From 8d0f0d7f869faf0aab3e521c227db81fa191cd75 Mon Sep 17 00:00:00 2001
From: Andrew Lewycky <alewycky@tenstorrent.com>
Date: Wed, 6 Dec 2023 15:05:35 -0500
Subject: [PATCH 01/10] Change version to 1.30-bh

---
 README.md | 5 ++---
 dkms.conf | 2 +-
 module.c  | 2 +-
 3 files changed, 4 insertions(+), 5 deletions(-)

diff --git a/README.md b/README.md
index 83afee6..04ea22a 100644
--- a/README.md
+++ b/README.md
@@ -18,7 +18,7 @@ The driver registers device files named `/dev/tenstorrent/%d`, one for each enum
     * `dnf install epel-release && dnf install dkms` (Enterprise Linux based)
 ```
 sudo dkms add .
-sudo dkms install tenstorrent/1.30
+sudo dkms install tenstorrent/1.30-bh
 sudo modprobe tenstorrent
 ```
 (or reboot, driver will auto-load next boot)
@@ -26,6 +26,5 @@ sudo modprobe tenstorrent
 ### To uninstall:
 ```
 sudo modprobe -r tenstorrent
-sudo dkms remove tenstorrent/1.30 --all
+sudo dkms remove tenstorrent/1.30-bh --all
 ```
-
diff --git a/dkms.conf b/dkms.conf
index 73616bc..635f506 100644
--- a/dkms.conf
+++ b/dkms.conf
@@ -1,5 +1,5 @@
 PACKAGE_NAME="tenstorrent"
-PACKAGE_VERSION="1.30"
+PACKAGE_VERSION="1.30-bh"
 BUILT_MODULE_NAME="tenstorrent"
 
 DEST_MODULE_LOCATION="/kernel/extra"
diff --git a/module.c b/module.c
index 4ffba29..bb66383 100644
--- a/module.c
+++ b/module.c
@@ -9,7 +9,7 @@
 #include "chardev.h"
 #include "enumerate.h"
 
-#define TTDRIVER_VER "1.30"
+#define TTDRIVER_VER "1.30-bh"
 
 MODULE_LICENSE("GPL");
 MODULE_DESCRIPTION("Tenstorrent AI kernel driver");

From fd1194267d8336d604f070b7b1812cb1467578aa Mon Sep 17 00:00:00 2001
From: Andrew Lewycky <alewycky@tenstorrent.com>
Date: Wed, 6 Dec 2023 15:03:53 -0500
Subject: [PATCH 02/10] Minimal Black Hole support

No HW programming or ARC messages.
---
 Makefile    |  2 +-
 README.md   |  1 +
 blackhole.c | 37 +++++++++++++++++++++++++++++++++++++
 blackhole.h | 17 +++++++++++++++++
 enumerate.h |  1 +
 module.c    |  3 +++
 6 files changed, 60 insertions(+), 1 deletion(-)
 create mode 100644 blackhole.c
 create mode 100644 blackhole.h

diff --git a/Makefile b/Makefile
index 29993f9..bf48fa0 100644
--- a/Makefile
+++ b/Makefile
@@ -2,7 +2,7 @@
 # SPDX-License-Identifier: GPL-2.0-only
 
 obj-m += tenstorrent.o
-tenstorrent-y := module.o chardev.o enumerate.o interrupt.o grayskull.o wormhole.o pcie.o hwmon.o sg_helpers.o memory.o
+tenstorrent-y := module.o chardev.o enumerate.o interrupt.o grayskull.o wormhole.o blackhole.o pcie.o hwmon.o sg_helpers.o memory.o
 
 KDIR := /lib/modules/$(shell uname -r)/build
 KMAKE := $(MAKE) -C $(KDIR) M=$(CURDIR)
diff --git a/README.md b/README.md
index 04ea22a..999d68f 100644
--- a/README.md
+++ b/README.md
@@ -7,6 +7,7 @@
 ## Supported hardware:
 * Grayskull
 * Wormhole
+* Blackhole
 
 The driver registers device files named `/dev/tenstorrent/%d`, one for each enumerated device.
 
