zynqmp: add non-cacheable DMA window and hal_dma_set_noncached()

Author: dgarske

Makefile

@@ -727,6 +727,7 @@ $(LSCRIPT): $(LSCRIPT_IN) FORCE
 		sed -e "s/@WOLFBOOT_STAGE1_FLASH_ADDR@/$(WOLFBOOT_STAGE1_FLASH_ADDR)/g" | \
 		sed -e "s/@WOLFBOOT_STAGE1_BASE_ADDR@/$(WOLFBOOT_STAGE1_BASE_ADDR)/g" | \
 		sed -e "s/@WOLFBOOT_LOAD_BASE@/$(WOLFBOOT_LOAD_BASE)/g" | \
+		sed -e "s/@WOLFBOOT_DMA_BUFFER_ADDRESS@/$(WOLFBOOT_DMA_BUFFER_ADDRESS)/g" | \
 		sed -e "s/@BOOTLOADER_START@/$(BOOTLOADER_START)/g" | \
 		sed -e "s/@IMAGE_HEADER_SIZE@/$(IMAGE_HEADER_SIZE)/g" | \
 		sed -e "s/@WOLFBOOT_LOAD_ADDRESS@/$(WOLFBOOT_LOAD_ADDRESS)/g" | \

docs/HAL.md

@@ -146,6 +146,22 @@ implementation in new ports must return immediately without performing any actio
 if the content of the bootloader partition in the two banks already match.
 
 
+### Optional support for non-coherent DMA
+
+`int hal_dma_set_noncached(uintptr_t start, uintptr_t end)`
+
+Re-attribute `[start, end)` as non-cacheable. Only needed by ports that run with the MMU and D-cache enabled and that hand memory to a bus master which is not coherent with the CPU caches, such as an Ethernet MAC driven from a hook. Returns 0 on success, or negative if the port cannot satisfy the request.
+
+A weak default in `src/libwolfboot.c` returns an error, so the symbol always links. It deliberately does **not** succeed silently: a no-op would leave the caller sharing write-back memory with a non-coherent master, which is the failure this function exists to prevent.
+
+A port implementing it must:
+
+- Round the range **outward** to whatever granule its translation tables can express. On ZynqMP (`hal/zynq.c`) that is a 2MB block, so the caller has to give the region an aligned block of its own rather than placing it next to other data.
+- Clean and invalidate the affected range **before** changing the attribute, so a line still dirty at the switch cannot later be written back over what the bus master has since put there.
+- Push the modified table entries out and invalidate the TLB at the exception level that owns the translation.
+
+Callers place their buffers with a dedicated linker section. The ZynqMP port provides `.dma_buffers` in `hal/zynq.ld`, based at `WOLFBOOT_DMA_BUFFER_ADDRESS` (default `0x8200000`, clear of the kernel, the FIT staging area and the DTS) and exporting `_dma_buffers_start` / `_dma_buffers_end`.
+
 ### wolfHSM HAL extensions
 
 Refer to [wolfHSM.md](wolfHSM.md) for the wolfHSM-specific HAL functions and an overview of wolfHSM compatibility.

docs/Targets.md

@@ -4184,6 +4184,10 @@ Key configuration options:
 
 Opt-in (off by default), wolfBoot can replay a board's U-Boot Ethernet PHY register sequence over the GEM MDIO management plane so the PHY is ready before the OS runs. Enable with `CFLAGS_EXTRA+=-DWOLFBOOT_ZYNQMP_PHY_INIT`. The default targets the ZCU102 on-board PHY (TI DP83867 at MDIO `0x0C` on GEM3, `0xFF0E0000`) and just reads the PHY ID as a diagnostic (printed with `DEBUG_UART=1`). A board supplies its own sequence by keeping its values in a small header selected with one line, `CFLAGS_EXTRA+=-DZYNQMP_PHY_INIT_HEADER='"myboard_phy.h"'`, where that header `#define`s any of `ZYNQMP_GEM_BASE`, `ZYNQMP_PHY_ADDR`, `ZYNQMP_PHY_GPIO_ADDR`, `ZYNQMP_GEM_MDC_DIV`, and the `{op, arg0, arg1}` step array `ZYNQMP_PHY_INIT_STEPS`; scalars can also be set directly with `-D`, and anything omitted falls back to the ZCU102 defaults (see `hal/zynq.h` and the commented example in `config/examples/zynqmp.config`). Where the PHY is behind the PL, the boot image must include the FPGA bitstream (bootgen `[destination_device=pl] system.bit`) or the transactions are no-ops.
 
