Skip to contentSkip to conversation

Add support for cyclic function execution infrastruture

PATCH In review Stefan Roese 3 participants 15 messages Sep 2, 2022 – Sep 15, 2022

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

v6 cover Add support for cyclic function execution infrastruture

This patchset adds the basic infrastructure to periodically execute code, e.g. all 100ms. Examples for such functions might be LED blinking etc. The functions that are hooked into this cyclic list should be small timewise as otherwise the execution of the other code that relies on a high frequent polling (e.g. UART rx char ready check) might be delayed too much. This patch also adds the Kconfig option CONFIG_CYCLIC_MAX_CPU_TIME_US, which configures the max allowed time for such a cyclic function. If it's execution time exceeds this time, this cyclic function will get removed from the cyclic list.

How is this cyclic functionality executed?
This patchset integrates the main function responsible for calling all registered cyclic functions cyclic_run() into the common WATCHDOG_RESET macro. This guarantees that cyclic_run() is executed very often, which is necessary for the cyclic functions to get scheduled and executed at their configured periods.

This cyclic infrastructure will be used by a board specific function on the NIC23 MIPS Octeon board, which needs to check periodically, if a PCIe FLR has occurred.

Ideas how to continue:
One idea is to rename WATCHDOG_RESET to something like SCHEDULE and move the watchdog_reset call into this cyclic infrastructure as well. Or to perhaps move the shell UART RX ready polling to a cyclic function.

It's also possible to extend the "cyclic" command, to support the creation of periodically executed shell commands (for testing etc).

Here the Azure build, without any issues:
https://dev.azure.com/sr0718/u-boot/_build/results?buildId=256&view=results

Aaron Williams (1):
mips: octeon_nic23: Add PCIe FLR fixup via cyclic infrastructure

Stefan Roese (7):
time: Import time_after64() and friends from Linux
cyclic: Add basic support for cyclic function execution infrastruture
cyclic: Integrate cyclic infrastructure into WATCHDOG_RESET
cyclic: Integrate cyclic functionality at bootup in board_r/f
cyclic: Add 'cyclic list' and 'cyclic demo' commands
cyclic: Add documentation
cyclic: Add a simple test

 MAINTAINERS                        |   7 +
 board/Marvell/octeon_nic23/board.c | 197 +++++++++++++++++++++++++++++
 cmd/Kconfig                        |  16 +++
 cmd/Makefile                       |   1 +
 cmd/cyclic.c                       |  84 ++++++++++++
 common/Kconfig                     |  20 +++
 common/Makefile                    |   1 +
 common/board_f.c                   |   2 +
 common/board_r.c                   |   2 +
 common/cyclic.c                    | 123 ++++++++++++++++++
 configs/octeon_nic23_defconfig     |   3 +
 doc/develop/cyclic.rst             |  50 ++++++++
 doc/develop/index.rst              |   1 +
 doc/usage/cmd/cyclic.rst           |  45 +++++++
 doc/usage/index.rst                |   1 +
 fs/cramfs/uncompress.c             |   2 +-
 include/asm-generic/global_data.h  |   7 +
 include/cyclic.h                   | 138 ++++++++++++++++++++
 include/time.h                     |  19 +++
 include/watchdog.h                 |  23 +++-
 test/common/Makefile               |   1 +
 test/common/cyclic.c               |  35 +++++
 test/test-main.c                   |   3 +
 23 files changed, 777 insertions(+), 4 deletions(-)
 create mode 100644 cmd/cyclic.c
 create mode 100644 common/cyclic.c
 create mode 100644 doc/develop/cyclic.rst
 create mode 100644 doc/usage/cmd/cyclic.rst
 create mode 100644 include/cyclic.h
 create mode 100644 test/common/cyclic.c
Signature · 1 line
2.37.3
#1 lore raw

PATCH v6 1/8 time: Import time_after64() and friends from Linux

When using us times it makes sense to use 64bit variables for storage. The currently implemented time_after() and friends functions only handle 32bit variables. This patch now includes the 64bit variants as well from Linux. This will be used by the upcoming generic cyclic function infrastructure.

These macros were copied from include/linux/jiffies.h of Linux 5.18.

