[PATCH v6 6/8] mips: octeon_nic23: Add PCIe FLR fixup via cyclic infrastructure
Stefan Roese
sr at denx.de
Fri Sep 2 13:57:52 CEST 2022
From: Aaron Williams <awilliams at marvell.com>
This patch adds a fixup function related to a PCIe FLR (Function Level
Reset) problem on the NIC23 PCIe board. This function is imported from
the Marvell Octeon 2013 U-Boot version as a (nearly) verbatim copy. It
uses the newly introduced cyclic infrastructure, so that this function
gets called every 100us, which is needed to detect this FLR issue.
Signed-off-by: Aaron Williams <awilliams at marvell.com>
Signed-off-by: Stefan Roese <sr at denx.de>
---
v6:
- No change
v5:
- No change
v4:
- Rename cyclic_struct to cyclic_info
v3:
- No change
v2:
- No change
board/Marvell/octeon_nic23/board.c | 197 +++++++++++++++++++++++++++++
configs/octeon_nic23_defconfig | 3 +
2 files changed, 200 insertions(+)
diff --git a/board/Marvell/octeon_nic23/board.c b/board/Marvell/octeon_nic23/board.c
index 3e2c54444397..08b1aa4b6efe 100644
--- a/board/Marvell/octeon_nic23/board.c
+++ b/board/Marvell/octeon_nic23/board.c
@@ -3,8 +3,10 @@
* Copyright (C) 2021-2022 Stefan Roese <sr at denx.de>
*/
+#include <cyclic.h>
#include <dm.h>
#include <ram.h>
+#include <time.h>
#include <asm/gpio.h>
#include <mach/octeon_ddr.h>
@@ -15,11 +17,90 @@
#include <mach/cvmx-helper-cfg.h>
#include <mach/cvmx-helper-util.h>
#include <mach/cvmx-bgxx-defs.h>
+#include <mach/cvmx-dtx-defs.h>
#include "board_ddr.h"
+/**
+ * cvmx_spem#_cfg_rd
+ *
+ * This register allows read access to the configuration in the PCIe core.
+ *
+ */
+union cvmx_spemx_cfg_rd {
+ u64 u64;
+ struct cvmx_spemx_cfg_rd_s {
+ u64 data : 32;
+ u64 addr : 32;
+ } s;
+ struct cvmx_spemx_cfg_rd_s cn73xx;
+};
+
+/**
+ * cvmx_spem#_cfg_wr
+ *
+ * This register allows write access to the configuration in the PCIe core.
+ *
+ */
+union cvmx_spemx_cfg_wr {
+ u64 u64;
+ struct cvmx_spemx_cfg_wr_s {
+ u64 data : 32;
+ u64 addr : 32;
+ } s;
+ struct cvmx_spemx_cfg_wr_s cn73xx;
+};
+
+/**
+ * cvmx_spem#_flr_pf_stopreq
+ *
+ * PF function level reset stop outbound requests register.
+ * Hardware automatically sets the STOPREQ bit for the PF when it enters a
+ * function level reset (FLR). Software is responsible for clearing the STOPREQ
+ * bit but must not do so prior to hardware taking down the FLR, which could be
+ * as long as 100ms. It may be appropriate for software to wait longer before clearing
+ * STOPREQ, software may need to drain deep DPI queues for example.
+ * Whenever SPEM receives a PF or child VF request mastered by CNXXXX over S2M (i.e. P or NP),
+ * when STOPREQ is set for the function, SPEM will discard the outgoing request
+ * before sending it to the PCIe core. If a NP, SPEM will schedule an immediate
+ * SWI_RSP_ERROR completion for the request - no timeout is required.
+ * In both cases, SPEM()_DBG_PF()_INFO[P()_BMD_E] will be set and a error
+ * interrupt is generated.
+ *
+ * STOPREQ mimics the behavior of PCIEEP()_CFG001[ME] for outbound requests that will
+ * master the PCIe bus (P and NP).
+ *
+ * STOPREQ will have no effect on completions returned by CNXXXX over the S2M,
+ * nor on M2S traffic.
+ *
+ * When a PF()_STOPREQ is set, none of the associated
+ * PEM()_FLR_PF()_VF_STOPREQ[VF_STOPREQ] will be set.
+ *
+ * STOPREQ is reset when the MAC is reset, and is not reset after a chip soft reset.
