bis: Pull latest lockpick driver and refactor it

- Refactor various variables and functions
- Flush whole cache when full
- Allow cache to be disabled
- Add support for raw emuMMC in nyx contenxt
- Use partition names for keys (to avoid issues with different ordering)
- Add deinit function that flushes the whole cache
- Change bis lookup address
- Halve cache size to 256MB in order to support 512MB ramdisk also.

Co-Authored-By: shchmue <7903403+shchmue@users.noreply.github.com>

bdk/memory_map.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2019 CTCaer
+ * Copyright (c) 2019-2021 CTCaer
  *
  * This program is free software; you can redistribute it and/or modify it
  * under the terms and conditions of the GNU General Public License,
@@ -50,6 +50,13 @@
 // Virtual disk / Chainloader buffers.
 #define RAM_DISK_ADDR 0xA4000000
 #define  RAM_DISK_SZ  0x41000000 // 1040MB.
+#define  RAM_DISK2_SZ 0x21000000 //  528MB.
+
+// NX BIS driver sector cache.
+#define NX_BIS_CACHE_ADDR  0xC5000000
+#define  NX_BIS_CACHE_SZ   0x10020000 // 256MB.
+#define NX_BIS_LOOKUP_ADDR 0xD6000000
+#define  NX_BIS_LOOKUP_SZ   0xF000000 // 240MB.
 
 // L4T Kernel Panic Storage (PSTORE).
 #define PSTORE_ADDR   0xB0000000
@@ -93,10 +100,6 @@
 /* ---   Hole: 129MB 0xF6A00000 - 0xFEB3FFFF --- */
 #define DRAM_START2       0xFEB40000
 
-// NX BIS driver sector cache.
-#define NX_BIS_CACHE_ADDR 0xFEE00000
-#define  NX_BIS_CACHE_SZ    0x100000
-
 // USB buffers.
 #define USBD_ADDR                 0xFEF00000
 #define USB_DESCRIPTOR_ADDR       0xFEF40000

nyx/nyx_gui/frontend/gui_info.c

@@ -413,8 +413,10 @@ t210b01:;
 	LIST_INIT(gpt);
 	nx_emmc_gpt_parse(&gpt, &emmc_storage);
 	emmc_part_t *cal0_part = nx_emmc_part_find(&gpt, "PRODINFO"); // check if null
-	nx_emmc_bis_init(cal0_part);
+	nx_emmc_bis_init(cal0_part, false, 0);
 	nx_emmc_bis_read(0, 0x40, cal0_buf);
+	nx_emmc_bis_end();
+	nx_emmc_gpt_free(&gpt);
 
 	// Clear BIS keys slots and reinstate SBK.
 	hos_bis_keys_clear();

nyx/nyx_gui/storage/nx_emmc_bis.c

@@ -1,8 +1,8 @@
 /*
  * eMMC BIS driver for Nintendo Switch
  *
- * Copyright (c) 2019 shchmue
+ * Copyright (c) 2019-2020 shchmue
- * Copyright (c) 2019-2020 CTCaer
+ * Copyright (c) 2019-2021 CTCaer
  *
  * This program is free software; you can redistribute it and/or modify it
  * under the terms and conditions of the GNU General Public License,
@@ -21,48 +21,64 @@
 
 #include <memory_map.h>
 
+#include <mem/heap.h>
 #include <sec/se.h>
+#include <sec/se_t210.h>
 #include "../storage/nx_emmc.h"
+#include <storage/nx_sd.h>
 #include <storage/sdmmc.h>
 #include <utils/types.h>
 
-#define MAX_SEC_CACHE_ENTRIES 1500
+#define BIS_CLUSTER_SECTORS   32
+#define BIS_CLUSTER_SIZE      16384
+#define BIS_CACHE_MAX_ENTRIES 16384
+#define BIS_CACHE_LOOKUP_TBL_EMPTY_ENTRY -1
 