diff --git a/blackhole.c b/blackhole.c
new file mode 100644
index 0000000..f70c65a
--- /dev/null
+++ b/blackhole.c
@@ -0,0 +1,37 @@
+// SPDX-FileCopyrightText: © 2024 Tenstorrent Inc.
+// SPDX-License-Identifier: GPL-2.0-only
+
+#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
+#include <linux/types.h>
+
+#include "blackhole.h"
+#include "pcie.h"
+#include "module.h"
+
+static bool blackhole_init(struct tenstorrent_device *tt_dev) {
+	return true;
+}
+
+static bool blackhole_init_hardware(struct tenstorrent_device *tt_dev) {
+	return true;
+}
+
+static bool blackhole_post_hardware_init(struct tenstorrent_device *tt_dev) {
+	return true;
+}
+
+static void blackhole_cleanup_hardware(struct tenstorrent_device *tt_dev) {
+}
+
+static void blackhole_cleanup(struct tenstorrent_device *tt_dev) {
+}
+
+struct tenstorrent_device_class blackhole_class = {
+	.name = "Blackhole",
+	.instance_size = sizeof(struct blackhole_device),
+	.init_device = blackhole_init,
+	.init_hardware = blackhole_init_hardware,
+	.post_hardware_init = blackhole_post_hardware_init,
+	.cleanup_hardware = blackhole_cleanup_hardware,
+	.cleanup_device = blackhole_cleanup,
+};
diff --git a/blackhole.h b/blackhole.h
new file mode 100644
index 0000000..2c9d561
--- /dev/null
+++ b/blackhole.h
@@ -0,0 +1,17 @@
+// SPDX-FileCopyrightText: © 2023 Tenstorrent Inc.
+// SPDX-License-Identifier: GPL-2.0-only
+
+#ifndef TTDRIVER_BLACKHOLE_H_INCLUDED
+#define TTDRIVER_BLACKHOLE_H_INCLUDED
+
+#include <linux/types.h>
+#include "device.h"
+
+struct blackhole_device {
+	struct tenstorrent_device tt;
+};
+
+#define tt_dev_to_wh_dev(ttdev) \
+	container_of((tt_dev), struct blackhole_device, tt)
+
+#endif
diff --git a/enumerate.h b/enumerate.h
index 0d58dc9..d485531 100644
--- a/enumerate.h
+++ b/enumerate.h
@@ -14,6 +14,7 @@
 #define PCI_VENDOR_ID_TENSTORRENT 0x1E52
 #define PCI_DEVICE_ID_GRAYSKULL	0xFACA
 #define PCI_DEVICE_ID_WORMHOLE	0x401E
+#define PCI_DEVICE_ID_BLACKHOLE	0xB140
 
 struct pci_dev;
 struct cdev;
diff --git a/module.c b/module.c
index bb66383..eed4f4e 100644
--- a/module.c
+++ b/module.c
@@ -94,12 +94,15 @@ MODULE_PARM_DESC(auto_reset_timeout, "Timeout duration in seconds for M3 auto re
 struct tenstorrent_device_class;
 extern struct tenstorrent_device_class grayskull_class;
 extern struct tenstorrent_device_class wormhole_class;
+extern struct tenstorrent_device_class blackhole_class;
 
 const struct pci_device_id tenstorrent_ids[] = {
 	{ PCI_DEVICE(PCI_VENDOR_ID_TENSTORRENT, PCI_DEVICE_ID_GRAYSKULL),
 	  .driver_data=(kernel_ulong_t)&grayskull_class },
 	{ PCI_DEVICE(PCI_VENDOR_ID_TENSTORRENT, PCI_DEVICE_ID_WORMHOLE),
 	  .driver_data=(kernel_ulong_t)&wormhole_class },
+	{ PCI_DEVICE(PCI_VENDOR_ID_TENSTORRENT, PCI_DEVICE_ID_BLACKHOLE),
+	  .driver_data=(kernel_ulong_t)&blackhole_class },
 	{ 0 },
 };
 

From 59804d53b23cecff282fbf5a1753b62528bb1672 Mon Sep 17 00:00:00 2001
From: Andrew Lewycky <alewycky@tenstorrent.com>
Date: Thu, 23 May 2024 11:51:22 -0400
Subject: [PATCH 03/10] Directly test for mapping regions so large that they
 wrap around

