Skip to contentSkip to conversation

thunderbolt: Intel Ice Lake support

PATCH Superseded Mika Westerberg 6 participants 32 messages Jul 5, 2019 – Aug 12, 2019

Subsystems (2)

v1 of 3 Superseded Latest is v3 →

lore
Share & tools
lore
#0 lore raw

cover thunderbolt: Intel Ice Lake support

Hi all,

With the exception of the first patch which is fix, this series enables Thunderbolt on Intel Ice Lake. Biggest difference from the previous controllers is that the Thunderbolt controller is now integrated as part of the SoC. The firmware messages pretty much follow Titan Ridge but there are some differences as well (such as the new RTD3 veto notification). Also Ice Lake does not implement security levels so DMA protection is handled by IOMMU.

This is v5.4 material but I'm sending it out now because I will be on vacation next 4 weeks mostly without internet access. When I get back I'll gather all the comments and update the series accordingly.

Thanks!

Mika Westerberg (8):
thunderbolt: Correct path indices for PCIe tunnel
thunderbolt: Move NVM upgrade support flag to struct icm
thunderbolt: Use 32-bit writes when writing ring producer/consumer
thunderbolt: Do not fail adding switch if some port is not implemented
thunderbolt: Hide switch attributes that are not set
thunderbolt: Expose active parts of NVM even if upgrade is not supported
thunderbolt: Add support for Intel Ice Lake
ACPI / property: Add two new Thunderbolt property GUIDs to the list

 drivers/acpi/property.c        |   6 +
 drivers/thunderbolt/ctl.c      |  23 ++-
 drivers/thunderbolt/icm.c      | 169 +++++++++++++++++--
 drivers/thunderbolt/nhi.c      | 300 ++++++++++++++++++++++++++++++++-
 drivers/thunderbolt/nhi.h      |   2 +
 drivers/thunderbolt/nhi_regs.h |  25 +++
 drivers/thunderbolt/switch.c   |  52 ++++--
 drivers/thunderbolt/tb_msgs.h  |  16 +-
 drivers/thunderbolt/tunnel.c   |   4 +-
 include/linux/thunderbolt.h    |   2 +
 10 files changed, 553 insertions(+), 46 deletions(-)
Signature · 1 line
2.20.1
#1 lore raw

PATCH 1/8 thunderbolt: Correct path indices for PCIe tunnel

PCIe tunnel path indices got mixed up when we added support for tunnels between switches that are not adjacent. This did not affect the functionality as it is just an index but fix it now nevertheless to make the code easier to understand.

Fixes 8c7acaaf020f ("thunderbolt: Extend tunnel creation to more than 2 adjacent switches")
---
 drivers/thunderbolt/tunnel.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
drivers/thunderbolt/tunnel.c +2 -2
index 31d0234837e4..5a99234826e7 100644
@@ -211,7 +211,7 @@ struct tb_tunnel *tb_tunnel_alloc_pci(struct tb *tb, struct tb_port *up,
 		return NULL;
 	}
 	tb_pci_init_path(path);
-	tunnel->paths[TB_PCI_PATH_UP] = path;
+	tunnel->paths[TB_PCI_PATH_DOWN] = path;
 
 	path = tb_path_alloc(tb, up, TB_PCI_HOPID, down, TB_PCI_HOPID, 0,
 			     "PCIe Up");
@@ -220,7 +220,7 @@ struct tb_tunnel *tb_tunnel_alloc_pci(struct tb *tb, struct tb_port *up,
 		return NULL;
 	}
 	tb_pci_init_path(path);
-	tunnel->paths[TB_PCI_PATH_DOWN] = path;
+	tunnel->paths[TB_PCI_PATH_UP] = path;
 
 	return tunnel;
 }
Signature · 1 line
2.20.1
#2 lore raw

PATCH 2/8 thunderbolt: Move NVM upgrade support flag to struct icm

In reply to Mika Westerberg

This is depends on the controller and on the platform/CPU we are running. Move it to struct icm so we can set it per controller.

---
 drivers/thunderbolt/icm.c | 18 ++++++++++++------
 1 file changed, 12 insertions(+), 6 deletions(-)
drivers/thunderbolt/icm.c +12 -6
index fbdcef56a676..2a56d9478b34 100644
@@ -55,6 +55,7 @@
  * @safe_mode: ICM is in safe mode
  * @max_boot_acl: Maximum number of preboot ACL entries (%0 if not supported)
  * @rpm: Does the controller support runtime PM (RTD3)
+ * @can_upgrade_nvm: Can the NVM firmware be upgrade on this controller
  * @is_supported: Checks if we can support ICM on this controller
  * @cio_reset: Trigger CIO reset
  * @get_mode: Read and return the ICM firmware mode (optional)
@@ -74,6 +75,7 @@ struct icm {
 	int vnd_cap;
 	bool safe_mode;
 	bool rpm;
+	bool can_upgrade_nvm;
 	bool (*is_supported)(struct tb *tb);
 	int (*cio_reset)(struct tb *tb);
 	int (*get_mode)(struct tb *tb);
@@ -1913,12 +1915,7 @@ static int icm_start(struct tb *tb)
 	if (IS_ERR(tb->root_switch))
 		return PTR_ERR(tb->root_switch);
 
-	/*
-	 * NVM upgrade has not been tested on Apple systems and they
-	 * don't provide images publicly either. To be on the safe side
-	 * prevent root switch NVM upgrade on Macs for now.
-	 */
-	tb->root_switch->no_nvm_upgrade = x86_apple_machine;
+	tb->root_switch->no_nvm_upgrade = !icm->can_upgrade_nvm;
 	tb->root_switch->rpm = icm->rpm;
 
 	ret = tb_switch_add(tb->root_switch);
@@ -2021,6 +2018,7 @@ struct tb *icm_probe(struct tb_nhi *nhi)
 	switch (nhi->pdev->device) {
 	case PCI_DEVICE_ID_INTEL_FALCON_RIDGE_2C_NHI:
 	case PCI_DEVICE_ID_INTEL_FALCON_RIDGE_4C_NHI:
+		icm->can_upgrade_nvm = true;
 		icm->is_supported = icm_fr_is_supported;
 		icm->get_route = icm_fr_get_route;
 		icm->save_devices = icm_fr_save_devices;
@@ -2038,6 +2036,13 @@ struct tb *icm_probe(struct tb_nhi *nhi)
 	case PCI_DEVICE_ID_INTEL_ALPINE_RIDGE_C_4C_NHI:
 	case PCI_DEVICE_ID_INTEL_ALPINE_RIDGE_C_2C_NHI:
 		icm->max_boot_acl = ICM_AR_PREBOOT_ACL_ENTRIES;
+		/*
+		 * NVM upgrade has not been tested on Apple systems and
+		 * they don't provide images publicly either. To be on
+		 * the safe side prevent root switch NVM upgrade on Macs
+		 * for now.
+		 */
+		icm->can_upgrade_nvm = !x86_apple_machine;
 		icm->is_supported = icm_ar_is_supported;
 		icm->cio_reset = icm_ar_cio_reset;
 		icm->get_mode = icm_ar_get_mode;
@@ -2054,6 +2059,7 @@ struct tb *icm_probe(struct tb_nhi *nhi)
 	case PCI_DEVICE_ID_INTEL_TITAN_RIDGE_2C_NHI:
 	case PCI_DEVICE_ID_INTEL_TITAN_RIDGE_4C_NHI:
 		icm->max_boot_acl = ICM_AR_PREBOOT_ACL_ENTRIES;
+		icm->can_upgrade_nvm = true;
 		icm->is_supported = icm_ar_is_supported;
 		icm->cio_reset = icm_tr_cio_reset;
 		icm->get_mode = icm_ar_get_mode;
Signature · 1 line
2.20.1
maintainer
#3 lore raw

Re: [PATCH 2/8] thunderbolt: Move NVM upgrade support flag to struct icm

Quoted 38 lines · Mika Westerberg
drivers/thunderbolt/icm.c +9 -6
@@ -1913,12 +1915,7 @@ static int icm_start(struct tb *tb)
        if (IS_ERR(tb->root_switch))
                return PTR_ERR(tb->root_switch);
 
-       /*
-        * NVM upgrade has not been tested on Apple systems and they
-        * don't provide images publicly either. To be on the safe side
-        * prevent root switch NVM upgrade on Macs for now.
-        */
-       tb->root_switch->no_nvm_upgrade = x86_apple_machine;
+       tb->root_switch->no_nvm_upgrade = !icm->can_upgrade_nvm;
        tb->root_switch->rpm = icm->rpm;
 
        ret = tb_switch_add(tb->root_switch);
@@ -2021,6 +2018,7 @@ struct tb *icm_probe(struct tb_nhi *nhi)
        switch (nhi->pdev->device) {
        case PCI_DEVICE_ID_INTEL_FALCON_RIDGE_2C_NHI:
        case PCI_DEVICE_ID_INTEL_FALCON_RIDGE_4C_NHI:
+               icm->can_upgrade_nvm = true;
                icm->is_supported = icm_fr_is_supported;
                icm->get_route = icm_fr_get_route;
                icm->save_devices = icm_fr_save_devices;
@@ -2038,6 +2036,13 @@ struct tb *icm_probe(struct tb_nhi *nhi)
        case PCI_DEVICE_ID_INTEL_ALPINE_RIDGE_C_4C_NHI:
        case PCI_DEVICE_ID_INTEL_ALPINE_RIDGE_C_2C_NHI:
                icm->max_boot_acl = ICM_AR_PREBOOT_ACL_ENTRIES;
+               /*
+                * NVM upgrade has not been tested on Apple systems and
+                * they don't provide images publicly either. To be on
+                * the safe side prevent root switch NVM upgrade on Macs
+                * for now.
+                */
+               icm->can_upgrade_nvm = !x86_apple_machine;
                icm->is_supported = icm_ar_is_supported;
                icm->cio_reset = icm_ar_cio_reset;
                icm->get_mode = icm_ar_get_mode;

drivers/thunderbolt/icm.c struct tb *icm_probe(struct tb_nhi *nhi) · line 2062

Quoted 4 lines
drivers/thunderbolt/icm.c +1 -0
@@ -2054,6 +2059,7 @@ struct tb *icm_probe(struct tb_nhi *nhi)
        case PCI_DEVICE_ID_INTEL_TITAN_RIDGE_2C_NHI:
        case PCI_DEVICE_ID_INTEL_TITAN_RIDGE_4C_NHI:
                icm->max_boot_acl = ICM_AR_PREBOOT_ACL_ENTRIES;
+               icm->can_upgrade_nvm = true;

Shouldn't this be also !x86_apple_machine just like AR?
(For FR, we don't use ICM on Apple machines, as much as I remember, so it's fine to enable it there unconditionally for ICM code path.)

Quoted 6 lines
drivers/thunderbolt/icm.c +0 -1
                icm->is_supported = icm_ar_is_supported;
                icm->cio_reset = icm_tr_cio_reset;
                icm->get_mode = icm_ar_get_mode;
--

2.20.1

#4 lore raw

Re: [PATCH 2/8] thunderbolt: Move NVM upgrade support flag to struct icm

Quoted 9 lines · Yehezkel Bernat

drivers/thunderbolt/icm.c struct tb *icm_probe(struct tb_nhi *nhi) · line 2062

Quoted 5 lines
drivers/thunderbolt/icm.c +1 -0
@@ -2054,6 +2059,7 @@ struct tb *icm_probe(struct tb_nhi *nhi)
        case PCI_DEVICE_ID_INTEL_TITAN_RIDGE_2C_NHI:
        case PCI_DEVICE_ID_INTEL_TITAN_RIDGE_4C_NHI:
                icm->max_boot_acl = ICM_AR_PREBOOT_ACL_ENTRIES;
+               icm->can_upgrade_nvm = true;

Shouldn't this be also !x86_apple_machine just like AR?
(For FR, we don't use ICM on Apple machines, as much as I remember, so it's fine to enable it there unconditionally for ICM code path.)