-typedef struct _sector_cache_t
+typedef struct _cluster_cache_t
 {
-	u32 sector;
+	u32  cluster_idx;            // Index of the cluster in the partition.
-	u32 visit_cnt;
+	bool dirty;                  // Has been modified without write-back flag.
-	u8  tweak[0x10];
+	u8   data[BIS_CLUSTER_SIZE]; // The cached cluster itself. Aligned to 8 bytes for DMA engine.
-	u8  data[0x200];
+} cluster_cache_t;
-	u8  align[8];
+
-} sector_cache_t;
+typedef struct _bis_cache_t
+{
+	bool full;
+	bool enabled;
+	u32  dirty_cnt;
+	u32  top_idx;
+	u8   dma_buff[BIS_CLUSTER_SIZE]; // Aligned to 8 bytes for DMA engine.
+	cluster_cache_t clusters[];
+} bis_cache_t;
 
 static u8  ks_crypt = 0;
 static u8  ks_tweak = 0;
-static u32 sector_cache_cnt = 0;
+static u32 emu_offset = 0;
 static emmc_part_t *system_part = NULL;
-static sector_cache_t *sector_cache = (sector_cache_t *)NX_BIS_CACHE_ADDR;
+static u32 *cache_lookup_tbl = (u32 *)NX_BIS_LOOKUP_ADDR;
+static bis_cache_t *bis_cache = (bis_cache_t *)NX_BIS_CACHE_ADDR;
 
-static void _gf256_mul_x_le(u8 *block)
+static void _gf256_mul_x_le(void *block)
 {
-	u8 *pdata = (u8 *)block;
+	u32 *pdata = (u32 *)block;
 	u32 carry = 0;
 
-	for (u32 i = 0; i < 0x10; i++)
+	for (u32 i = 0; i < 4; i++)
 	{
-		u8 b = pdata[i];
+		u32 b = pdata[i];
 		pdata[i] = (b << 1) | carry;
-		carry = b >> 7;
+		carry = b >> 31;
 	}
 
 	if (carry)
 		pdata[0x0] ^= 0x87;
 }
 
-static int _nx_aes_xts_crypt_sec(u32 ks1, u32 ks2, u32 enc, u8 *tweak, bool regen_tweak, u32 tweak_exp, u64 sec, void *dst, void *src, u32 sec_size)
+static int _nx_aes_xts_crypt_sec(u32 tweak_ks, u32 crypt_ks, u32 enc, u8 *tweak, bool regen_tweak, u32 tweak_exp, u64 sec, void *dst, void *src, u32 sec_size)
 {
-	u8 *pdst = (u8 *)dst;
+	u32 *pdst = (u32 *)dst;
-	u8 *psrc = (u8 *)src;
+	u32 *psrc = (u32 *)src;
+	u32 *ptweak = (u32 *)tweak;
 
 	if (regen_tweak)
 	{
@@ -71,155 +87,298 @@ static int _nx_aes_xts_crypt_sec(u32 ks1, u32 ks2, u32 enc, u8 *tweak, bool rege
 			tweak[i] = sec & 0xFF;
 			sec >>= 8;
 		}
-		if (!se_aes_crypt_block_ecb(ks1, 1, tweak, tweak))
+		if (!se_aes_crypt_block_ecb(tweak_ks, 1, tweak, tweak))
 			return 0;
 	}
 
+	// tweak_exp allows us to use a saved tweak to reduce _gf256_mul_x_le calls.
 	for (u32 i = 0; i < (tweak_exp << 5); i++)
 		_gf256_mul_x_le(tweak);
 
-	u8 tmp_tweak[0x10];
+	u8 orig_tweak[SE_KEY_128_SIZE] __attribute__((aligned(4)));
-	memcpy(tmp_tweak, tweak, 0x10);
+	memcpy(orig_tweak, tweak, SE_KEY_128_SIZE);
 
