>> +++++++++++++++++++++++++++++-
>>> drivers/baseband/acc100/rte_acc100_pmd.h | 3 +
>>> 3 files changed, 1630 insertions(+), 6 deletions(-)
>>>
>>> diff --git a/doc/guides/bbdevs/features/acc100.ini
>> b/doc/guides/bbdevs/features/acc100.ini
>>> index c89a4d7..40c7adc 100644
>>> --- a/doc/guides/bbdevs/features/acc100.ini
>>> +++ b/doc/guides/bbdevs/features/acc100.ini
>>> @@ -6,9 +6,9 @@
>>> [Features]
>>> Turbo Decoder (4G) = N
>>> Turbo Encoder (4G) = N
>>> -LDPC Decoder (5G) = N
>>> -LDPC Encoder (5G) = N
>>> -LLR/HARQ Compression = N
>>> -External DDR Access = N
>>> +LDPC Decoder (5G) = Y
>>> +LDPC Encoder (5G) = Y
>>> +LLR/HARQ Compression = Y
>>> +External DDR Access = Y
>>> HW Accelerated = Y
>>> BBDEV API = Y
>>> diff --git a/drivers/baseband/acc100/rte_acc100_pmd.c
>> b/drivers/baseband/acc100/rte_acc100_pmd.c
>>> index 7a21c57..b223547 100644
>>> --- a/drivers/baseband/acc100/rte_acc100_pmd.c
>>> +++ b/drivers/baseband/acc100/rte_acc100_pmd.c
>>> @@ -15,6 +15,9 @@
>>> #include <rte_hexdump.h>
>>> #include <rte_pci.h>
>>> #include <rte_bus_pci.h>
>>> +#ifdef RTE_BBDEV_OFFLOAD_COST
>>> +#include <rte_cycles.h>
>>> +#endif
>>>
>>> #include <rte_bbdev.h>
>>> #include <rte_bbdev_pmd.h>
>>> @@ -449,7 +452,6 @@
>>> return 0;
>>> }
>>>
>>> -
>>> /**
>>> * Report a ACC100 queue index which is free
>>> * Return 0 to 16k for a valid queue_idx or -1 when no queue is available
>>> @@ -634,6 +636,46 @@
>>> struct acc100_device *d = dev->data->dev_private;
>>>
>>> static const struct rte_bbdev_op_cap bbdev_capabilities[] = {
>>> + {
>>> + .type = RTE_BBDEV_OP_LDPC_ENC,
>>> + .cap.ldpc_enc = {
>>> + .capability_flags =
>>> + RTE_BBDEV_LDPC_RATE_MATCH |
>>> +
>> RTE_BBDEV_LDPC_CRC_24B_ATTACH |
>>> +
>> RTE_BBDEV_LDPC_INTERLEAVER_BYPASS,
>>> + .num_buffers_src =
>>> +
>> RTE_BBDEV_LDPC_MAX_CODE_BLOCKS,
>>> + .num_buffers_dst =
>>> +
>> RTE_BBDEV_LDPC_MAX_CODE_BLOCKS,
>>> + }
>>> + },
>>> + {
>>> + .type = RTE_BBDEV_OP_LDPC_DEC,
>>> + .cap.ldpc_dec = {
>>> + .capability_flags =
>>> + RTE_BBDEV_LDPC_CRC_TYPE_24B_CHECK |
>>> + RTE_BBDEV_LDPC_CRC_TYPE_24B_DROP |
>>> +
>> RTE_BBDEV_LDPC_HQ_COMBINE_IN_ENABLE |
>>> +
>> RTE_BBDEV_LDPC_HQ_COMBINE_OUT_ENABLE |
>>> +#ifdef ACC100_EXT_MEM
>> This is unconditionally defined in rte_acc100_pmd.h but it seems
>>
>> like it could be a hw config. Please add a comment in the *.h
>>
> It is not really an HW config, just a potential alternate way to run
> the device notably for troubleshooting.
> I can add a comment though
>
>> Could also change to
>>
>> #if ACC100_EXT_MEM
>>
>> and change the #define ACC100_EXT_MEM 1
> ok
>
>>> +
>> RTE_BBDEV_LDPC_INTERNAL_HARQ_MEMORY_IN_ENABLE |
>>> +
>> RTE_BBDEV_LDPC_INTERNAL_HARQ_MEMORY_OUT_ENABLE |
>>> +#endif
>>> +
>> RTE_BBDEV_LDPC_ITERATION_STOP_ENABLE |
>>> + RTE_BBDEV_LDPC_DEINTERLEAVER_BYPASS
>> |
>>> + RTE_BBDEV_LDPC_DECODE_BYPASS |
>>> + RTE_BBDEV_LDPC_DEC_SCATTER_GATHER |
>>> +
>> RTE_BBDEV_LDPC_HARQ_6BIT_COMPRESSION |
>>> + RTE_BBDEV_LDPC_LLR_COMPRESSION,
>>> + .llr_size = 8,
>>> + .llr_decimals = 1,
>>> + .num_buffers_src =
>>> +
>> RTE_BBDEV_LDPC_MAX_CODE_BLOCKS,
>>> + .num_buffers_hard_out =
>>> +
>> RTE_BBDEV_LDPC_MAX_CODE_BLOCKS,
>>> + .num_buffers_soft_out = 0,
>>> + }
>>> + },
>>> RTE_BBDEV_END_OF_CAPABILITIES_LIST()
>>> };
>>>
>>> @@ -669,9 +711,14 @@
>>> dev_info->cpu_flag_reqs = NULL;
>>> dev_info->min_alignment = 64;
>>> dev_info->capabilities = bbdev_capabilities;
>>> +#ifdef ACC100_EXT_MEM
>>> dev_info->harq_buffer_size = d->ddr_size;
>>> +#else
>>> + dev_info->harq_buffer_size = 0;
>>> +#endif
>>> }
>>>
>>> +
>>> static const struct rte_bbdev_ops acc100_bbdev_ops = {
>>> .setup_queues = acc100_setup_queues,
>>> .close = acc100_dev_close,
>>> @@ -696,6 +743,1577 @@
>>> {.device_id = 0},
>>> };
>>>
>>> +/* Read flag value 0/1 from bitmap */
>>> +static inline bool
>>> +check_bit(uint32_t bitmap, uint32_t bitmask)
>>> +{
>>> + return bitmap & bitmask;
>>> +}
>>> +
>> All the bbdev have this function, its pretty trival but it would be good if
>> common bbdev
>>
>> functions got moved to a common place.
> Noted for future change affecting all PMDs outside of that serie.
ok.
>
>>> +static inline char *
>>> +mbuf_append(struct rte_mbuf *m_head, struct rte_mbuf *m, uint16_t
>> len)
>>> +{
>>> + if (unlikely(len > rte_pktmbuf_tailroom(m)))
>>> + return NULL;
>>> +
>>> + char *tail = (char *)m->buf_addr + m->data_off + m->data_len;
>>> + m->data_len = (uint16_t)(m->data_len + len);
>>> + m_head->pkt_len = (m_head->pkt_len + len);
>>> + return tail;
>>> +}
>>> +
>>> +/* Compute value of k0.
>>> + * Based on 3GPP 38.212 Table 5.4.2.1-2
>>> + * Starting position of different redundancy versions, k0
>>> + */
>>> +static inline uint16_t
>>> +get_k0(uint16_t n_cb, uint16_t z_c, uint8_t bg, uint8_t rv_index)
>>> +{
>>> + if (rv_index == 0)
>>> + return 0;
>>> + uint16_t n = (bg == 1 ? N_ZC_1 : N_ZC_2) * z_c;
>>> + if (n_cb == n) {
>>> + if (rv_index == 1)
>>> + return (bg == 1 ? K0_1_1 : K0_1_2) * z_c;
>>> + else if (rv_index == 2)
>>> + return (bg == 1 ? K0_2_1 : K0_2_2) * z_c;
>>> + else
>>> + return (bg == 1 ? K0_3_1 : K0_3_2) * z_c;
>>> + }
>>> + /* LBRM case - includes a division by N */
>>> + if (rv_index == 1)
>>> + return (((bg == 1 ? K0_1_1 : K0_1_2) * n_cb)
>>> + / n) * z_c;
>>> + else if (rv_index == 2)
>>> + return (((bg == 1 ? K0_2_1 : K0_2_2) * n_cb)
>>> + / n) * z_c;
>>> + else
>>> + return (((bg == 1 ? K0_3_1 : K0_3_2) * n_cb)
>>> + / n) * z_c;
>>> +}
>>> +
>>> +/* Fill in a frame control word for LDPC encoding. */
>>> +static inline void
>>> +acc100_fcw_le_fill(const struct rte_bbdev_enc_op *op,
>>> + struct acc100_fcw_le *fcw, int num_cb)
>>> +{
>>> + fcw->qm = op->ldpc_enc.q_m;
>>> + fcw->nfiller = op->ldpc_enc.n_filler;
>>> + fcw->BG = (op->ldpc_enc.basegraph - 1);
>>> + fcw->Zc = op->ldpc_enc.z_c;
>>> + fcw->ncb = op->ldpc_enc.n_cb;
>>> + fcw->k0 = get_k0(fcw->ncb, fcw->Zc, op->ldpc_enc.basegraph,
>>> + op->ldpc_enc.rv_index);
>>> + fcw->rm_e = op->ldpc_enc.cb_params.e;
>>> + fcw->crc_select = check_bit(op->ldpc_enc.op_flags,
>>> + RTE_BBDEV_LDPC_CRC_24B_ATTACH);
>>> + fcw->bypass_intlv = check_bit(op->ldpc_enc.op_flags,
>>> + RTE_BBDEV_LDPC_INTERLEAVER_BYPASS);
>>> + fcw->mcb_count = num_cb;
>>> +}
>>> +
>>> +/* Fill in a frame control word for LDPC decoding. */
>>> +static inline void
>>> +acc100_fcw_ld_fill(const struct rte_bbdev_dec_op *op, struct
>> acc100_fcw_ld *fcw,
>>> + union acc100_harq_layout_data *harq_layout)
>>> +{
>>> + uint16_t harq_out_length, harq_in_length, ncb_p, k0_p,
>> parity_offset;
>>> + uint16_t harq_index;
>>> + uint32_t l;
>>> + bool harq_prun = false;
>>> +
>>> + fcw->qm = op->ldpc_dec.q_m;
>>> + fcw->nfiller = op->ldpc_dec.n_filler;
>>> + fcw->BG = (op->ldpc_dec.basegraph - 1);
>>> + fcw->Zc = op->ldpc_dec.z_c;
>>> + fcw->ncb = op->ldpc_dec.n_cb;
>>> + fcw->k0 = get_k0(fcw->ncb, fcw->Zc, op->ldpc_dec.basegraph,
>>> + op->ldpc_dec.rv_index);
>>> + if (op->ldpc_dec.code_block_mode == 1)
>> 1 is magic, consider a #define
> This would be a changed not related to that PMD, but noted and agreed.
>
>>> + fcw->rm_e = op->ldpc_dec.cb_params.e;
>>> + else
>>> + fcw->rm_e = (op->ldpc_dec.tb_params.r <
>>> + op->ldpc_dec.tb_params.cab) ?