Yes, good point. I'll fix it up.

#5 lore raw

RE: [PATCH 2/8] thunderbolt: Move NVM upgrade support flag to struct icm

Quoted 23 lines

-----Original Message-----
From: Mika Westerberg <[email protected]>
Sent: Friday, July 5, 2019 5:58 AM
To: Yehezkel Bernat
Cc: LKML; Andreas Noever; Michael Jamet; Rafael J . Wysocki; Len Brown; Lukas Wunner; Limonciello, Mario; Anthony Wong; [email protected]
Subject: Re: [PATCH 2/8] thunderbolt: Move NVM upgrade support flag to struct icm

[EXTERNAL EMAIL]

Quoted 9 lines · Yehezkel Bernat

drivers/thunderbolt/icm.c struct tb *icm_probe(struct tb_nhi *nhi) · line 2062

Quoted 5 lines
drivers/thunderbolt/icm.c +1 -0
@@ -2054,6 +2059,7 @@ struct tb *icm_probe(struct tb_nhi *nhi)
        case PCI_DEVICE_ID_INTEL_TITAN_RIDGE_2C_NHI:
        case PCI_DEVICE_ID_INTEL_TITAN_RIDGE_4C_NHI:
                icm->max_boot_acl = ICM_AR_PREBOOT_ACL_ENTRIES;
+               icm->can_upgrade_nvm = true;

Shouldn't this be also !x86_apple_machine just like AR?
(For FR, we don't use ICM on Apple machines, as much as I remember, so it's fine to enable it there unconditionally for ICM code path.)

Quoted 1 line

Yes, good point. I'll fix it up.

Another thought - does the TR or AR ID's setting can_upgrade_nvm to !x86_apple_machine show up in anything like a dock or is it only host controllers? If it's in docks, then it might be worth only blocking on apple if it's a host.

#6 lore raw

Re: [PATCH 2/8] thunderbolt: Move NVM upgrade support flag to struct icm

Quoted 23 lines · On Tue, Jul 09, 2019 at 03:11:15PM +0000, [email protected] wrote:
Quoted 23 lines

-----Original Message-----
From: Mika Westerberg <[email protected]>
Sent: Friday, July 5, 2019 5:58 AM
To: Yehezkel Bernat
Cc: LKML; Andreas Noever; Michael Jamet; Rafael J . Wysocki; Len Brown; Lukas Wunner; Limonciello, Mario; Anthony Wong; [email protected]
Subject: Re: [PATCH 2/8] thunderbolt: Move NVM upgrade support flag to struct icm

[EXTERNAL EMAIL]

Quoted 9 lines · Yehezkel Bernat

drivers/thunderbolt/icm.c struct tb *icm_probe(struct tb_nhi *nhi) · line 2062

Quoted 5 lines
drivers/thunderbolt/icm.c +1 -0
@@ -2054,6 +2059,7 @@ struct tb *icm_probe(struct tb_nhi *nhi)
        case PCI_DEVICE_ID_INTEL_TITAN_RIDGE_2C_NHI:
        case PCI_DEVICE_ID_INTEL_TITAN_RIDGE_4C_NHI:
                icm->max_boot_acl = ICM_AR_PREBOOT_ACL_ENTRIES;
+               icm->can_upgrade_nvm = true;

Shouldn't this be also !x86_apple_machine just like AR?
(For FR, we don't use ICM on Apple machines, as much as I remember, so it's fine to enable it there unconditionally for ICM code path.)

Quoted 5 lines
Quoted 1 line

Yes, good point. I'll fix it up.

Another thought - does the TR or AR ID's setting can_upgrade_nvm to !x86_apple_machine show up in anything like a dock or is it only host controllers? If it's in docks, then it might be worth only blocking on apple if it's a host.

It affects only hosts so on Apple system you can't upgrade host NVM but docks and other devices you can.

#7 lore raw

PATCH 3/8 thunderbolt: Use 32-bit writes when writing ring producer/consumer

In reply to Mika Westerberg

The register access should be using 32-bit reads/writes according to the datasheet. With the previous generation hardware 16-bit writes have been working but starting with ICL this is not the case anymore so fix producer/consumer register update to use correct width register address.

---
 drivers/thunderbolt/nhi.c | 26 ++++++++++++++++++++++----
 1 file changed, 22 insertions(+), 4 deletions(-)
drivers/thunderbolt/nhi.c +22 -4
index 27fbe62c7ddd..09242653da67 100644
@@ -143,9 +143,24 @@ static void __iomem *ring_options_base(struct tb_ring *ring)
 	return io;
 }
 
-static void ring_iowrite16desc(struct tb_ring *ring, u32 value, u32 offset)
+static void ring_iowrite_prod(struct tb_ring *ring, u16 prod)
 {
-	iowrite16(value, ring_desc_base(ring) + offset);
+	u32 val;
+
+	val = ioread32(ring_desc_base(ring) + 8);
+	val &= 0x0000ffff;
+	val |= prod << 16;
+	iowrite32(val, ring_desc_base(ring) + 8);
+}
+
+static void ring_iowrite_cons(struct tb_ring *ring, u16 cons)
+{
+	u32 val;
+
+	val = ioread32(ring_desc_base(ring) + 8);
+	val &= 0xffff0000;
+	val |= cons;
+	iowrite32(val, ring_desc_base(ring) + 8);
 }
 
 static void ring_iowrite32desc(struct tb_ring *ring, u32 value, u32 offset)
@@ -197,7 +212,10 @@ static void ring_write_descriptors(struct tb_ring *ring)
 			descriptor->sof = frame->sof;
 		}
 		ring->head = (ring->head + 1) % ring->size;
-		ring_iowrite16desc(ring, ring->head, ring->is_tx ? 10 : 8);
+		if (ring->is_tx)
+			ring_iowrite_prod(ring, ring->head);
+		else
+			ring_iowrite_cons(ring, ring->head);
 	}
 }
 
@@ -662,7 +680,7 @@ void tb_ring_stop(struct tb_ring *ring)
 
 	ring_iowrite32options(ring, 0, 0);
 	ring_iowrite64desc(ring, 0, 0);
-	ring_iowrite16desc(ring, 0, ring->is_tx ? 10 : 8);
+	ring_iowrite32desc(ring, 0, 8);
 	ring_iowrite32desc(ring, 0, 12);
 	ring->head = 0;
 	ring->tail = 0;
Signature · 1 line
2.20.1
maintainer
#8 lore raw

Re: [PATCH 3/8] thunderbolt: Use 32-bit writes when writing ring producer/consumer

Quoted 43 lines · Mika Westerberg

The register access should be using 32-bit reads/writes according to the datasheet. With the previous generation hardware 16-bit writes have been working but starting with ICL this is not the case anymore so fix producer/consumer register update to use correct width register address.

---
 drivers/thunderbolt/nhi.c | 26 ++++++++++++++++++++++----
 1 file changed, 22 insertions(+), 4 deletions(-)
drivers/thunderbolt/nhi.c +17 -2
index 27fbe62c7ddd..09242653da67 100644
@@ -143,9 +143,24 @@ static void __iomem *ring_options_base(struct tb_ring *ring)
        return io;
 }
 
-static void ring_iowrite16desc(struct tb_ring *ring, u32 value, u32 offset)
+static void ring_iowrite_prod(struct tb_ring *ring, u16 prod)
 {
-       iowrite16(value, ring_desc_base(ring) + offset);
+       u32 val;
+
+       val = ioread32(ring_desc_base(ring) + 8);
+       val &= 0x0000ffff;
+       val |= prod << 16;
+       iowrite32(val, ring_desc_base(ring) + 8);
+}
+
+static void ring_iowrite_cons(struct tb_ring *ring, u16 cons)
+{
+       u32 val;
+
+       val = ioread32(ring_desc_base(ring) + 8);
+       val &= 0xffff0000;
+       val |= cons;
+       iowrite32(val, ring_desc_base(ring) + 8);
 }
 
 static void ring_iowrite32desc(struct tb_ring *ring, u32 value, u32 offset)

drivers/thunderbolt/nhi.c static void ring_write_descriptors(struct tb_ring *ring) · line 218

Quoted 8 lines
drivers/thunderbolt/nhi.c +4 -1
@@ -197,7 +212,10 @@ static void ring_write_descriptors(struct tb_ring *ring)
                        descriptor->sof = frame->sof;
                }
                ring->head = (ring->head + 1) % ring->size;
-               ring_iowrite16desc(ring, ring->head, ring->is_tx ? 10 : 8);
+               if (ring->is_tx)
+                       ring_iowrite_prod(ring, ring->head);
+               else
+                       ring_iowrite_cons(ring, ring->head);

Really a matter of taste, but maybe you want to consider having a single function, with a 3rd parameter, bool is_tx.
The calls here will be unified to:

        ring_iowrite(ring, ring->head, ring->is_tx);

(No condition is needed here).

The implementation uses the new parameter to decide which part of the register to mask, reducing the code duplication (in my eyes):

        val = ioread32(ring_desc_base(ring) + 8);
        if (is_tx) {
                val &= 0x0000ffff;
                val |= value << 16;
        } else {
                val &= 0xffff0000;
                val |= value;
        }
        iowrite32(val, ring_desc_base(ring) + 8);

I'm not sure if it improves the readability or makes it worse. Your call.

#9 lore raw

Re: [PATCH 3/8] thunderbolt: Use 32-bit writes when writing ring producer/consumer

Quoted 57 lines · Yehezkel Bernat
Quoted 43 lines · Mika Westerberg

The register access should be using 32-bit reads/writes according to the datasheet. With the previous generation hardware 16-bit writes have been working but starting with ICL this is not the case anymore so fix producer/consumer register update to use correct width register address.

---
 drivers/thunderbolt/nhi.c | 26 ++++++++++++++++++++++----
 1 file changed, 22 insertions(+), 4 deletions(-)
drivers/thunderbolt/nhi.c +17 -2
index 27fbe62c7ddd..09242653da67 100644
@@ -143,9 +143,24 @@ static void __iomem *ring_options_base(struct tb_ring *ring)
        return io;
 }
 
-static void ring_iowrite16desc(struct tb_ring *ring, u32 value, u32 offset)
+static void ring_iowrite_prod(struct tb_ring *ring, u16 prod)
 {
-       iowrite16(value, ring_desc_base(ring) + offset);
+       u32 val;
+
+       val = ioread32(ring_desc_base(ring) + 8);
+       val &= 0x0000ffff;
+       val |= prod << 16;
+       iowrite32(val, ring_desc_base(ring) + 8);
+}
+
+static void ring_iowrite_cons(struct tb_ring *ring, u16 cons)
+{
+       u32 val;
+
+       val = ioread32(ring_desc_base(ring) + 8);
+       val &= 0xffff0000;
+       val |= cons;
+       iowrite32(val, ring_desc_base(ring) + 8);
 }
 
 static void ring_iowrite32desc(struct tb_ring *ring, u32 value, u32 offset)

drivers/thunderbolt/nhi.c static void ring_write_descriptors(struct tb_ring *ring) · line 218

Quoted 8 lines
drivers/thunderbolt/nhi.c +4 -1
@@ -197,7 +212,10 @@ static void ring_write_descriptors(struct tb_ring *ring)
                        descriptor->sof = frame->sof;
                }
                ring->head = (ring->head + 1) % ring->size;
-               ring_iowrite16desc(ring, ring->head, ring->is_tx ? 10 : 8);
+               if (ring->is_tx)
+                       ring_iowrite_prod(ring, ring->head);
+               else
+                       ring_iowrite_cons(ring, ring->head);

Really a matter of taste, but maybe you want to consider having a single function, with a 3rd parameter, bool is_tx.
The calls here will be unified to:

Quoted 2 lines
        ring_iowrite(ring, ring->head, ring->is_tx);

(No condition is needed here).