-	// We are assuming a 0x10-aligned sector size in this implementation.
+	// We are assuming a 16 sector aligned size in this implementation.
 	for (u32 i = 0; i < (sec_size >> 4); i++)
 	{
-		for (u32 j = 0; j < 0x10; j++)
+		for (u32 j = 0; j < 4; j++)
-			pdst[j] = psrc[j] ^ tweak[j];
+			pdst[j] = psrc[j] ^ ptweak[j];
 
 		_gf256_mul_x_le(tweak);
-		psrc += 0x10;
+		psrc += 4;
-		pdst += 0x10;
+		pdst += 4;
 	}
 
-	se_aes_crypt_ecb(ks2, enc, dst, sec_size, src, sec_size);
+	if (!se_aes_crypt_ecb(crypt_ks, enc, dst, sec_size, dst, sec_size))
+		return 0;
 
-	memcpy(tweak, tmp_tweak, 0x10);
+	pdst = (u32 *)dst;
-
+	ptweak = (u32 *)orig_tweak;
-	pdst = (u8 *)dst;
 	for (u32 i = 0; i < (sec_size >> 4); i++)
 	{
-		for (u32 j = 0; j < 0x10; j++)
+		for (u32 j = 0; j < 4; j++)
-			pdst[j] = pdst[j] ^ tweak[j];
+			pdst[j] = pdst[j] ^ ptweak[j];
 
-		_gf256_mul_x_le(tweak);
+		_gf256_mul_x_le(orig_tweak);
-		pdst += 0x10;
+		pdst += 4;
 	}
 
