blackbeard420/esp-idf
1
0
Fork 0
mirror of https://github.com/espressif/esp-idf synced 2025-03-09 09:09:10 -04:00
Code Issues Projects Releases Wiki Activity

fix(dma): also consider buffer alignment when calculating the DMA nodes

Browse Source 
Closes https://github.com/espressif/esp-idf/issues/15228
This commit is contained in:
morris 2025-01-24 18:02:57 +08:00
parent 5c0611cb3b
commit e121a77c92
8 changed files with 66 additions and 37 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

17
components/esp_driver_parlio/src/parlio_tx.c
View File

@ -37,6 +37,7 @@
#include "esp_private/esp_clk_tree_common.h"
#include "esp_private/gdma.h"
#include "esp_private/gdma_link.h"
#include "esp_private/esp_dma_utils.h"
static const char *TAG = "parlio-tx";
@ -53,7 +54,6 @@ typedef struct parlio_tx_unit_t {
esp_pm_lock_handle_t pm_lock; // power management lock
gdma_channel_handle_t dma_chan; // DMA channel
gdma_link_list_handle_t dma_link; // DMA link list handle
size_t dma_nodes_num; // number of DMA descriptor nodes
size_t int_mem_align; // Alignment for internal memory
size_t ext_mem_align; // Alignment for external memory
#if CONFIG_PM_ENABLE
@ -217,14 +217,12 @@ static esp_err_t parlio_tx_unit_init_dma(parlio_tx_unit_t *tx_unit, const parlio
gdma_get_alignment_constraints(tx_unit->dma_chan, &tx_unit->int_mem_align, &tx_unit->ext_mem_align);
// create DMA link list
size_t dma_nodes_num = tx_unit->dma_nodes_num;
size_t buffer_alignment = MAX(tx_unit->int_mem_align, tx_unit->ext_mem_align);
size_t num_dma_nodes = esp_dma_calculate_node_count(config->max_transfer_size, buffer_alignment, DMA_DESCRIPTOR_BUFFER_MAX_SIZE);
gdma_link_list_config_t dma_link_config = {
.buffer_alignment = 1,
.buffer_alignment = buffer_alignment,
.item_alignment = PARLIO_DMA_DESC_ALIGNMENT,
.num_items = dma_nodes_num,
.flags = {
.check_owner = true,
},
.num_items = num_dma_nodes,
};
// throw the error to the caller
@ -318,11 +316,6 @@ esp_err_t parlio_new_tx_unit(const parlio_tx_unit_config_t *config, parlio_tx_un
unit = heap_caps_calloc(1, sizeof(parlio_tx_unit_t) + sizeof(parlio_tx_trans_desc_t) * config->trans_queue_depth, mem_caps);
ESP_GOTO_ON_FALSE(unit, ESP_ERR_NO_MEM, err, TAG, "no memory for tx unit");
// create DMA descriptors
// DMA descriptors must be placed in internal SRAM
size_t dma_nodes_num = config->max_transfer_size / DMA_DESCRIPTOR_BUFFER_MAX_SIZE + 1;
unit->dma_nodes_num = dma_nodes_num;
unit->max_transfer_bits = config->max_transfer_size * 8;
unit->base.dir = PARLIO_DIR_TX;
unit->data_width = data_width;