---
v6:
- No change

v5:
- No change

v4:
- Added Simon's RB tag

v3:
- No change

v2:
- No change

 include/time.h | 19 +++++++++++++++++++
 1 file changed, 19 insertions(+)
include/time.h +19 -0
index 9deb2cf61cc4..3b2ba0912470 100644
@@ -83,6 +83,25 @@ uint64_t usec_to_tick(unsigned long usec);
 	(time_after_eq(a,b) && \
 	 time_before(a,c))
 
+/* Same as above, but does so with platform independent 64bit types.
+ * These must be used when utilizing jiffies_64 (i.e. return value of
+ * get_jiffies_64() */
+#define time_after64(a,b)	\
+	(typecheck(__u64, a) &&	\
+	 typecheck(__u64, b) && \
+	 ((__s64)((b) - (a)) < 0))
+#define time_before64(a,b)	time_after64(b,a)
+
+#define time_after_eq64(a,b)	\
+	(typecheck(__u64, a) && \
+	 typecheck(__u64, b) && \
+	 ((__s64)((a) - (b)) >= 0))
+#define time_before_eq64(a,b)	time_after_eq64(b,a)
+
+#define time_in_range64(a, b, c) \
+	(time_after_eq64(a, b) && \
+	 time_before_eq64(a, c))
+
 /**
  * usec2ticks() - Convert microseconds to internal ticks
  *
Signature · 1 line
2.37.3
#2 lore raw

PATCH v6 2/8 cyclic: Add basic support for cyclic function execution infrastruture

In reply to Stefan Roese

Add the basic infrastructure to periodically execute code, e.g. all 100ms. Examples for such functions might be LED blinking etc. The functions that are hooked into this cyclic list should be small timewise as otherwise the execution of the other code that relies on a high frequent polling (e.g. UART rx char ready check) might be delayed too much. This patch also adds the Kconfig option CONFIG_CYCLIC_MAX_CPU_TIME_US, which configures the max allowed time for such a cyclic function. If it's execution time exceeds this time, this cyclic function will get removed from the cyclic list.

How is this cyclic functionality executed?
The following patch integrates the main function responsible for calling all registered cyclic functions cyclic_run() into the common WATCHDOG_RESET macro. This guarantees that cyclic_run() is executed very often, which is necessary for the cyclic functions to get scheduled and executed at their configured periods.

This cyclic infrastructure will be used by a board specific function on the NIC23 MIPS Octeon board, which needs to check periodically, if a PCIe FLR has occurred.

---
v6:
- Move static and global variables into struct in GD
- Add cyclic_get_list() to retieve pointer to the cyclic list
- Add move documentation for structs and function prototypes

v5:
- No change

v4:
- Rename cyclic_struct to cyclic_info and document it
- Set 'cyclic_ready' to false in cyclic_uninit()
- Added Simon's RB tag

v3:
- No change

v2:
- Also add cyclic_register() and cyclic_unregister() as empty functions when CONFIG_CYCLIC is not defined
- Misc minor changes

 MAINTAINERS                       |   6 ++
 common/Kconfig                    |  20 +++++
 common/Makefile                   |   1 +
 common/cyclic.c                   | 123 ++++++++++++++++++++++++++
 include/asm-generic/global_data.h |   7 ++
 include/cyclic.h                  | 138 ++++++++++++++++++++++++++++++
 6 files changed, 295 insertions(+)
 create mode 100644 common/cyclic.c
 create mode 100644 include/cyclic.h
MAINTAINERS +6 -0
index 36a2b69fcb97..7255b8286ae2 100644
@@ -788,6 +788,12 @@ T:	git https://source.denx.de/u-boot/custodians/u-boot-coldfire.git
 F:	arch/m68k/
 F:	doc/arch/m68k.rst
 
+CYCLIC
+M:	Stefan Roese <[email protected]>
+S:	Maintained
+F:	common/cyclic.c
+F:	include/cyclic.h
+
 DFU
 M:	Lukasz Majewski <[email protected]>
 S:	Maintained
common/Kconfig +20 -0
index e7914ca750a3..a6a94ab8dfbf 100644
@@ -545,6 +545,26 @@ config DISPLAY_BOARDINFO_LATE
 
 menu "Start-up hooks"
 