>>> + op->ldpc_dec.tb_params.ea :
>>> + op->ldpc_dec.tb_params.eb;
>>> +
>>> + fcw->hcin_en = check_bit(op->ldpc_dec.op_flags,
>>> + RTE_BBDEV_LDPC_HQ_COMBINE_IN_ENABLE);
>>> + fcw->hcout_en = check_bit(op->ldpc_dec.op_flags,
>>> + RTE_BBDEV_LDPC_HQ_COMBINE_OUT_ENABLE);
>>> + fcw->crc_select = check_bit(op->ldpc_dec.op_flags,
>>> + RTE_BBDEV_LDPC_CRC_TYPE_24B_CHECK);
>>> + fcw->bypass_dec = check_bit(op->ldpc_dec.op_flags,
>>> + RTE_BBDEV_LDPC_DECODE_BYPASS);
>>> + fcw->bypass_intlv = check_bit(op->ldpc_dec.op_flags,
>>> + RTE_BBDEV_LDPC_DEINTERLEAVER_BYPASS);
>>> + if (op->ldpc_dec.q_m == 1) {
>>> + fcw->bypass_intlv = 1;
>>> + fcw->qm = 2;
>>> + }
>> similar magic.
> Qm is an integer number defined in 3GPP, not a magic number. This literally
> means qm = 2.
ok
>>> + fcw->hcin_decomp_mode = check_bit(op->ldpc_dec.op_flags,
>>> + RTE_BBDEV_LDPC_HARQ_6BIT_COMPRESSION);
>>> + fcw->hcout_comp_mode = check_bit(op->ldpc_dec.op_flags,
>>> + RTE_BBDEV_LDPC_HARQ_6BIT_COMPRESSION);
>>> + fcw->llr_pack_mode = check_bit(op->ldpc_dec.op_flags,
>>> + RTE_BBDEV_LDPC_LLR_COMPRESSION);
>>> + harq_index = op->ldpc_dec.harq_combined_output.offset /
>>> + ACC100_HARQ_OFFSET;
>>> +#ifdef ACC100_EXT_MEM
>>> + /* Limit cases when HARQ pruning is valid */
>>> + harq_prun = ((op->ldpc_dec.harq_combined_output.offset %
>>> + ACC100_HARQ_OFFSET) == 0) &&
>>> + (op->ldpc_dec.harq_combined_output.offset <=
>> UINT16_MAX
>>> + * ACC100_HARQ_OFFSET);
>>> +#endif
>>> + if (fcw->hcin_en > 0) {
>>> + harq_in_length = op-
>>> ldpc_dec.harq_combined_input.length;
>>> + if (fcw->hcin_decomp_mode > 0)
>>> + harq_in_length = harq_in_length * 8 / 6;
>>> + harq_in_length = RTE_ALIGN(harq_in_length, 64);
>>> + if ((harq_layout[harq_index].offset > 0) & harq_prun) {
>>> + rte_bbdev_log_debug("HARQ IN offset unexpected
>> for now\n");
>>> + fcw->hcin_size0 = harq_layout[harq_index].size0;
>>> + fcw->hcin_offset = harq_layout[harq_index].offset;
>>> + fcw->hcin_size1 = harq_in_length -
>>> + harq_layout[harq_index].offset;
>>> + } else {
>>> + fcw->hcin_size0 = harq_in_length;
>>> + fcw->hcin_offset = 0;
>>> + fcw->hcin_size1 = 0;
>>> + }
>>> + } else {
>>> + fcw->hcin_size0 = 0;
>>> + fcw->hcin_offset = 0;
>>> + fcw->hcin_size1 = 0;
>>> + }
>>> +
>>> + fcw->itmax = op->ldpc_dec.iter_max;
>>> + fcw->itstop = check_bit(op->ldpc_dec.op_flags,
>>> + RTE_BBDEV_LDPC_ITERATION_STOP_ENABLE);
>>> + fcw->synd_precoder = fcw->itstop;
>>> + /*
>>> + * These are all implicitly set
>>> + * fcw->synd_post = 0;
>>> + * fcw->so_en = 0;
>>> + * fcw->so_bypass_rm = 0;
>>> + * fcw->so_bypass_intlv = 0;
>>> + * fcw->dec_convllr = 0;
>>> + * fcw->hcout_convllr = 0;
>>> + * fcw->hcout_size1 = 0;
>>> + * fcw->so_it = 0;
>>> + * fcw->hcout_offset = 0;
>>> + * fcw->negstop_th = 0;
>>> + * fcw->negstop_it = 0;
>>> + * fcw->negstop_en = 0;
>>> + * fcw->gain_i = 1;
>>> + * fcw->gain_h = 1;
>>> + */
>>> + if (fcw->hcout_en > 0) {
>>> + parity_offset = (op->ldpc_dec.basegraph == 1 ? 20 : 8)
>>> + * op->ldpc_dec.z_c - op->ldpc_dec.n_filler;
>>> + k0_p = (fcw->k0 > parity_offset) ?
>>> + fcw->k0 - op->ldpc_dec.n_filler : fcw->k0;
>>> + ncb_p = fcw->ncb - op->ldpc_dec.n_filler;
>>> + l = k0_p + fcw->rm_e;
>>> + harq_out_length = (uint16_t) fcw->hcin_size0;
>>> + harq_out_length = RTE_MIN(RTE_MAX(harq_out_length, l),
>> ncb_p);
>>> + harq_out_length = (harq_out_length + 0x3F) & 0xFFC0;
>>> + if ((k0_p > fcw->hcin_size0 +
>> ACC100_HARQ_OFFSET_THRESHOLD) &&
>>> + harq_prun) {
>>> + fcw->hcout_size0 = (uint16_t) fcw->hcin_size0;
>>> + fcw->hcout_offset = k0_p & 0xFFC0;
>>> + fcw->hcout_size1 = harq_out_length - fcw-
>>> hcout_offset;
>>> + } else {
>>> + fcw->hcout_size0 = harq_out_length;
>>> + fcw->hcout_size1 = 0;
>>> + fcw->hcout_offset = 0;
>>> + }
>>> + harq_layout[harq_index].offset = fcw->hcout_offset;
>>> + harq_layout[harq_index].size0 = fcw->hcout_size0;
>>> + } else {
>>> + fcw->hcout_size0 = 0;
>>> + fcw->hcout_size1 = 0;
>>> + fcw->hcout_offset = 0;
>>> + }
>>> +}
>>> +
>>> +/**
>>> + * Fills descriptor with data pointers of one block type.
>>> + *
>>> + * @param desc
>>> + * Pointer to DMA descriptor.
>>> + * @param input
>>> + * Pointer to pointer to input data which will be encoded. It can be
>> changed
>>> + * and points to next segment in scatter-gather case.
>>> + * @param offset
>>> + * Input offset in rte_mbuf structure. It is used for calculating the
>>> point
>>> + * where data is starting.
>>> + * @param cb_len
>>> + * Length of currently processed Code Block
>>> + * @param seg_total_left
>>> + * It indicates how many bytes still left in segment (mbuf) for further
>>> + * processing.
>>> + * @param op_flags
>>> + * Store information about device capabilities
>>> + * @param next_triplet
>>> + * Index for ACC100 DMA Descriptor triplet
>>> + *
>>> + * @return
>>> + * Returns index of next triplet on success, other value if lengths of
>>> + * pkt and processed cb do not match.
>>> + *
>>> + */
>>> +static inline int
>>> +acc100_dma_fill_blk_type_in(struct acc100_dma_req_desc *desc,
>>> + struct rte_mbuf **input, uint32_t *offset, uint32_t cb_len,
>>> + uint32_t *seg_total_left, int next_triplet)
>>> +{
>>> + uint32_t part_len;
>>> + struct rte_mbuf *m = *input;
>>> +
>>> + part_len = (*seg_total_left < cb_len) ? *seg_total_left : cb_len;
>>> + cb_len -= part_len;
>>> + *seg_total_left -= part_len;
>>> +
>>> + desc->data_ptrs[next_triplet].address =
>>> + rte_pktmbuf_iova_offset(m, *offset);
>>> + desc->data_ptrs[next_triplet].blen = part_len;
>>> + desc->data_ptrs[next_triplet].blkid = ACC100_DMA_BLKID_IN;
>>> + desc->data_ptrs[next_triplet].last = 0;
>>> + desc->data_ptrs[next_triplet].dma_ext = 0;
>>> + *offset += part_len;
>>> + next_triplet++;
>>> +
>>> + while (cb_len > 0) {
>> Since cb_len is unsigned, a better check would be
>>
>> while (cb_len != 0)
> Why would this be better?
It is unsigned it will never be < 0.
!= 0 reflects that.
>
>>> + if (next_triplet < ACC100_DMA_MAX_NUM_POINTERS &&
>>> + m->next != NULL) {
>>> +
>>> + m = m->next;
>>> + *seg_total_left = rte_pktmbuf_data_len(m);
>>> + part_len = (*seg_total_left < cb_len) ?
>>> + *seg_total_left :
>>> + cb_len;
>>> + desc->data_ptrs[next_triplet].address =
>>> + rte_pktmbuf_iova_offset(m, 0);
>>> + desc->data_ptrs[next_triplet].blen = part_len;
>>> + desc->data_ptrs[next_triplet].blkid =
>>> + ACC100_DMA_BLKID_IN;
>>> + desc->data_ptrs[next_triplet].last = 0;
>>> + desc->data_ptrs[next_triplet].dma_ext = 0;
>>> + cb_len -= part_len;
>>> + *seg_total_left -= part_len;
>> when *sec_total_left goes to zero here, there will be a lot of iterations
>> doing
>> nothing.
>>
>> should stop early.
> Not really, it would pick next m anyway and keep adding buffer descriptor
> pointer.
ok
>
>
>>> + /* Initializing offset for next segment (mbuf) */
>>> + *offset = part_len;
>>> + next_triplet++;
>>> + } else {
>>> + rte_bbdev_log(ERR,
>>> + "Some data still left for processing: "
>>> + "data_left: %u, next_triplet: %u, next_mbuf:
>> %p",
>>> + cb_len, next_triplet, m->next);
>>> + return -EINVAL;
>>> + }
>>> + }
>>> + /* Storing new mbuf as it could be changed in scatter-gather case*/
>>> + *input = m;
>>> +
>>> + return next_triplet;
>> callers, after checking, dec the return.
>>
>> Maybe change return to next_triplet-- and save the callers from doing it.
> I miss your point
Looking at how the callers of this function use the return,
a fair number decrement it to get to the current_triplet.
So maybe returning the current_triplet would be better.
Something to think about, not required.
>>> +}
>>> +
>>> +/* Fills descriptor with data pointers of one block type.
>>> + * Returns index of next triplet on success, other value if lengths of
>>> + * output data and processed mbuf do not match.
>>> + */
>>> +static inline int
>>> +acc100_dma_fill_blk_type_out(struct acc100_dma_req_desc *desc,
>>> + struct rte_mbuf *output, uint32_t out_offset,
>>> + uint32_t output_len, int next_triplet, int blk_id)
>>> +{
>>> + desc->data_ptrs[next_triplet].address =
>>> + rte_pktmbuf_iova_offset(output, out_offset);
>>> + desc->data_ptrs[next_triplet].blen = output_len;
>>> + desc->data_ptrs[next_triplet].blkid = blk_id;
>>> + desc->data_ptrs[next_triplet].last = 0;
>>> + desc->data_ptrs[next_triplet].dma_ext = 0;
>>> + next_triplet++;
>> Callers check return is < 0, like above but there is no similar logic to
>>
>> check the bounds of next_triplet to return -EINVAL
>>
>> so add this check here or remove the is < 0 checks by the callers.
>>
> fair enough thanks.