+### Non-cacheable DMA window
+
+The ZynqMP GEM is not coherent with the CPU caches, so a hook that drives it needs non-cacheable memory. `hal/zynq.ld` reserves a 2MB-aligned `.dma_buffers` region at `WOLFBOOT_DMA_BUFFER_ADDRESS` (default `0x8200000`) and `hal_dma_set_noncached()` re-attributes it at runtime; see [HAL.md](./HAL.md).
+
 ### Building with Xilinx tools (Vitis IDE)
 
 See [IDE/XilinxSDK/README.md](/IDE/XilinxSDK/README.md) for using Xilinx IDE

hal/zynq.c

@@ -3063,4 +3063,93 @@ void sdhci_platform_dma_complete(void *buf, uint32_t sz, int is_write)
 #endif /* DISK_SDCARD || DISK_EMMC */
 
 
+#if defined(MMU) && defined(__WOLFBOOT)
+/* wolfBoot maps DDR write-back in 2MB blocks from the static MMUTableL2. A
+ * non-coherent DMA master cannot share 8-byte descriptors through that:
+ * several fall in one cache line, so cleaning one clobbers its neighbours'
+ * ownership bits. Re-attribute the whole block Normal-NC instead.
+ *
+ * Valid block, AttrIndx=0 (MAIR[0] is Normal-NC), AF set, never executable.
+ * Shareability is ignored for Non-Cacheable memory. */
+#define ZYNQMP_L2_BLOCK_NORMAL_NC \
+    (0x401ULL | (1ULL << 53) | (1ULL << 54))
+
+
+/* From src/boot_aarch64_start.S: four contiguous 512-entry tables off
+ * L1[0..3] mapping 0x0-0xFFFFFFFF, so the index is addr >> 21. */
+extern uint64_t MMUTableL2[];
+
+/* Mark every 2MB block overlapping [start,end) Normal-NC. Cleans first: a
+ * line still dirty at the change could land on what the master since wrote. */
+/* The walker may not snoop, so push entries out before invalidating. */
+static void zynqmp_mmu_publish(uint64_t first, uint64_t last)
+{
+    uint64_t i;
+
+    __asm__ volatile("dsb ishst" : : : "memory");
+    for (i = first; i <= last; i++) {
+        __asm__ volatile("dc civac, %0"
+            : : "r"((uintptr_t)&MMUTableL2[i]) : "memory");
+    }
+    __asm__ volatile("dsb sy" : : : "memory");
+}
+
+/* wolfBoot runs at EL3 as an FSBL replacement, at EL2 under BL31. */
+static void zynqmp_mmu_tlbi(void)
+{
+    switch (current_el()) {
+        case 3:
+            __asm__ volatile("tlbi alle3" : : : "memory");
+            break;
+        case 2:
+            __asm__ volatile("tlbi alle2" : : : "memory");
+            break;
+        default:
+            __asm__ volatile("tlbi vmalle1" : : : "memory");
+            break;
+    }
+    __asm__ volatile("dsb sy" : : : "memory");
+    __asm__ volatile("isb" : : : "memory");
+}
+
+int hal_dma_set_noncached(uintptr_t start, uintptr_t end)
+{
+    uintptr_t addr;
+    uint64_t first, last, i;
+
+    if (zynqmp_l2_block_range((uint64_t)start, (uint64_t)end, &first, &last)
+            != 0) {
+        return -1;
+    }
+
+    /* Whole blocks: the attribute applies per block. */
+    for (addr = (uintptr_t)(first << ZYNQMP_L2_BLOCK_SHIFT);
+         addr < (uintptr_t)((last + 1) << ZYNQMP_L2_BLOCK_SHIFT);
+         addr += CACHE_LINE_SIZE) {
+        __asm__ volatile("dc civac, %0" : : "r"(addr) : "memory");
+    }
+    __asm__ volatile("dsb sy" : : : "memory");
+
+    /* Break-before-make: valid -> valid memory-type changes are CONSTRAINED
+     * UNPREDICTABLE on ARMv8-A. Nothing may touch the range while it is
+     * unmapped; the caller owns a dedicated region and wolfBoot's own code
+     * and data are in a different block. */
+    for (i = first; i <= last; i++) {
+        MMUTableL2[i] = 0;
+    }
+    zynqmp_mmu_publish(first, last);
+    zynqmp_mmu_tlbi();
+
+    for (i = first; i <= last; i++) {
+        MMUTableL2[i] = (i << ZYNQMP_L2_BLOCK_SHIFT)
+            | ZYNQMP_L2_BLOCK_NORMAL_NC;
+    }
+    zynqmp_mmu_publish(first, last);
+    zynqmp_mmu_tlbi();
+
+    return 0;
+}
+#endif /* MMU && __WOLFBOOT */
+
+
 #endif /* TARGET_zynq */