I like the idea. We could even make it

	ring_iowrite_prod_cons(ring);

and then use ring->head and ring->is_tx inside.

#10 lore raw

RE: [PATCH 3/8] thunderbolt: Use 32-bit writes when writing ring producer/consumer

In reply to Yehezkel Bernat

From: Yehezkel Bernat

Quoted 74 lines

Sent: 05 July 2019 12:10

Quoted 43 lines · Mika Westerberg

The register access should be using 32-bit reads/writes according to the datasheet. With the previous generation hardware 16-bit writes have been working but starting with ICL this is not the case anymore so fix producer/consumer register update to use correct width register address.

---
 drivers/thunderbolt/nhi.c | 26 ++++++++++++++++++++++----
 1 file changed, 22 insertions(+), 4 deletions(-)
drivers/thunderbolt/nhi.c +17 -2
index 27fbe62c7ddd..09242653da67 100644
@@ -143,9 +143,24 @@ static void __iomem *ring_options_base(struct tb_ring *ring)
        return io;
 }
 
-static void ring_iowrite16desc(struct tb_ring *ring, u32 value, u32 offset)
+static void ring_iowrite_prod(struct tb_ring *ring, u16 prod)
 {
-       iowrite16(value, ring_desc_base(ring) + offset);
+       u32 val;
+
+       val = ioread32(ring_desc_base(ring) + 8);
+       val &= 0x0000ffff;
+       val |= prod << 16;
+       iowrite32(val, ring_desc_base(ring) + 8);
+}
+
+static void ring_iowrite_cons(struct tb_ring *ring, u16 cons)
+{
+       u32 val;
+
+       val = ioread32(ring_desc_base(ring) + 8);
+       val &= 0xffff0000;
+       val |= cons;
+       iowrite32(val, ring_desc_base(ring) + 8);
 }
 
 static void ring_iowrite32desc(struct tb_ring *ring, u32 value, u32 offset)

drivers/thunderbolt/nhi.c static void ring_write_descriptors(struct tb_ring *ring) · line 218

Quoted 8 lines
drivers/thunderbolt/nhi.c +4 -1
@@ -197,7 +212,10 @@ static void ring_write_descriptors(struct tb_ring *ring)
                        descriptor->sof = frame->sof;
                }
                ring->head = (ring->head + 1) % ring->size;
-               ring_iowrite16desc(ring, ring->head, ring->is_tx ? 10 : 8);
+               if (ring->is_tx)
+                       ring_iowrite_prod(ring, ring->head);
+               else
+                       ring_iowrite_cons(ring, ring->head);

Really a matter of taste, but maybe you want to consider having a single function, with a 3rd parameter, bool is_tx.
The calls here will be unified to:

        ring_iowrite(ring, ring->head, ring->is_tx);

(No condition is needed here).

The implementation uses the new parameter to decide which part of the register to mask, reducing the code duplication (in my eyes):

        val = ioread32(ring_desc_base(ring) + 8);
        if (is_tx) {
                val &= 0x0000ffff;
                val |= value << 16;
        } else {
                val &= 0xffff0000;
                val |= value;
        }
        iowrite32(val, ring_desc_base(ring) + 8);
Quoted 1 line

I'm not sure if it improves the readability or makes it worse. Your call.

Gah, that is all horrid beyond belief.
If a 32bit write is valid then the hardware must not be updating the other 16 bits.
In which case the driver knows what they should be.
So it can do a single 32bit write of the required value.

David

-
Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK Registration No: 1397386 (Wales)

#11 lore raw

Re: [PATCH 3/8] thunderbolt: Use 32-bit writes when writing ring producer/consumer

Quoted 19 lines · David Laight
Quoted 19 lines

Really a matter of taste, but maybe you want to consider having a single function, with a 3rd parameter, bool is_tx.
The calls here will be unified to:

        ring_iowrite(ring, ring->head, ring->is_tx);

(No condition is needed here).

The implementation uses the new parameter to decide which part of the register to mask, reducing the code duplication (in my eyes):

        val = ioread32(ring_desc_base(ring) + 8);
        if (is_tx) {
                val &= 0x0000ffff;
                val |= value << 16;
        } else {
                val &= 0xffff0000;
                val |= value;
        }
        iowrite32(val, ring_desc_base(ring) + 8);
Quoted 7 lines
Quoted 1 line

I'm not sure if it improves the readability or makes it worse. Your call.

Gah, that is all horrid beyond belief.
If a 32bit write is valid then the hardware must not be updating the other 16 bits.
In which case the driver knows what they should be.
So it can do a single 32bit write of the required value.

I'm not entirely sure I understand what you say above. Can you shed some light on this by a concrete example how it should look like? :-)

#12 lore raw

RE: [PATCH 3/8] thunderbolt: Use 32-bit writes when writing ring producer/consumer

From: Mika Westerberg

Quoted 25 lines

Sent: 07 August 2019 17:14
To: David Laight

Quoted 21 lines · David Laight
Quoted 20 lines

Really a matter of taste, but maybe you want to consider having a single function, with a 3rd parameter, bool is_tx.
The calls here will be unified to:

        ring_iowrite(ring, ring->head, ring->is_tx);

(No condition is needed here).

The implementation uses the new parameter to decide which part of the register to mask, reducing the code duplication (in my eyes):

        val = ioread32(ring_desc_base(ring) + 8);
        if (is_tx) {
                val &= 0x0000ffff;
                val |= value << 16;
        } else {
                val &= 0xffff0000;
                val |= value;
        }
        iowrite32(val, ring_desc_base(ring) + 8);

I'm not sure if it improves the readability or makes it worse. Your call.

Quoted 8 lines
Quoted 5 lines

Gah, that is all horrid beyond belief.
If a 32bit write is valid then the hardware must not be updating the other 16 bits.
In which case the driver knows what they should be.
So it can do a single 32bit write of the required value.

I'm not entirely sure I understand what you say above. Can you shed some light on this by a concrete example how it should look like? :-)

The driver must know both the tx and rx ring values, so:

	iowrite32(tx_val << 16 | rx_val, ring_desc_base(ring) + 8);

The ioread32() is likely to be very slow - you only want to do them if absolutely necessary.
The speed of the iowrite32() doesn't matter (much) since it is almost certainly 'posted' and execution continues while the bus cycle is in progress.

David

-
Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK Registration No: 1397386 (Wales)

#13 lore raw

Re: [PATCH 3/8] thunderbolt: Use 32-bit writes when writing ring producer/consumer

Quoted 32 lines · David Laight

From: Mika Westerberg

Quoted 31 lines

Sent: 07 August 2019 17:14
To: David Laight

Quoted 26 lines · David Laight
Quoted 19 lines

Really a matter of taste, but maybe you want to consider having a single function, with a 3rd parameter, bool is_tx.
The calls here will be unified to:

        ring_iowrite(ring, ring->head, ring->is_tx);

(No condition is needed here).

The implementation uses the new parameter to decide which part of the register to mask, reducing the code duplication (in my eyes):

        val = ioread32(ring_desc_base(ring) + 8);
        if (is_tx) {
                val &= 0x0000ffff;
                val |= value << 16;
        } else {
                val &= 0xffff0000;
                val |= value;
        }
        iowrite32(val, ring_desc_base(ring) + 8);
Quoted 1 line

I'm not sure if it improves the readability or makes it worse. Your call.

Gah, that is all horrid beyond belief.
If a 32bit write is valid then the hardware must not be updating the other 16 bits.
In which case the driver knows what they should be.
So it can do a single 32bit write of the required value.

Quoted 6 lines
Quoted 2 lines

I'm not entirely sure I understand what you say above. Can you shed some light on this by a concrete example how it should look like? :-)

The driver must know both the tx and rx ring values, so:

	iowrite32(tx_val << 16 | rx_val, ring_desc_base(ring) + 8);

I see. However, prod or cons side gets updated by the hardware as it processes buffers and other side is only updated by the driver. I'm not sure the above works here.

Quoted 5 lines

The ioread32() is likely to be very slow - you only want to do them if absolutely necessary.
The speed of the iowrite32() doesn't matter (much) since it is almost certainly 'posted' and execution continues while the bus cycle is in progress.

OK thanks for the explanation.

#14 lore raw

RE: [PATCH 3/8] thunderbolt: Use 32-bit writes when writing ring producer/consumer

From: 'Mika Westerberg' [mailto:[email protected]]

Quoted 38 lines

Sent: 07 August 2019 17:36

Quoted 35 lines · David Laight

From: Mika Westerberg

Quoted 33 lines

Sent: 07 August 2019 17:14
To: David Laight

Quoted 19 lines · David Laight
Quoted 19 lines

Really a matter of taste, but maybe you want to consider having a single function, with a 3rd parameter, bool is_tx.
The calls here will be unified to:

        ring_iowrite(ring, ring->head, ring->is_tx);

(No condition is needed here).

The implementation uses the new parameter to decide which part of the register to mask, reducing the code duplication (in my eyes):

        val = ioread32(ring_desc_base(ring) + 8);
        if (is_tx) {
                val &= 0x0000ffff;
                val |= value << 16;
        } else {
                val &= 0xffff0000;
                val |= value;
        }
        iowrite32(val, ring_desc_base(ring) + 8);
Quoted 7 lines
Quoted 1 line

I'm not sure if it improves the readability or makes it worse. Your call.

Gah, that is all horrid beyond belief.
If a 32bit write is valid then the hardware must not be updating the other 16 bits.
In which case the driver knows what they should be.
So it can do a single 32bit write of the required value.

I'm not entirely sure I understand what you say above. Can you shed some light on this by a concrete example how it should look like? :-)

Quoted 7 lines
Quoted 3 lines

The driver must know both the tx and rx ring values, so:

	iowrite32(tx_val << 16 | rx_val, ring_desc_base(ring) + 8);

I see. However, prod or cons side gets updated by the hardware as it processes buffers and other side is only updated by the driver. I'm not sure the above works here.

If the hardware updates the other half of the 32bit word it doesn't ever work. In that case you must do 16bit writes.
If the hardware is ignoring the byte-enables it is broken and unusable.

David

-
Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK Registration No: 1397386 (Wales)

#15 lore raw

Re: [PATCH 3/8] thunderbolt: Use 32-bit writes when writing ring producer/consumer

Quoted 43 lines · David Laight

From: 'Mika Westerberg' [mailto:[email protected]]

Quoted 42 lines

Sent: 07 August 2019 17:36

Quoted 38 lines · David Laight

From: Mika Westerberg

Quoted 25 lines

Sent: 07 August 2019 17:14
To: David Laight

Quoted 21 lines · David Laight
Quoted 20 lines

Really a matter of taste, but maybe you want to consider having a single function, with a 3rd parameter, bool is_tx.
The calls here will be unified to:

        ring_iowrite(ring, ring->head, ring->is_tx);

(No condition is needed here).

The implementation uses the new parameter to decide which part of the register to mask, reducing the code duplication (in my eyes):

        val = ioread32(ring_desc_base(ring) + 8);
        if (is_tx) {
                val &= 0x0000ffff;
                val |= value << 16;
        } else {
                val &= 0xffff0000;
                val |= value;
        }
        iowrite32(val, ring_desc_base(ring) + 8);

I'm not sure if it improves the readability or makes it worse. Your call.

Quoted 8 lines
Quoted 5 lines

Gah, that is all horrid beyond belief.
If a 32bit write is valid then the hardware must not be updating the other 16 bits.
In which case the driver knows what they should be.
So it can do a single 32bit write of the required value.

I'm not entirely sure I understand what you say above. Can you shed some light on this by a concrete example how it should look like? :-)

The driver must know both the tx and rx ring values, so:

	iowrite32(tx_val << 16 | rx_val, ring_desc_base(ring) + 8);
Quoted 8 lines
Quoted 3 lines

I see. However, prod or cons side gets updated by the hardware as it processes buffers and other side is only updated by the driver. I'm not sure the above works here.