>
>>> +
>>> + return next_triplet;
>>> +}
>>> +
>>> +static inline int
>>> +acc100_dma_desc_le_fill(struct rte_bbdev_enc_op *op,
>>> + struct acc100_dma_req_desc *desc, struct rte_mbuf **input,
>>> + struct rte_mbuf *output, uint32_t *in_offset,
>>> + uint32_t *out_offset, uint32_t *out_length,
>>> + uint32_t *mbuf_total_left, uint32_t *seg_total_left)
>>> +{
>>> + int next_triplet = 1; /* FCW already done */
>>> + uint16_t K, in_length_in_bits, in_length_in_bytes;
>>> + struct rte_bbdev_op_ldpc_enc *enc = &op->ldpc_enc;
>>> +
>>> + desc->word0 = ACC100_DMA_DESC_TYPE;
>>> + desc->word1 = 0; /**< Timestamp could be disabled */
>>> + desc->word2 = 0;
>>> + desc->word3 = 0;
>>> + desc->numCBs = 1;
>>> +
>>> + K = (enc->basegraph == 1 ? 22 : 10) * enc->z_c;
>>> + in_length_in_bits = K - enc->n_filler;
>> can this overflow ? enc->n_filler > K ?
> I would not add such checks in the time critical function. For valid scenario
> it can't.
> It could be added to the validate_ldpc_dec_op() which is only run in debug
> mode.
>
>>> + if ((enc->op_flags & RTE_BBDEV_LDPC_CRC_24A_ATTACH) ||
>>> + (enc->op_flags &
>> RTE_BBDEV_LDPC_CRC_24B_ATTACH))
>>> + in_length_in_bits -= 24;
>>> + in_length_in_bytes = in_length_in_bits >> 3;
>>> +
>>> + if (unlikely((*mbuf_total_left == 0) ||
>> This check is covered by the next and can be removed.
> Not necessaraly, would keep as is.
only if in_length_in_bytes was negative
>
>>> + (*mbuf_total_left < in_length_in_bytes))) {
>>> + rte_bbdev_log(ERR,
>>> + "Mismatch between mbuf length and
>> included CB sizes: mbuf len %u, cb len %u",
>>> + *mbuf_total_left, in_length_in_bytes);
>>> + return -1;
>>> + }
>>> +
>>> + next_triplet = acc100_dma_fill_blk_type_in(desc, input, in_offset,
>>> + in_length_in_bytes,
>>> + seg_total_left, next_triplet);
>>> + if (unlikely(next_triplet < 0)) {
>>> + rte_bbdev_log(ERR,
>>> + "Mismatch between data to process and
>> mbuf data length in bbdev_op: %p",
>>> + op);
>>> + return -1;
>>> + }
>>> + desc->data_ptrs[next_triplet - 1].last = 1;
>>> + desc->m2dlen = next_triplet;
>>> + *mbuf_total_left -= in_length_in_bytes;
>> Updating output pointers should be deferred until the the call is known to
>> be successful.
>>
>> Otherwise caller is left in a bad, unknown state.
> We already had to touch them by that point.
ugh.
>
>>> +
>>> + /* Set output length */
>>> + /* Integer round up division by 8 */
>>> + *out_length = (enc->cb_params.e + 7) >> 3;
>>> +
>>> + next_triplet = acc100_dma_fill_blk_type_out(desc, output,
>> *out_offset,
>>> + *out_length, next_triplet,
>> ACC100_DMA_BLKID_OUT_ENC);
>>> + if (unlikely(next_triplet < 0)) {
>>> + rte_bbdev_log(ERR,
>>> + "Mismatch between data to process and
>> mbuf data length in bbdev_op: %p",
>>> + op);
>>> + return -1;
>>> + }
>>> + op->ldpc_enc.output.length += *out_length;
>>> + *out_offset += *out_length;
>>> + desc->data_ptrs[next_triplet - 1].last = 1;
>>> + desc->data_ptrs[next_triplet - 1].dma_ext = 0;
>>> + desc->d2mlen = next_triplet - desc->m2dlen;
>>> +
>>> + desc->op_addr = op;
>>> +
>>> + return 0;
>>> +}
>>> +
>>> +static inline int
>>> +acc100_dma_desc_ld_fill(struct rte_bbdev_dec_op *op,
>>> + struct acc100_dma_req_desc *desc,
>>> + struct rte_mbuf **input, struct rte_mbuf *h_output,
>>> + uint32_t *in_offset, uint32_t *h_out_offset,
>>> + uint32_t *h_out_length, uint32_t *mbuf_total_left,
>>> + uint32_t *seg_total_left,
>>> + struct acc100_fcw_ld *fcw)
>>> +{
>>> + struct rte_bbdev_op_ldpc_dec *dec = &op->ldpc_dec;
>>> + int next_triplet = 1; /* FCW already done */
>>> + uint32_t input_length;
>>> + uint16_t output_length, crc24_overlap = 0;
>>> + uint16_t sys_cols, K, h_p_size, h_np_size;
>>> + bool h_comp = check_bit(dec->op_flags,
>>> + RTE_BBDEV_LDPC_HARQ_6BIT_COMPRESSION);
>>> +
>>> + desc->word0 = ACC100_DMA_DESC_TYPE;
>>> + desc->word1 = 0; /**< Timestamp could be disabled */
>>> + desc->word2 = 0;
>>> + desc->word3 = 0;
>>> + desc->numCBs = 1;
>> This seems to be a common setup logic, maybe use a macro or inline
>> function.
> fair enough
>
>>> +
>>> + if (check_bit(op->ldpc_dec.op_flags,
>>> + RTE_BBDEV_LDPC_CRC_TYPE_24B_DROP))
>>> + crc24_overlap = 24;
>>> +
>>> + /* Compute some LDPC BG lengths */
>>> + input_length = dec->cb_params.e;
>>> + if (check_bit(op->ldpc_dec.op_flags,
>>> + RTE_BBDEV_LDPC_LLR_COMPRESSION))
>>> + input_length = (input_length * 3 + 3) / 4;
>>> + sys_cols = (dec->basegraph == 1) ? 22 : 10;
>>> + K = sys_cols * dec->z_c;
>>> + output_length = K - dec->n_filler - crc24_overlap;
>>> +
>>> + if (unlikely((*mbuf_total_left == 0) ||
>> similar to above, this check can be removed.
> same comment
>
>>> + (*mbuf_total_left < input_length))) {
>>> + rte_bbdev_log(ERR,
>>> + "Mismatch between mbuf length and
>> included CB sizes: mbuf len %u, cb len %u",
>>> + *mbuf_total_left, input_length);
>>> + return -1;
>>> + }
>>> +
>>> + next_triplet = acc100_dma_fill_blk_type_in(desc, input,
>>> + in_offset, input_length,
>>> + seg_total_left, next_triplet);
>>> +
>>> + if (unlikely(next_triplet < 0)) {
>>> + rte_bbdev_log(ERR,
>>> + "Mismatch between data to process and
>> mbuf data length in bbdev_op: %p",
>>> + op);
>>> + return -1;
>>> + }
>>> +
>>> + if (check_bit(op->ldpc_dec.op_flags,
>>> +
>> RTE_BBDEV_LDPC_HQ_COMBINE_IN_ENABLE)) {
>>> + h_p_size = fcw->hcin_size0 + fcw->hcin_size1;
>>> + if (h_comp)
>>> + h_p_size = (h_p_size * 3 + 3) / 4;
>>> + desc->data_ptrs[next_triplet].address =
>>> + dec->harq_combined_input.offset;
>>> + desc->data_ptrs[next_triplet].blen = h_p_size;
>>> + desc->data_ptrs[next_triplet].blkid =
>> ACC100_DMA_BLKID_IN_HARQ;
>>> + desc->data_ptrs[next_triplet].dma_ext = 1;
>>> +#ifndef ACC100_EXT_MEM
>>> + acc100_dma_fill_blk_type_out(
>>> + desc,
>>> + op->ldpc_dec.harq_combined_input.data,
>>> + op->ldpc_dec.harq_combined_input.offset,
>>> + h_p_size,
>>> + next_triplet,
>>> + ACC100_DMA_BLKID_IN_HARQ);
>>> +#endif
>>> + next_triplet++;
>>> + }
>>> +
>>> + desc->data_ptrs[next_triplet - 1].last = 1;
>>> + desc->m2dlen = next_triplet;
>>> + *mbuf_total_left -= input_length;
>>> +
>>> + next_triplet = acc100_dma_fill_blk_type_out(desc, h_output,
>>> + *h_out_offset, output_length >> 3, next_triplet,
>>> + ACC100_DMA_BLKID_OUT_HARD);
>>> + if (unlikely(next_triplet < 0)) {
>>> + rte_bbdev_log(ERR,
>>> + "Mismatch between data to process and
>> mbuf data length in bbdev_op: %p",
>>> + op);
>>> + return -1;
>>> + }
>>> +
>>> + if (check_bit(op->ldpc_dec.op_flags,
>>> +
>> RTE_BBDEV_LDPC_HQ_COMBINE_OUT_ENABLE)) {
>>> + /* Pruned size of the HARQ */
>>> + h_p_size = fcw->hcout_size0 + fcw->hcout_size1;
>>> + /* Non-Pruned size of the HARQ */
>>> + h_np_size = fcw->hcout_offset > 0 ?
>>> + fcw->hcout_offset + fcw->hcout_size1 :
>>> + h_p_size;
>>> + if (h_comp) {
>>> + h_np_size = (h_np_size * 3 + 3) / 4;
>>> + h_p_size = (h_p_size * 3 + 3) / 4;
>> * 4 -1 ) / 4
>>
>> may produce better assembly.
> that is not the same arithmetic
?
>>> + }
>>> + dec->harq_combined_output.length = h_np_size;
>>> + desc->data_ptrs[next_triplet].address =
>>> + dec->harq_combined_output.offset;
>>> + desc->data_ptrs[next_triplet].blen = h_p_size;
>>> + desc->data_ptrs[next_triplet].blkid =
>> ACC100_DMA_BLKID_OUT_HARQ;
>>> + desc->data_ptrs[next_triplet].dma_ext = 1;
>>> +#ifndef ACC100_EXT_MEM
>>> + acc100_dma_fill_blk_type_out(
>>> + desc,
>>> + dec->harq_combined_output.data,
>>> + dec->harq_combined_output.offset,
>>> + h_p_size,
>>> + next_triplet,
>>> + ACC100_DMA_BLKID_OUT_HARQ);
>>> +#endif
>>> + next_triplet++;
>>> + }
>>> +
>>> + *h_out_length = output_length >> 3;
>>> + dec->hard_output.length += *h_out_length;
>>> + *h_out_offset += *h_out_length;
>>> + desc->data_ptrs[next_triplet - 1].last = 1;
>>> + desc->d2mlen = next_triplet - desc->m2dlen;
>>> +
>>> + desc->op_addr = op;
>>> +
>>> + return 0;
>>> +}
>>> +
>>> +static inline void
>>> +acc100_dma_desc_ld_update(struct rte_bbdev_dec_op *op,
>>> + struct acc100_dma_req_desc *desc,
>>> + struct rte_mbuf *input, struct rte_mbuf *h_output,
>>> + uint32_t *in_offset, uint32_t *h_out_offset,
>>> + uint32_t *h_out_length,
>>> + union acc100_harq_layout_data *harq_layout)
>>> +{
>>> + int next_triplet = 1; /* FCW already done */
>>> + desc->data_ptrs[next_triplet].address =
>>> + rte_pktmbuf_iova_offset(input, *in_offset);
>>> + next_triplet++;
>> No overflow checks on next_triplet
>>
>> This is a general problem.