 	return 1;
 }
 
+static int nx_emmc_bis_write_block(u32 sector, u32 count, void *buff, bool flush)
+{
+	if (!system_part)
+		return 3; // Not ready.
+
+	int res;
+	u8   tweak[SE_KEY_128_SIZE] __attribute__((aligned(4)));
+	u32  cluster = sector / BIS_CLUSTER_SECTORS;
+	u32  aligned_sector = cluster * BIS_CLUSTER_SECTORS;
+	u32  sector_in_cluster = sector % BIS_CLUSTER_SECTORS;
+	u32  lookup_idx = cache_lookup_tbl[cluster];
+	bool is_cached = lookup_idx != BIS_CACHE_LOOKUP_TBL_EMPTY_ENTRY;
+
+	// Write to cached cluster.
+	if (is_cached)
+	{
+		if (buff)
+			memcpy(bis_cache->clusters[lookup_idx].data + sector_in_cluster * NX_EMMC_BLOCKSIZE, buff, count * NX_EMMC_BLOCKSIZE);
+		else
+			buff = bis_cache->clusters[lookup_idx].data;
+		if (!bis_cache->clusters[lookup_idx].dirty)
+			bis_cache->dirty_cnt++;
+		bis_cache->clusters[lookup_idx].dirty = true;
+
+		if (!flush)
+			return 0; // Success.
+
+		// Reset args to trigger a full cluster flush to emmc.
+		sector_in_cluster = 0;
+		sector = aligned_sector;
+		count = BIS_CLUSTER_SECTORS;
+	}
+
+	// Encrypt cluster.
+	if (!_nx_aes_xts_crypt_sec(ks_tweak, ks_crypt, 1, tweak, true, sector_in_cluster, cluster, bis_cache->dma_buff, buff, count * NX_EMMC_BLOCKSIZE))
+		return 1; // Encryption error.
+
+	// If not reading from cache, do a regular read and decrypt.
+	if (!emu_offset)
+		res = nx_emmc_part_write(&emmc_storage, system_part, sector, count, bis_cache->dma_buff);
+	else
+		res = sdmmc_storage_read(&sd_storage, emu_offset + system_part->lba_start + sector, count, bis_cache->dma_buff);
+	if (!res)
+		return 1; // R/W error.
+
+	// Mark cache entry not dirty if write succeeds.
+	if (is_cached)
+	{
+		bis_cache->clusters[lookup_idx].dirty = false;
+		bis_cache->dirty_cnt--;
+	}
+
+	return 0; // Success.
+}
+
+static void _nx_emmc_bis_cluster_cache_init(bool enable_cache)
+{
+	u32 cache_lookup_tbl_size = (system_part->lba_end - system_part->lba_start + 1) / BIS_CLUSTER_SECTORS * sizeof(*cache_lookup_tbl);
+
+	// Clear cache header.
+	memset(bis_cache, 0, sizeof(bis_cache_t));
+
+	// Clear cluster lookup table.
+	memset(cache_lookup_tbl, BIS_CACHE_LOOKUP_TBL_EMPTY_ENTRY, cache_lookup_tbl_size);
+
+	// Enable cache.
+	bis_cache->enabled = enable_cache;
+}
+
+static void _nx_emmc_bis_flush_cache()
+{
+	if (!bis_cache->enabled || !bis_cache->dirty_cnt)
+		return;
+
+	for (u32 i = 0; i < bis_cache->top_idx && bis_cache->dirty_cnt; i++)
+	{
+		if (bis_cache->clusters[i].dirty) {
+			nx_emmc_bis_write_block(bis_cache->clusters[i].cluster_idx * BIS_CLUSTER_SECTORS, BIS_CLUSTER_SECTORS, NULL, true);
+			bis_cache->dirty_cnt--;
+		}
+	}
+
+	_nx_emmc_bis_cluster_cache_init(true);
+}
+
+static int nx_emmc_bis_read_block_normal(u32 sector, u32 count, void *buff)
+{
+	static u32 prev_cluster = -1;
+	static u32 prev_sector = 0;
+	static u8  tweak[SE_KEY_128_SIZE] __attribute__((aligned(4)));
+
+	int  res;
+	bool regen_tweak = true;
+	u32  tweak_exp = 0;
+	u32  cluster = sector / BIS_CLUSTER_SECTORS;
+	u32  sector_in_cluster = sector % BIS_CLUSTER_SECTORS;
+
+	// If not reading from cache, do a regular read and decrypt.
+	if (!emu_offset)
+		res = nx_emmc_part_read(&emmc_storage, system_part, sector, count, bis_cache->dma_buff);
+	else
+		res = sdmmc_storage_read(&sd_storage, emu_offset + system_part->lba_start + sector, count, bis_cache->dma_buff);
+	if (!res)
+		return 1; // R/W error.
+
+	if (prev_cluster != cluster) // Sector in different cluster than last read.
+	{
+		prev_cluster = cluster;
+		tweak_exp = sector_in_cluster;
+	}
+	else if (sector > prev_sector) // Sector in same cluster and past last sector.
+	{
+		// Calculates the new tweak using the saved one, reducing expensive _gf256_mul_x_le calls.
+		tweak_exp = sector - prev_sector - 1;
+		regen_tweak = false;
+	}
+	else // Sector in same cluster and before or same as last sector.
+		tweak_exp = sector_in_cluster;
+
+	// Maximum one cluster (1 XTS crypto block 16KB).
+	if (!_nx_aes_xts_crypt_sec(ks_tweak, ks_crypt, 0, tweak, regen_tweak, tweak_exp, prev_cluster, buff, bis_cache->dma_buff, count * NX_EMMC_BLOCKSIZE))
+		return 1; // R/W error.
+
+	prev_sector = sector + count - 1;
+
+	return 0; // Success.
+}
+
+static int nx_emmc_bis_read_block_cached(u32 sector, u32 count, void *buff)
+{
+	int res;
+	u8  cache_tweak[SE_KEY_128_SIZE] __attribute__((aligned(4)));
+	u32 cluster = sector / BIS_CLUSTER_SECTORS;
+	u32 cluster_sector = cluster * BIS_CLUSTER_SECTORS;
+	u32 sector_in_cluster = sector % BIS_CLUSTER_SECTORS;
+	u32 lookup_idx = cache_lookup_tbl[cluster];
+
+	// Read from cached cluster.
+	if (lookup_idx != BIS_CACHE_LOOKUP_TBL_EMPTY_ENTRY)
+	{
+		memcpy(buff, bis_cache->clusters[lookup_idx].data + sector_in_cluster * NX_EMMC_BLOCKSIZE, count * NX_EMMC_BLOCKSIZE);
+
+		return 0; // Success.
+	}
+
+	// Flush cache if full.
+	if (bis_cache->top_idx >= BIS_CACHE_MAX_ENTRIES)
+		_nx_emmc_bis_flush_cache();
+
+	// Set new cached cluster parameters.
+	bis_cache->clusters[bis_cache->top_idx].cluster_idx = cluster;
+	bis_cache->clusters[bis_cache->top_idx].dirty = false;
+	cache_lookup_tbl[cluster] = bis_cache->top_idx;
+
+	// Read the whole cluster the sector resides in.
+	if (!emu_offset)
+		res = nx_emmc_part_read(&emmc_storage, system_part, cluster_sector, BIS_CLUSTER_SECTORS, bis_cache->dma_buff);
+	else
+		res = sdmmc_storage_read(&sd_storage, emu_offset + system_part->lba_start + cluster_sector, BIS_CLUSTER_SECTORS, bis_cache->dma_buff);
+	if (!res)
+		return 1; // R/W error.
+
+	// Decrypt cluster.
+	if (!_nx_aes_xts_crypt_sec(ks_tweak, ks_crypt, 0, cache_tweak, true, 0, cluster, bis_cache->dma_buff, bis_cache->dma_buff, BIS_CLUSTER_SIZE))
+		return 1; // Decryption error.
+
+	// Copy to cluster cache.
+	memcpy(bis_cache->clusters[bis_cache->top_idx].data, bis_cache->dma_buff, BIS_CLUSTER_SIZE);
+	memcpy(buff, bis_cache->dma_buff + sector_in_cluster * NX_EMMC_BLOCKSIZE, count * NX_EMMC_BLOCKSIZE);
+
+	// Increment cache count.
+	bis_cache->top_idx++;
+
+	return 0; // Success.
+}
+
 static int nx_emmc_bis_read_block(u32 sector, u32 count, void *buff)
 {
 	if (!system_part)
 		return 3; // Not ready.
 