hal/zynq.h

@@ -689,6 +689,31 @@
 #define CRL_APB_DBG_LPD_CTRL (CRL_APB_BASE + 0x00B0U)
 #define CRL_APB_RST_LPD_DBG  (CRL_APB_BASE + 0x0240U)
 
-
+#ifndef __ASSEMBLER__
+#include <stdint.h>
+#include <stddef.h>
+
+/* 2MB: the smallest granule this translation table can re-attribute. */
+#define ZYNQMP_L2_BLOCK_SHIFT 21
+/* Four contiguous 512-entry tables covering 0x0-0xFFFFFFFF. */
+#define ZYNQMP_L2_ENTRIES     2048
+
+/* [start,end) to the inclusive 2MB block indices covering it. Split out so
+ * it can be unit tested: these indices decide which physical blocks get
+ * re-attributed. Returns 0, or -1 for an empty range or one past 4GB. */
+static inline int zynqmp_l2_block_range(uint64_t start, uint64_t end,
+    uint64_t* first, uint64_t* last)
+{
+    if (end <= start || first == NULL || last == NULL) {
+        return -1;
+    }
+    *first = start >> ZYNQMP_L2_BLOCK_SHIFT;
+    *last = (end - 1) >> ZYNQMP_L2_BLOCK_SHIFT;
+    if (*last >= ZYNQMP_L2_ENTRIES) {
+        return -1;
+    }
+    return 0;
+}
+#endif /* !__ASSEMBLER__ */
 
 #endif /* _ZYNQMP_H_ */

hal/zynq.ld

@@ -19,6 +19,11 @@ MEMORY
     *   Must match WOLFBOOT_ORIGIN in the target .config.
     */
    psu_ddr_0_MEM_0 : ORIGIN = 0x8000000, LENGTH = 0x200000
+   /* Non-cacheable DDR for bus-master DMA. Its own 2MB-aligned block: 2MB
+    * is the smallest granule the translation table can re-attribute, and it
+    * must stay clear of .text so the image hash is not run uncached.
+    * NOLOAD, so an unused region costs nothing. */
+   psu_ddr_dma_MEM_0 : ORIGIN = @WOLFBOOT_DMA_BUFFER_ADDRESS@, LENGTH = 0x200000
    psu_ddr_1_MEM_0 : ORIGIN = 0x800000000, LENGTH = 0x80000000
    psu_ocm_ram_0_MEM_0 : ORIGIN = 0xFFFC0000, LENGTH = 0x40000
    psu_qspi_linear_0_MEM_0 : ORIGIN = 0xC0000000, LENGTH = 0x20000000
@@ -293,6 +298,17 @@ _SDA_BASE_ = __sdata_start + ((__sbss_end - __sdata_start) / 2 );
 
 _SDA2_BASE_ = __sdata2_start + ((__sbss2_end - __sdata2_start) / 2 );
 
