hal/arm: pmap_switch full MPU reconfiguration

This commit introduces full MPU regions reconfiguration on context switch,
allowing for more flexibile configuration of memory maps on MPU targets.
Performed tests show no memory coherence problems and minor improvements
in pmap_switch performance. According to ARM documentation, cache maintenance
is not required, as long as memory maps are not overlapping, and that
assumption is already present in Phoenix-RTOS.

Changes include
* additional hal_syspage_prog_t structure, initialized in loader, containing
  program configuration of MPU regions in form of ready-to-copy register values
* pmap_t structure contain pointer to above structure instead of regions bitmask
* pmap_switch disables MPU and performs full reconfiguration, optimized with
  LDMIA/STMIA assembly operations
* handling of process's kernel-code access is moved to loader
* hal/pmap.c functions update to the new pmap_t structure

JIRA: RTOS-1149

Author: etiaro

hal/aarch64/pmap.c

@@ -265,7 +265,7 @@ static void _pmap_cacheOpAfterChange(descr_t newEntry, ptr_t vaddr, unsigned int
 
 
 /* Function creates empty page table */
-int pmap_create(pmap_t *pmap, pmap_t *kpmap, page_t *p, void *vaddr)
+int pmap_create(pmap_t *pmap, pmap_t *kpmap, page_t *p, const syspage_prog_t *prog, void *vaddr)
 {
 	pmap->ttl1 = vaddr;
 	pmap->addr = p->addr;

hal/armv7a/pmap.c

@@ -175,7 +175,7 @@ static void _pmap_asidDealloc(pmap_t *pmap)
 
 
 /* Function creates empty page table */
-int pmap_create(pmap_t *pmap, pmap_t *kpmap, page_t *p, void *vaddr)
+int pmap_create(pmap_t *pmap, pmap_t *kpmap, page_t *p, const syspage_prog_t *prog, void *vaddr)
 {
 	pmap->pdir = vaddr;
 	pmap->addr = p->addr;

hal/armv7m/arch/pmap.h

@@ -17,6 +17,7 @@
 #define _PH_HAL_PMAP_ARMV7M_H_
 
 #include "hal/types.h"
+#include "syspage.h"
 
 /* Architecture dependent page attributes - used for mapping */
 #define PGHD_PRESENT    0x01
@@ -55,7 +56,7 @@ typedef struct _page_t {
 typedef struct _pmap_t {
 	void *start;
 	void *end;
-	u32 regions;
+	const hal_syspage_prog_t *hal;
 } pmap_t;
 
 #endif

hal/armv7m/pmap.c

@@ -17,9 +17,14 @@
 #include "config.h"
 #include "syspage.h"
 #include "halsyspage.h"
+#include "lib/lib.h"
 #include <arch/cpu.h>
 #include <arch/spinlock.h>
 
+
+#define MPU_BASE ((volatile u32 *)0xe000ed90)
+
+
 /* clang-format off */
 enum { mpu_type, mpu_ctrl, mpu_rnr, mpu_rbar, mpu_rasr, mpu_rbar_a1, mpu_rasr_a1, mpu_rbar_a2, mpu_rasr_a2,
 	mpu_rbar_a3, mpu_rasr_a3 };
@@ -35,13 +40,19 @@ static struct {
 	volatile u32 *mpu;
 	unsigned int kernelCodeRegion;
 	spinlock_t lock;
+	int last_mpu_count;
 } pmap_common;
 
 
 /* Function creates empty page table */
-int pmap_create(pmap_t *pmap, pmap_t *kpmap, page_t *p, void *vaddr)
+int pmap_create(pmap_t *pmap, pmap_t *kpmap, page_t *p, const syspage_prog_t *prog, void *vaddr)
 {
-	pmap->regions = pmap_common.kernelCodeRegion;
+	if (prog != NULL) {
+		pmap->hal = &prog->hal;
+	}
+	else {
+		pmap->hal = NULL;
+	}
 	return 0;
 }
 