-	static u32 prev_cluster = -1;
+	if (bis_cache->enabled)
-	static u32 prev_sector = 0;
+		return nx_emmc_bis_read_block_cached(sector, count, buff);
-	static u8 tweak[0x10];
+	else
-
+		return nx_emmc_bis_read_block_normal(sector, count, buff);
-	u32 cache_idx = 0;
-	u32 tweak_exp = 0;
-	bool regen_tweak = true;
-	bool cache_sector = false;
-
-	if (count == 1)
-	{
-		for ( ; cache_idx < sector_cache_cnt; cache_idx++)
-		{
-			if (sector_cache[cache_idx].sector == sector)
-			{
-				sector_cache[cache_idx].visit_cnt++;
-				memcpy(buff, sector_cache[cache_idx].data, 0x200);
-				memcpy(tweak, sector_cache[cache_idx].tweak, 0x10);
-				prev_sector = sector;
-				prev_cluster = sector >> 5;
-
-				return 0;
-			}
-		}
-		// add to cache
-		if (cache_idx == sector_cache_cnt && cache_idx < MAX_SEC_CACHE_ENTRIES)
-		{
-			sector_cache[cache_idx].sector = sector;
-			sector_cache[cache_idx].visit_cnt++;
-			cache_sector = true;
-			sector_cache_cnt++;
-		}
-	}
-
-	if (nx_emmc_part_read(&emmc_storage, system_part, sector, count, buff))
-	{
-		if (prev_cluster != sector >> 5) // Sector in different cluster than last read.
-		{
-			prev_cluster = sector >> 5;
-			tweak_exp = sector % 0x20;
-		}
-		else if (sector > prev_sector) // Sector in same cluster and past last sector.
-		{
-			tweak_exp = sector - prev_sector - 1;
-			regen_tweak = false;
-		}
-		else // Sector in same cluster and before or same as last sector.
-			tweak_exp = sector % 0x20;
-
-		// Maximum one cluster (1 XTS crypto block 16KB).
-		_nx_aes_xts_crypt_sec(ks_tweak, ks_crypt, 0, tweak, regen_tweak, tweak_exp, prev_cluster, buff, buff, count << 9);
-		if (cache_sector)
-		{
-			memcpy(sector_cache[cache_idx].data, buff, 0x200);
-			memcpy(sector_cache[cache_idx].tweak, tweak, 0x10);
-		}
-		prev_sector = sector + count - 1;
-
-		return 0;
-	}
-
-	// Error occurred.
-	return 1;
 }
 