If the hardware updates the other half of the 32bit word it doesn't ever work.

In that case you must do 16bit writes.
If the hardware is ignoring the byte-enables it is broken and unusable.

It is quite usable as I'm running this code on real hardware ;-) but 32-bit access is needed.

The low or high 16-bits are read-only depending on the ring (Tx or Rx) and updated by the hardware. It ignores writes to that part so we could do this for Tx ring:

	iowrite32(prod << 16, ring_desc_base(ring) + 8);

and this for Rx ring:

	iowrite32(cons, ring_desc_base(ring) + 8);

Do you see any issues with this?

#16 lore raw

RE: [PATCH 3/8] thunderbolt: Use 32-bit writes when writing ring producer/consumer

From: 'Mika Westerberg'

Quoted 61 lines

Sent: 08 August 2019 10:58

Quoted 43 lines · David Laight

From: 'Mika Westerberg' [mailto:[email protected]]

Quoted 42 lines

Sent: 07 August 2019 17:36

Quoted 38 lines · David Laight

From: Mika Westerberg

Quoted 25 lines

Sent: 07 August 2019 17:14
To: David Laight

Quoted 21 lines · David Laight
Quoted 20 lines

Really a matter of taste, but maybe you want to consider having a single function, with a 3rd parameter, bool is_tx.
The calls here will be unified to:

        ring_iowrite(ring, ring->head, ring->is_tx);

(No condition is needed here).

The implementation uses the new parameter to decide which part of the register to mask, reducing the code duplication (in my eyes):

        val = ioread32(ring_desc_base(ring) + 8);
        if (is_tx) {
                val &= 0x0000ffff;
                val |= value << 16;
        } else {
                val &= 0xffff0000;
                val |= value;
        }
        iowrite32(val, ring_desc_base(ring) + 8);

I'm not sure if it improves the readability or makes it worse. Your call.

Quoted 8 lines
Quoted 5 lines

Gah, that is all horrid beyond belief.
If a 32bit write is valid then the hardware must not be updating the other 16 bits.
In which case the driver knows what they should be.
So it can do a single 32bit write of the required value.

I'm not entirely sure I understand what you say above. Can you shed some light on this by a concrete example how it should look like? :-)

The driver must know both the tx and rx ring values, so:

	iowrite32(tx_val << 16 | rx_val, ring_desc_base(ring) + 8);
Quoted 8 lines
Quoted 3 lines

I see. However, prod or cons side gets updated by the hardware as it processes buffers and other side is only updated by the driver. I'm not sure the above works here.

If the hardware updates the other half of the 32bit word it doesn't ever work.

In that case you must do 16bit writes.
If the hardware is ignoring the byte-enables it is broken and unusable.

It is quite usable as I'm running this code on real hardware ;-) but 32-bit access is needed.

The low or high 16-bits are read-only depending on the ring (Tx or Rx) and updated by the hardware. It ignores writes to that part so we could do this for Tx ring:

Quoted 7 lines
	iowrite32(prod << 16, ring_desc_base(ring) + 8);

and this for Rx ring:

	iowrite32(cons, ring_desc_base(ring) + 8);

Do you see any issues with this?

No, just comment that the hardware ignores the write to the other bytes.

You probably want separate functions - to remove the mispredicted branch.

David

-
Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK Registration No: 1397386 (Wales)

#17 lore raw

PATCH 4/8 thunderbolt: Do not fail adding switch if some port is not implemented

In reply to Mika Westerberg

There are two ways to mark a port as unimplemented. Typical way is to return port type as TB_TYPE_INACTIVE when its config space is read. Alternatively if the port is not physically present (such as ports 10 and 11 in ICL) reading from port config space returns TB_CFG_ERROR_INVALID_CONFIG_SPACE instead. Currently the driver bails out from adding the switch if it receives any error during port inititialization which is wrong.

Handle this properly and just leave the port as TB_TYPE_INACTIVE before continuing to the next port.

---
 drivers/thunderbolt/ctl.c    | 23 +++++++++++++++++++----
 drivers/thunderbolt/switch.c |  8 +++++++-
 2 files changed, 26 insertions(+), 5 deletions(-)
drivers/thunderbolt/ctl.c +19 -4
index 2427d73be731..2ec1af8f7968 100644
@@ -930,6 +930,23 @@ struct tb_cfg_result tb_cfg_write_raw(struct tb_ctl *ctl, const void *buffer,
 	return res;
 }
 
+static int tb_cfg_get_error(struct tb_ctl *ctl, enum tb_cfg_space space,
+			    const struct tb_cfg_result *res)
+{
+	/*
+	 * For unimplemented ports access to port config space may return
+	 * TB_CFG_ERROR_INVALID_CONFIG_SPACE (alternatively their type is
+	 * set to TB_TYPE_INACTIVE). In the former case return -ENODEV so
+	 * that the caller can mark the port as disabled.
+	 */
+	if (space == TB_CFG_PORT &&
+	    res->tb_error == TB_CFG_ERROR_INVALID_CONFIG_SPACE)
+		return -ENODEV;
+
+	tb_cfg_print_error(ctl, res);
+	return -EIO;
+}
+
 int tb_cfg_read(struct tb_ctl *ctl, void *buffer, u64 route, u32 port,
 		enum tb_cfg_space space, u32 offset, u32 length)
 {
@@ -942,8 +959,7 @@ int tb_cfg_read(struct tb_ctl *ctl, void *buffer, u64 route, u32 port,
 
 	case 1:
 		/* Thunderbolt error, tb_error holds the actual number */
-		tb_cfg_print_error(ctl, &res);
-		return -EIO;
+		return tb_cfg_get_error(ctl, space, &res);
 
 	case -ETIMEDOUT:
 		tb_ctl_warn(ctl, "timeout reading config space %u from %#x\n",
@@ -969,8 +985,7 @@ int tb_cfg_write(struct tb_ctl *ctl, const void *buffer, u64 route, u32 port,
 
 	case 1:
 		/* Thunderbolt error, tb_error holds the actual number */
-		tb_cfg_print_error(ctl, &res);
-		return -EIO;
+		return tb_cfg_get_error(ctl, space, &res);
 
 	case -ETIMEDOUT:
 		tb_ctl_warn(ctl, "timeout writing config space %u to %#x\n",
drivers/thunderbolt/switch.c +7 -1
index 10b56c66fec3..eac62ff1b85c 100644
@@ -611,8 +611,14 @@ static int tb_init_port(struct tb_port *port)
 	int cap;
 
 	res = tb_port_read(port, &port->config, TB_CFG_PORT, 0, 8);
-	if (res)
+	if (res) {
+		if (res == -ENODEV) {
+			tb_dbg(port->sw->tb, " Port %d: not implemented\n",
+			       port->port);
+			return 0;
+		}
 		return res;
+	}
 
 	/* Port 0 is the switch itself and has no PHY. */
 	if (port->config.type == TB_TYPE_PORT && port->port != 0) {
Signature · 1 line
2.20.1
#18 lore raw

Re: [PATCH 4/8] thunderbolt: Do not fail adding switch if some port is not implemented

Quoted 10 lines · Mika Westerberg

There are two ways to mark a port as unimplemented. Typical way is to return port type as TB_TYPE_INACTIVE when its config space is read. Alternatively if the port is not physically present (such as ports 10 and 11 in ICL) reading from port config space returns TB_CFG_ERROR_INVALID_CONFIG_SPACE instead. Currently the driver bails out from adding the switch if it receives any error during port inititialization which is wrong.

Handle this properly and just leave the port as TB_TYPE_INACTIVE before continuing to the next port.

Your patch may also cause this snippet in eeprom.c to become obsolete:

		/* Port 5 is inaccessible on this gen 1 controller */
		if (sw->config.device_id == PCI_DEVICE_ID_INTEL_LIGHT_RIDGE)
			sw->ports[5].disabled = true;

To verify this hypothesis, one needs to comment out the call to tb_drom_copy_efi() as well as the above-quoted snippet and boot on a Mac with Light Ridge. The driver should hit an error without your patch and it may work correctly with your patch.

Thanks,

Lukas

#19 lore raw

Re: [PATCH 4/8] thunderbolt: Do not fail adding switch if some port is not implemented

Quoted 14 lines · Lukas Wunner
Quoted 10 lines · Mika Westerberg

There are two ways to mark a port as unimplemented. Typical way is to return port type as TB_TYPE_INACTIVE when its config space is read. Alternatively if the port is not physically present (such as ports 10 and 11 in ICL) reading from port config space returns TB_CFG_ERROR_INVALID_CONFIG_SPACE instead. Currently the driver bails out from adding the switch if it receives any error during port inititialization which is wrong.

Handle this properly and just leave the port as TB_TYPE_INACTIVE before continuing to the next port.

Your patch may also cause this snippet in eeprom.c to become obsolete:

Quoted 8 lines
		/* Port 5 is inaccessible on this gen 1 controller */
		if (sw->config.device_id == PCI_DEVICE_ID_INTEL_LIGHT_RIDGE)
			sw->ports[5].disabled = true;

To verify this hypothesis, one needs to comment out the call to tb_drom_copy_efi() as well as the above-quoted snippet and boot on a Mac with Light Ridge. The driver should hit an error without your patch and it may work correctly with your patch.

Indeed. I'll check this - as I have Mac with LR, and update the patch accordingly (e.g drop the check if the error does not appear).

#20 lore raw

PATCH 5/8 thunderbolt: Hide switch attributes that are not set

In reply to Mika Westerberg

Thunderbolt host routers may not always contain DROM that includes device identification information. This is mostly needed for Ice Lake systems but some Falcon Ridge controllers on PCs also do not have DROM.

In that case hide the identification attributes.

---
 drivers/thunderbolt/switch.c | 14 +++++++++++++-
 1 file changed, 13 insertions(+), 1 deletion(-)
drivers/thunderbolt/switch.c +13 -1
index eac62ff1b85c..e84067084dcd 100644
@@ -1337,7 +1337,19 @@ static umode_t switch_attr_is_visible(struct kobject *kobj,
 	struct device *dev = container_of(kobj, struct device, kobj);
 	struct tb_switch *sw = tb_to_switch(dev);
 
-	if (attr == &dev_attr_key.attr) {
+	if (attr == &dev_attr_device.attr) {
+		if (!sw->device)
+			return 0;
+	} else if (attr == &dev_attr_device_name.attr) {
+		if (!sw->device_name)
+			return 0;
+	} else if (attr == &dev_attr_vendor.attr)  {
+		if (!sw->vendor)
+			return 0;
+	} else if (attr == &dev_attr_vendor_name.attr)  {
+		if (!sw->vendor_name)
+			return 0;
+	} else if (attr == &dev_attr_key.attr) {
 		if (tb_route(sw) &&
 		    sw->tb->security_level == TB_SECURITY_SECURE &&
 		    sw->security_level == TB_SECURITY_SECURE)
Signature · 1 line
2.20.1
#21 lore raw

PATCH 6/8 thunderbolt: Expose active parts of NVM even if upgrade is not supported

In reply to Mika Westerberg

Ice Lake Thunderbolt controller NVM firmware is part of the BIOS image which means it is not writable through the DMA port anymore. However, we can still read it so we can keep nvm_version and active parts of NVM. This way users still can find out the active NVM version and other potentially useful information directly from Linux.

