| Message ID | 20220526143715.277342-10-pan@semihalf.com |
|---|---|
| State | Superseded |
| Delegated to: | Simon Goldschmidt |
| Headers | show |
| Series | Add Chameleon v3 support | expand |
On Thu, 26 May 2022 at 07:38, Paweł Anikiel <pan@semihalf.com> wrote: > > Apply some optimizations to speed up bitstream loading > (both for full and split periph/core bitstreams): > > * Change the size of the first fs read, so that all the subsequent > reads are aligned to a specific value (called MAX_FIRST_LOAD_SIZE). > This value was chosen so that in subsequent reads the fat fs driver > doesn't have to allocate a temporary buffer in get_contents > (assuming 8KiB clusters). > > * Change the buffer size to a larger value when reading to ddr > (but not too large, because large transfers cause a stack overflow > in the dwmmc driver). When the size is too large, where exactly does that stack overflow happen? > > Signed-off-by: Paweł Anikiel <pan@semihalf.com> > --- > drivers/fpga/socfpga_arria10.c | 20 ++++++++++++++++++-- > 1 file changed, 18 insertions(+), 2 deletions(-) > Reviewed-by: Simon Glass <sjg@chromium.org>
On Fri, May 27, 2022 at 5:55 PM Simon Glass <sjg@chromium.org> wrote: > > On Thu, 26 May 2022 at 07:38, Paweł Anikiel <pan@semihalf.com> wrote: > > > > Apply some optimizations to speed up bitstream loading > > (both for full and split periph/core bitstreams): > > > > * Change the size of the first fs read, so that all the subsequent > > reads are aligned to a specific value (called MAX_FIRST_LOAD_SIZE). > > This value was chosen so that in subsequent reads the fat fs driver > > doesn't have to allocate a temporary buffer in get_contents > > (assuming 8KiB clusters). > > > > * Change the buffer size to a larger value when reading to ddr > > (but not too large, because large transfers cause a stack overflow > > in the dwmmc driver). > > When the size is too large, where exactly does that stack overflow happen? In dwmci_send_cmd (at drivers/mmc/dw_mmc.c:243). It stack-allocates a buffer of size sizeof(struct dwmci_idmac) * (data->blocks / 8). Since loading the bitstream is done from SPL (which is still in sram), we only have about 100K of stack, which is not enough to load an 11MB file in one go.
diff --git a/drivers/fpga/socfpga_arria10.c b/drivers/fpga/socfpga_arria10.c index 798e3a3f90..07bfe3060e 100644 --- a/drivers/fpga/socfpga_arria10.c +++ b/drivers/fpga/socfpga_arria10.c @@ -30,6 +30,14 @@ #define FPGA_TIMEOUT_MSEC 1000 /* timeout in ms */ #define FPGA_TIMEOUT_CNT 0x1000000 #define DEFAULT_DDR_LOAD_ADDRESS 0x400 +#define DDR_BUFFER_SIZE 0x100000 + +/* When reading bitstream from a filesystem, the size of the first read is + * changed so that the subsequent reads are aligned to this value. This value + * was chosen so that in subsequent reads the fat fs driver doesn't have to + * allocate a temporary buffer in get_contents (assuming 8KiB clusters). + */ +#define MAX_FIRST_LOAD_SIZE 0x2000 DECLARE_GLOBAL_DATA_PTR; @@ -526,7 +534,8 @@ static void get_rbf_image_info(struct rbf_info *rbf, u16 *buffer) #ifdef CONFIG_FS_LOADER static int first_loading_rbf_to_buffer(struct udevice *dev, struct fpga_loadfs_info *fpga_loadfs, - u32 *buffer, size_t *buffer_bsize) + u32 *buffer, size_t *buffer_bsize, + size_t *buffer_bsize_ori) { u32 *buffer_p = (u32 *)*buffer; u32 *loadable = buffer_p; @@ -674,6 +683,7 @@ static int first_loading_rbf_to_buffer(struct udevice *dev, } buffer_size = rbf_size; + *buffer_bsize_ori = DDR_BUFFER_SIZE; } debug("FPGA: External data: offset = 0x%x, size = 0x%x.\n", @@ -686,11 +696,16 @@ static int first_loading_rbf_to_buffer(struct udevice *dev, * chunk by chunk transfer is required due to smaller buffer size * compare to bitstream */ + + if (buffer_size > MAX_FIRST_LOAD_SIZE) + buffer_size = MAX_FIRST_LOAD_SIZE; + if (rbf_size <= buffer_size) { /* Loading whole bitstream into buffer */ buffer_size = rbf_size; fpga_loadfs->remaining = 0; } else { + buffer_size -= rbf_offset % buffer_size; fpga_loadfs->remaining -= buffer_size; } @@ -806,7 +821,8 @@ int socfpga_loadfs(fpga_fs_info *fpga_fsinfo, const void *buf, size_t bsize, * function below. */ ret = first_loading_rbf_to_buffer(dev, &fpga_loadfs, &buffer, - &buffer_sizebytes); + &buffer_sizebytes, + &buffer_sizebytes_ori); if (ret == 1) { printf("FPGA: Skipping configuration ...\n"); return 0;
Apply some optimizations to speed up bitstream loading (both for full and split periph/core bitstreams): * Change the size of the first fs read, so that all the subsequent reads are aligned to a specific value (called MAX_FIRST_LOAD_SIZE). This value was chosen so that in subsequent reads the fat fs driver doesn't have to allocate a temporary buffer in get_contents (assuming 8KiB clusters). * Change the buffer size to a larger value when reading to ddr (but not too large, because large transfers cause a stack overflow in the dwmmc driver). Signed-off-by: Paweł Anikiel <pan@semihalf.com> --- drivers/fpga/socfpga_arria10.c | 20 ++++++++++++++++++-- 1 file changed, 18 insertions(+), 2 deletions(-)