> I dont see the overflow risk.
A lot of places increments without checking the bounds.
To me, it seems like we are getting lucky that data_ptrs[] is big enough.
>
>>> +
>>> + if (check_bit(op->ldpc_dec.op_flags,
>>> +
>> RTE_BBDEV_LDPC_HQ_COMBINE_IN_ENABLE)) {
>>> + struct rte_bbdev_op_data hi = op-
>>> ldpc_dec.harq_combined_input;
>>> + desc->data_ptrs[next_triplet].address = hi.offset;
>>> +#ifndef ACC100_EXT_MEM
>>> + desc->data_ptrs[next_triplet].address =
>>> + rte_pktmbuf_iova_offset(hi.data, hi.offset);
>>> +#endif
>>> + next_triplet++;
>>> + }
>>> +
>>> + desc->data_ptrs[next_triplet].address =
>>> + rte_pktmbuf_iova_offset(h_output, *h_out_offset);
>>> + *h_out_length = desc->data_ptrs[next_triplet].blen;
>>> + next_triplet++;
>>> +
>>> + if (check_bit(op->ldpc_dec.op_flags,
>>> +
>> RTE_BBDEV_LDPC_HQ_COMBINE_OUT_ENABLE)) {
>>> + desc->data_ptrs[next_triplet].address =
>>> + op->ldpc_dec.harq_combined_output.offset;
>>> + /* Adjust based on previous operation */
>>> + struct rte_bbdev_dec_op *prev_op = desc->op_addr;
>>> + op->ldpc_dec.harq_combined_output.length =
>>> + prev_op-
>>> ldpc_dec.harq_combined_output.length;
>>> + int16_t hq_idx = op-
>>> ldpc_dec.harq_combined_output.offset /
>>> + ACC100_HARQ_OFFSET;
>>> + int16_t prev_hq_idx =
>>> + prev_op-
>>> ldpc_dec.harq_combined_output.offset
>>> + / ACC100_HARQ_OFFSET;
>>> + harq_layout[hq_idx].val = harq_layout[prev_hq_idx].val;
>>> +#ifndef ACC100_EXT_MEM
>>> + struct rte_bbdev_op_data ho =
>>> + op->ldpc_dec.harq_combined_output;
>>> + desc->data_ptrs[next_triplet].address =
>>> + rte_pktmbuf_iova_offset(ho.data, ho.offset);
>>> +#endif
>>> + next_triplet++;
>>> + }
>>> +
>>> + op->ldpc_dec.hard_output.length += *h_out_length;
>>> + desc->op_addr = op;
>>> +}
>>> +
>>> +
>>> +/* Enqueue a number of operations to HW and update software rings */
>>> +static inline void
>>> +acc100_dma_enqueue(struct acc100_queue *q, uint16_t n,
>>> + struct rte_bbdev_stats *queue_stats)
>>> +{
>>> + union acc100_enqueue_reg_fmt enq_req;
>>> +#ifdef RTE_BBDEV_OFFLOAD_COST
>>> + uint64_t start_time = 0;
>>> + queue_stats->acc_offload_cycles = 0;
>>> + RTE_SET_USED(queue_stats);
>>> +#else
>>> + RTE_SET_USED(queue_stats);
>>> +#endif
>> RTE_SET_UNUSED(... is common in the #ifdef/#else
>>
>> so it should be moved out.
> ok
>
>>> +
>>> + enq_req.val = 0;
>>> + /* Setting offset, 100b for 256 DMA Desc */
>>> + enq_req.addr_offset = ACC100_DESC_OFFSET;
>>> +
>> should n != 0 be checked here ?
> This is all checked before that point.
ok
>
>>> + /* Split ops into batches */
>>> + do {
>>> + union acc100_dma_desc *desc;
>>> + uint16_t enq_batch_size;
>>> + uint64_t offset;
>>> + rte_iova_t req_elem_addr;
>>> +
>>> + enq_batch_size = RTE_MIN(n, MAX_ENQ_BATCH_SIZE);
>>> +
>>> + /* Set flag on last descriptor in a batch */
>>> + desc = q->ring_addr + ((q->sw_ring_head + enq_batch_size -
>> 1) &
>>> + q->sw_ring_wrap_mask);
>>> + desc->req.last_desc_in_batch = 1;
>>> +
>>> + /* Calculate the 1st descriptor's address */
>>> + offset = ((q->sw_ring_head & q->sw_ring_wrap_mask) *
>>> + sizeof(union acc100_dma_desc));
>>> + req_elem_addr = q->ring_addr_phys + offset;
>>> +
>>> + /* Fill enqueue struct */
>>> + enq_req.num_elem = enq_batch_size;
>>> + /* low 6 bits are not needed */
>>> + enq_req.req_elem_addr = (uint32_t)(req_elem_addr >> 6);
>>> +
>>> +#ifdef RTE_LIBRTE_BBDEV_DEBUG
>>> + rte_memdump(stderr, "Req sdone", desc, sizeof(*desc));
>>> +#endif
>>> + rte_bbdev_log_debug(
>>> + "Enqueue %u reqs (phys %#"PRIx64") to reg
>> %p",
>>> + enq_batch_size,
>>> + req_elem_addr,
>>> + (void *)q->mmio_reg_enqueue);
>>> +
>>> + rte_wmb();
>>> +
>>> +#ifdef RTE_BBDEV_OFFLOAD_COST
>>> + /* Start time measurement for enqueue function offload. */
>>> + start_time = rte_rdtsc_precise();
>>> +#endif
>>> + rte_bbdev_log(DEBUG, "Debug : MMIO Enqueue");
>> logging time will be tracked with the mmio_write
>>
>> so logging should be moved above the start_time setting
> Not required. Running with debug traces is expected to make real time offload
> measurement irrelevant.
I disagree, if logging has a pagefault or writes to disk even sometimes
there will be huge spike in the time that would make the accumulated
acc_offload_cycles meaningless. It would be ok if the write time is of
the same order of magnitude as disk access.
>>> + mmio_write(q->mmio_reg_enqueue, enq_req.val);
>>> +
>>> +#ifdef RTE_BBDEV_OFFLOAD_COST
>>> + queue_stats->acc_offload_cycles +=
>>> + rte_rdtsc_precise() - start_time;
>>> +#endif
>>> +
>>> + q->aq_enqueued++;
>>> + q->sw_ring_head += enq_batch_size;
>>> + n -= enq_batch_size;
>>> +
>>> + } while (n);
>>> +
>>> +
>>> +}
>>> +
>>> +/* Enqueue one encode operations for ACC100 device in CB mode */
>>> +static inline int
>>> +enqueue_ldpc_enc_n_op_cb(struct acc100_queue *q, struct
>> rte_bbdev_enc_op **ops,
>>> + uint16_t total_enqueued_cbs, int16_t num)
>>> +{
>>> + union acc100_dma_desc *desc = NULL;
>>> + uint32_t out_length;
>>> + struct rte_mbuf *output_head, *output;
>>> + int i, next_triplet;
>>> + uint16_t in_length_in_bytes;
>>> + struct rte_bbdev_op_ldpc_enc *enc = &ops[0]->ldpc_enc;
>>> +
>>> + uint16_t desc_idx = ((q->sw_ring_head + total_enqueued_cbs)
>>> + & q->sw_ring_wrap_mask);
>>> + desc = q->ring_addr + desc_idx;
>>> + acc100_fcw_le_fill(ops[0], &desc->req.fcw_le, num);
>>> +
>>> + /** This could be done at polling */
>>> + desc->req.word0 = ACC100_DMA_DESC_TYPE;
>>> + desc->req.word1 = 0; /**< Timestamp could be disabled */
>>> + desc->req.word2 = 0;
>>> + desc->req.word3 = 0;
>>> + desc->req.numCBs = num;
>>> +
>>> + in_length_in_bytes = ops[0]->ldpc_enc.input.data->data_len;
>>> + out_length = (enc->cb_params.e + 7) >> 3;
>>> + desc->req.m2dlen = 1 + num;
>>> + desc->req.d2mlen = num;
>>> + next_triplet = 1;
>>> +
>>> + for (i = 0; i < num; i++) {
>> i is not needed here, it is next_triplet - 1
> would impact readability as these refer to different concepts (code blocks
> and bdescs).
> Would keep as is
ok
>
>>> + desc->req.data_ptrs[next_triplet].address =
>>> + rte_pktmbuf_iova_offset(ops[i]-
>>> ldpc_enc.input.data, 0);
>>> + desc->req.data_ptrs[next_triplet].blen = in_length_in_bytes;
>>> + next_triplet++;
>>> + desc->req.data_ptrs[next_triplet].address =
>>> + rte_pktmbuf_iova_offset(
>>> + ops[i]->ldpc_enc.output.data, 0);
>>> + desc->req.data_ptrs[next_triplet].blen = out_length;
>>> + next_triplet++;
>>> + ops[i]->ldpc_enc.output.length = out_length;
>>> + output_head = output = ops[i]->ldpc_enc.output.data;
>>> + mbuf_append(output_head, output, out_length);
>>> + output->data_len = out_length;
>>> + }
>>> +
>>> + desc->req.op_addr = ops[0];
>>> +
>>> +#ifdef RTE_LIBRTE_BBDEV_DEBUG
>>> + rte_memdump(stderr, "FCW", &desc->req.fcw_le,
>>> + sizeof(desc->req.fcw_le) - 8);
>>> + rte_memdump(stderr, "Req Desc.", desc, sizeof(*desc));
>>> +#endif
>>> +
>>> + /* One CB (one op) was successfully prepared to enqueue */
>>> + return num;
>> caller does not use num, only check if < 0
>>
>> So could change to return 0
> would keep as is for debug
ok
>
>>> +}
>>> +
>>> +/* Enqueue one encode operations for ACC100 device in CB mode */
>>> +static inline int
>>> +enqueue_ldpc_enc_one_op_cb(struct acc100_queue *q, struct
>> rte_bbdev_enc_op *op,
>>> + uint16_t total_enqueued_cbs)
>> rte_fpga_5gnr_fec.c has this same function. It would be good if common
>> functions could be collected and used to stabilize the internal bbdev
>> interface.
>>
>> This is general issue
> This is true for some part of the code and noted.
> In that very case they are distinct implementation with HW specifics
> But agreed to look into such refactory later on.
ok
>
>>> +{
>>> + union acc100_dma_desc *desc = NULL;
>>> + int ret;
>>> + uint32_t in_offset, out_offset, out_length, mbuf_total_left,
>>> + seg_total_left;
>>> + struct rte_mbuf *input, *output_head, *output;
>>> +
>>> + uint16_t desc_idx = ((q->sw_ring_head + total_enqueued_cbs)
>>> + & q->sw_ring_wrap_mask);
>>> + desc = q->ring_addr + desc_idx;
>>> + acc100_fcw_le_fill(op, &desc->req.fcw_le, 1);
>>> +
>>> + input = op->ldpc_enc.input.data;
>>> + output_head = output = op->ldpc_enc.output.data;
>>> + in_offset = op->ldpc_enc.input.offset;
>>> + out_offset = op->ldpc_enc.output.offset;
>>> + out_length = 0;
>>> + mbuf_total_left = op->ldpc_enc.input.length;
>>> + seg_total_left = rte_pktmbuf_data_len(op->ldpc_enc.input.data)
>>> + - in_offset;
>>> +
>>> + ret = acc100_dma_desc_le_fill(op, &desc->req, &input, output,
>>> + &in_offset, &out_offset, &out_length,
>> &mbuf_total_left,
>>> + &seg_total_left);
>>> +
>>> + if (unlikely(ret < 0))
>>> + return ret;
>>> +
>>> + mbuf_append(output_head, output, out_length);
>>> +
>>> +#ifdef RTE_LIBRTE_BBDEV_DEBUG
>>> + rte_memdump(stderr, "FCW", &desc->req.fcw_le,
>>> + sizeof(desc->req.fcw_le) - 8);
>>> + rte_memdump(stderr, "Req Desc.", desc, sizeof(*desc));
>>> +
>>> + /* Check if any data left after processing one CB */
>>> + if (mbuf_total_left != 0) {
>>> + rte_bbdev_log(ERR,
>>> + "Some date still left after processing one CB:
>> mbuf_total_left = %u",
>>> + mbuf_total_left);
>>> + return -EINVAL;
>>> + }
>>> +#endif
>>> + /* One CB (one op) was successfully prepared to enqueue */
>>> + return 1;
>> Another case where caller only check for < 0
>>
>> Consider changes all similar to return 0 on success.