+config CYCLIC
+	bool "General-purpose cyclic execution mechanism"
+	help
+	  This enables a general-purpose cyclic execution infrastructure,
+	  to allow "small" (run-time wise) functions to be executed at
+	  a specified frequency. Things like LED blinking or watchdog
+	  triggering are examples for such tasks.
+
+if CYCLIC
+
+config CYCLIC_MAX_CPU_TIME_US
+	int "Sets the max allowed time for a cyclic function in us"
+	default 1000
+	help
+	  The max allowed time for a cyclic function in us. If a functions
+	  takes longer than this duration this function will get unregistered
+	  automatically.
+
+endif # CYCLIC
+
 config EVENT
 	bool "General-purpose event-handling mechanism"
 	default y if SANDBOX
common/Makefile +1 -0
index 2ed8672c3ac1..1d56c9f2895a 100644
@@ -84,6 +84,7 @@ obj-y += malloc_simple.o
 endif
 endif
 
+obj-$(CONFIG_CYCLIC) += cyclic.o
 obj-$(CONFIG_$(SPL_TPL_)EVENT) += event.o
 
 obj-$(CONFIG_$(SPL_TPL_)HASH) += hash.o
common/cyclic.c new file +123 -0

123 lines not on this page · raw message

include/asm-generic/global_data.h +7 -0
index 2a747d91e16f..4aeb61f08c40 100644
@@ -20,6 +20,7 @@
  */
 
 #ifndef __ASSEMBLY__
+#include <cyclic.h>
 #include <event_internal.h>
 #include <fdtdec.h>
 #include <membuff.h>
@@ -477,6 +478,12 @@ struct global_data {
 	 * @event_state: Points to the current state of events
 	 */
 	struct event_state event_state;
+#endif
+#ifdef CONFIG_CYCLIC
+	/**
+	 * @cyclic: cyclic driver data
+	 */
+	struct cyclic_drv *cyclic;
 #endif
 	/**
 	 * @dmtag_list: List of DM tags
include/cyclic.h new file +138 -0

138 lines not on this page · raw message

Signature · 1 line
2.37.3
#3 lore raw

PATCH v6 3/8 cyclic: Integrate cyclic infrastructure into WATCHDOG_RESET

In reply to Stefan Roese

This patch integrates the main function responsible for calling all registered cyclic functions cyclic_run() into the common WATCHDOG_RESET macro. This guarantees that cyclic_run() is executed very often, which is necessary for the cyclic functions to get scheduled and executed at their configured periods.

If CONFIG_WATCHDOG is not enabled, only cyclic_run() without calling watchdog_reset(). This guarantees that the cyclic functionality does not rely on CONFIG_WATCHDOG being enabled.

---
v6:
- No change

v5:
- No change

v4:
- Added Simon's RB tag

v3:
- No change

v2:
- No change

 fs/cramfs/uncompress.c |  2 +-
 include/watchdog.h     | 23 ++++++++++++++++++++---
 2 files changed, 21 insertions(+), 4 deletions(-)
fs/cramfs/uncompress.c +1 -1
index f431cc46c1f7..38e10e2e4422 100644
@@ -62,7 +62,7 @@ int cramfs_uncompress_init (void)
 	stream.avail_in = 0;
 
 #if defined(CONFIG_HW_WATCHDOG) || defined(CONFIG_WATCHDOG)
-	stream.outcb = (cb_func) WATCHDOG_RESET;
+	stream.outcb = (cb_func)watchdog_reset_func;
 #else
 	stream.outcb = Z_NULL;
 #endif /* CONFIG_HW_WATCHDOG */
include/watchdog.h +20 -3
index 813cc8f2a5d3..0a9777edcbad 100644
@@ -11,6 +11,8 @@
 #define _WATCHDOG_H_
 
 #if !defined(__ASSEMBLY__)
+#include <cyclic.h>
+
 /*
  * Reset the watchdog timer, always returns 0
  *
@@ -60,11 +62,16 @@ int init_func_watchdog_reset(void);
 			/* Don't require the watchdog to be enabled in SPL */
 			#if defined(CONFIG_SPL_BUILD) &&		\
 				!defined(CONFIG_SPL_WATCHDOG)
-				#define WATCHDOG_RESET() {}
+				#define WATCHDOG_RESET() { \
+					cyclic_run(); \
+				}
 			#else
 				extern void watchdog_reset(void);
 