---
 drivers/thunderbolt/switch.c | 28 +++++++++++++++++++---------
 1 file changed, 19 insertions(+), 9 deletions(-)
drivers/thunderbolt/switch.c +19 -9
index e84067084dcd..e318e9714a38 100644
@@ -364,12 +364,14 @@ static int tb_switch_nvm_add(struct tb_switch *sw)
 		nvm->active = nvm_dev;
 	}
 
-	nvm_dev = register_nvmem(sw, nvm->id, NVM_MAX_SIZE, false);
-	if (IS_ERR(nvm_dev)) {
-		ret = PTR_ERR(nvm_dev);
-		goto err_nvm_active;
+	if (!sw->no_nvm_upgrade) {
+		nvm_dev = register_nvmem(sw, nvm->id, NVM_MAX_SIZE, false);
+		if (IS_ERR(nvm_dev)) {
+			ret = PTR_ERR(nvm_dev);
+			goto err_nvm_active;
+		}
+		nvm->non_active = nvm_dev;
 	}
-	nvm->non_active = nvm_dev;
 
 	sw->nvm = nvm;
 	return 0;
@@ -398,7 +400,8 @@ static void tb_switch_nvm_remove(struct tb_switch *sw)
 	if (!nvm->authenticating)
 		nvm_clear_auth_status(sw);
 
-	nvmem_unregister(nvm->non_active);
+	if (nvm->non_active)
+		nvmem_unregister(nvm->non_active);
 	if (nvm->active)
 		nvmem_unregister(nvm->active);
 	ida_simple_remove(&nvm_ida, nvm->id);
@@ -1355,8 +1358,11 @@ static umode_t switch_attr_is_visible(struct kobject *kobj,
 		    sw->security_level == TB_SECURITY_SECURE)
 			return attr->mode;
 		return 0;
-	} else if (attr == &dev_attr_nvm_authenticate.attr ||
-		   attr == &dev_attr_nvm_version.attr) {
+	} else if (attr == &dev_attr_nvm_authenticate.attr) {
+		if (sw->dma_port && !sw->no_nvm_upgrade)
+			return attr->mode;
+		return 0;
+	} else if (attr == &dev_attr_nvm_version.attr) {
 		if (sw->dma_port)
 			return attr->mode;
 		return 0;
@@ -1707,13 +1713,17 @@ static int tb_switch_add_dma_port(struct tb_switch *sw)
 		break;
 	}
 
-	if (sw->no_nvm_upgrade)
+	/* Root switch DMA port requires running firmware */
+	if (!tb_route(sw) && sw->config.enabled)
 		return 0;
 
 	sw->dma_port = dma_port_alloc(sw);
 	if (!sw->dma_port)
 		return 0;
 
+	if (sw->no_nvm_upgrade)
+		return 0;
+
 	/*
 	 * Check status of the previous flash authentication. If there
 	 * is one we need to power cycle the switch in any case to make
Signature · 1 line
2.20.1
#22 lore raw

PATCH 7/8 thunderbolt: Add support for Intel Ice Lake

In reply to Mika Westerberg

The Thunderbolt controller is integrated into the Ice Lake CPU itself and requires special flows to power it on and off using force power bit in NHI VSEC registers. Runtime PM (RTD3) and Sx flows also differ from the discrete solutions. Now the firmware notifies the driver whether RTD3 entry or exit are possible. The driver is responsible of sending Go2Sx command through link controller mailbox when system enters Sx states (suspend-to-mem/disk). Rest of the ICM firwmare flows follow Titan Ridge.

---
 drivers/thunderbolt/icm.c      | 151 ++++++++++++++++--
 drivers/thunderbolt/nhi.c      | 274 ++++++++++++++++++++++++++++++++-
 drivers/thunderbolt/nhi.h      |   2 +
 drivers/thunderbolt/nhi_regs.h |  25 +++
 drivers/thunderbolt/switch.c   |   2 +
 drivers/thunderbolt/tb_msgs.h  |  16 +-
 include/linux/thunderbolt.h    |   2 +
 7 files changed, 453 insertions(+), 19 deletions(-)
drivers/thunderbolt/icm.c +141 -10
index 2a56d9478b34..e882e79b32a4 100644
@@ -56,16 +56,19 @@
  * @max_boot_acl: Maximum number of preboot ACL entries (%0 if not supported)
  * @rpm: Does the controller support runtime PM (RTD3)
  * @can_upgrade_nvm: Can the NVM firmware be upgrade on this controller
+ * @veto: Is RTD3 veto in effect
  * @is_supported: Checks if we can support ICM on this controller
  * @cio_reset: Trigger CIO reset
  * @get_mode: Read and return the ICM firmware mode (optional)
  * @get_route: Find a route string for given switch
  * @save_devices: Ask ICM to save devices to ACL when suspending (optional)
  * @driver_ready: Send driver ready message to ICM
+ * @set_uuid: Set UUID for the root switch (optional)
  * @device_connected: Handle device connected ICM message
  * @device_disconnected: Handle device disconnected ICM message
  * @xdomain_connected - Handle XDomain connected ICM message
  * @xdomain_disconnected - Handle XDomain disconnected ICM message
+ * @rtd3_veto: Handle RTD3 veto notification ICM message
  */
 struct icm {
 	struct mutex request_lock;
@@ -76,6 +79,7 @@ struct icm {
 	bool safe_mode;
 	bool rpm;
 	bool can_upgrade_nvm;
+	bool veto;
 	bool (*is_supported)(struct tb *tb);
 	int (*cio_reset)(struct tb *tb);
 	int (*get_mode)(struct tb *tb);
@@ -84,6 +88,7 @@ struct icm {
 	int (*driver_ready)(struct tb *tb,
 			    enum tb_security_level *security_level,
 			    size_t *nboot_acl, bool *rpm);
+	void (*set_uuid)(struct tb *tb);
 	void (*device_connected)(struct tb *tb,
 				 const struct icm_pkg_header *hdr);
 	void (*device_disconnected)(struct tb *tb,
@@ -92,6 +97,7 @@ struct icm {
 				  const struct icm_pkg_header *hdr);
 	void (*xdomain_disconnected)(struct tb *tb,
 				     const struct icm_pkg_header *hdr);
+	void (*rtd3_veto)(struct tb *tb, const struct icm_pkg_header *hdr);
 };
 
 struct icm_notification {
@@ -519,14 +525,16 @@ static int icm_fr_disconnect_xdomain_paths(struct tb *tb, struct tb_xdomain *xd)
 	return 0;
 }
 
-static void add_switch(struct tb_switch *parent_sw, u64 route,
-		       const uuid_t *uuid, const u8 *ep_name,
-		       size_t ep_name_size, u8 connection_id, u8 connection_key,
-		       u8 link, u8 depth, enum tb_security_level security_level,
-		       bool authorized, bool boot)
+static struct tb_switch *add_switch(struct tb_switch *parent_sw, u64 route,
+				    const uuid_t *uuid, const u8 *ep_name,
+				    size_t ep_name_size, u8 connection_id,
+				    u8 connection_key, u8 link, u8 depth,
+				    enum tb_security_level security_level,
+				    bool authorized, bool boot)
 {
 	const struct intel_vss *vss;
 	struct tb_switch *sw;
+	int ret;
 
 	pm_runtime_get_sync(&parent_sw->dev);
 
@@ -557,14 +565,18 @@ static void add_switch(struct tb_switch *parent_sw, u64 route,
 	tb_port_at(route, parent_sw)->remote = tb_upstream_port(sw);
 	tb_upstream_port(sw)->remote = tb_port_at(route, parent_sw);
 
-	if (tb_switch_add(sw)) {
+	ret = tb_switch_add(sw);
+	if (ret) {
 		tb_port_at(tb_route(sw), parent_sw)->remote = NULL;
 		tb_switch_put(sw);
+		sw = ERR_PTR(ret);
 	}
 
 out:
 	pm_runtime_mark_last_busy(&parent_sw->dev);
 	pm_runtime_put_autosuspend(&parent_sw->dev);
+
+	return sw;
 }
 
 static void update_switch(struct tb_switch *parent_sw, struct tb_switch *sw,
@@ -1086,7 +1098,8 @@ static int icm_tr_disconnect_xdomain_paths(struct tb *tb, struct tb_xdomain *xd)
 }
 
 static void
-icm_tr_device_connected(struct tb *tb, const struct icm_pkg_header *hdr)
+__icm_tr_device_connected(struct tb *tb, const struct icm_pkg_header *hdr,
+			  bool force_rtd3)
 {
 	const struct icm_tr_event_device_connected *pkg =
 		(const struct icm_tr_event_device_connected *)hdr;
@@ -1151,13 +1164,21 @@ icm_tr_device_connected(struct tb *tb, const struct icm_pkg_header *hdr)
 		return;
 	}
 
-	add_switch(parent_sw, route, &pkg->ep_uuid, (const u8 *)pkg->ep_name,
-		   sizeof(pkg->ep_name), pkg->connection_id,
-		   0, 0, 0, security_level, authorized, boot);
+	sw = add_switch(parent_sw, route, &pkg->ep_uuid, (const u8 *)pkg->ep_name,
+			sizeof(pkg->ep_name), pkg->connection_id, 0, 0, 0,
+			security_level, authorized, boot);
+	if (!IS_ERR(sw) && force_rtd3)
+		sw->rpm = true;
 
 	tb_switch_put(parent_sw);
 }
 
+static void
+icm_tr_device_connected(struct tb *tb, const struct icm_pkg_header *hdr)
+{
+	__icm_tr_device_connected(tb, hdr, false);
+}
+
 static void
 icm_tr_device_disconnected(struct tb *tb, const struct icm_pkg_header *hdr)
 {
@@ -1468,6 +1489,72 @@ static int icm_ar_set_boot_acl(struct tb *tb, const uuid_t *uuids,
 	return 0;
 }
 
+static int
+icm_icl_driver_ready(struct tb *tb, enum tb_security_level *security_level,
+		    size_t *nboot_acl, bool *rpm)
+{
+	struct icm_tr_pkg_driver_ready_response reply;
+	struct icm_pkg_driver_ready request = {
+		.hdr.code = ICM_DRIVER_READY,
+	};
+	int ret;
+
+	memset(&reply, 0, sizeof(reply));
+	ret = icm_request(tb, &request, sizeof(request), &reply, sizeof(reply),
+			  1, 20000);
+	if (ret)
+		return ret;
+
+	/* Ice Lake always supports RTD3 */
+	if (rpm)
+		*rpm = true;
+
+	return 0;
+}
+
+static void icm_icl_set_uuid(struct tb *tb)
+{
+	struct tb_nhi *nhi = tb->nhi;
+	u32 uuid[4];
+
+	pci_read_config_dword(nhi->pdev, VS_CAP_10, &uuid[0]);
+	pci_read_config_dword(nhi->pdev, VS_CAP_11, &uuid[1]);
+	uuid[2] = 0xffffffff;
+	uuid[3] = 0xffffffff;
+
+	tb->root_switch->uuid = kmemdup(uuid, sizeof(uuid), GFP_KERNEL);
+}
+
+static void
+icm_icl_device_connected(struct tb *tb, const struct icm_pkg_header *hdr)
+{
+	__icm_tr_device_connected(tb, hdr, true);
+}
+
+static void icm_icl_rtd3_veto(struct tb *tb, const struct icm_pkg_header *hdr)
+{
+	const struct icm_icl_event_rtd3_veto *pkg =
+		(const struct icm_icl_event_rtd3_veto *)hdr;
+	struct icm *icm = tb_priv(tb);
+
+	tb_dbg(tb, "ICM rtd3 veto=0x%08x\n", pkg->veto_reason);
+
+	if (pkg->veto_reason) {
+		if (!icm->veto) {
+			icm->veto = true;
+			/* Keep the domain powered while veto is in effect */
+			pm_runtime_get(&tb->dev);
+		}
+	} else {
+		if (icm->veto) {
+			icm->veto = false;
+			/* Allow the domain suspend now */
+			pm_runtime_mark_last_busy(&tb->dev);
+			pm_runtime_put_autosuspend(&tb->dev);
+		}
+	}
+}
+
 static void icm_handle_notification(struct work_struct *work)
 {
 	struct icm_notification *n = container_of(work, typeof(*n), work);
@@ -1495,6 +1582,9 @@ static void icm_handle_notification(struct work_struct *work)
 		case ICM_EVENT_XDOMAIN_DISCONNECTED:
 			icm->xdomain_disconnected(tb, n->pkg);
 			break;
+		case ICM_EVENT_RTD3_VETO:
+			icm->rtd3_veto(tb, n->pkg);
+			break;
 		}
 	}
 