> same comment as above, would keep as is.
>
>>> +}
>>> +
>>> +/** Enqueue one decode operations for ACC100 device in CB mode */
>>> +static inline int
>>> +enqueue_ldpc_dec_one_op_cb(struct acc100_queue *q, struct
>> rte_bbdev_dec_op *op,
>>> + uint16_t total_enqueued_cbs, bool same_op)
>>> +{
>>> + int ret;
>>> +
>>> + union acc100_dma_desc *desc;
>>> + uint16_t desc_idx = ((q->sw_ring_head + total_enqueued_cbs)
>>> + & q->sw_ring_wrap_mask);
>>> + desc = q->ring_addr + desc_idx;
>>> + struct rte_mbuf *input, *h_output_head, *h_output;
>>> + uint32_t in_offset, h_out_offset, mbuf_total_left, h_out_length = 0;
>>> + input = op->ldpc_dec.input.data;
>>> + h_output_head = h_output = op->ldpc_dec.hard_output.data;
>>> + in_offset = op->ldpc_dec.input.offset;
>>> + h_out_offset = op->ldpc_dec.hard_output.offset;
>>> + mbuf_total_left = op->ldpc_dec.input.length;
>>> +#ifdef RTE_LIBRTE_BBDEV_DEBUG
>>> + if (unlikely(input == NULL)) {
>>> + rte_bbdev_log(ERR, "Invalid mbuf pointer");
>>> + return -EFAULT;
>>> + }
>>> +#endif
>>> + union acc100_harq_layout_data *harq_layout = q->d->harq_layout;
>>> +
>>> + if (same_op) {
>>> + union acc100_dma_desc *prev_desc;
>>> + desc_idx = ((q->sw_ring_head + total_enqueued_cbs - 1)
>>> + & q->sw_ring_wrap_mask);
>>> + prev_desc = q->ring_addr + desc_idx;
>>> + uint8_t *prev_ptr = (uint8_t *) prev_desc;
>>> + uint8_t *new_ptr = (uint8_t *) desc;
>>> + /* Copy first 4 words and BDESCs */
>>> + rte_memcpy(new_ptr, prev_ptr, 16);
>>> + rte_memcpy(new_ptr + 36, prev_ptr + 36, 40);
>> These magic numbers should be #defines
> yes
>
>>> + desc->req.op_addr = prev_desc->req.op_addr;
>>> + /* Copy FCW */
>>> + rte_memcpy(new_ptr + ACC100_DESC_FCW_OFFSET,
>>> + prev_ptr + ACC100_DESC_FCW_OFFSET,
>>> + ACC100_FCW_LD_BLEN);
>>> + acc100_dma_desc_ld_update(op, &desc->req, input,
>> h_output,
>>> + &in_offset, &h_out_offset,
>>> + &h_out_length, harq_layout);
>>> + } else {
>>> + struct acc100_fcw_ld *fcw;
>>> + uint32_t seg_total_left;
>>> + fcw = &desc->req.fcw_ld;
>>> + acc100_fcw_ld_fill(op, fcw, harq_layout);
>>> +
>>> + /* Special handling when overusing mbuf */
>>> + if (fcw->rm_e < MAX_E_MBUF)
>>> + seg_total_left = rte_pktmbuf_data_len(input)
>>> + - in_offset;
>>> + else
>>> + seg_total_left = fcw->rm_e;
>>> +
>>> + ret = acc100_dma_desc_ld_fill(op, &desc->req, &input,
>> h_output,
>>> + &in_offset, &h_out_offset,
>>> + &h_out_length, &mbuf_total_left,
>>> + &seg_total_left, fcw);
>>> + if (unlikely(ret < 0))
>>> + return ret;
>>> + }
>>> +
>>> + /* Hard output */
>>> + mbuf_append(h_output_head, h_output, h_out_length);
>>> +#ifndef ACC100_EXT_MEM
>>> + if (op->ldpc_dec.harq_combined_output.length > 0) {
>>> + /* Push the HARQ output into host memory */
>>> + struct rte_mbuf *hq_output_head, *hq_output;
>>> + hq_output_head = op-
>>> ldpc_dec.harq_combined_output.data;
>>> + hq_output = op->ldpc_dec.harq_combined_output.data;
>>> + mbuf_append(hq_output_head, hq_output,
>>> + op-
>>> ldpc_dec.harq_combined_output.length);
>>> + }
>>> +#endif
>>> +
>>> +#ifdef RTE_LIBRTE_BBDEV_DEBUG
>>> + rte_memdump(stderr, "FCW", &desc->req.fcw_ld,
>>> + sizeof(desc->req.fcw_ld) - 8);
>>> + rte_memdump(stderr, "Req Desc.", desc, sizeof(*desc));
>>> +#endif
>>> +
>>> + /* One CB (one op) was successfully prepared to enqueue */
>>> + return 1;
>>> +}
>>> +
>>> +
>>> +/* Enqueue one decode operations for ACC100 device in TB mode */
>>> +static inline int
>>> +enqueue_ldpc_dec_one_op_tb(struct acc100_queue *q, struct
>> rte_bbdev_dec_op *op,
>>> + uint16_t total_enqueued_cbs, uint8_t cbs_in_tb)
>>> +{
>>> + union acc100_dma_desc *desc = NULL;
>>> + int ret;
>>> + uint8_t r, c;
>>> + uint32_t in_offset, h_out_offset,
>>> + h_out_length, mbuf_total_left, seg_total_left;
>>> + struct rte_mbuf *input, *h_output_head, *h_output;
>>> + uint16_t current_enqueued_cbs = 0;
>>> +
>>> + uint16_t desc_idx = ((q->sw_ring_head + total_enqueued_cbs)
>>> + & q->sw_ring_wrap_mask);
>>> + desc = q->ring_addr + desc_idx;
>>> + uint64_t fcw_offset = (desc_idx << 8) + ACC100_DESC_FCW_OFFSET;
>>> + union acc100_harq_layout_data *harq_layout = q->d->harq_layout;
>>> + acc100_fcw_ld_fill(op, &desc->req.fcw_ld, harq_layout);
>>> +
>>> + input = op->ldpc_dec.input.data;
>>> + h_output_head = h_output = op->ldpc_dec.hard_output.data;
>>> + in_offset = op->ldpc_dec.input.offset;
>>> + h_out_offset = op->ldpc_dec.hard_output.offset;
>>> + h_out_length = 0;
>>> + mbuf_total_left = op->ldpc_dec.input.length;
>>> + c = op->ldpc_dec.tb_params.c;
>>> + r = op->ldpc_dec.tb_params.r;
>>> +
>>> + while (mbuf_total_left > 0 && r < c) {
>>> +
>>> + seg_total_left = rte_pktmbuf_data_len(input) - in_offset;
>>> +
>>> + /* Set up DMA descriptor */
>>> + desc = q->ring_addr + ((q->sw_ring_head +
>> total_enqueued_cbs)
>>> + & q->sw_ring_wrap_mask);
>>> + desc->req.data_ptrs[0].address = q->ring_addr_phys +
>> fcw_offset;
>>> + desc->req.data_ptrs[0].blen = ACC100_FCW_LD_BLEN;
>>> + ret = acc100_dma_desc_ld_fill(op, &desc->req, &input,
>>> + h_output, &in_offset, &h_out_offset,
>>> + &h_out_length,
>>> + &mbuf_total_left, &seg_total_left,
>>> + &desc->req.fcw_ld);
>>> +
>>> + if (unlikely(ret < 0))
>>> + return ret;
>>> +
>>> + /* Hard output */
>>> + mbuf_append(h_output_head, h_output, h_out_length);
>>> +
>>> + /* Set total number of CBs in TB */
>>> + desc->req.cbs_in_tb = cbs_in_tb;
>>> +#ifdef RTE_LIBRTE_BBDEV_DEBUG
>>> + rte_memdump(stderr, "FCW", &desc->req.fcw_td,
>>> + sizeof(desc->req.fcw_td) - 8);
>>> + rte_memdump(stderr, "Req Desc.", desc, sizeof(*desc));
>>> +#endif
>>> +
>>> + if (seg_total_left == 0) {
>>> + /* Go to the next mbuf */
>>> + input = input->next;
>>> + in_offset = 0;
>>> + h_output = h_output->next;
>>> + h_out_offset = 0;
>>> + }
>>> + total_enqueued_cbs++;
>>> + current_enqueued_cbs++;
>>> + r++;
>>> + }
>>> +
>>> + if (unlikely(desc == NULL))
>> How is this possible ? desc has be dereferenced already.
> related to static code analysis, arguably a false alarm
>
>>> + return current_enqueued_cbs;
>>> +
>>> +#ifdef RTE_LIBRTE_BBDEV_DEBUG
>>> + /* Check if any CBs left for processing */
>>> + if (mbuf_total_left != 0) {
>>> + rte_bbdev_log(ERR,
>>> + "Some date still left for processing:
>> mbuf_total_left = %u",
>>> + mbuf_total_left);
>>> + return -EINVAL;
>>> + }
>>> +#endif
>>> + /* Set SDone on last CB descriptor for TB mode */
>>> + desc->req.sdone_enable = 1;
>>> + desc->req.irq_enable = q->irq_enable;
>>> +
>>> + return current_enqueued_cbs;
>>> +}
>>> +
>>> +
>>> +/* Calculates number of CBs in processed encoder TB based on 'r' and
>> input
>>> + * length.
>>> + */
>>> +static inline uint8_t
>>> +get_num_cbs_in_tb_enc(struct rte_bbdev_op_turbo_enc *turbo_enc)
>>> +{
>>> + uint8_t c, c_neg, r, crc24_bits = 0;
>>> + uint16_t k, k_neg, k_pos;
>>> + uint8_t cbs_in_tb = 0;
>>> + int32_t length;
>>> +
>>> + length = turbo_enc->input.length;
>>> + r = turbo_enc->tb_params.r;
>>> + c = turbo_enc->tb_params.c;
>>> + c_neg = turbo_enc->tb_params.c_neg;
>>> + k_neg = turbo_enc->tb_params.k_neg;
>>> + k_pos = turbo_enc->tb_params.k_pos;
>>> + crc24_bits = 0;
>>> + if (check_bit(turbo_enc->op_flags,
>> RTE_BBDEV_TURBO_CRC_24B_ATTACH))
>>> + crc24_bits = 24;
>>> + while (length > 0 && r < c) {
>>> + k = (r < c_neg) ? k_neg : k_pos;
>>> + length -= (k - crc24_bits) >> 3;
>>> + r++;
>>> + cbs_in_tb++;
>>> + }
>>> +
>>> + return cbs_in_tb;
>>> +}
>>> +
>>> +/* Calculates number of CBs in processed decoder TB based on 'r' and
>> input
>>> + * length.