7
components/esp_hw_support/dma/async_memcpy_cp_dma.c
View File

@ -20,6 +20,7 @@
#include "esp_async_memcpy.h"
#include "esp_async_memcpy_priv.h"
#include "esp_private/gdma_link.h"
#include "esp_private/esp_dma_utils.h"
#include "hal/cp_dma_hal.h"
#include "hal/cp_dma_ll.h"
@ -211,11 +212,13 @@ static esp_err_t mcp_cpdma_memcpy(async_memcpy_context_t *ctx, void *dst, void *
trans->rx_link_list = NULL;
}
size_t num_dma_nodes = esp_dma_calculate_node_count(n, 1, MCP_DMA_DESCRIPTOR_BUFFER_MAX_SIZE);
// allocate gdma TX link
gdma_link_list_config_t tx_link_cfg = {
.buffer_alignment = 1, // CP_DMA doesn't have alignment requirement for internal memory
.item_alignment = 4, // CP_DMA requires 4 bytes alignment for each descriptor
.num_items = n / MCP_DMA_DESCRIPTOR_BUFFER_MAX_SIZE + 1,
.num_items = num_dma_nodes,
.flags = {
.check_owner = true,
.items_in_ext_mem = false,
@ -239,7 +242,7 @@ static esp_err_t mcp_cpdma_memcpy(async_memcpy_context_t *ctx, void *dst, void *
gdma_link_list_config_t rx_link_cfg = {
.buffer_alignment = 1, // CP_DMA doesn't have alignment requirement for internal memory
.item_alignment = 4, // CP_DMA requires 4 bytes alignment for each descriptor
.num_items = n / MCP_DMA_DESCRIPTOR_BUFFER_MAX_SIZE + 1,
.num_items = num_dma_nodes,
.flags = {
.check_owner = true,
.items_in_ext_mem = false,

15
components/esp_hw_support/dma/async_memcpy_gdma.c
View File

@ -342,11 +342,16 @@ static esp_err_t mcp_gdma_memcpy(async_memcpy_context_t *ctx, void *dst, void *s
trans->stash_buffer = NULL;
}
size_t buffer_alignment = 0;
size_t num_dma_nodes = 0;
// allocate gdma TX link
buffer_alignment = esp_ptr_internal(src) ? mcp_gdma->tx_int_mem_alignment : mcp_gdma->tx_ext_mem_alignment;
num_dma_nodes = esp_dma_calculate_node_count(n, buffer_alignment, MCP_DMA_DESCRIPTOR_BUFFER_MAX_SIZE);
gdma_link_list_config_t tx_link_cfg = {
.buffer_alignment = esp_ptr_internal(src) ? mcp_gdma->tx_int_mem_alignment : mcp_gdma->tx_ext_mem_alignment,
.buffer_alignment = buffer_alignment,
.item_alignment = dma_link_item_alignment,
.num_items = n / MCP_DMA_DESCRIPTOR_BUFFER_MAX_SIZE + 1,
.num_items = num_dma_nodes,
.flags = {
.check_owner = true,
.items_in_ext_mem = false, // TODO: if the memcopy size is too large, we may need to allocate the link list items from external memory
@ -381,10 +386,12 @@ static esp_err_t mcp_gdma_memcpy(async_memcpy_context_t *ctx, void *dst, void *s
}
// allocate gdma RX link
buffer_alignment = esp_ptr_internal(dst) ? mcp_gdma->rx_int_mem_alignment : mcp_gdma->rx_ext_mem_alignment;
num_dma_nodes = esp_dma_calculate_node_count(n, buffer_alignment, MCP_DMA_DESCRIPTOR_BUFFER_MAX_SIZE);
gdma_link_list_config_t rx_link_cfg = {
.buffer_alignment = esp_ptr_internal(dst) ? mcp_gdma->rx_int_mem_alignment : mcp_gdma->rx_ext_mem_alignment,
.buffer_alignment = buffer_alignment,
.item_alignment = dma_link_item_alignment,
.num_items = n / MCP_DMA_DESCRIPTOR_BUFFER_MAX_SIZE + 3,
.num_items = num_dma_nodes + 3, // add 3 extra items for the cache aligned buffers
.flags = {
.check_owner = true,
.items_in_ext_mem = false, // TODO: if the memcopy size is too large, we may need to allocate the link list items from external memory

10
components/esp_hw_support/dma/esp_dma_utils.c
View File

@ -134,6 +134,16 @@ esp_err_t esp_dma_merge_aligned_rx_buffers(dma_buffer_split_array_t *align_array
return ESP_OK;
}
size_t esp_dma_calculate_node_count(size_t buffer_size, size_t buffer_alignment, size_t max_buffer_size_per_node)
{
// buffer_alignment should be power of 2
ESP_RETURN_ON_FALSE(buffer_alignment && ((buffer_alignment & (buffer_alignment - 1)) == 0), 0, TAG, "invalid buffer alignment");
// align down the max_buffer_size_per_node
max_buffer_size_per_node = max_buffer_size_per_node & ~(buffer_alignment - 1);
// calculate the number of nodes
return (buffer_size + max_buffer_size_per_node - 1) / max_buffer_size_per_node;
}
esp_err_t esp_dma_capable_malloc(size_t size, const esp_dma_mem_info_t *dma_mem_info, void **out_ptr, size_t *actual_size)
{
ESP_RETURN_ON_FALSE_ISR(dma_mem_info && out_ptr, ESP_ERR_INVALID_ARG, TAG, "null pointer");