+ */
+union cvmx_spemx_flr_pf_stopreq {
+ u64 u64;
+ struct cvmx_spemx_flr_pf_stopreq_s {
+ u64 reserved_3_63 : 61;
+ u64 pf2_stopreq : 1;
+ u64 pf1_stopreq : 1;
+ u64 pf0_stopreq : 1;
+ } s;
+ struct cvmx_spemx_flr_pf_stopreq_s cn73xx;
+};
+
+#define CVMX_SPEMX_CFG_WR(offset) 0x00011800C0000028ull
+#define CVMX_SPEMX_CFG_RD(offset) 0x00011800C0000030ull
+#define CVMX_SPEMX_FLR_PF_STOPREQ(offset) 0x00011800C0000218ull
+
+#define DTX_SELECT_LTSSM 0x0
+#define DTX_SELECT_LTSSM_ENA 0x3ff
+#define LTSSM_L0 0x11
+
#define NIC23_DEF_DRAM_FREQ 800
+static u32 pci_cfgspace_reg0[2] = { 0, 0 };
+
static u8 octeon_nic23_cfg0_spd_values[512] = {
OCTEON_NIC23_CFG0_SPD_VALUES
};
@@ -145,8 +226,118 @@ void board_configure_qlms(void)
cvmx_qlm_measure_clock(4), cvmx_qlm_measure_clock(5));
}
+/**
+ * If there is a PF FLR then the PCI EEPROM is not re-read. In this case
+ * we need to re-program the vendor and device ID immediately after hardware
+ * completes FLR.
+ *
+ * PCI spec requires FLR to be completed within 100ms. The user who triggered
+ * FLR expects hardware to finish FLR within 100ms, otherwise the user will
+ * end up reading DEVICE_ID incorrectly from the reset value.
+ * CN23XX exits FLR at any point between 66 and 99ms, so U-Boot has to wait
+ * 99ms to let hardware finish its part, then finish reprogramming the
+ * correct device ID before the end of 100ms.
+ *
+ * Note: this solution only works properly when there is no other activity
+ * within U-Boot for 100ms from the time FLR is triggered.
+ *
+ * This function gets called every 100usec. If FLR happens during any
+ * other activity like bootloader/image update then it is possible that
+ * this function does not get called for more than the FLR period which will
+ * cause the PF device ID restore to happen after whoever initiated the FLR to
+ * read the incorrect device ID 0x9700 (reset value) instead of 0x9702
+ * (restored value).
+ */
+static void octeon_board_restore_pf(void *ctx)
+{
+ union cvmx_spemx_flr_pf_stopreq stopreq;
+ static bool start_initialized[2] = {false, false};
+ bool pf0_flag, pf1_flag;
+ u64 ltssm_bits;
+ const u64 pf_flr_wait_usecs = 99700;
+ u64 elapsed_usecs;
+ union cvmx_spemx_cfg_wr cfg_wr;
+ union cvmx_spemx_cfg_rd cfg_rd;
+ static u64 start_us[2];
+ int pf_num;
+
+ csr_wr(CVMX_DTX_SPEM_SELX(0), DTX_SELECT_LTSSM);
+ csr_rd(CVMX_DTX_SPEM_SELX(0));
+ csr_wr(CVMX_DTX_SPEM_ENAX(0), DTX_SELECT_LTSSM_ENA);
+ csr_rd(CVMX_DTX_SPEM_ENAX(0));
+ ltssm_bits = csr_rd(CVMX_DTX_SPEM_DATX(0));
+ if (((ltssm_bits >> 3) & 0x3f) != LTSSM_L0)
+ return;
+
+ stopreq.u64 = csr_rd(CVMX_SPEMX_FLR_PF_STOPREQ(0));
+ pf0_flag = stopreq.s.pf0_stopreq;
+ pf1_flag = stopreq.s.pf1_stopreq;
+ /* See if PF interrupt happened */
+ if (!(pf0_flag || pf1_flag))
+ return;
+