+/* Its own MEMORY region so hal_dma_set_noncached() can mark the enclosing
+ * 2MB block Normal-NC without touching wolfBoot's code or data. */
+.dma_buffers (NOLOAD) : {
+   . = ALIGN(64);
+   _dma_buffers_start = .;
+   *(.dma_buffers)
+   *(.dma_buffers.*)
+   . = ALIGN(64);
+   _dma_buffers_end = .;
+} > psu_ddr_dma_MEM_0
+
 /* Generate Stack and Heap definitions */
 
 .heap (NOLOAD) : {

include/hal.h

@@ -60,7 +60,8 @@ void hal_deinit();
 void hal_init(void);
 
 /* Timer functions (platform-specific, used for benchmarking) */
-#if defined(WOLFBOOT_UPDATE_DISK) || defined(BOOT_BENCHMARK)
+#if defined(WOLFBOOT_UPDATE_DISK) || defined(BOOT_BENCHMARK) || \
+    defined(PREBOOT_NETCHECK)
 uint64_t hal_get_timer_us(void);
 #endif
 
@@ -113,6 +114,12 @@ void hal_cache_invalidate(void);
 int hal_flash_protect(haladdr_t address, int len);
 void hal_prepare_boot(void);
 
+/* Re-attribute [start,end) non-cacheable, for memory shared with a
+ * non-coherent bus master. Returns 0, or negative if the port cannot; see
+ * docs/HAL.md. The weak default fails rather than doing nothing, so a caller
+ * never silently runs DMA through write-back memory. */
+int hal_dma_set_noncached(uintptr_t start, uintptr_t end);
+
 #ifdef DUALBANK_SWAP
     void hal_flash_dualbank_swap(void);
 #endif

options.mk

@@ -1683,6 +1683,11 @@ ifeq ($(WOLFHSM_SERVER),1)
 
 endif
 
+# Non-cacheable DDR carve-out for bus-master DMA (hal/zynq.ld). Substituted
+# into the linker script only; code uses the _dma_buffers_start/_end symbols
+# the script exports rather than this address.
+WOLFBOOT_DMA_BUFFER_ADDRESS?=0x8200000
+
 # wolfBoot hooks framework
 # WOLFBOOT_HOOKS_FILE: path to a single .c file containing hook definitions
 WOLFBOOT_HOOKS_ENABLED :=

src/libwolfboot.c

@@ -252,6 +252,16 @@ void WEAKFUNCTION hal_cache_invalidate(void)
     /* if cache flushing is required implement in hal */
 }
 
+/* Weak default; a port with the MMU and D-cache on overrides it. Fails
+ * rather than succeeding: a silent no-op would leave the caller sharing
+ * write-back memory with a non-coherent master. */
+int WEAKFUNCTION hal_dma_set_noncached(uintptr_t start, uintptr_t end)
+{
+    (void)start;
+    (void)end;
+    return -1;
+}
+
 #ifdef NVM_FLASH_WRITEONCE
 /* Some internal FLASH memory models don't allow
  * multiple writes after erase in the same page/area.

tools/unit-tests/Makefile

@@ -125,6 +125,7 @@ TESTS+=unit-rp2350-flash-write
 TESTS+=unit-fwtpm-rsp-overrun
 TESTS+=unit-fwtpm-cmd-toctou
 TESTS+=unit-fdt-memrsv-wrap
+TESTS+=unit-zynq-dma-range
 TESTS+=unit-pkcs11_store-stalecache
 TESTS+=unit-aurix-erased-fill
 TESTS+=unit-aurix-erased-fill-invert
@@ -473,6 +474,10 @@ unit-fwtpm-cmd-toctou: ../../include/target.h unit-fwtpm-cmd-toctou.c
 	gcc -o $@ $^ $(CFLAGS) -I$(WOLFBOOT_LIB_WOLFTPM) \
 		-DWOLFTPM_USER_SETTINGS $(LDFLAGS)
 
+# unit-zynq-dma-range: block-index math behind hal_dma_set_noncached()
+unit-zynq-dma-range: unit-zynq-dma-range.c
+	gcc -o $@ $^ $(CFLAGS) $(LDFLAGS)
+
 # unit-fdt-memrsv-wrap: layout validation in front of fdt_add_mem_rsv()
 # (F-11045). Links the real parser rather than extracting one function.
 unit-fdt-memrsv-wrap:CFLAGS+=-DWOLFBOOT_FDT