10
components/esp_hw_support/dma/include/esp_private/esp_dma_utils.h
View File

@ -83,6 +83,16 @@ esp_err_t esp_dma_split_rx_buffer_to_cache_aligned(void *rx_buffer, size_t buffe
*/
esp_err_t esp_dma_merge_aligned_rx_buffers(dma_buffer_split_array_t *align_buf_array);
/**
* @brief Calculate the number of DMA linked list nodes required for a given buffer size
*
* @param[in] buffer_size Total size of the buffer
* @param[in] buffer_alignment Alignment requirement for the buffer
* @param[in] max_buffer_size_per_node Maximum buffer size that each node can handle
* @return Number of DMA linked list nodes required
*/
size_t esp_dma_calculate_node_count(size_t buffer_size, size_t buffer_alignment, size_t max_buffer_size_per_node);
#ifdef __cplusplus
}
#endif

13
components/esp_lcd/i80/esp_lcd_panel_io_i2s.c
View File

@ -39,6 +39,7 @@
#include "esp_private/periph_ctrl.h"
#include "esp_private/i2s_platform.h"
#include "esp_private/gdma_link.h"
#include "esp_private/esp_dma_utils.h"
#include "esp_private/gpio.h"
#include "soc/lcd_periph.h"
#include "hal/i2s_hal.h"
@ -71,7 +72,7 @@ struct esp_lcd_i80_bus_t {
int wr_gpio_num; // GPIO used for WR line
intr_handle_t intr; // LCD peripheral interrupt handle
esp_pm_lock_handle_t pm_lock; // lock APB frequency when necessary
size_t num_dma_nodes; // Number of DMA descriptors
size_t max_transfer_bytes; // Maximum number of bytes that can be transferred in one transaction
gdma_link_list_handle_t dma_link; // DMA link list handle
uint8_t *format_buffer;// The driver allocates an internal buffer for DMA to do data format transformer
unsigned long resolution_hz; // LCD_CLK resolution, determined by selected clock source
@ -141,7 +142,7 @@ esp_err_t esp_lcd_new_i80_bus(const esp_lcd_i80_bus_config_t *bus_config, esp_lc
// allocate i80 bus memory
bus = heap_caps_calloc(1, sizeof(esp_lcd_i80_bus_t), LCD_I80_MEM_ALLOC_CAPS);
ESP_GOTO_ON_FALSE(bus, ESP_ERR_NO_MEM, err, TAG, "no mem for i80 bus");
size_t num_dma_nodes = max_transfer_bytes / LCD_DMA_DESCRIPTOR_BUFFER_MAX_SIZE + 1;
size_t num_dma_nodes = esp_dma_calculate_node_count(max_transfer_bytes, 1, LCD_DMA_DESCRIPTOR_BUFFER_MAX_SIZE);
// create DMA link list
gdma_link_list_config_t dma_link_config = {
.buffer_alignment = 1, // no special buffer alignment for LCD TX buffer
@ -153,7 +154,7 @@ esp_err_t esp_lcd_new_i80_bus(const esp_lcd_i80_bus_config_t *bus_config, esp_lc