VerifyNoOverlap will catch this too, but this is more obvious.
---
 test/query_mappings.cpp | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

diff --git a/test/query_mappings.cpp b/test/query_mappings.cpp
index df95ca4..bb30294 100644
--- a/test/query_mappings.cpp
+++ b/test/query_mappings.cpp
@@ -1,4 +1,4 @@
-// SPDX-FileCopyrightText: © 2023 Tenstorrent Inc.
+// SPDX-FileCopyrightText: © 2024 Tenstorrent Inc.
 // SPDX-License-Identifier: GPL-2.0-only
 
 // Verify that all resource IDs are known to us.
@@ -143,6 +143,10 @@ void VerifySizes(const std::vector<tenstorrent_mapping> &mappings)
     if (std::any_of(mappings.begin(), mappings.end(),
                     [=](const auto &m) { return m.mapping_id != TENSTORRENT_MAPPING_UNUSED && m.mapping_base % pagesize != 0; }))
         THROW_TEST_FAILURE("Mapping base is not a multiple of page size in QUERY_MAPPINGS results.");
+
+    if (std::any_of(mappings.begin(), mappings.end(),
+                    [](const auto &m) { return m.mapping_size > std::numeric_limits<std::uint64_t>::max() - m.mapping_base; }))
+        THROW_TEST_FAILURE("Mapping region wraps around.");
 }
 
 void PrintMappings(const std::vector<tenstorrent_mapping>& mappings)

From 199831c4b3d2bce18fea8bf43f9419a92e3883f0 Mon Sep 17 00:00:00 2001
From: Andrew Lewycky <alewycky@tenstorrent.com>
Date: Thu, 23 May 2024 12:05:52 -0400
Subject: [PATCH 04/10] Adjust MMAP_OFFSET_DMA_BUF to fit all DMA buffers in
 32-bit mmap

---
 memory.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/memory.c b/memory.c
index c3c7719..391eccf 100644
--- a/memory.c
+++ b/memory.c
@@ -132,7 +132,7 @@ static void unpin_user_pages_dirty_lock(struct page **pages, unsigned long npage
 // tenstorrent_allocate_dma_buf_in.buf_index is u8 so that sets a limit of
 // U8_MAX DMA buffers per fd. 32-bit mmap offsets are divided by PAGE_SIZE,
 // so PAGE_SIZE << 32 is the largest possible offset.
-#define MMAP_OFFSET_DMA_BUF		((u64)(PAGE_SIZE-U8_MAX) << 32)
+#define MMAP_OFFSET_DMA_BUF		((u64)(PAGE_SIZE-U8_MAX-1) << 32)
 
 #define MMAP_SIZE_DMA_BUF (U64_C(1) << 32)
 

From f2c88fd68867e1fcd78ab515b6161b109b120f61 Mon Sep 17 00:00:00 2001
From: Andrew Lewycky <alewycky@tenstorrent.com>
Date: Thu, 23 May 2024 12:12:11 -0400
Subject: [PATCH 05/10] Test for 32-bit mmap offset limit

---
 test/query_mappings.cpp | 7 +++++++
 1 file changed, 7 insertions(+)

diff --git a/test/query_mappings.cpp b/test/query_mappings.cpp
index bb30294..0721485 100644
--- a/test/query_mappings.cpp
+++ b/test/query_mappings.cpp
@@ -128,6 +128,7 @@ void VerifyNoOverlap(const std::vector<tenstorrent_mapping> &mappings)
 }
 
 // Verify that size > 0. Verify that base & size are multiples of the page size.
+// Verify that the size is not too large and that mapping_base is not too high.
 void VerifySizes(const std::vector<tenstorrent_mapping> &mappings)
 {
     if (std::any_of(mappings.begin(), mappings.end(),
@@ -147,6 +148,12 @@ void VerifySizes(const std::vector<tenstorrent_mapping> &mappings)
     if (std::any_of(mappings.begin(), mappings.end(),
                     [](const auto &m) { return m.mapping_size > std::numeric_limits<std::uint64_t>::max() - m.mapping_base; }))
         THROW_TEST_FAILURE("Mapping region wraps around.");
+
+    std::uint64_t mmap_offset_limit_for_32b = (std::uint64_t)1 << 44; // 32 + log(PAGE_SIZE)
+
+    if (std::any_of(mappings.begin(), mappings.end(),
+                    [=](const auto &m) { return m.mapping_size + m.mapping_base >= mmap_offset_limit_for_32b; }))
+        THROW_TEST_FAILURE("Mapping base/size do not fit into 32-bit mmap offset.");
 }
 
 void PrintMappings(const std::vector<tenstorrent_mapping>& mappings)