-				#define WATCHDOG_RESET watchdog_reset
+				#define WATCHDOG_RESET() { \
+					watchdog_reset(); \
+					cyclic_run(); \
+				}
 			#endif
 		#endif
 	#else
@@ -74,11 +81,21 @@ int init_func_watchdog_reset(void);
 		#if defined(__ASSEMBLY__)
 			#define WATCHDOG_RESET /*XXX DO_NOT_DEL_THIS_COMMENT*/
 		#else
-			#define WATCHDOG_RESET() {}
+			#define WATCHDOG_RESET() { \
+				cyclic_run(); \
+			}
 		#endif /* __ASSEMBLY__ */
 	#endif /* CONFIG_WATCHDOG && !__ASSEMBLY__ */
 #endif /* CONFIG_HW_WATCHDOG */
 
+#if !defined(__ASSEMBLY__)
+/* Currently only needed for fs/cramfs/uncompress.c */
+static inline void watchdog_reset_func(void)
+{
+	WATCHDOG_RESET();
+}
+#endif
+
 /*
  * Prototypes from $(CPU)/cpu.c.
  */
Signature · 1 line
2.37.3
#4 lore raw

PATCH v6 4/8 cyclic: Integrate cyclic functionality at bootup in board_r/f

In reply to Stefan Roese

This patch adds a call to cyclic_init() to board_f/r.c, enabling the common cyclic infrastructure. After this it's possible to add cyclic functions via cyclic_register().

---
v6:
- Move cyclic_init() in board_r.c after initr_malloc, as we now need malloc to work

v5:
- No change

v4:
- Added Simon's RB tag

v3:
- No change

v2:
- No change

 common/board_f.c | 2 ++
 common/board_r.c | 2 ++
 2 files changed, 4 insertions(+)
common/board_f.c +2 -0
index 18e2246733b0..deb46be18227 100644
@@ -16,6 +16,7 @@
 #include <console.h>
 #include <cpu.h>
 #include <cpu_func.h>
+#include <cyclic.h>
 #include <display_options.h>
 #include <dm.h>
 #include <env.h>
@@ -828,6 +829,7 @@ static const init_fnc_t init_sequence_f[] = {
 	initf_malloc,
 	log_init,
 	initf_bootstage,	/* uses its own timer, so does not need DM */
+	cyclic_init,
 	event_init,
 #ifdef CONFIG_BLOBLIST
 	bloblist_init,
common/board_r.c +2 -0
index 56eb60fa2753..062bc3e688c6 100644
@@ -13,6 +13,7 @@
 #include <api.h>
 #include <bootstage.h>
 #include <cpu_func.h>
+#include <cyclic.h>
 #include <display_options.h>
 #include <exports.h>
 #ifdef CONFIG_MTD_NOR_FLASH
@@ -611,6 +612,7 @@ static init_fnc_t init_sequence_r[] = {
 #endif
 	initr_barrier,
 	initr_malloc,
+	cyclic_init,
 	log_init,
 	initr_bootstage,	/* Needs malloc() but has its own timer */
 #if defined(CONFIG_CONSOLE_RECORD)
Signature · 1 line
2.37.3
#5 lore raw

PATCH v6 5/8 cyclic: Add 'cyclic list' and 'cyclic demo' commands

In reply to Stefan Roese

This patch adds the cyclic command, which currently only supports the 'list' subcommand, to list all currently registered cyclic functions. Here an example:

=> cyclic list
function: cyclic_demo, cpu-time: 7010 us, frequency: 99.80 times/s
function: cyclic_demo2, cpu-time: 1 us, frequency: 1.13 times/s

As you can see, the cpu-time is accounted, so that cyclic functions that take too long might be discovered. Additionally the frequency is logged.

The 'cyclic demo' commands registers the cyclic_demo() function to be executed all 'cycletime_ms' milliseconds. The only thing this function does is delaying by 'delay_us' microseconds.

---
v6:
- Use cyclic_get_list()
- Use div64 macros for division and remainder

v5:
- Add 'cyclic demo' command

v4:
- Add 'default y' in Kconfig and extend help text
- Minor spelling fix
- Added Simon's RB tag

v3:
- No change

v2:
- Add depends on CYCLIC in Kconfig

 MAINTAINERS  |  1 +
 cmd/Kconfig  | 16 ++++++++++
 cmd/Makefile |  1 +
 cmd/cyclic.c | 84 ++++++++++++++++++++++++++++++++++++++++++++++++++++
 4 files changed, 102 insertions(+)
 create mode 100644 cmd/cyclic.c
MAINTAINERS +1 -0
index 7255b8286ae2..87eef2f1b259 100644
@@ -791,6 +791,7 @@ F:	doc/arch/m68k.rst
 CYCLIC
 M:	Stefan Roese <[email protected]>
 S:	Maintained
+F:	cmd/cyclic.c
 F:	common/cyclic.c
 F:	include/cyclic.h
 
cmd/Kconfig +16 -0
index 8ea064b8d2fb..0649ffcbdf8e 100644
@@ -2505,6 +2505,22 @@ config CMD_CBSYSINFO
 	  memory by coreboot before jumping to U-Boot. It can be useful for
 	  debugging the beaaviour of coreboot or U-Boot.
 