};
ESP_GOTO_ON_ERROR(gdma_new_link_list(&dma_link_config, &bus->dma_link), err, TAG, "create DMA link list failed");
bus->bus_id = -1;
bus->num_dma_nodes = num_dma_nodes;
bus->max_transfer_bytes = max_transfer_bytes;
#if SOC_I2S_TRANS_SIZE_ALIGN_WORD
// transform format for LCD commands, parameters and color data, so we need a big buffer
bus->format_buffer = heap_caps_calloc(1, max_transfer_bytes, MALLOC_CAP_INTERNAL | MALLOC_CAP_8BIT | MALLOC_CAP_DMA);
@ -217,7 +218,7 @@ esp_err_t esp_lcd_new_i80_bus(const esp_lcd_i80_bus_config_t *bus_config, esp_lc
bus->dc_gpio_num = bus_config->dc_gpio_num;
bus->wr_gpio_num = bus_config->wr_gpio_num;
*ret_bus = bus;
ESP_LOGD(TAG, "new i80 bus(%d) @%p, %zu dma nodes, resolution %luHz", bus->bus_id, bus, num_dma_nodes, bus->resolution_hz);
ESP_LOGD(TAG, "new i80 bus(%d) @%p, resolution %luHz", bus->bus_id, bus, bus->resolution_hz);
return ESP_OK;
err:
@ -510,7 +511,7 @@ static esp_err_t panel_io_i80_tx_param(esp_lcd_panel_io_t *io, int lcd_cmd, cons
esp_lcd_i80_bus_t *bus = next_device->bus;
lcd_panel_io_i80_t *cur_device = bus->cur_device;
lcd_i80_trans_descriptor_t *trans_desc = NULL;
assert(param_size <= (bus->num_dma_nodes * LCD_DMA_DESCRIPTOR_BUFFER_MAX_SIZE) && "parameter bytes too long, enlarge max_transfer_bytes");
assert(param_size <= bus->max_transfer_bytes && "parameter bytes too long, enlarge max_transfer_bytes");
assert(param_size <= LCD_I80_IO_FORMAT_BUF_SIZE && "format buffer too small, increase LCD_I80_IO_FORMAT_BUF_SIZE");
size_t num_trans_inflight = next_device->num_trans_inflight;
// before issue a polling transaction, need to wait queued transactions finished
@ -587,7 +588,7 @@ static esp_err_t panel_io_i80_tx_color(esp_lcd_panel_io_t *io, int lcd_cmd, cons
esp_lcd_i80_bus_t *bus = next_device->bus;
lcd_panel_io_i80_t *cur_device = bus->cur_device;
lcd_i80_trans_descriptor_t *trans_desc = NULL;
assert(color_size <= (bus->num_dma_nodes * LCD_DMA_DESCRIPTOR_BUFFER_MAX_SIZE) && "color bytes too long, enlarge max_transfer_bytes");
assert(color_size <= bus->max_transfer_bytes && "color bytes too long, enlarge max_transfer_bytes");
size_t num_trans_inflight = next_device->num_trans_inflight;
// before issue a polling transaction, need to wait queued transactions finished
for (size_t i = 0; i < num_trans_inflight; i++) {