From 552e1fc07a5925129a1af766dbfcc0c410c9508e Mon Sep 17 00:00:00 2001
From: Andrew Lewycky <alewycky@tenstorrent.com>
Date: Mon, 27 May 2024 12:25:39 -0400
Subject: [PATCH 06/10] Adjust mmap offsets to allow for 32GB between each

This allows for a single resource mapping to be 32GB. These values are reported
through query mapping so the ABI has not changed.
---
 memory.c | 12 ++++++------
 1 file changed, 6 insertions(+), 6 deletions(-)

diff --git a/memory.c b/memory.c
index 391eccf..a9d7c22 100644
--- a/memory.c
+++ b/memory.c
@@ -122,12 +122,12 @@ static void unpin_user_pages_dirty_lock(struct page **pages, unsigned long npage
 // These are the mmap offsets for various resources. In the user-kernel
 // interface they are dynamic (TENSTORRENT_IOCTL_QUERY_MAPPINGS and
 // TENSTORRENT_IOCTL_ALLOCATE_DMA_BUF), but they are actually hard-coded.
-#define MMAP_OFFSET_RESOURCE0_UC	(U64_C(0) << 32)
-#define MMAP_OFFSET_RESOURCE0_WC	(U64_C(1) << 32)
-#define MMAP_OFFSET_RESOURCE1_UC	(U64_C(2) << 32)
-#define MMAP_OFFSET_RESOURCE1_WC	(U64_C(3) << 32)
-#define MMAP_OFFSET_RESOURCE2_UC	(U64_C(4) << 32)
-#define MMAP_OFFSET_RESOURCE2_WC	(U64_C(5) << 32)
+#define MMAP_OFFSET_RESOURCE0_UC	(U64_C(0) << 35)
+#define MMAP_OFFSET_RESOURCE0_WC	(U64_C(1) << 35)
+#define MMAP_OFFSET_RESOURCE1_UC	(U64_C(2) << 35)
+#define MMAP_OFFSET_RESOURCE1_WC	(U64_C(3) << 35)
+#define MMAP_OFFSET_RESOURCE2_UC	(U64_C(4) << 35)
+#define MMAP_OFFSET_RESOURCE2_WC	(U64_C(5) << 35)
 
 // tenstorrent_allocate_dma_buf_in.buf_index is u8 so that sets a limit of
 // U8_MAX DMA buffers per fd. 32-bit mmap offsets are divided by PAGE_SIZE,

From 9e58aed95297556a88c6bb2ed14e2f6c73e36566 Mon Sep 17 00:00:00 2001
From: nxu <nxu@tenstorrent.com>
Date: Wed, 10 Jul 2024 16:01:19 +0000
Subject: [PATCH 07/10] added pcie timer interface interrupt sequence

---
 chardev.c |  4 ++--
 ioctl.h   |  1 +
 pcie.c    | 13 +++++++++++++
 pcie.h    |  1 +
 4 files changed, 17 insertions(+), 2 deletions(-)

diff --git a/chardev.c b/chardev.c
index 1329fae..cd7e860 100644
--- a/chardev.c
+++ b/chardev.c
@@ -186,8 +186,8 @@ static long ioctl_reset_device(struct chardev_private *priv,
 		else
 			ok = false;
 