+ if (pf0_flag && !start_initialized[0]) {
+ start_initialized[0] = true;
+ start_us[0] = get_timer_us(0);
+ }
+
+ /* Store programmed PCIe DevID SPEM0 PF0 */
+ if (pf0_flag && !pci_cfgspace_reg0[0]) {
+ cfg_rd.s.addr = (0 << 24) | 0x0;
+ csr_wr(CVMX_SPEMX_CFG_RD(0), cfg_rd.u64);
+ cfg_rd.u64 = csr_rd(CVMX_SPEMX_CFG_RD(0));
+ pci_cfgspace_reg0[0] = cfg_rd.s.data;
+ }
+
+ if (pf1_flag && !start_initialized[1]) {
+ start_initialized[1] = true;
+ start_us[1] = get_timer_us(0);
+ }
+
+ /* Store programmed PCIe DevID SPEM0 PF1 */
+ if (pf1_flag && !pci_cfgspace_reg0[1]) {
+ cfg_rd.s.addr = (1 << 24) | 0x0;
+ csr_wr(CVMX_SPEMX_CFG_RD(0), cfg_rd.u64);
+ cfg_rd.u64 = csr_rd(CVMX_SPEMX_CFG_RD(0));
+ pci_cfgspace_reg0[1] = cfg_rd.s.data;
+ }
+
+ /* For PF, rewrite pci config space reg 0 */
+ for (pf_num = 0; pf_num < 2; pf_num++) {
+ if (!start_initialized[pf_num])
+ continue;
+
+ elapsed_usecs = get_timer_us(0) - start_us[pf_num];
+
+ if (elapsed_usecs > pf_flr_wait_usecs) {
+ /* Here, our measured FLR duration has passed;
+ * check if device ID has been reset,
+ * which indicates FLR completion (per MA team).
+ */
+ cfg_rd.s.addr = (pf_num << 24) | 0x0;
+ csr_wr(CVMX_SPEMX_CFG_RD(0), cfg_rd.u64);
+ cfg_rd.u64 = csr_rd(CVMX_SPEMX_CFG_RD(0));
+ /* if DevID has NOT been reset, FLR is not yet
+ * complete
+ */
+ if (cfg_rd.s.data != pci_cfgspace_reg0[pf_num]) {
+ stopreq.s.pf0_stopreq = (pf_num == 0) ? 1 : 0;
+ stopreq.s.pf1_stopreq = (pf_num == 1) ? 1 : 0;
+ csr_wr(CVMX_SPEMX_FLR_PF_STOPREQ(0), stopreq.u64);
+
+ cfg_wr.u64 = 0;
+ cfg_wr.s.addr = (pf_num << 24) | 0;
+ cfg_wr.s.data = pci_cfgspace_reg0[pf_num];
+ csr_wr(CVMX_SPEMX_CFG_WR(0), cfg_wr.u64);
+ start_initialized[pf_num] = false;
+ }
+ }
+ }
+}
+
int board_late_init(void)
{
+ struct cyclic_info *cyclic;
struct gpio_desc gpio = {};
ofnode node;
@@ -164,6 +355,12 @@ int board_late_init(void)
board_configure_qlms();
+ /* Register cyclic function for PCIe FLR fixup */
+ cyclic = cyclic_register(octeon_board_restore_pf, 100,
+ "pcie_flr_fix", NULL);
+ if (!cyclic)
+ printf("Registering of cyclic function failed\n");
+
return 0;
}
diff --git a/configs/octeon_nic23_defconfig b/configs/octeon_nic23_defconfig
index 95e98c1161db..098831e0b1ed 100644
--- a/configs/octeon_nic23_defconfig
+++ b/configs/octeon_nic23_defconfig
@@ -20,6 +20,8 @@ CONFIG_AHCI=y
CONFIG_OF_BOARD_FIXUP=y
CONFIG_SYS_CONSOLE_ENV_OVERWRITE=y
# CONFIG_SYS_DEVICE_NULLDEV is not set
+CONFIG_CYCLIC=y
+CONFIG_CYCLIC_MAX_CPU_TIME_US=5000
CONFIG_ARCH_MISC_INIT=y
CONFIG_BOARD_EARLY_INIT_F=y
CONFIG_BOARD_LATE_INIT=y
@@ -41,6 +43,7 @@ CONFIG_CMD_TIME=y
CONFIG_CMD_EXT4=y
CONFIG_CMD_FAT=y
CONFIG_CMD_FS_GENERIC=y
+CONFIG_CMD_CYCLIC=y
CONFIG_EFI_PARTITION=y
CONFIG_ENV_IS_IN_SPI_FLASH=y
CONFIG_TFTP_TSIZE=y
--
2.37.3
More information about the U-Boot
mailing list