16
components/esp_lcd/i80/esp_lcd_panel_io_i80.c
View File

@ -33,6 +33,7 @@
#include "esp_private/gpio.h"
#include "esp_private/gdma.h"
#include "esp_private/gdma_link.h"
#include "esp_private/esp_dma_utils.h"
#include "esp_private/periph_ctrl.h"
#include "esp_lcd_common.h"
#include "soc/lcd_periph.h"
@ -87,7 +88,7 @@ struct esp_lcd_i80_bus_t {
uint8_t *format_buffer; // The driver allocates an internal buffer for DMA to do data format transformer
uint8_t *format_buffer_nc; // Non-cacheable version of format buffer
size_t resolution_hz; // LCD_CLK resolution, determined by selected clock source
size_t num_dma_nodes; // Number of DMA nodes (descriptors)
size_t max_transfer_bytes; // Maximum number of bytes that can be transferred in one transaction
gdma_channel_handle_t dma_chan; // DMA channel handle
gdma_link_list_handle_t dma_link; // DMA link list handle
size_t int_mem_align; // Alignment for internal memory
@ -199,6 +200,7 @@ esp_err_t esp_lcd_new_i80_bus(const esp_lcd_i80_bus_config_t *bus_config, esp_lc
lcd_ll_enable_interrupt(bus->hal.dev, LCD_LL_EVENT_TRANS_DONE, false); // disable all interrupts
lcd_ll_clear_interrupt_status(bus->hal.dev, UINT32_MAX); // clear pending interrupt
// install DMA service
bus->max_transfer_bytes = bus_config->max_transfer_bytes;
ret = lcd_i80_init_dma_link(bus, bus_config);
ESP_GOTO_ON_ERROR(ret, err, TAG, "install DMA failed");
// disable RGB-LCD mode
@ -223,7 +225,7 @@ esp_err_t esp_lcd_new_i80_bus(const esp_lcd_i80_bus_config_t *bus_config, esp_lc
LIST_INIT(&bus->device_list); // initialize device list head
bus->spinlock = (portMUX_TYPE)portMUX_INITIALIZER_UNLOCKED;
*ret_bus = bus;
ESP_LOGD(TAG, "new i80 bus(%d) @%p, %zu dma nodes", bus_id, bus, bus->num_dma_nodes);
ESP_LOGD(TAG, "new i80 bus(%d) @%p", bus_id, bus);
return ESP_OK;
err:
@ -449,7 +451,7 @@ static esp_err_t panel_io_i80_tx_param(esp_lcd_panel_io_t *io, int lcd_cmd, cons
esp_lcd_i80_bus_t *bus = next_device->bus;
lcd_panel_io_i80_t *cur_device = bus->cur_device;
lcd_i80_trans_descriptor_t *trans_desc = NULL;
assert(param_size <= (bus->num_dma_nodes * LCD_DMA_DESCRIPTOR_BUFFER_MAX_SIZE) && "parameter bytes too long, enlarge max_transfer_bytes");
assert(param_size <= bus->max_transfer_bytes && "parameter bytes too long, enlarge max_transfer_bytes");
assert(param_size <= LCD_I80_IO_FORMAT_BUF_SIZE && "format buffer too small, increase LCD_I80_IO_FORMAT_BUF_SIZE");
uint32_t cmd_cycles = next_device->lcd_cmd_bits / bus->bus_width;
// in case bus_width=16 and cmd_bits=8, we still need 1 cmd_cycle
@ -514,7 +516,7 @@ static esp_err_t panel_io_i80_tx_color(esp_lcd_panel_io_t *io, int lcd_cmd, cons
lcd_panel_io_i80_t *i80_device = __containerof(io, lcd_panel_io_i80_t, base);
esp_lcd_i80_bus_t *bus = i80_device->bus;
lcd_i80_trans_descriptor_t *trans_desc = NULL;
assert(color_size <= (bus->num_dma_nodes * LCD_DMA_DESCRIPTOR_BUFFER_MAX_SIZE) && "color bytes too long, enlarge max_transfer_bytes");
assert(color_size <= bus->max_transfer_bytes && "color bytes too long, enlarge max_transfer_bytes");
uint32_t cache_line_size = 0;
if (esp_ptr_external_ram(color)) {
// check alignment
@ -610,10 +612,11 @@ static esp_err_t lcd_i80_init_dma_link(esp_lcd_i80_bus_handle_t bus, const esp_l
ESP_RETURN_ON_ERROR(gdma_config_transfer(bus->dma_chan, &trans_cfg), TAG, "config DMA transfer failed");
gdma_get_alignment_constraints(bus->dma_chan, &bus->int_mem_align, &bus->ext_mem_align);
size_t num_dma_nodes = bus_config->max_transfer_bytes / LCD_DMA_DESCRIPTOR_BUFFER_MAX_SIZE + 1;
size_t buffer_alignment = MAX(bus->int_mem_align, bus->ext_mem_align);
size_t num_dma_nodes = esp_dma_calculate_node_count(bus->max_transfer_bytes, buffer_alignment, LCD_DMA_DESCRIPTOR_BUFFER_MAX_SIZE);
// create DMA link list
gdma_link_list_config_t dma_link_config = {
.buffer_alignment = MAX(bus->int_mem_align, bus->ext_mem_align),
.buffer_alignment = buffer_alignment,
.item_alignment = LCD_GDMA_DESCRIPTOR_ALIGN,
.num_items = num_dma_nodes,
.flags = {
@ -621,7 +624,6 @@ static esp_err_t lcd_i80_init_dma_link(esp_lcd_i80_bus_handle_t bus, const esp_l
},
};
ESP_RETURN_ON_ERROR(gdma_new_link_list(&dma_link_config, &bus->dma_link), TAG, "create DMA link list failed");
bus->num_dma_nodes = num_dma_nodes;
return ESP_OK;
}