-	} else if (in.flags == TENSTORRENT_RESET_DEVICE_RESET_PCIE_LINK) {
-		ok = pcie_hot_reset_and_restore_state(pdev);
+	} else if (in.flags == TENSTORRENT_RESET_DEVICE_CONFIG_WRITE) {
+		ok = pcie_timer_interrupt(pdev);
 	} else {
 		return -EINVAL;
 	}
diff --git a/ioctl.h b/ioctl.h
index e33890f..d56e1f6 100644
--- a/ioctl.h
+++ b/ioctl.h
@@ -125,6 +125,7 @@ struct tenstorrent_get_driver_info {
 // tenstorrent_reset_device_in.flags
 #define TENSTORRENT_RESET_DEVICE_RESTORE_STATE 0
 #define TENSTORRENT_RESET_DEVICE_RESET_PCIE_LINK 1
+#define TENSTORRENT_RESET_DEVICE_CONFIG_WRITE 2
 
 struct tenstorrent_reset_device_in {
 	__u32 output_size_bytes;
diff --git a/pcie.c b/pcie.c
index 84e1db0..95a5dce 100644
--- a/pcie.c
+++ b/pcie.c
@@ -13,6 +13,12 @@
 #include "grayskull.h"
 
 #define FW_MSG_PCIE_RETRAIN 0xB6
+#define INTERFACE_TIMER_CONTROL_OFF 0x930
+#define INTERFACE_TIMER_TARGET_OFF 0x934
+
+#define INTERFACE_TIMER_TARGET 0x1
+#define INTERFACE_TIMER_EN 0x1
+#define INTERFACE_FORCE_PENDING 0x10
 
 static bool poll_pcie_link_up(struct pci_dev *pdev, u32 timeout_ms) {
 	u16 tt_vendor_id;
@@ -116,3 +122,10 @@ bool complete_pcie_init(struct tenstorrent_device *tt_dev, u8 __iomem* reset_uni
 
 	return false;
 }
+
+bool pcie_timer_interrupt(struct pci_dev *pdev)
+{
+	pci_write_config_dword(pdev, INTERFACE_TIMER_TARGET_OFF, INTERFACE_TIMER_TARGET);
+	pci_write_config_dword(pdev, INTERFACE_TIMER_CONTROL_OFF, INTERFACE_TIMER_EN | INTERFACE_FORCE_PENDING);
+	return true;
+}
diff --git a/pcie.h b/pcie.h
index bede693..7f6e707 100644
--- a/pcie.h
+++ b/pcie.h
@@ -9,5 +9,6 @@
 bool safe_pci_restore_state(struct pci_dev *pdev);
 bool complete_pcie_init(struct tenstorrent_device *tt_dev, u8 __iomem* reset_unit_regs);
 bool pcie_hot_reset_and_restore_state(struct pci_dev *pdev);
+bool pcie_timer_interrupt(struct pci_dev *pdev);
 
 #endif

From 99e3b3538cc3da4dcb09f85efd437048e304271d Mon Sep 17 00:00:00 2001
From: Andrew Lewycky <alewycky@tenstorrent.com>
Date: Thu, 27 Jul 2023 11:22:54 -0400
Subject: [PATCH 08/10] Get DMA address limit from device class

---
 blackhole.c | 1 +
 device.h    | 1 +
 enumerate.c | 3 ++-
 grayskull.c | 1 +
 wormhole.c  | 1 +
 5 files changed, 6 insertions(+), 1 deletion(-)

diff --git a/blackhole.c b/blackhole.c
index f70c65a..44811f3 100644
--- a/blackhole.c
+++ b/blackhole.c
@@ -29,6 +29,7 @@ static void blackhole_cleanup(struct tenstorrent_device *tt_dev) {
 struct tenstorrent_device_class blackhole_class = {
 	.name = "Blackhole",
 	.instance_size = sizeof(struct blackhole_device),
+	.dma_address_bits = 58,
 	.init_device = blackhole_init,
 	.init_hardware = blackhole_init_hardware,
 	.post_hardware_init = blackhole_post_hardware_init,
diff --git a/device.h b/device.h
index f09bcce..437a014 100644
--- a/device.h
+++ b/device.h
@@ -42,6 +42,7 @@ struct tenstorrent_device {
 struct tenstorrent_device_class {
 	const char *name;
 	u32 instance_size;
+	u32 dma_address_bits;
 	bool (*init_device)(struct tenstorrent_device *ttdev);
 	bool (*init_hardware)(struct tenstorrent_device *ttdev);
 	bool (*post_hardware_init)(struct tenstorrent_device *ttdev);
diff --git a/enumerate.c b/enumerate.c
index 131bb9d..487c553 100644
--- a/enumerate.c
+++ b/enumerate.c
@@ -77,7 +77,8 @@ static int tenstorrent_pci_probe(struct pci_dev *dev, const struct pci_device_id
 
 	mutex_init(&tt_dev->chardev_mutex);
 