@@ -1853,6 +1943,18 @@ static void icm_complete(struct tb *tb)
 	if (tb->nhi->going_away)
 		return;
 
+	/*
+	 * If RTD3 was vetoed before we entered system suspend allow it
+	 * again now before driver ready is sent. Firmware sends a new RTD3
+	 * veto if it is still the case after we have sent it driver ready
+	 * command.
+	 */
+	if (icm->veto) {
+		icm->veto = false;
+		pm_runtime_mark_last_busy(&tb->dev);
+		pm_runtime_put_autosuspend(&tb->dev);
+	}
+
 	icm_unplug_children(tb->root_switch);
 
 	/*
@@ -1918,6 +2020,9 @@ static int icm_start(struct tb *tb)
 	tb->root_switch->no_nvm_upgrade = !icm->can_upgrade_nvm;
 	tb->root_switch->rpm = icm->rpm;
 
+	if (icm->set_uuid)
+		icm->set_uuid(tb);
+
 	ret = tb_switch_add(tb->root_switch);
 	if (ret) {
 		tb_switch_put(tb->root_switch);
@@ -2002,6 +2107,19 @@ static const struct tb_cm_ops icm_tr_ops = {
 	.disconnect_xdomain_paths = icm_tr_disconnect_xdomain_paths,
 };
 
+/* Ice Lake */
+static const struct tb_cm_ops icm_icl_ops = {
+	.driver_ready = icm_driver_ready,
+	.start = icm_start,
+	.stop = icm_stop,
+	.complete = icm_complete,
+	.runtime_suspend = icm_runtime_suspend,
+	.runtime_resume = icm_runtime_resume,
+	.handle_event = icm_handle_event,
+	.approve_xdomain_paths = icm_tr_approve_xdomain_paths,
+	.disconnect_xdomain_paths = icm_tr_disconnect_xdomain_paths,
+};
+
 struct tb *icm_probe(struct tb_nhi *nhi)
 {
 	struct icm *icm;
@@ -2070,6 +2188,19 @@ struct tb *icm_probe(struct tb_nhi *nhi)
 		icm->xdomain_disconnected = icm_tr_xdomain_disconnected;
 		tb->cm_ops = &icm_tr_ops;
 		break;
+
+	case PCI_DEVICE_ID_INTEL_ICL_NHI0:
+	case PCI_DEVICE_ID_INTEL_ICL_NHI1:
+		icm->is_supported = icm_ar_is_supported;
+		icm->driver_ready = icm_icl_driver_ready;
+		icm->set_uuid = icm_icl_set_uuid;
+		icm->device_connected = icm_icl_device_connected;
+		icm->device_disconnected = icm_tr_device_disconnected;
+		icm->xdomain_connected = icm_tr_xdomain_connected;
+		icm->xdomain_disconnected = icm_tr_xdomain_disconnected;
+		icm->rtd3_veto = icm_icl_rtd3_veto;
+		tb->cm_ops = &icm_icl_ops;
+		break;
 	}
 
 	if (!icm->is_supported || !icm->is_supported(tb)) {
drivers/thunderbolt/nhi.c +269 -5

379 lines not on this page · raw message

drivers/thunderbolt/nhi.h +2 -0
index 1b5d47ecd3ed..dad691d1b704 100644
@@ -51,5 +51,7 @@ enum nhi_fw_mode nhi_mailbox_mode(struct tb_nhi *nhi);
 #define PCI_DEVICE_ID_INTEL_TITAN_RIDGE_4C_BRIDGE	0x15ea
 #define PCI_DEVICE_ID_INTEL_TITAN_RIDGE_4C_NHI		0x15eb
 #define PCI_DEVICE_ID_INTEL_TITAN_RIDGE_DD_BRIDGE	0x15ef
+#define PCI_DEVICE_ID_INTEL_ICL_NHI1			0x8a0d
+#define PCI_DEVICE_ID_INTEL_ICL_NHI0			0x8a17
 
 #endif
drivers/thunderbolt/nhi_regs.h +25 -0
index a60bd98c1d04..8754e7dd63d5 100644
@@ -124,4 +124,29 @@ struct ring_desc {
 #define REG_FW_STS_ICM_EN_INVERT	BIT(1)
 #define REG_FW_STS_ICM_EN		BIT(0)
 
+/* ICL NHI VSEC registers */
+
+/* FW ready */
+#define VS_CAP_9			0xc8
+#define VS_CAP_9_FW_READY		BIT(31)
+/* UUID */
+#define VS_CAP_10			0xcc
+#define VS_CAP_11			0xd0
+/* LTR */
+#define VS_CAP_15			0xe0
+#define VS_CAP_16			0xe4
+/* TBT2PCIe */
+#define VS_CAP_18			0xec
+#define VS_CAP_18_DONE			BIT(0)
+/* PCIe2TBT */
+#define VS_CAP_19			0xf0
+#define VS_CAP_19_VALID			BIT(0)
+#define VS_CAP_19_CMD_SHIFT		1
+#define VS_CAP_19_CMD_MASK		GENMASK(7, 1)
+/* Force power */
+#define VS_CAP_22			0xfc
+#define VS_CAP_22_FORCE_POWER		BIT(1)
+#define VS_CAP_22_DMA_DELAY_MASK	GENMASK(31, 24)
+#define VS_CAP_22_DMA_DELAY_SHIFT	24
+
 #endif
drivers/thunderbolt/switch.c +2 -0
index e318e9714a38..b896aa78fa9f 100644
@@ -1470,6 +1470,8 @@ static int tb_switch_get_generation(struct tb_switch *sw)
 	case PCI_DEVICE_ID_INTEL_TITAN_RIDGE_2C_BRIDGE:
 	case PCI_DEVICE_ID_INTEL_TITAN_RIDGE_4C_BRIDGE:
 	case PCI_DEVICE_ID_INTEL_TITAN_RIDGE_DD_BRIDGE:
+	case PCI_DEVICE_ID_INTEL_ICL_NHI0:
+	case PCI_DEVICE_ID_INTEL_ICL_NHI1:
 		return 3;
 
 	default:
drivers/thunderbolt/tb_msgs.h +12 -4
index afbe1d29bb03..4b641e4ee0c5 100644
@@ -104,10 +104,11 @@ enum icm_pkg_code {
 };
 
 enum icm_event_code {
-	ICM_EVENT_DEVICE_CONNECTED = 3,
-	ICM_EVENT_DEVICE_DISCONNECTED = 4,
-	ICM_EVENT_XDOMAIN_CONNECTED = 6,
-	ICM_EVENT_XDOMAIN_DISCONNECTED = 7,
+	ICM_EVENT_DEVICE_CONNECTED = 0x3,
+	ICM_EVENT_DEVICE_DISCONNECTED = 0x4,
+	ICM_EVENT_XDOMAIN_CONNECTED = 0x6,
+	ICM_EVENT_XDOMAIN_DISCONNECTED = 0x7,
+	ICM_EVENT_RTD3_VETO = 0xa,
 };
 
 struct icm_pkg_header {
@@ -463,6 +464,13 @@ struct icm_tr_pkg_disconnect_xdomain_response {
 	uuid_t remote_uuid;
 };
 
+/* Ice Lake messages */
+
+struct icm_icl_event_rtd3_veto {
+	struct icm_pkg_header hdr;
+	u32 veto_reason;
+};
+
 /* XDomain messages */
 
 struct tb_xdomain_header {
include/linux/thunderbolt.h +2 -0
index 2d7e012db03f..ece782ef5466 100644
@@ -429,6 +429,7 @@ static inline struct tb_xdomain *tb_service_parent(struct tb_service *svc)
  * @lock: Must be held during ring creation/destruction. Is acquired by
  *	  interrupt_work when dispatching interrupts to individual rings.
  * @pdev: Pointer to the PCI device
+ * @ops: NHI specific optional ops
  * @iobase: MMIO space of the NHI
  * @tx_rings: All Tx rings available on this host controller
  * @rx_rings: All Rx rings available on this host controller
@@ -442,6 +443,7 @@ static inline struct tb_xdomain *tb_service_parent(struct tb_service *svc)
 struct tb_nhi {
 	spinlock_t lock;
 	struct pci_dev *pdev;
+	const struct tb_nhi_ops *ops;
 	void __iomem *iobase;
 	struct tb_ring **tx_rings;
 	struct tb_ring **rx_rings;
Signature · 1 line
2.20.1
maintainer
#23 lore raw

Re: [PATCH 7/8] thunderbolt: Add support for Intel Ice Lake

Quoted 11 lines · Mika Westerberg
drivers/thunderbolt/icm.c +10 -0
+static void icm_icl_rtd3_veto(struct tb *tb, const struct icm_pkg_header *hdr)
+{
+       const struct icm_icl_event_rtd3_veto *pkg =
+               (const struct icm_icl_event_rtd3_veto *)hdr;
+       struct icm *icm = tb_priv(tb);
+
+       tb_dbg(tb, "ICM rtd3 veto=0x%08x\n", pkg->veto_reason);
+
+       if (pkg->veto_reason) {
+               if (!icm->veto) {

drivers/thunderbolt/icm.c static int icm_ar_set_boot_acl(struct tb *tb, const uuid_t *uuids, · line 1553

Quoted 10 lines
drivers/thunderbolt/icm.c +10 -0
+                       icm->veto = true;
+                       /* Keep the domain powered while veto is in effect */
+                       pm_runtime_get(&tb->dev);
+               }
+       } else {
+               if (icm->veto) {
+                       icm->veto = false;
+                       /* Allow the domain suspend now */
+                       pm_runtime_mark_last_busy(&tb->dev);
+                       pm_runtime_put_autosuspend(&tb->dev);

Handling the removal of the veto is duplicated below. Worth introducing as a helper function?

Quoted 4 lines
drivers/thunderbolt/icm.c +4 -0
+               }
+       }
+}
+

...

Quoted 6 lines
drivers/thunderbolt/icm.c +2 -0
@@ -1853,6 +1943,18 @@ static void icm_complete(struct tb *tb)
        if (tb->nhi->going_away)
                return;
 
+       /*
+        * If RTD3 was vetoed before we entered system suspend allow it

drivers/thunderbolt/icm.c static void icm_complete(struct tb *tb) · line 1957

Quoted 10 lines
drivers/thunderbolt/icm.c +10 -0
+        * again now before driver ready is sent. Firmware sends a new RTD3
+        * veto if it is still the case after we have sent it driver ready
+        * command.
+        */
+       if (icm->veto) {
+               icm->veto = false;
+               pm_runtime_mark_last_busy(&tb->dev);
+               pm_runtime_put_autosuspend(&tb->dev);
+       }
+

Here is the duplication.

Quoted 9 lines
drivers/thunderbolt/nhi.c +9 -0
+static int nhi_suspend_power_down(struct tb *tb)
+{
+       int ret;
+
+       /*
+        * If there is no device connected we need to perform an additional
+        * handshake through LC mailbox and force power down before
+        * entering D3.
+        */

drivers/thunderbolt/nhi.c static irqreturn_t nhi_msi(int irq, void *data) · line 945

Quoted 10 lines
drivers/thunderbolt/nhi.c +10 -0
+       ret = device_for_each_child(&tb->root_switch->dev, NULL,
+                                   nhi_device_connected);
+       if (!ret) {
+               lc_mailbox_cmd(tb->nhi, LC_PREPARE_FOR_RESET);
+               ret = lc_mailbox_cmd_complete(tb->nhi,
+                                             LC_MAILBOX_TIMEOUT);
+               if (ret)
+                       return ret;
+
+               return nhi_power_down(tb->nhi);

Just to be sure: unforce power is done only if no device is connected? My understanding of the comment above was that unforce power should be done anyway (so it should be outside of this if block), and the difference between the cases is only about the additional LC mailbox message. I guess I misread it.