15
components/esp_lcd/rgb/esp_lcd_panel_rgb.c
View File

@ -32,6 +32,7 @@
#include "esp_private/esp_clk_tree_common.h"
#include "esp_private/gdma.h"
#include "esp_private/gdma_link.h"
#include "esp_private/esp_dma_utils.h"
#include "esp_private/periph_ctrl.h"
#include "esp_private/gpio.h"
#include "esp_psram.h"
@ -918,9 +919,10 @@ static esp_err_t lcd_rgb_panel_init_trans_link(esp_rgb_panel_t *rgb_panel)
#endif
if (rgb_panel->bb_size) {
// DMA is used to convey the bounce buffer
size_t num_dma_nodes_per_bounce_buffer = (rgb_panel->bb_size + LCD_DMA_DESCRIPTOR_BUFFER_MAX_SIZE - 1) / LCD_DMA_DESCRIPTOR_BUFFER_MAX_SIZE;
size_t buffer_alignment = rgb_panel->int_mem_align;
size_t num_dma_nodes_per_bounce_buffer = esp_dma_calculate_node_count(rgb_panel->bb_size, buffer_alignment, LCD_DMA_DESCRIPTOR_BUFFER_MAX_SIZE);
gdma_link_list_config_t link_cfg = {
.buffer_alignment = rgb_panel->int_mem_align,
.buffer_alignment = buffer_alignment,
.item_alignment = LCD_GDMA_DESCRIPTOR_ALIGN,
.num_items = num_dma_nodes_per_bounce_buffer * RGB_LCD_PANEL_BOUNCE_BUF_NUM,
.flags = {
@ -940,7 +942,7 @@ static esp_err_t lcd_rgb_panel_init_trans_link(esp_rgb_panel_t *rgb_panel)
#if RGB_LCD_NEEDS_SEPARATE_RESTART_LINK
// create restart link
gdma_link_list_config_t restart_link_cfg = {
.buffer_alignment = rgb_panel->int_mem_align,
.buffer_alignment = buffer_alignment,
.item_alignment = LCD_GDMA_DESCRIPTOR_ALIGN,
.num_items = 1, // the restart link only contains one node
.flags = {
@ -960,9 +962,10 @@ static esp_err_t lcd_rgb_panel_init_trans_link(esp_rgb_panel_t *rgb_panel)
#endif
} else {
// DMA is used to convey the frame buffer
size_t num_dma_nodes = (rgb_panel->fb_size + LCD_DMA_DESCRIPTOR_BUFFER_MAX_SIZE - 1) / LCD_DMA_DESCRIPTOR_BUFFER_MAX_SIZE;
size_t buffer_alignment = rgb_panel->flags.fb_in_psram ? rgb_panel->ext_mem_align : rgb_panel->int_mem_align;
uint32_t num_dma_nodes = esp_dma_calculate_node_count(rgb_panel->fb_size, buffer_alignment, LCD_DMA_DESCRIPTOR_BUFFER_MAX_SIZE);
gdma_link_list_config_t link_cfg = {
.buffer_alignment = rgb_panel->flags.fb_in_psram ? rgb_panel->ext_mem_align : rgb_panel->int_mem_align,
.buffer_alignment = buffer_alignment,
.item_alignment = LCD_GDMA_DESCRIPTOR_ALIGN,
.num_items = num_dma_nodes,
.flags = {
@ -986,7 +989,7 @@ static esp_err_t lcd_rgb_panel_init_trans_link(esp_rgb_panel_t *rgb_panel)
#if RGB_LCD_NEEDS_SEPARATE_RESTART_LINK
// create restart link
gdma_link_list_config_t restart_link_cfg = {
.buffer_alignment = rgb_panel->flags.fb_in_psram ? rgb_panel->ext_mem_align : rgb_panel->int_mem_align,
.buffer_alignment = buffer_alignment,
.item_alignment = LCD_GDMA_DESCRIPTOR_ALIGN,
.num_items = 1, // the restart link only contains one node
.flags = {