+config CMD_CYCLIC
+	bool "cyclic - Show information about cyclic functions"
+	depends on CYCLIC
+	default y
+	help
+	  This enables the 'cyclic' command which provides information about
+	  cyclic execution functions. This infrastructure allows registering
+	  functions to be executed cyclically, e.g. every 100ms. These commands
+	  are supported:
+
+	    cyclic list - list cyclic functions
+	    cyclic cyclic demo <cycletime_ms> <delay_us> - register cyclic
+		demo function
+
+	  See doc/develop/cyclic.rst for more details.
+
 config CMD_DIAG
 	bool "diag - Board diagnostics"
 	help
cmd/Makefile +1 -0
index 6e87522b62e8..a0abfc623a2b 100644
@@ -56,6 +56,7 @@ obj-$(CONFIG_CMD_DIAG) += diag.o
 endif
 obj-$(CONFIG_CMD_ADTIMG) += adtimg.o
 obj-$(CONFIG_CMD_ABOOTIMG) += abootimg.o
+obj-$(CONFIG_CMD_CYCLIC) += cyclic.o
 obj-$(CONFIG_CMD_EVENT) += event.o
 obj-$(CONFIG_CMD_EXTENSION) += extension_board.o
 obj-$(CONFIG_CMD_ECHO) += echo.o
cmd/cyclic.c new file +84 -0

84 lines not on this page · raw message

Signature · 1 line
2.37.3
#6 lore raw

PATCH v6 6/8 mips: octeon_nic23: Add PCIe FLR fixup via cyclic infrastructure

In reply to Stefan Roese

From: Aaron Williams <[email protected]>

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.

---
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(+)
board/Marvell/octeon_nic23/board.c +197 -0
index 3e2c54444397..08b1aa4b6efe 100644
@@ -3,8 +3,10 @@
  * Copyright (C) 2021-2022 Stefan Roese <[email protected]>
  */
 
+#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;
 }
 
configs/octeon_nic23_defconfig +3 -0
index 95e98c1161db..098831e0b1ed 100644
@@ -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
Signature · 1 line
2.37.3
#7 lore raw

PATCH v6 7/8 cyclic: Add documentation

In reply to Stefan Roese

Add documentation for the cyclic function infrastructure, including the cyclic command.

---
v6:
- No change

v5:
- No change

v4:
- Minor spelling fix
- Added Simon's RB tag

v3:
- New patch

 doc/develop/cyclic.rst   | 50 ++++++++++++++++++++++++++++++++++++++++
 doc/develop/index.rst    |  1 +
 doc/usage/cmd/cyclic.rst | 45 ++++++++++++++++++++++++++++++++++++
 doc/usage/index.rst      |  1 +
 4 files changed, 97 insertions(+)
 create mode 100644 doc/develop/cyclic.rst
 create mode 100644 doc/usage/cmd/cyclic.rst
doc/develop/cyclic.rst new file +50 -0
new file mode 100644
index 000000000000..43bedacb9f88
@@ -0,0 +1,50 @@
+.. SPDX-License-Identifier: GPL-2.0+
+
+Cyclic functions
+================
+
+The cyclic function execution infrastruture provides a way to periodically
+execute code, e.g. every 100ms. Examples for such functions might be LED
+blinking etc. The functions that are hooked into this cyclic list should
+be small timewise as otherwise the execution of the other code that relies
+on a high frequent polling (e.g. UART rx char ready check) might be
+delayed too much. To detect cyclic functions with a too long execution
+time, the Kconfig option `CONFIG_CYCLIC_MAX_CPU_TIME_US` is introduced,
+which configures the max allowed time for such a cyclic function. If it's
+execution time exceeds this time, this cyclic function will get removed
+from the cyclic list.
+
+Registering a cyclic function
+-----------------------------
+
+To register a cyclic function, use something like this::
+
+    static void cyclic_demo(void *ctx)
+    {
+        /* Just a small dummy delay here */
+        udelay(10);
+    }
+    
+    int board_init(void)
+    {
+        struct cyclic_info *cyclic;
+        
+        /* Register demo cyclic function */
+        cyclic = cyclic_register(cyclic_demo, 10 * 1000, "cyclic_demo", NULL);
+        if (!cyclic)
+        printf("Registering of cyclic_demo failed\n");
+        
+        return 0;
+    }
+
+This will register the function `cyclic_demo()` to be periodically
+executed all 10ms.
+
+How is this cyclic functionality integrated /  executed?
+--------------------------------------------------------
+
+The cyclic infrastructure integrates the main function responsible for
+calling all registered cyclic functions cyclic_run() into the common
+WATCHDOG_RESET macro. This guarantees that cyclic_run() is executed
+very often, which is necessary for the cyclic functions to get scheduled
+and executed at their configured periods.
doc/develop/index.rst +1 -0
index f7ee09db2467..ce6b38e57629 100644
@@ -27,6 +27,7 @@ Implementation
    ci_testing
    commands
    config_binding
+   cyclic
    devicetree/index
    distro
    driver-model/index
doc/usage/cmd/cyclic.rst new file +45 -0