@@ -52,55 +63,40 @@ addr_t pmap_destroy(pmap_t *pmap, int *i)
 }
 
 
-static unsigned int pmap_map2region(unsigned int map)
+void pmap_switch(pmap_t *pmap)
 {
-	int i;
-	unsigned int mask = 0;
-
-	for (i = 0; i < sizeof(syspage->hs.mpu.map) / sizeof(*syspage->hs.mpu.map); ++i) {
-		if (map == syspage->hs.mpu.map[i]) {
-			mask |= (1 << i);
-		}
-	}
-
-	return mask;
-}
-
+	static const volatile u32 *RBAR_ADDR = MPU_BASE + mpu_rbar;
+	unsigned int allocCnt;
+	spinlock_ctx_t sc;
+	unsigned int i;
+	const u32 *tableCurrent;
 
-int pmap_addMap(pmap_t *pmap, unsigned int map)
-{
-	unsigned int rmask = pmap_map2region(map);
-	if (rmask == 0) {
-		return -1;
-	}
+	if (pmap != NULL && pmap->hal != NULL) {
+		hal_spinlockSet(&pmap_common.lock, &sc);
 
-	pmap->regions |= rmask;
+		allocCnt = pmap->hal->mpu.allocCnt;
+		tableCurrent = &pmap->hal->mpu.table[0].rbar;
 
-	return 0;
-}
+		/* Disable MPU */
+		hal_cpuDataMemoryBarrier();
+		*(pmap_common.mpu + mpu_ctrl) &= ~1;
+
+		for (i = 0; i < max(allocCnt, pmap_common.last_mpu_count); i += 4) {
+			/* RNR update is done by writes to RBAR */
+			__asm__ volatile(
+					"ldmia %[tableCurrent]!, {r3-r8, r10, r11} \n\t" /* Load 4 regions (rbar/rasr pairs) from table, update table pointer */
+					"stmia %[mpu_rbar], {r3-r8, r10, r11}      \n\t" /* Write 4 regions via RBAR/RASR and aliases */
+					: [tableCurrent] "+&r"(tableCurrent)
+					: [mpu_rbar] "r"(RBAR_ADDR)
+					: "r3", "r4", "r5", "r6", "r7", "r8", "r10", "r11");
+		}
 
+		/* Enable MPU */
+		*(pmap_common.mpu + mpu_ctrl) |= 1;
+		hal_cpuDataSyncBarrier();
 
-void pmap_switch(pmap_t *pmap)
-{
-	unsigned int i, cnt = syspage->hs.mpu.allocCnt;
-	spinlock_ctx_t sc;
+		pmap_common.last_mpu_count = allocCnt;
 
-	if (pmap != NULL) {
-		hal_spinlockSet(&pmap_common.lock, &sc);
-		for (i = 0; i < cnt; ++i) {
-			/* Select region */
-			*(pmap_common.mpu + mpu_rnr) = i;
-			hal_cpuDataMemoryBarrier();
-
-			/* Enable/disable region according to the mask */
-			if ((pmap->regions & (1 << i)) != 0) {
-				*(pmap_common.mpu + mpu_rasr) |= 1;
-			}
-			else {
-				*(pmap_common.mpu + mpu_rasr) &= ~1;
-			}
-			hal_cpuDataMemoryBarrier();
-		}
 		hal_spinlockClear(&pmap_common.lock, &sc);
 	}
 }
@@ -127,13 +123,21 @@ addr_t pmap_resolve(pmap_t *pmap, void *vaddr)
 int pmap_isAllowed(pmap_t *pmap, const void *vaddr, size_t size)
 {
 	const syspage_map_t *map = syspage_mapAddrResolve((addr_t)vaddr);
-	unsigned int rmask;
 	if (map == NULL) {
 		return 0;
 	}
-	rmask = pmap_map2region(map->id);
 
-	return ((pmap->regions & rmask) == 0) ? 0 : 1;
+	if (pmap->hal == NULL) {
+		/* Kernel pmap has access to everything */
+		return 1;
+	}
+
+	for (int i = 0; i < pmap->hal->mpu.allocCnt; ++i) {
+		if (pmap->hal->mpu.map[i] == map->id) {
+			return 1;
+		}
+	}
+	return 0;
 }
 