Quoted 5 lines
drivers/thunderbolt/nhi.c +5 -0
+       }
+
+       return 0;
+}
+
#24 lore raw

Re: [PATCH 7/8] thunderbolt: Add support for Intel Ice Lake

Quoted 38 lines · Yehezkel Bernat
Quoted 11 lines · Mika Westerberg
drivers/thunderbolt/icm.c +10 -0
+static void icm_icl_rtd3_veto(struct tb *tb, const struct icm_pkg_header *hdr)
+{
+       const struct icm_icl_event_rtd3_veto *pkg =
+               (const struct icm_icl_event_rtd3_veto *)hdr;
+       struct icm *icm = tb_priv(tb);
+
+       tb_dbg(tb, "ICM rtd3 veto=0x%08x\n", pkg->veto_reason);
+
+       if (pkg->veto_reason) {
+               if (!icm->veto) {

drivers/thunderbolt/icm.c static int icm_ar_set_boot_acl(struct tb *tb, const uuid_t *uuids, · line 1553

Quoted 10 lines
drivers/thunderbolt/icm.c +10 -0
+                       icm->veto = true;
+                       /* Keep the domain powered while veto is in effect */
+                       pm_runtime_get(&tb->dev);
+               }
+       } else {
+               if (icm->veto) {
+                       icm->veto = false;
+                       /* Allow the domain suspend now */
+                       pm_runtime_mark_last_busy(&tb->dev);
+                       pm_runtime_put_autosuspend(&tb->dev);

Handling the removal of the veto is duplicated below. Worth introducing as a helper function?

Quoted 4 lines
drivers/thunderbolt/icm.c +4 -0
+               }
+       }
+}
+

...

Quoted 4 lines
drivers/thunderbolt/icm.c +0 -0
@@ -1853,6 +1943,18 @@ static void icm_complete(struct tb *tb)
        if (tb->nhi->going_away)
                return;
 
Quoted 14 lines

drivers/thunderbolt/icm.c static void icm_complete(struct tb *tb) · line 1957

Quoted 12 lines
drivers/thunderbolt/icm.c +12 -0
+       /*
+        * If RTD3 was vetoed before we entered system suspend allow it
+        * again now before driver ready is sent. Firmware sends a new RTD3
+        * veto if it is still the case after we have sent it driver ready
+        * command.
+        */
+       if (icm->veto) {
+               icm->veto = false;
+               pm_runtime_mark_last_busy(&tb->dev);
+               pm_runtime_put_autosuspend(&tb->dev);
+       }
+

Here is the duplication.

Indeed, I'll put it to a helper function.

Quoted 16 lines
Quoted 16 lines
drivers/thunderbolt/nhi.c +16 -0
+static int nhi_suspend_power_down(struct tb *tb)
+{
+       int ret;
+
+       /*
+        * If there is no device connected we need to perform an additional
+        * handshake through LC mailbox and force power down before
+        * entering D3.
+        */
+       ret = device_for_each_child(&tb->root_switch->dev, NULL,
+                                   nhi_device_connected);
+       if (!ret) {
+               lc_mailbox_cmd(tb->nhi, LC_PREPARE_FOR_RESET);
+               ret = lc_mailbox_cmd_complete(tb->nhi,
+                                             LC_MAILBOX_TIMEOUT);
+               if (ret)
Quoted 8 lines

drivers/thunderbolt/nhi.c static irqreturn_t nhi_msi(int irq, void *data) · line 945

Quoted 3 lines
drivers/thunderbolt/nhi.c +3 -0
+                       return ret;
+
+               return nhi_power_down(tb->nhi);

Just to be sure: unforce power is done only if no device is connected? My understanding of the comment above was that unforce power should be done anyway (so it should be outside of this if block), and the difference between the cases is only about the additional LC mailbox message. I guess I misread it.

nhi_power_down() should be only called if no device was connected so it should be in correct place. I can try to clarify the comment a bit, though.

maintainer
#25 lore raw

Re: [PATCH 7/8] thunderbolt: Add support for Intel Ice Lake

Quoted 26 lines · Mika Westerberg
Quoted 17 lines
Quoted 17 lines
drivers/thunderbolt/nhi.c +16 -0
+static int nhi_suspend_power_down(struct tb *tb)
+{
+       int ret;
+
+       /*
+        * If there is no device connected we need to perform an additional
+        * handshake through LC mailbox and force power down before
+        * entering D3.
+        */
+       ret = device_for_each_child(&tb->root_switch->dev, NULL,
+                                   nhi_device_connected);
+       if (!ret) {
+               lc_mailbox_cmd(tb->nhi, LC_PREPARE_FOR_RESET);
+               ret = lc_mailbox_cmd_complete(tb->nhi,
+                                             LC_MAILBOX_TIMEOUT);
+               if (ret)
Quoted 8 lines

drivers/thunderbolt/nhi.c static irqreturn_t nhi_msi(int irq, void *data) · line 945

Quoted 3 lines
drivers/thunderbolt/nhi.c +3 -0
+                       return ret;
+
+               return nhi_power_down(tb->nhi);

Just to be sure: unforce power is done only if no device is connected? My understanding of the comment above was that unforce power should be done anyway (so it should be outside of this if block), and the difference between the cases is only about the additional LC mailbox message. I guess I misread it.

Quoted 3 lines

nhi_power_down() should be only called if no device was connected so it should be in correct place. I can try to clarify the comment a bit, though.

Maybe adding the word "both" ("to perform both an additional") will make it clearer. Maybe removing the "additional" (which to my ears sounds like "an additional operation besides the normal one, to unforce power") is enough. Again, your call. I'm not sure it's strictly needed, maybe it's just me.

Thanks!

#26 lore raw

Re: [PATCH 7/8] thunderbolt: Add support for Intel Ice Lake

In reply to Mika Westerberg

drivers/thunderbolt/nhi.c static int nhi_resume_noirq(struct device *dev) · line 1025

Quoted 7 lines · Mika Westerberg
drivers/thunderbolt/nhi.c +3 -0
@@ -891,16 +1020,23 @@ static int nhi_resume_noirq(struct device *dev)
 {
 	struct pci_dev *pdev = to_pci_dev(dev);
 	struct tb *tb = pci_get_drvdata(pdev);
+	int ret;
+
+	ret = nhi_power_up(tb->nhi);

Some general thoughts:

You're enabling Force Power during the resume_noirq phase of the NHI.

If TBT devices were already connected before system sleep, the PCI core walks down the hierarchy during resume_noirq, puts each device into D0 and restores config space. For this to work, the NHI must have set up tunnels to attached devices. But it can hardly do that until Force Power is enabled. Yet I cannot see any provision here which causes the hotplug bridges to wait for the NHI to finish its resume_noirq phase.

Don't you need that?

On Macs we have quirk_apple_wait_for_thunderbolt() to resolve this issue.

And a commit has been rotting on my development branch for years which replaces quirk_apple_wait_for_thunderbolt() with device links:

https://github.com/l1k/linux/commit/8cbb1d589660

Also, you recently posted a patch stating that on Ice Lake, the NHI and hotplug bridges share an ACPI Power Resource:

https://patchwork.kernel.org/patch/11015233/

How does that relate to the Force Power and Go2Sx dance implemented in the present patch? Does the ACPI Power Resource toggle the same Force Power and Go2Sx registers as this patch here? Or is that an additional power management mechanism? If Force Power is off on the NHI, are the hotplug bridges' config and mmio spaces accessible? All of this should be documented in code comments or at least the commit message.

There's a quirk for Macs with Cactus Ridge to perform a Go2Sx dance during the suspend_noirq phase, quirk_apple_poweroff_thunderbolt(). Can this be reimplemented using the infrastructure you're adding in this patch?

On Macs with Thunderbolt 1 and 2, there are ACPI methods to toggle Force Power in the NHI's namespace. However disabling Force Power not only cuts power to the NHI but also to the integrated PCIe switch. For this reason, I've implemented power management on these Macs using a struct dev_pm_domain which is assigned to the upstream bridge of the integrated PCIe switch. That way, power isn't cut until the Thunderbolt controller's top-most device in the hierarchy is runtime suspended:

https://github.com/l1k/linux/commit/4db7f0b1f5c9

I'm wondering if it would make sense to similarly assign a dev_pm_domain to the NHI on Ice Lake.

One thing that bothers me a bit with the present patch is that nhi.c is cluttered up with code specific to Ice Lake. That doesn't seem sustainable going forward as I expect we'll have to add plenty of other quirks once USB4 chipsets become available. Ideally the file should contain only (or mostly) generic code and quirks should be contained in separate files which need not be compiled in on unrelated arches. That's why in the above-linked commit to add Apple-specific power management, I put all the code in a pm_apple.c and nhi.c only calls tb_pm_apple_init() / _fini(), which become empty inline stubs if CONFIG_ACPI or CONFIG_PM isn't enabled.

drivers/thunderbolt/nhi.c line 49

Quoted 13 lines
drivers/thunderbolt/nhi.c +7 -0
@@ -38,6 +40,60 @@
 #define MSIX_MAX_VECS		16
 
 #define NHI_MAILBOX_TIMEOUT	500 /* ms */
+#define LC_MAILBOX_TIMEOUT	500 /* ms */
+
+enum lc_mailbox_cmd {
+	LC_GO2SX = 0x02,
+	LC_GO2SX_NO_WAKE = 0x03,
+	LC_PREPARE_FOR_RESET = 0x21,
+};

Shouldn't these live together with the other LC macros in tb_regs.h?

Quoted 4 lines
drivers/thunderbolt/nhi.c +4 -0
+	ret = device_for_each_child(&tb->root_switch->dev, NULL,
+				    nhi_device_connected);
+	if (!ret) {
+		lc_mailbox_cmd(tb->nhi, LC_PREPARE_FOR_RESET);

drivers/thunderbolt/nhi.c static irqreturn_t nhi_msi(int irq, void *data) · line 949

Quoted 10 lines
drivers/thunderbolt/nhi.c +10 -0
+		ret = lc_mailbox_cmd_complete(tb->nhi,
+					      LC_MAILBOX_TIMEOUT);
+		if (ret)
+			return ret;
+
+		return nhi_power_down(tb->nhi);
+	}
+
+	return 0;
+}

Why not:

	if (ret)
		return 0;

and that way reduce the indentation of the code in your if-block by 1 tab?

Thanks,

Lukas

#27 lore raw

Re: [PATCH 7/8] thunderbolt: Add support for Intel Ice Lake

Quoted 13 lines · Lukas Wunner

drivers/thunderbolt/nhi.c static int nhi_resume_noirq(struct device *dev) · line 1025

Quoted 7 lines · Mika Westerberg
drivers/thunderbolt/nhi.c +3 -0
@@ -891,16 +1020,23 @@ static int nhi_resume_noirq(struct device *dev)
 {
 	struct pci_dev *pdev = to_pci_dev(dev);
 	struct tb *tb = pci_get_drvdata(pdev);
+	int ret;
+
+	ret = nhi_power_up(tb->nhi);

Some general thoughts:

You're enabling Force Power during the resume_noirq phase of the NHI.

Quoted 8 lines

If TBT devices were already connected before system sleep, the PCI core walks down the hierarchy during resume_noirq, puts each device into D0 and restores config space. For this to work, the NHI must have set up tunnels to attached devices. But it can hardly do that until Force Power is enabled. Yet I cannot see any provision here which causes the hotplug bridges to wait for the NHI to finish its resume_noirq phase.

Don't you need that?