45 lines not on this page · raw message

doc/usage/index.rst +1 -0
index 28f9683a3e6f..1a0665ec1b78 100644
@@ -33,6 +33,7 @@ Shell commands
    cmd/bootz
    cmd/cbsysinfo
    cmd/conitrace
+   cmd/cyclic
    cmd/dm
    cmd/echo
    cmd/env
Signature · 1 line
2.37.3
#8 lore raw

PATCH v6 8/8 cyclic: Add a simple test

In reply to Stefan Roese

Add a test for cyclic function registration and activation.

---
v6:
- No change

v5:
- Set 'cyclic_active' to false before registering the test
- Use WATCHDOG_RESET() instead of mdelay()
- Use COMMON_TEST()
- Added Simon's RB tag

v4:
- New patch

 test/common/Makefile |  1 +
 test/common/cyclic.c | 35 +++++++++++++++++++++++++++++++++++
 test/test-main.c     |  3 +++
 3 files changed, 39 insertions(+)
 create mode 100644 test/common/cyclic.c
test/common/Makefile +1 -0
index 9087788ba6a8..cc918f64e544 100644
@@ -1,4 +1,5 @@
 # SPDX-License-Identifier: GPL-2.0+
 obj-y += cmd_ut_common.o
 obj-$(CONFIG_AUTOBOOT) += test_autoboot.o
+obj-$(CONFIG_CYCLIC) += cyclic.o
 obj-$(CONFIG_EVENT) += event.o
test/common/cyclic.c new file +35 -0
new file mode 100644
index 000000000000..a5c4e78989b9
@@ -0,0 +1,35 @@
+// SPDX-License-Identifier: GPL-2.0+
+/*
+ * Copyright (C) 2022 Stefan Roese <[email protected]>
+ */
+
+#include <common.h>
+#include <cyclic.h>
+#include <dm.h>
+#include <test/common.h>
+#include <test/test.h>
+#include <test/ut.h>
+#include <watchdog.h>
+#include <linux/delay.h>
+
+/* Test that cyclic function is called */
+static bool cyclic_active = false;
+
+static void cyclic_test(void *ctx)
+{
+	cyclic_active = true;
+}
+
+static int dm_test_cyclic_running(struct unit_test_state *uts)
+{
+	cyclic_active = false;
+	ut_assertnonnull(cyclic_register(cyclic_test, 10 * 1000, "cyclic_demo",
+					 NULL));
+
+	/* Execute all registered cyclic functions */
+	WATCHDOG_RESET();
+	ut_asserteq(true, cyclic_active);
+
+	return 0;
+}
+COMMON_TEST(dm_test_cyclic_running, 0);
test/test-main.c +3 -0
index 31837e57a8fb..8a609a8a2fce 100644
@@ -6,6 +6,7 @@
 