>>> + */
>>> +static inline uint16_t
>>> +get_num_cbs_in_tb_dec(struct rte_bbdev_op_turbo_dec *turbo_dec)
>>> +{
>>> + uint8_t c, c_neg, r = 0;
>>> + uint16_t kw, k, k_neg, k_pos, cbs_in_tb = 0;
>>> + int32_t length;
>>> +
>>> + length = turbo_dec->input.length;
>>> + r = turbo_dec->tb_params.r;
>>> + c = turbo_dec->tb_params.c;
>>> + c_neg = turbo_dec->tb_params.c_neg;
>>> + k_neg = turbo_dec->tb_params.k_neg;
>>> + k_pos = turbo_dec->tb_params.k_pos;
>>> + while (length > 0 && r < c) {
>>> + k = (r < c_neg) ? k_neg : k_pos;
>>> + kw = RTE_ALIGN_CEIL(k + 4, 32) * 3;
>>> + length -= kw;
>>> + r++;
>>> + cbs_in_tb++;
>>> + }
>>> +
>>> + return cbs_in_tb;
>>> +}
>>> +
>>> +/* Calculates number of CBs in processed decoder TB based on 'r' and
>> input
>>> + * length.
>>> + */
>>> +static inline uint16_t
>>> +get_num_cbs_in_tb_ldpc_dec(struct rte_bbdev_op_ldpc_dec *ldpc_dec)
>>> +{
>>> + uint16_t r, cbs_in_tb = 0;
>>> + int32_t length = ldpc_dec->input.length;
>>> + r = ldpc_dec->tb_params.r;
>>> + while (length > 0 && r < ldpc_dec->tb_params.c) {
>>> + length -= (r < ldpc_dec->tb_params.cab) ?
>>> + ldpc_dec->tb_params.ea :
>>> + ldpc_dec->tb_params.eb;
>>> + r++;
>>> + cbs_in_tb++;
>>> + }
>>> + return cbs_in_tb;
>>> +}
>>> +
>>> +/* Check we can mux encode operations with common FCW */
>>> +static inline bool
>>> +check_mux(struct rte_bbdev_enc_op **ops, uint16_t num) {
>>> + uint16_t i;
>>> + if (num == 1)
>>> + return false;
>> likely should strengthen check to num <= 1
> no impact, but doesnt hurt to change ok.
>
>>> + for (i = 1; i < num; ++i) {
>>> + /* Only mux compatible code blocks */
>>> + if (memcmp((uint8_t *)(&ops[i]->ldpc_enc) + ENC_OFFSET,
>>> + (uint8_t *)(&ops[0]->ldpc_enc) +
>> ENC_OFFSET,
>> ops[0]->ldpc_enc should be hoisted out of loop as it is invariant.
> compiler takes care of this I believe
hopefully, yes.
>
>>> + CMP_ENC_SIZE) != 0)
>>> + return false;
>>> + }
>>> + return true;
>>> +}
>>> +
>>> +/** Enqueue encode operations for ACC100 device in CB mode. */
>>> +static inline uint16_t
>>> +acc100_enqueue_ldpc_enc_cb(struct rte_bbdev_queue_data *q_data,
>>> + struct rte_bbdev_enc_op **ops, uint16_t num)
>>> +{
>>> + struct acc100_queue *q = q_data->queue_private;
>>> + int32_t avail = q->sw_ring_depth + q->sw_ring_tail - q-
>>> sw_ring_head;
>>> + uint16_t i = 0;
>>> + union acc100_dma_desc *desc;
>>> + int ret, desc_idx = 0;
>>> + int16_t enq, left = num;
>>> +
>>> + while (left > 0) {
>>> + if (unlikely(avail - 1 < 0))
>>> + break;
>>> + avail--;
>>> + enq = RTE_MIN(left, MUX_5GDL_DESC);
>>> + if (check_mux(&ops[i], enq)) {
>>> + ret = enqueue_ldpc_enc_n_op_cb(q, &ops[i],
>>> + desc_idx, enq);
>>> + if (ret < 0)
>>> + break;
>>> + i += enq;
>>> + } else {
>>> + ret = enqueue_ldpc_enc_one_op_cb(q, ops[i],
>> desc_idx);
>>> + if (ret < 0)
>>> + break;
>> failure is not handled well, what happens if this is one of serveral
> the aim is to flag the error and move on
>
>
>>> + i++;
>>> + }
>>> + desc_idx++;
>>> + left = num - i;
>>> + }
>>> +
>>> + if (unlikely(i == 0))
>>> + return 0; /* Nothing to enqueue */
>> this does not look correct for all cases
> I miss your point
I was thinking this was an error handler and needed beefing up.
>>> +
>>> + /* Set SDone in last CB in enqueued ops for CB mode*/
>>> + desc = q->ring_addr + ((q->sw_ring_head + desc_idx - 1)
>>> + & q->sw_ring_wrap_mask);
>>> + desc->req.sdone_enable = 1;
>>> + desc->req.irq_enable = q->irq_enable;
>>> +
>>> + acc100_dma_enqueue(q, desc_idx, &q_data->queue_stats);
>>> +
>>> + /* Update stats */
>>> + q_data->queue_stats.enqueued_count += i;
>>> + q_data->queue_stats.enqueue_err_count += num - i;
>>> +
>>> + return i;
>>> +}
>>> +
>>> +/* Enqueue encode operations for ACC100 device. */
>>> +static uint16_t
>>> +acc100_enqueue_ldpc_enc(struct rte_bbdev_queue_data *q_data,
>>> + struct rte_bbdev_enc_op **ops, uint16_t num)
>>> +{
>>> + if (unlikely(num == 0))
>>> + return 0;
>> Handling num == 0 should be in acc100_enqueue_ldpc_enc_cb
> Why would this be better not to catch early from user api call?
ok, because it was 'static' i was unsure
>
>>> + return acc100_enqueue_ldpc_enc_cb(q_data, ops, num);
>>> +}
>>> +
>>> +/* Check we can mux encode operations with common FCW */
>>> +static inline bool
>>> +cmp_ldpc_dec_op(struct rte_bbdev_dec_op **ops) {
>>> + /* Only mux compatible code blocks */
>>> + if (memcmp((uint8_t *)(&ops[0]->ldpc_dec) + DEC_OFFSET,
>>> + (uint8_t *)(&ops[1]->ldpc_dec) +
>>> + DEC_OFFSET, CMP_DEC_SIZE) != 0) {
>>> + return false;
>>> + } else
>> do not need the else, there are no other statements.
> debatable. Not considering change except if that becomes a DPDK
> coding guideline.
fine.
>>> + return true;
>>> +}
>>> +
>>> +
>>> +/* Enqueue decode operations for ACC100 device in TB mode */
>>> +static uint16_t
>>> +acc100_enqueue_ldpc_dec_tb(struct rte_bbdev_queue_data *q_data,
>>> + struct rte_bbdev_dec_op **ops, uint16_t num)
>>> +{
>>> + struct acc100_queue *q = q_data->queue_private;
>>> + int32_t avail = q->sw_ring_depth + q->sw_ring_tail - q-
>>> sw_ring_head;
>>> + uint16_t i, enqueued_cbs = 0;
>>> + uint8_t cbs_in_tb;
>>> + int ret;
>>> +
>>> + for (i = 0; i < num; ++i) {
>>> + cbs_in_tb = get_num_cbs_in_tb_ldpc_dec(&ops[i]-
>>> ldpc_dec);
>>> + /* Check if there are available space for further processing */
>>> + if (unlikely(avail - cbs_in_tb < 0))
>>> + break;
>>> + avail -= cbs_in_tb;
>>> +
>>> + ret = enqueue_ldpc_dec_one_op_tb(q, ops[i],
>>> + enqueued_cbs, cbs_in_tb);
>>> + if (ret < 0)
>>> + break;
>>> + enqueued_cbs += ret;
>>> + }
>>> +
>>> + acc100_dma_enqueue(q, enqueued_cbs, &q_data->queue_stats);
>>> +
>>> + /* Update stats */
>>> + q_data->queue_stats.enqueued_count += i;
>>> + q_data->queue_stats.enqueue_err_count += num - i;
>>> + return i;
>>> +}
>>> +
>>> +/* Enqueue decode operations for ACC100 device in CB mode */
>>> +static uint16_t
>>> +acc100_enqueue_ldpc_dec_cb(struct rte_bbdev_queue_data *q_data,
>>> + struct rte_bbdev_dec_op **ops, uint16_t num)
>>> +{
>>> + struct acc100_queue *q = q_data->queue_private;
>>> + int32_t avail = q->sw_ring_depth + q->sw_ring_tail - q-
>>> sw_ring_head;
>>> + uint16_t i;
>>> + union acc100_dma_desc *desc;
>>> + int ret;
>>> + bool same_op = false;
>>> + for (i = 0; i < num; ++i) {
>>> + /* Check if there are available space for further processing */
>>> + if (unlikely(avail - 1 < 0))
>> change to (avail < 1)
>>
>> Generally.
> ok
>
>>> + break;
>>> + avail -= 1;
>>> +
>>> + if (i > 0)
>>> + same_op = cmp_ldpc_dec_op(&ops[i-1]);
>>> + rte_bbdev_log(INFO, "Op %d %d %d %d %d %d %d %d %d %d
>> %d %d\n",
>>> + i, ops[i]->ldpc_dec.op_flags, ops[i]-
>>> ldpc_dec.rv_index,
>>> + ops[i]->ldpc_dec.iter_max, ops[i]-
>>> ldpc_dec.iter_count,
>>> + ops[i]->ldpc_dec.basegraph, ops[i]->ldpc_dec.z_c,
>>> + ops[i]->ldpc_dec.n_cb, ops[i]->ldpc_dec.q_m,
>>> + ops[i]->ldpc_dec.n_filler, ops[i]-
>>> ldpc_dec.cb_params.e,
>>> + same_op);
>>> + ret = enqueue_ldpc_dec_one_op_cb(q, ops[i], i, same_op);
>>> + if (ret < 0)
>>> + break;
>>> + }
>>> +
>>> + if (unlikely(i == 0))
>>> + return 0; /* Nothing to enqueue */
>>> +
>>> + /* Set SDone in last CB in enqueued ops for CB mode*/
>>> + desc = q->ring_addr + ((q->sw_ring_head + i - 1)
>>> + & q->sw_ring_wrap_mask);
>>> +
>>> + desc->req.sdone_enable = 1;
>>> + desc->req.irq_enable = q->irq_enable;
>>> +
>>> + acc100_dma_enqueue(q, i, &q_data->queue_stats);
>>> +
>>> + /* Update stats */
>>> + q_data->queue_stats.enqueued_count += i;
>>> + q_data->queue_stats.enqueue_err_count += num - i;
>>> + return i;
>>> +}
>>> +
>>> +/* Enqueue decode operations for ACC100 device. */
>>> +static uint16_t
>>> +acc100_enqueue_ldpc_dec(struct rte_bbdev_queue_data *q_data,
>>> + struct rte_bbdev_dec_op **ops, uint16_t num)
>>> +{
>>> + struct acc100_queue *q = q_data->queue_private;
>>> + int32_t aq_avail = q->aq_depth +
>>> + (q->aq_dequeued - q->aq_enqueued) / 128;
>>> +
>>> + if (unlikely((aq_avail == 0) || (num == 0)))
>>> + return 0;
>>> +
>>> + if (ops[0]->ldpc_dec.code_block_mode == 0)
>>> + return acc100_enqueue_ldpc_dec_tb(q_data, ops, num);
>>> + else
>>> + return acc100_enqueue_ldpc_dec_cb(q_data, ops, num);
>>> +}
>>> +
>>> +
>>> +/* Dequeue one encode operations from ACC100 device in CB mode */
>>> +static inline int
>>> +dequeue_enc_one_op_cb(struct acc100_queue *q, struct
>> rte_bbdev_enc_op **ref_op,
>>> + uint16_t total_dequeued_cbs, uint32_t *aq_dequeued)
>>> +{
>>> + union acc100_dma_desc *desc, atom_desc;
>>> + union acc100_dma_rsp_desc rsp;
>>> + struct rte_bbdev_enc_op *op;
>>> + int i;
>>> +
>>> + desc = q->ring_addr + ((q->sw_ring_tail + total_dequeued_cbs)
>>> + & q->sw_ring_wrap_mask);
>>> + atom_desc.atom_hdr = __atomic_load_n((uint64_t *)desc,
>>> + __ATOMIC_RELAXED);
>>> +
>>> + /* Check fdone bit */
>>> + if (!(atom_desc.rsp.val & ACC100_FDONE))
>>> + return -1;
>>> +
>>> + rsp.val = atom_desc.rsp.val;
>>> + rte_bbdev_log_debug("Resp. desc %p: %x", desc, rsp.val);
>>> +
>>> + /* Dequeue */
>>> + op = desc->req.op_addr;
>>> +
>>> + /* Clearing status, it will be set based on response */
>>> + op->status = 0;
>>> +
>>> + op->status |= ((rsp.input_err)
>>> + ? (1 << RTE_BBDEV_DATA_ERROR) : 0);
>> can remove the = 0, if |= is changed to =
> yes in principle, but easy to break by mistake, so would keep.