For the FW CM we don't because there is some magic ASL as part of the ACPI power resources (shared between root ports and NHI) that causes the firmware to establish tunnels as part of _ON() method.

Quoted 11 lines

On Macs we have quirk_apple_wait_for_thunderbolt() to resolve this issue.

And a commit has been rotting on my development branch for years which replaces quirk_apple_wait_for_thunderbolt() with device links:

https://github.com/l1k/linux/commit/8cbb1d589660

Also, you recently posted a patch stating that on Ice Lake, the NHI and hotplug bridges share an ACPI Power Resource:

Quoted 6 lines

https://patchwork.kernel.org/patch/11015233/

How does that relate to the Force Power and Go2Sx dance implemented in the present patch? Does the ACPI Power Resource toggle the same Force Power and Go2Sx registers as this patch here? Or is that an additional power management mechanism?

It is an additional mechanism (but I think it also touches force power at some point). Force power here is not the same as with the discrete chips. AFAIK it pretty much only loads and starts the firmware whereas in discrete it actually controls power to the chip.

Quoted 2 lines

If Force Power is off on the NHI, are the hotplug bridges' config and mmio spaces accessible?

Yes, they are all accessible.

Quoted 2 lines

All of this should be documented in code comments or at least the commit message.

Will do.

Quoted 4 lines

There's a quirk for Macs with Cactus Ridge to perform a Go2Sx dance during the suspend_noirq phase, quirk_apple_poweroff_thunderbolt(). Can this be reimplemented using the infrastructure you're adding in this patch?

It might be possible to implement ->force_power for Cactus Ridge and do the ACPI magic there instead. I'll try and see what it looks like.

Quoted 9 lines

On Macs with Thunderbolt 1 and 2, there are ACPI methods to toggle Force Power in the NHI's namespace. However disabling Force Power not only cuts power to the NHI but also to the integrated PCIe switch. For this reason, I've implemented power management on these Macs using a struct dev_pm_domain which is assigned to the upstream bridge of the integrated PCIe switch. That way, power isn't cut until the Thunderbolt controller's top-most device in the hierarchy is runtime suspended:

Quoted 4 lines

https://github.com/l1k/linux/commit/4db7f0b1f5c9

I'm wondering if it would make sense to similarly assign a dev_pm_domain to the NHI on Ice Lake.

Using PM domain here does not feel right to me. The force power thing here is not really about power (but I'm calling it that because that's the what we use internally) so I think it should be part of normal driver flows.

Quoted 7 lines

One thing that bothers me a bit with the present patch is that nhi.c is cluttered up with code specific to Ice Lake. That doesn't seem sustainable going forward as I expect we'll have to add plenty of other quirks once USB4 chipsets become available. Ideally the file should contain only (or mostly) generic code and quirks should be contained in separate files which need not be compiled in on unrelated arches.

I originally intented to have ICL support in icm.c but I realized that the same steps are most likely needed on Apple systems as well (I don't have any so can't tell if that's really the case) and since these flows are for the NHI I figured that's the correct place for it.

I agree with you that moving chip specific things to separate files makes sense (such as icl.c and so on) in the long run to keep the driver maintainable.

Quoted 11 lines

That's why in the above-linked commit to add Apple-specific power management, I put all the code in a pm_apple.c and nhi.c only calls tb_pm_apple_init() / _fini(), which become empty inline stubs if CONFIG_ACPI or CONFIG_PM isn't enabled.

Quoted 5 lines
drivers/thunderbolt/nhi.c +0 -0
@@ -38,6 +40,60 @@
 #define MSIX_MAX_VECS		16
 
Quoted 10 lines

drivers/thunderbolt/nhi.c line 49

Quoted 8 lines
drivers/thunderbolt/nhi.c +7 -0
 #define NHI_MAILBOX_TIMEOUT	500 /* ms */
+#define LC_MAILBOX_TIMEOUT	500 /* ms */
+
+enum lc_mailbox_cmd {
+	LC_GO2SX = 0x02,
+	LC_GO2SX_NO_WAKE = 0x03,
+	LC_PREPARE_FOR_RESET = 0x21,
+};

Shouldn't these live together with the other LC macros in tb_regs.h?

I've used tb_regs.h to keep registers and definitions that belong to Thunderbolt fabric. The other LC macros are there because they are used to access LC over TBT.

Here we access it through NHI and only use it in nhi.c.

No strong feelings moving it to tb_regs.h if you still think that's more suitable.

Quoted 17 lines
Quoted 4 lines
drivers/thunderbolt/nhi.c +4 -0
+	ret = device_for_each_child(&tb->root_switch->dev, NULL,
+				    nhi_device_connected);
+	if (!ret) {
+		lc_mailbox_cmd(tb->nhi, LC_PREPARE_FOR_RESET);

drivers/thunderbolt/nhi.c static irqreturn_t nhi_msi(int irq, void *data) · line 949

Quoted 10 lines
drivers/thunderbolt/nhi.c +10 -0
+		ret = lc_mailbox_cmd_complete(tb->nhi,
+					      LC_MAILBOX_TIMEOUT);
+		if (ret)
+			return ret;
+
+		return nhi_power_down(tb->nhi);
+	}
+
+	return 0;
+}

Why not:

Quoted 4 lines
	if (ret)
		return 0;

and that way reduce the indentation of the code in your if-block by 1 tab?

Indeed. Will do.

#28 lore raw

PATCH 8/8 ACPI / property: Add two new Thunderbolt property GUIDs to the list

In reply to Mika Westerberg

Ice Lake Thunderbolt controller includes two new device property compatible properties that we need to be able to extract in the driver so add them to the growing array of GUIDs.

---
 drivers/acpi/property.c | 6 ++++++
 1 file changed, 6 insertions(+)
drivers/acpi/property.c +6 -0
index da3ced297f19..07cbacbab861 100644
@@ -39,6 +39,12 @@ static const guid_t prp_guids[] = {
 	/* External facing port GUID: efcc06cc-73ac-4bc3-bff0-76143807c389 */
 	GUID_INIT(0xefcc06cc, 0x73ac, 0x4bc3,
 		  0xbf, 0xf0, 0x76, 0x14, 0x38, 0x07, 0xc3, 0x89),
+	/* TBT GUID for IMR_VALID: c44d002f-69f9-4e7d-a904-a7baabdf43f7 */
+	GUID_INIT(0xc44d002f, 0x69f9, 0x4e7d,
+		  0xa9, 0x04, 0xa7, 0xba, 0xab, 0xdf, 0x43, 0xf7),
+	/* TBT GUID for WAKE_SUPPORTED: 6c501103-c189-4296-ba72-9bf5a26ebe5d */
+	GUID_INIT(0x6c501103, 0xc189, 0x4296,
+		  0xba, 0x72, 0x9b, 0xf5, 0xa2, 0x6e, 0xbe, 0x5d),
 };
 
 /* ACPI _DSD data subnodes GUID: dbb8e3e6-5886-4ba6-8795-1319f52a966b */
Signature · 1 line
2.20.1
#29 lore raw

Re: [PATCH 8/8] ACPI / property: Add two new Thunderbolt property GUIDs to the list

Quoted 15 lines · Mika Westerberg

Ice Lake Thunderbolt controller includes two new device property compatible properties that we need to be able to extract in the driver so add them to the growing array of GUIDs.

---
 drivers/acpi/property.c | 6 ++++++
 1 file changed, 6 insertions(+)
drivers/acpi/property.c +0 -0
index da3ced297f19..07cbacbab861 100644

drivers/acpi/property.c static const guid_t prp_guids[] = { · line 42

Quoted 4 lines
drivers/acpi/property.c +1 -0
@@ -39,6 +39,12 @@ static const guid_t prp_guids[] = {
        /* External facing port GUID: efcc06cc-73ac-4bc3-bff0-76143807c389 */
        GUID_INIT(0xefcc06cc, 0x73ac, 0x4bc3,
                  0xbf, 0xf0, 0x76, 0x14, 0x38, 0x07, 0xc3, 0x89),
+       /* TBT GUID for IMR_VALID: c44d002f-69f9-4e7d-a904-a7baabdf43f7 */

I'd prefer Thunderbolt to be spelled in full here (and below).

LGTM otherwise.

Quoted 11 lines
drivers/acpi/property.c +5 -1
+       GUID_INIT(0xc44d002f, 0x69f9, 0x4e7d,
+                 0xa9, 0x04, 0xa7, 0xba, 0xab, 0xdf, 0x43, 0xf7),
+       /* TBT GUID for WAKE_SUPPORTED: 6c501103-c189-4296-ba72-9bf5a26ebe5d */
+       GUID_INIT(0x6c501103, 0xc189, 0x4296,
+                 0xba, 0x72, 0x9b, 0xf5, 0xa2, 0x6e, 0xbe, 0x5d),
 };
 
 /* ACPI _DSD data subnodes GUID: dbb8e3e6-5886-4ba6-8795-1319f52a966b */
--

2.20.1

#30 lore raw
In reply to Mika Westerberg
Quoted 15 lines · Mika Westerberg

Hi all,

With the exception of the first patch which is fix, this series enables Thunderbolt on Intel Ice Lake. Biggest difference from the previous controllers is that the Thunderbolt controller is now integrated as part of the SoC. The firmware messages pretty much follow Titan Ridge but there are some differences as well (such as the new RTD3 veto notification). Also Ice Lake does not implement security levels so DMA protection is handled by IOMMU.

This is v5.4 material but I'm sending it out now because I will be on vacation next 4 weeks mostly without internet access. When I get back I'll gather all the comments and update the series accordingly.

Thanks!

Quoted 9 lines

Mika Westerberg (8):
thunderbolt: Correct path indices for PCIe tunnel
thunderbolt: Move NVM upgrade support flag to struct icm
thunderbolt: Use 32-bit writes when writing ring producer/consumer
thunderbolt: Do not fail adding switch if some port is not implemented
thunderbolt: Hide switch attributes that are not set
thunderbolt: Expose active parts of NVM even if upgrade is not supported
thunderbolt: Add support for Intel Ice Lake
ACPI / property: Add two new Thunderbolt property GUIDs to the list

Forgot to Cc Raanan and Raj, now added. Sorry about that. The patch series can also be viewed here:

https://lore.kernel.org/lkml/[email protected]/T/#m9cb5a393dfc79f1c2212d0787b6bad5b689db6bd

maintainer Reviewed-by
#31 lore raw
Quoted 27 lines · Mika Westerberg
Quoted 15 lines · Mika Westerberg

Hi all,

With the exception of the first patch which is fix, this series enables Thunderbolt on Intel Ice Lake. Biggest difference from the previous controllers is that the Thunderbolt controller is now integrated as part of the SoC. The firmware messages pretty much follow Titan Ridge but there are some differences as well (such as the new RTD3 veto notification). Also Ice Lake does not implement security levels so DMA protection is handled by IOMMU.

This is v5.4 material but I'm sending it out now because I will be on vacation next 4 weeks mostly without internet access. When I get back I'll gather all the comments and update the series accordingly.

Thanks!

Quoted 9 lines

Mika Westerberg (8):
thunderbolt: Correct path indices for PCIe tunnel
thunderbolt: Move NVM upgrade support flag to struct icm
thunderbolt: Use 32-bit writes when writing ring producer/consumer
thunderbolt: Do not fail adding switch if some port is not implemented
thunderbolt: Hide switch attributes that are not set
thunderbolt: Expose active parts of NVM even if upgrade is not supported
thunderbolt: Add support for Intel Ice Lake
ACPI / property: Add two new Thunderbolt property GUIDs to the list

Quoted 4 lines

Forgot to Cc Raanan and Raj, now added. Sorry about that. The patch series can also be viewed here:

https://lore.kernel.org/lkml/[email protected]/T/#m9cb5a393dfc79f1c2212d0787b6bad5b689db6bd

Besides a few comments, LGTM.

For Thunderbolt patches,