 #include <common.h>
 #include <console.h>
+#include <cyclic.h>
 #include <dm.h>
 #include <event.h>
 #include <dm/root.h>
@@ -220,6 +221,7 @@ static int dm_test_restore(struct device_node *of_root)
 static int test_pre_run(struct unit_test_state *uts, struct unit_test *test)
 {
 	ut_assertok(event_init());
+	ut_assertok(cyclic_init());
 
 	if (test->flags & UT_TESTF_DM)
 		ut_assertok(dm_test_pre_run(uts));
@@ -265,6 +267,7 @@ static int test_post_run(struct unit_test_state *uts, struct unit_test *test)
 	ut_unsilence_console(uts);
 	if (test->flags & UT_TESTF_DM)
 		ut_assertok(dm_test_post_run(uts));
+	ut_assertok(cyclic_uninit());
 	ut_assertok(event_uninit());
 
 	return 0;
Signature · 1 line
2.37.3
#9 lore raw
In reply to Stefan Roese
Quoted 11 lines · Stefan Roese

This patchset adds the basic infrastructure to periodically execute code, e.g. all 100ms. Examples for such functions might be LED blinking etc. The functions that are hooked into this cyclic list should be small timewise as otherwise the execution of the other code that relies on a high frequent polling (e.g. UART rx char ready check) might be delayed too much. This patch also adds the Kconfig option CONFIG_CYCLIC_MAX_CPU_TIME_US, which configures the max allowed time for such a cyclic function. If it's execution time exceeds this time, this cyclic function will get removed from the cyclic list.

[...]

For the series, applied to u-boot/next, thanks!

Signature · 1 line
Tom
#10 lore raw
Quoted 12 lines · Tom Rini
Quoted 10 lines · Stefan Roese

This patchset adds the basic infrastructure to periodically execute code, e.g. all 100ms. Examples for such functions might be LED blinking etc. The functions that are hooked into this cyclic list should be small timewise as otherwise the execution of the other code that relies on a high frequent polling (e.g. UART rx char ready check) might be delayed too much. This patch also adds the Kconfig option CONFIG_CYCLIC_MAX_CPU_TIME_US, which configures the max allowed time for such a cyclic function. If it's execution time exceeds this time, this cyclic function will get removed from the cyclic list.

Quoted 3 lines
Quoted 1 line

[...]

For the series, applied to u-boot/next, thanks!

Thanks Tom.

This patch series should not introduce any functional changes. The next one though, which migrates the WATCHDOG_RESET into the cyclic function execution IF needs a quite some testing. Perhaps it makes sense to pull this other series as well quite early in the release cycle, so that it receives a lot of "attention"? So that we have time to fix potential problems here.

What do you think?

Thanks,
Stefan

#11 lore raw
Quoted 19 lines · Stefan Roese
Quoted 12 lines · Tom Rini
Quoted 10 lines · Stefan Roese

This patchset adds the basic infrastructure to periodically execute code, e.g. all 100ms. Examples for such functions might be LED blinking etc. The functions that are hooked into this cyclic list should be small timewise as otherwise the execution of the other code that relies on a high frequent polling (e.g. UART rx char ready check) might be delayed too much. This patch also adds the Kconfig option CONFIG_CYCLIC_MAX_CPU_TIME_US, which configures the max allowed time for such a cyclic function. If it's execution time exceeds this time, this cyclic function will get removed from the cyclic list.

Quoted 3 lines
Quoted 1 line

[...]

For the series, applied to u-boot/next, thanks!

Thanks Tom.

Quoted 8 lines

This patch series should not introduce any functional changes. The next one though, which migrates the WATCHDOG_RESET into the cyclic function execution IF needs a quite some testing. Perhaps it makes sense to pull this other series as well quite early in the release cycle, so that it receives a lot of "attention"? So that we have time to fix potential problems here.

What do you think?

Yes, the watchdog series should come to next, for v2023.01 as well. Are you comfortable with it now, or are you planning to v2 it?

Signature · 1 line
Tom
#12 lore raw
Quoted 27 lines · Tom Rini
Quoted 26 lines · Stefan Roese
Quoted 12 lines · Tom Rini
Quoted 10 lines · Stefan Roese

This patchset adds the basic infrastructure to periodically execute code, e.g. all 100ms. Examples for such functions might be LED blinking etc. The functions that are hooked into this cyclic list should be small timewise as otherwise the execution of the other code that relies on a high frequent polling (e.g. UART rx char ready check) might be delayed too much. This patch also adds the Kconfig option CONFIG_CYCLIC_MAX_CPU_TIME_US, which configures the max allowed time for such a cyclic function. If it's execution time exceeds this time, this cyclic function will get removed from the cyclic list.