ok
>>> + op->status |= ((rsp.dma_err) ? (1 << RTE_BBDEV_DRV_ERROR) : 0);
>>> + op->status |= ((rsp.fcw_err) ? (1 << RTE_BBDEV_DRV_ERROR) : 0);
>>> +
>>> + if (desc->req.last_desc_in_batch) {
>>> + (*aq_dequeued)++;
>>> + desc->req.last_desc_in_batch = 0;
>>> + }
>>> + desc->rsp.val = ACC100_DMA_DESC_TYPE;
>>> + desc->rsp.add_info_0 = 0; /*Reserved bits */
>>> + desc->rsp.add_info_1 = 0; /*Reserved bits */
>>> +
>>> + /* Flag that the muxing cause loss of opaque data */
>>> + op->opaque_data = (void *)-1;
>> as a ptr, shouldn't opaque_data be poisoned with '0' ?
> more obvious this way I think.
the idiom (ptr == NULL) would need to be changed.
as a non standard poison, it is likely that someone will trip over this.
>>> + for (i = 0 ; i < desc->req.numCBs; i++)
>>> + ref_op[i] = op;
>>> +
>>> + /* One CB (op) was successfully dequeued */
>>> + return desc->req.numCBs;
>>> +}
>>> +
>>> +/* Dequeue one encode operations from ACC100 device in TB mode */
>>> +static inline int
>>> +dequeue_enc_one_op_tb(struct acc100_queue *q, struct
>> rte_bbdev_enc_op **ref_op,
>>> + uint16_t total_dequeued_cbs, uint32_t *aq_dequeued)
>>> +{
>>> + union acc100_dma_desc *desc, *last_desc, atom_desc;
>>> + union acc100_dma_rsp_desc rsp;
>>> + struct rte_bbdev_enc_op *op;
>>> + uint8_t i = 0;
>>> + uint16_t current_dequeued_cbs = 0, cbs_in_tb;
>>> +
>>> + desc = q->ring_addr + ((q->sw_ring_tail + total_dequeued_cbs)
>>> + & q->sw_ring_wrap_mask);
>>> + atom_desc.atom_hdr = __atomic_load_n((uint64_t *)desc,
>>> + __ATOMIC_RELAXED);
>>> +
>>> + /* Check fdone bit */
>>> + if (!(atom_desc.rsp.val & ACC100_FDONE))
>>> + return -1;
>>> +
>>> + /* Get number of CBs in dequeued TB */
>>> + cbs_in_tb = desc->req.cbs_in_tb;
>>> + /* Get last CB */
>>> + last_desc = q->ring_addr + ((q->sw_ring_tail
>>> + + total_dequeued_cbs + cbs_in_tb - 1)
>>> + & q->sw_ring_wrap_mask);
>>> + /* Check if last CB in TB is ready to dequeue (and thus
>>> + * the whole TB) - checking sdone bit. If not return.
>>> + */
>>> + atom_desc.atom_hdr = __atomic_load_n((uint64_t *)last_desc,
>>> + __ATOMIC_RELAXED);
>>> + if (!(atom_desc.rsp.val & ACC100_SDONE))
>>> + return -1;
>>> +
>>> + /* Dequeue */
>>> + op = desc->req.op_addr;
>>> +
>>> + /* Clearing status, it will be set based on response */
>>> + op->status = 0;
>>> +
>>> + while (i < cbs_in_tb) {
>>> + desc = q->ring_addr + ((q->sw_ring_tail
>>> + + total_dequeued_cbs)
>>> + & q->sw_ring_wrap_mask);
>>> + atom_desc.atom_hdr = __atomic_load_n((uint64_t *)desc,
>>> + __ATOMIC_RELAXED);
>>> + rsp.val = atom_desc.rsp.val;
>>> + rte_bbdev_log_debug("Resp. desc %p: %x", desc,
>>> + rsp.val);
>>> +
>>> + op->status |= ((rsp.input_err)
>>> + ? (1 << RTE_BBDEV_DATA_ERROR) : 0);
>>> + op->status |= ((rsp.dma_err) ? (1 <<
>> RTE_BBDEV_DRV_ERROR) : 0);
>>> + op->status |= ((rsp.fcw_err) ? (1 << RTE_BBDEV_DRV_ERROR)
>> : 0);
>>> +
>>> + if (desc->req.last_desc_in_batch) {
>>> + (*aq_dequeued)++;
>>> + desc->req.last_desc_in_batch = 0;
>>> + }
>>> + desc->rsp.val = ACC100_DMA_DESC_TYPE;
>>> + desc->rsp.add_info_0 = 0;
>>> + desc->rsp.add_info_1 = 0;
>>> + total_dequeued_cbs++;
>>> + current_dequeued_cbs++;
>>> + i++;
>>> + }
>>> +
>>> + *ref_op = op;
>>> +
>>> + return current_dequeued_cbs;
>>> +}
>>> +
>>> +/* Dequeue one decode operation from ACC100 device in CB mode */
>>> +static inline int
>>> +dequeue_dec_one_op_cb(struct rte_bbdev_queue_data *q_data,
>>> + struct acc100_queue *q, struct rte_bbdev_dec_op **ref_op,
>>> + uint16_t dequeued_cbs, uint32_t *aq_dequeued)
>>> +{
>>> + union acc100_dma_desc *desc, atom_desc;
>>> + union acc100_dma_rsp_desc rsp;
>>> + struct rte_bbdev_dec_op *op;
>>> +
>>> + desc = q->ring_addr + ((q->sw_ring_tail + dequeued_cbs)
>>> + & q->sw_ring_wrap_mask);
>>> + atom_desc.atom_hdr = __atomic_load_n((uint64_t *)desc,
>>> + __ATOMIC_RELAXED);
>>> +
>>> + /* Check fdone bit */
>>> + if (!(atom_desc.rsp.val & ACC100_FDONE))
>>> + return -1;
>>> +
>>> + rsp.val = atom_desc.rsp.val;
>>> + rte_bbdev_log_debug("Resp. desc %p: %x", desc, rsp.val);
>>> +
>>> + /* Dequeue */
>>> + op = desc->req.op_addr;
>>> +
>>> + /* Clearing status, it will be set based on response */
>>> + op->status = 0;
>>> + op->status |= ((rsp.input_err)
>>> + ? (1 << RTE_BBDEV_DATA_ERROR) : 0);
>> similar to above, can remove the = 0
>>
>> This is a general issue.
> same comment above
>
>>> + op->status |= ((rsp.dma_err) ? (1 << RTE_BBDEV_DRV_ERROR) : 0);
>>> + op->status |= ((rsp.fcw_err) ? (1 << RTE_BBDEV_DRV_ERROR) : 0);
>>> + if (op->status != 0)
>>> + q_data->queue_stats.dequeue_err_count++;
>>> +
>>> + /* CRC invalid if error exists */
>>> + if (!op->status)
>>> + op->status |= rsp.crc_status << RTE_BBDEV_CRC_ERROR;
>>> + op->turbo_dec.iter_count = (uint8_t) rsp.iter_cnt / 2;
>>> + /* Check if this is the last desc in batch (Atomic Queue) */
>>> + if (desc->req.last_desc_in_batch) {
>>> + (*aq_dequeued)++;
>>> + desc->req.last_desc_in_batch = 0;
>>> + }
>>> + desc->rsp.val = ACC100_DMA_DESC_TYPE;
>>> + desc->rsp.add_info_0 = 0;
>>> + desc->rsp.add_info_1 = 0;
>>> + *ref_op = op;
>>> +
>>> + /* One CB (op) was successfully dequeued */
>>> + return 1;
>>> +}
>>> +
>>> +/* Dequeue one decode operations from ACC100 device in CB mode */
>>> +static inline int
>>> +dequeue_ldpc_dec_one_op_cb(struct rte_bbdev_queue_data *q_data,
>>> + struct acc100_queue *q, struct rte_bbdev_dec_op **ref_op,
>>> + uint16_t dequeued_cbs, uint32_t *aq_dequeued)
>>> +{
>>> + union acc100_dma_desc *desc, atom_desc;
>>> + union acc100_dma_rsp_desc rsp;
>>> + struct rte_bbdev_dec_op *op;
>>> +
>>> + desc = q->ring_addr + ((q->sw_ring_tail + dequeued_cbs)
>>> + & q->sw_ring_wrap_mask);
>>> + atom_desc.atom_hdr = __atomic_load_n((uint64_t *)desc,
>>> + __ATOMIC_RELAXED);
>>> +
>>> + /* Check fdone bit */
>>> + if (!(atom_desc.rsp.val & ACC100_FDONE))
>>> + return -1;
>>> +
>>> + rsp.val = atom_desc.rsp.val;
>>> +
>>> + /* Dequeue */
>>> + op = desc->req.op_addr;
>>> +
>>> + /* Clearing status, it will be set based on response */
>>> + op->status = 0;
>>> + op->status |= rsp.input_err << RTE_BBDEV_DATA_ERROR;
>>> + op->status |= rsp.dma_err << RTE_BBDEV_DRV_ERROR;
>>> + op->status |= rsp.fcw_err << RTE_BBDEV_DRV_ERROR;
>>> + if (op->status != 0)
>>> + q_data->queue_stats.dequeue_err_count++;
>>> +
>>> + op->status |= rsp.crc_status << RTE_BBDEV_CRC_ERROR;
>>> + if (op->ldpc_dec.hard_output.length > 0 && !rsp.synd_ok)
>>> + op->status |= 1 << RTE_BBDEV_SYNDROME_ERROR;
>>> + op->ldpc_dec.iter_count = (uint8_t) rsp.iter_cnt;
>>> +
>>> + /* Check if this is the last desc in batch (Atomic Queue) */
>>> + if (desc->req.last_desc_in_batch) {
>>> + (*aq_dequeued)++;
>>> + desc->req.last_desc_in_batch = 0;
>>> + }
>>> +
>>> + desc->rsp.val = ACC100_DMA_DESC_TYPE;
>>> + desc->rsp.add_info_0 = 0;
>>> + desc->rsp.add_info_1 = 0;
>>> +
>>> + *ref_op = op;
>>> +
>>> + /* One CB (op) was successfully dequeued */
>>> + return 1;
>>> +}
>>> +
>>> +/* Dequeue one decode operations from ACC100 device in TB mode. */
>>> +static inline int
>>> +dequeue_dec_one_op_tb(struct acc100_queue *q, struct
>> rte_bbdev_dec_op **ref_op,
>>> + uint16_t dequeued_cbs, uint32_t *aq_dequeued)
>>> +{
>> similar call as fpga_lte_fec
> distinct though as HW specific
>
>>> + union acc100_dma_desc *desc, *last_desc, atom_desc;
>>> + union acc100_dma_rsp_desc rsp;
>>> + struct rte_bbdev_dec_op *op;
>>> + uint8_t cbs_in_tb = 1, cb_idx = 0;
>>> +
>>> + desc = q->ring_addr + ((q->sw_ring_tail + dequeued_cbs)
>>> + & q->sw_ring_wrap_mask);
>>> + atom_desc.atom_hdr = __atomic_load_n((uint64_t *)desc,
>>> + __ATOMIC_RELAXED);
>>> +
>>> + /* Check fdone bit */
>>> + if (!(atom_desc.rsp.val & ACC100_FDONE))
>>> + return -1;
>>> +
>>> + /* Dequeue */
>>> + op = desc->req.op_addr;
>>> +
>>> + /* Get number of CBs in dequeued TB */
>>> + cbs_in_tb = desc->req.cbs_in_tb;
>>> + /* Get last CB */
>>> + last_desc = q->ring_addr + ((q->sw_ring_tail
>>> + + dequeued_cbs + cbs_in_tb - 1)
>>> + & q->sw_ring_wrap_mask);
>>> + /* Check if last CB in TB is ready to dequeue (and thus
>>> + * the whole TB) - checking sdone bit. If not return.