-	tt_dev->dma_capable = (dma_set_mask_and_coherent(&dev->dev, DMA_BIT_MASK(dma_address_bits ?: 32)) == 0);
+	tt_dev->dma_capable = (dma_set_mask_and_coherent(&dev->dev, DMA_BIT_MASK(dma_address_bits ?: device_class->dma_address_bits)) == 0);
+
 	// Max these to ensure the IOVA allocator will not split large pinned regions.
 	dma_set_max_seg_size(&dev->dev, UINT_MAX);
 	dma_set_seg_boundary(&dev->dev, ULONG_MAX);
diff --git a/grayskull.c b/grayskull.c
index 78244ed..08e4b9c 100644
--- a/grayskull.c
+++ b/grayskull.c
@@ -875,6 +875,7 @@ static void grayskull_last_release_handler(struct tenstorrent_device *tt_dev) {
 struct tenstorrent_device_class grayskull_class = {
 	.name = "Grayskull",
 	.instance_size = sizeof(struct grayskull_device),
+	.dma_address_bits = 32,
 	.init_device = grayskull_init,
 	.init_hardware = grayskull_init_hardware,
 	.post_hardware_init = grayskull_post_hardware_init,
diff --git a/wormhole.c b/wormhole.c
index 0c1d1cf..496c132 100644
--- a/wormhole.c
+++ b/wormhole.c
@@ -211,6 +211,7 @@ static void wormhole_cleanup(struct tenstorrent_device *tt_dev) {
 struct tenstorrent_device_class wormhole_class = {
 	.name = "Wormhole",
 	.instance_size = sizeof(struct wormhole_device),
+	.dma_address_bits = 32,
 	.init_device = wormhole_init,
 	.init_hardware = wormhole_init_hardware,
 	.post_hardware_init = wormhole_post_hardware_init,

From fe38bbc35b9173e79a58dd3815504c1ce4116e41 Mon Sep 17 00:00:00 2001
From: nxu <nxu@tenstorrent.com>
Date: Thu, 8 Aug 2024 23:03:04 +0000
Subject: [PATCH 09/10] increase MRRS to 4096B

---
 blackhole.c | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/blackhole.c b/blackhole.c
index 44811f3..cc4f651 100644
--- a/blackhole.c
+++ b/blackhole.c
@@ -8,11 +8,15 @@
 #include "pcie.h"
 #include "module.h"
 
+#define MAX_MRRS 4096
+
 static bool blackhole_init(struct tenstorrent_device *tt_dev) {
 	return true;
 }
 
 static bool blackhole_init_hardware(struct tenstorrent_device *tt_dev) {
+	struct pci_dev *pdev = tt_dev->pdev;
+	pcie_set_readrq(pdev, MAX_MRRS);
 	return true;
 }
 

From ddcf4f2b8c7e50210af3456aab090153cd833148 Mon Sep 17 00:00:00 2001
From: Daniel Rosen <drosen@tenstorrent.com>
Date: Tue, 24 Sep 2024 18:38:26 -0400
Subject: [PATCH 10/10] Readded support for WH reset

The condition for handling the RESET_PCIE_LINK flag was removed. This
commit readds it. This is required for handling the reset of WH chips.

Signed-off-by: Daniel Rosen <drosen@tenstorrent.com>
---
 chardev.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/chardev.c b/chardev.c
index cd7e860..22e2ece 100644
--- a/chardev.c
+++ b/chardev.c
@@ -185,7 +185,8 @@ static long ioctl_reset_device(struct chardev_private *priv,
 			ok = priv->device->dev_class->init_hardware(priv->device);
 		else
 			ok = false;
-
+	} else if (in.flags == TENSTORRENT_RESET_DEVICE_RESET_PCIE_LINK) {
+		ok = pcie_hot_reset_and_restore_state(pdev);
 	} else if (in.flags == TENSTORRENT_RESET_DEVICE_CONFIG_WRITE) {
 		ok = pcie_timer_interrupt(pdev);
 	} else {