 int nx_emmc_bis_read(u32 sector, u32 count, void *buff)
 {
-	int res = 1;
 	u8 *buf = (u8 *)buff;
 	u32 curr_sct = sector;
 
 	while (count)
 	{
-		u32 sct_cnt = MIN(count, 0x20);
+		u32 sct_cnt = MIN(count, BIS_CLUSTER_SECTORS);
-		res = nx_emmc_bis_read_block(curr_sct, sct_cnt, buf);
+		if (nx_emmc_bis_read_block(curr_sct, sct_cnt, buf))
-		if (res)
+			return 0;
-			return 1;
 
 		count    -= sct_cnt;
 		curr_sct += sct_cnt;
-		buf += 512 * sct_cnt;
+		buf      += sct_cnt * NX_EMMC_BLOCKSIZE;
 	}
 
-	return res;
+	return 1;
 }
 
-void nx_emmc_bis_init(emmc_part_t *part)
+int nx_emmc_bis_write(u32 sector, u32 count, void *buff)
+{
+	u8 *buf = (u8 *)buff;
+	u32 curr_sct = sector;
+
+	while (count)
+	{
+		u32 sct_cnt = MIN(count, BIS_CLUSTER_SECTORS);
+		if (nx_emmc_bis_write_block(curr_sct, sct_cnt, buf, false))
+			return 0;
+
+		count    -= sct_cnt;
+		curr_sct += sct_cnt;
+		buf      += sct_cnt * NX_EMMC_BLOCKSIZE;
+	}
+
+	return 1;
+}
+
+void nx_emmc_bis_init(emmc_part_t *part, bool enable_cache, u32 emummc_offset)
 {
 	system_part = part;
-	sector_cache_cnt = 0;
+	emu_offset = emummc_offset;
 
-	switch (part->index)
+	_nx_emmc_bis_cluster_cache_init(enable_cache);
+
+	if (!strcmp(part->name, "PRODINFO") || !strcmp(part->name, "PRODINFOF"))
 	{
-	case 0:  // PRODINFO.
-	case 1:  // PRODINFOF.
 		ks_crypt = 0;
 		ks_tweak = 1;
-		break;
+	}
-	case 8:  // SAFE.
+	else if (!strcmp(part->name, "SAFE"))
+	{
 		ks_crypt = 2;
 		ks_tweak = 3;
-		break;
+	}
-	case 9:  // SYSTEM.
+	else if (!strcmp(part->name, "SYSTEM") || !strcmp(part->name, "USER"))
-	case 10: // USER.
+	{
 		ks_crypt = 4;
 		ks_tweak = 5;
-		break;
 	}
+	else
+		system_part = NULL;
+}
+
+void nx_emmc_bis_end()
+{
+	_nx_emmc_bis_flush_cache();
+	system_part = NULL;
 }

nyx/nyx_gui/storage/nx_emmc_bis.h

@@ -224,6 +224,7 @@ typedef struct _nx_emmc_cal0_t
 } __attribute__((packed)) nx_emmc_cal0_t;
 
 int  nx_emmc_bis_read(u32 sector, u32 count, void *buff);
-void nx_emmc_bis_init(emmc_part_t *part);
+void nx_emmc_bis_init(emmc_part_t *part, bool enable_cache, u32 emummc_offset);
+void nx_emmc_bis_end();
 
 #endif