 
@@ -170,11 +174,8 @@ int pmap_segment(unsigned int i, void **vaddr, size_t *size, vm_prot_t *prot, vo
 
 void _pmap_init(pmap_t *pmap, void **vstart, void **vend)
 {
-	const syspage_map_t *ikmap;
-	unsigned int ikregion;
-	u32 t;
-	addr_t pc;
-	unsigned int i, cnt = syspage->hs.mpu.allocCnt;
+	unsigned int cnt = (syspage->hs.mpuType >> 8U) & 0xffU;
+	unsigned int i;
 
 	(*vstart) = (void *)(((ptr_t)_init_vectors + 7) & ~7);
 	(*vend) = (*((char **)vstart)) + SIZE_PAGE;
@@ -184,8 +185,11 @@ void _pmap_init(pmap_t *pmap, void **vstart, void **vend)
 	/* Initial size of kernel map */
 	pmap->end = (void *)((addr_t)&__bss_start + 32 * 1024);
 
+	pmap->hal = NULL;
+	pmap_common.last_mpu_count = cnt;
+
 	/* Configure MPU */
-	pmap_common.mpu = (void *)0xe000ed90;
+	pmap_common.mpu = MPU_BASE;
 
 	/* Disable MPU just in case */
 	*(pmap_common.mpu + mpu_ctrl) &= ~1;
@@ -196,52 +200,17 @@ void _pmap_init(pmap_t *pmap, void **vstart, void **vend)
 	hal_cpuDataMemoryBarrier();
 
 	for (i = 0; i < cnt; ++i) {
-		t = syspage->hs.mpu.table[i].rbar;
-		if ((t & (1 << 4)) == 0) {
-			continue;
-		}
-
-		*(pmap_common.mpu + mpu_rbar) = t;
-		hal_cpuDataMemoryBarrier();
+		/* Select region */
+		*(pmap_common.mpu + mpu_rnr) = i;
 
-		/* Disable regions for now */
-		t = syspage->hs.mpu.table[i].rasr & ~1;
-		*(pmap_common.mpu + mpu_rasr) = t;
+		/* Disable all regions for now */
+		*(pmap_common.mpu + mpu_rasr) = 0;
 		hal_cpuDataMemoryBarrier();
 	}
 
 	/* Enable MPU */
 	*(pmap_common.mpu + mpu_ctrl) |= 1;
 	hal_cpuDataMemoryBarrier();
 
-	/* FIXME HACK
-	 * allow all programs to execute (and read) kernel code map.
-	 * Needed because of hal_jmp, syscalls handler and signals handler.
-	 * In these functions we need to switch to the user mode when still
-	 * executing kernel code. This will cause memory management fault
-	 * if the application does not have access to the kernel instruction
-	 * map. Possible fix - place return to the user code in the separate
-	 * region and allow this region instead. */
-
-	/* Find kernel code region */
-	__asm__ volatile("\tmov %0, pc;" : "=r"(pc));
-	ikmap = syspage_mapAddrResolve(pc);
-	if (ikmap == NULL) {
-		hal_consolePrint(ATTR_BOLD, "pmap: Kernel code map not found. Bad system config\n");
-		for (;;) {
-			hal_cpuHalt();
-		}
-	}
-
-	ikregion = pmap_map2region(ikmap->id);
-	if (ikregion == 0) {
-		hal_consolePrint(ATTR_BOLD, "pmap: Kernel code map has no assigned region. Bad system config\n");
-		for (;;) {
-			hal_cpuHalt();
-		}
-	}
-
-	pmap_common.kernelCodeRegion = ikregion;
-
 	hal_spinlockCreate(&pmap_common.lock, "pmap");
 }

hal/armv7r/arch/pmap.h

@@ -17,6 +17,7 @@
 #define _PH_HAL_PMAP_ARMV7R_H_
 
 #include "hal/types.h"
+#include "syspage.h"
 
 #define PGHD_PRESENT    0x01U
 #define PGHD_USER       0x04U
@@ -54,7 +55,7 @@ typedef struct _page_t {
 typedef struct _pmap_t {
 	void *start;
 	void *end;
-	u32 regions;
+	const hal_syspage_prog_t *hal;
 } pmap_t;
 
 #endif