>>> + */
>>> + atom_desc.atom_hdr = __atomic_load_n((uint64_t *)last_desc,
>>> + __ATOMIC_RELAXED);
>>> + if (!(atom_desc.rsp.val & ACC100_SDONE))
>>> + return -1;
>>> +
>>> + /* Clearing status, it will be set based on response */
>>> + op->status = 0;
>>> +
>>> + /* Read remaining CBs if exists */
>>> + while (cb_idx < cbs_in_tb) {
>> Other similar calls use 'i' , 'cb_idx' is more meaningful, consider changing
>> the
>> other loops.
> More relevant here due to split of TB into CBs.
ok
>>> + desc = q->ring_addr + ((q->sw_ring_tail + dequeued_cbs)
>>> + & q->sw_ring_wrap_mask);
>>> + atom_desc.atom_hdr = __atomic_load_n((uint64_t *)desc,
>>> + __ATOMIC_RELAXED);
>>> + rsp.val = atom_desc.rsp.val;
>>> + rte_bbdev_log_debug("Resp. desc %p: %x", desc,
>>> + rsp.val);
>>> +
>>> + op->status |= ((rsp.input_err)
>>> + ? (1 << RTE_BBDEV_DATA_ERROR) : 0);
>>> + op->status |= ((rsp.dma_err) ? (1 <<
>> RTE_BBDEV_DRV_ERROR) : 0);
>>> + op->status |= ((rsp.fcw_err) ? (1 << RTE_BBDEV_DRV_ERROR)
>> : 0);
>>> +
>>> + /* CRC invalid if error exists */
>>> + if (!op->status)
>>> + op->status |= rsp.crc_status <<
>> RTE_BBDEV_CRC_ERROR;
>>> + op->turbo_dec.iter_count = RTE_MAX((uint8_t) rsp.iter_cnt,
>>> + op->turbo_dec.iter_count);
>>> +
>>> + /* Check if this is the last desc in batch (Atomic Queue) */
>>> + if (desc->req.last_desc_in_batch) {
>>> + (*aq_dequeued)++;
>>> + desc->req.last_desc_in_batch = 0;
>>> + }
>>> + desc->rsp.val = ACC100_DMA_DESC_TYPE;
>>> + desc->rsp.add_info_0 = 0;
>>> + desc->rsp.add_info_1 = 0;
>>> + dequeued_cbs++;
>>> + cb_idx++;
>>> + }
>>> +
>>> + *ref_op = op;
>>> +
>>> + return cb_idx;
>>> +}
>>> +
>>> +/* Dequeue LDPC encode operations from ACC100 device. */
>>> +static uint16_t
>>> +acc100_dequeue_ldpc_enc(struct rte_bbdev_queue_data *q_data,
>>> + struct rte_bbdev_enc_op **ops, uint16_t num)
>>> +{
>>> + struct acc100_queue *q = q_data->queue_private;
>>> + uint32_t avail = q->sw_ring_head - q->sw_ring_tail;
>>> + uint32_t aq_dequeued = 0;
>>> + uint16_t dequeue_num, i, dequeued_cbs = 0, dequeued_descs = 0;
>>> + int ret;
>>> +
>>> +#ifdef RTE_LIBRTE_BBDEV_DEBUG
>>> + if (unlikely(ops == 0 && q == NULL))
>>> + return 0;
>>> +#endif
>>> +
>>> + dequeue_num = (avail < num) ? avail : num;
>> Similar to RTE_MIN
>>
>> general issue
> ok, will check
>
>>> +
>>> + for (i = 0; i < dequeue_num; i++) {
>>> + ret = dequeue_enc_one_op_cb(q, &ops[dequeued_cbs],
>>> + dequeued_descs, &aq_dequeued);
>>> + if (ret < 0)
>>> + break;
>>> + dequeued_cbs += ret;
>>> + dequeued_descs++;
>>> + if (dequeued_cbs >= num)
>>> + break;
>> condition should be added to the for-loop
> unsure this would helps readability personnaly
ok
Tom
>>> + }
>>> +
>>> + q->aq_dequeued += aq_dequeued;
>>> + q->sw_ring_tail += dequeued_descs;
>>> +
>>> + /* Update enqueue stats */
>>> + q_data->queue_stats.dequeued_count += dequeued_cbs;
>>> +
>>> + return dequeued_cbs;
>>> +}
>>> +
>>> +/* Dequeue decode operations from ACC100 device. */
>>> +static uint16_t
>>> +acc100_dequeue_ldpc_dec(struct rte_bbdev_queue_data *q_data,
>>> + struct rte_bbdev_dec_op **ops, uint16_t num)
>>> +{
>>> + struct acc100_queue *q = q_data->queue_private;
>>> + uint16_t dequeue_num;
>>> + uint32_t avail = q->sw_ring_head - q->sw_ring_tail;
>>> + uint32_t aq_dequeued = 0;
>>> + uint16_t i;
>>> + uint16_t dequeued_cbs = 0;
>>> + struct rte_bbdev_dec_op *op;
>>> + int ret;
>>> +
>>> +#ifdef RTE_LIBRTE_BBDEV_DEBUG
>>> + if (unlikely(ops == 0 && q == NULL))
>>> + return 0;
>>> +#endif
>>> +
>>> + dequeue_num = (avail < num) ? avail : num;
>>> +
>>> + for (i = 0; i < dequeue_num; ++i) {
>>> + op = (q->ring_addr + ((q->sw_ring_tail + dequeued_cbs)
>>> + & q->sw_ring_wrap_mask))->req.op_addr;
>>> + if (op->ldpc_dec.code_block_mode == 0)
>> 0 should be a #define
> mentioned in previous review.
>
> Thanks
>
>> Tom
>>
>>> + ret = dequeue_dec_one_op_tb(q, &ops[i],
>> dequeued_cbs,
>>> + &aq_dequeued);
>>> + else
>>> + ret = dequeue_ldpc_dec_one_op_cb(
>>> + q_data, q, &ops[i], dequeued_cbs,
>>> + &aq_dequeued);
>>> +
>>> + if (ret < 0)
>>> + break;
>>> + dequeued_cbs += ret;
>>> + }
>>> +
>>> + q->aq_dequeued += aq_dequeued;
>>> + q->sw_ring_tail += dequeued_cbs;
>>> +
>>> + /* Update enqueue stats */
>>> + q_data->queue_stats.dequeued_count += i;
>>> +
>>> + return i;
>>> +}
>>> +
>>> /* Initialization Function */
>>> static void
>>> acc100_bbdev_init(struct rte_bbdev *dev, struct rte_pci_driver *drv)
>>> @@ -703,6 +2321,10 @@
>>> struct rte_pci_device *pci_dev = RTE_DEV_TO_PCI(dev->device);
>>>
>>> dev->dev_ops = &acc100_bbdev_ops;
>>> + dev->enqueue_ldpc_enc_ops = acc100_enqueue_ldpc_enc;
>>> + dev->enqueue_ldpc_dec_ops = acc100_enqueue_ldpc_dec;
>>> + dev->dequeue_ldpc_enc_ops = acc100_dequeue_ldpc_enc;
>>> + dev->dequeue_ldpc_dec_ops = acc100_dequeue_ldpc_dec;
>>>
>>> ((struct acc100_device *) dev->data->dev_private)->pf_device =
>>> !strcmp(drv->driver.name,
>>> @@ -815,4 +2437,3 @@ static int acc100_pci_remove(struct
>> rte_pci_device *pci_dev)
>>> RTE_PMD_REGISTER_PCI_TABLE(ACC100PF_DRIVER_NAME,
>> pci_id_acc100_pf_map);
>>> RTE_PMD_REGISTER_PCI(ACC100VF_DRIVER_NAME,
>> acc100_pci_vf_driver);
>>> RTE_PMD_REGISTER_PCI_TABLE(ACC100VF_DRIVER_NAME,
>> pci_id_acc100_vf_map);
>>> -
>>> diff --git a/drivers/baseband/acc100/rte_acc100_pmd.h
>> b/drivers/baseband/acc100/rte_acc100_pmd.h
>>> index 0e2b79c..78686c1 100644
>>> --- a/drivers/baseband/acc100/rte_acc100_pmd.h
>>> +++ b/drivers/baseband/acc100/rte_acc100_pmd.h
>>> @@ -88,6 +88,8 @@
>>> #define TMPL_PRI_3 0x0f0e0d0c
>>> #define QUEUE_ENABLE 0x80000000 /* Bit to mark Queue as Enabled */
>>> #define WORDS_IN_ARAM_SIZE (128 * 1024 / 4)
>>> +#define ACC100_FDONE 0x80000000
>>> +#define ACC100_SDONE 0x40000000
>>>
>>> #define ACC100_NUM_TMPL 32
>>> #define VF_OFFSET_QOS 16 /* offset in Memory Space specific to QoS
>> Mon */
>>> @@ -398,6 +400,7 @@ struct __rte_packed acc100_dma_req_desc {
>>> union acc100_dma_desc {
>>> struct acc100_dma_req_desc req;
>>> union acc100_dma_rsp_desc rsp;
>>> + uint64_t atom_hdr;
>>> };
>>>
>>>