Quoted 3 lines
Quoted 1 line

[...]

For the series, applied to u-boot/next, thanks!

Thanks Tom.

This patch series should not introduce any functional changes. The next one though, which migrates the WATCHDOG_RESET into the cyclic function execution IF needs a quite some testing. Perhaps it makes sense to pull this other series as well quite early in the release cycle, so that it receives a lot of "attention"? So that we have time to fix potential problems here.

Quoted 4 lines
Quoted 1 line

What do you think?

Yes, the watchdog series should come to next, for v2023.01 as well. Are you comfortable with it now, or are you planning to v2 it?

I've only tested it on an MIPS Octeon platform so far. Where I am a bit hesitant is, the very early watchdog integration and SPL. This definitely needs some extensive testing. I don't have any platform for this here ready right now.

Thanks,
Stefan

#13 lore raw

Hi Stefan,

Quoted 31 lines · Stefan Roese
Quoted 29 lines · Tom Rini
Quoted 27 lines · Stefan Roese
Quoted 12 lines · Tom Rini
Quoted 10 lines · Stefan Roese

This patchset adds the basic infrastructure to periodically execute code, e.g. all 100ms. Examples for such functions might be LED blinking etc. The functions that are hooked into this cyclic list should be small timewise as otherwise the execution of the other code that relies on a high frequent polling (e.g. UART rx char ready check) might be delayed too much. This patch also adds the Kconfig option CONFIG_CYCLIC_MAX_CPU_TIME_US, which configures the max allowed time for such a cyclic function. If it's execution time exceeds this time, this cyclic function will get removed from the cyclic list.

Quoted 3 lines
Quoted 1 line

[...]

For the series, applied to u-boot/next, thanks!

Thanks Tom.

This patch series should not introduce any functional changes. The next one though, which migrates the WATCHDOG_RESET into the cyclic function execution IF needs a quite some testing. Perhaps it makes sense to pull this other series as well quite early in the release cycle, so that it receives a lot of "attention"? So that we have time to fix potential problems here.

What do you think?

Quoted 7 lines
Quoted 2 lines

Yes, the watchdog series should come to next, for v2023.01 as well. Are you comfortable with it now, or are you planning to v2 it?

I've only tested it on an MIPS Octeon platform so far. Where I am a bit hesitant is, the very early watchdog integration and SPL. This definitely needs some extensive testing. I don't have any platform for this here ready right now.

Once it is in -next, people will test it and we'll flush out any problems.

Regards,
Simon

#14 lore raw

Hi Simon,

Quoted 37 lines · Simon Glass

Hi Stefan,

Quoted 34 lines · Stefan Roese
Quoted 31 lines · Tom Rini
Quoted 19 lines · Stefan Roese
Quoted 12 lines · Tom Rini
Quoted 10 lines · Stefan Roese

This patchset adds the basic infrastructure to periodically execute code, e.g. all 100ms. Examples for such functions might be LED blinking etc. The functions that are hooked into this cyclic list should be small timewise as otherwise the execution of the other code that relies on a high frequent polling (e.g. UART rx char ready check) might be delayed too much. This patch also adds the Kconfig option CONFIG_CYCLIC_MAX_CPU_TIME_US, which configures the max allowed time for such a cyclic function. If it's execution time exceeds this time, this cyclic function will get removed from the cyclic list.

Quoted 3 lines
Quoted 1 line

[...]

For the series, applied to u-boot/next, thanks!

Thanks Tom.

Quoted 8 lines

This patch series should not introduce any functional changes. The next one though, which migrates the WATCHDOG_RESET into the cyclic function execution IF needs a quite some testing. Perhaps it makes sense to pull this other series as well quite early in the release cycle, so that it receives a lot of "attention"? So that we have time to fix potential problems here.

What do you think?

Yes, the watchdog series should come to next, for v2023.01 as well. Are you comfortable with it now, or are you planning to v2 it?

Quoted 6 lines
Quoted 4 lines

I've only tested it on an MIPS Octeon platform so far. Where I am a bit hesitant is, the very early watchdog integration and SPL. This definitely needs some extensive testing. I don't have any platform for this here ready right now.

Once it is in -next, people will test it and we'll flush out any problems.

Yes, agreed. This is the way. ;)

Thanks,
Stefan