Add support for ANS2 NVMe on Apple A11 SoC.

This version of ANS2 is less quirky than the one in M1, and does not have
NVMMU or Linear SQ. However, it still requires a non-standard 128-byte
SQE.

Signed-off-by: Nick Chan <towinchenmi@gmail.com>
---
 drivers/nvme/host/apple.c | 196 ++++++++++++++++++++++++++++++++--------------
 1 file changed, 136 insertions(+), 60 deletions(-)

diff --git a/drivers/nvme/host/apple.c b/drivers/nvme/host/apple.c
index 1286c31320e630cb012009d6b962526e0553869f..d899ea3684af1d02daa236c77c973f5831d8f593 100644
--- a/drivers/nvme/host/apple.c
+++ b/drivers/nvme/host/apple.c
@@ -35,7 +35,6 @@
 #include "nvme.h"
 
 #define APPLE_ANS_BOOT_TIMEOUT	  USEC_PER_SEC
-#define APPLE_ANS_MAX_QUEUE_DEPTH 64
 
 #define APPLE_ANS_COPROC_CPU_CONTROL	 0x44
 #define APPLE_ANS_COPROC_CPU_CONTROL_RUN BIT(4)
@@ -75,6 +74,8 @@
 #define APPLE_NVME_AQ_DEPTH	   2
 #define APPLE_NVME_AQ_MQ_TAG_DEPTH (APPLE_NVME_AQ_DEPTH - 1)
 
+#define APPLE_NVME_IOSQES	7
+
 /*
  * These can be higher, but we need to ensure that any command doesn't
  * require an sg allocation that needs more than a page of data.
@@ -142,6 +143,7 @@ struct apple_nvme_queue {
 	u32 __iomem *sq_db;
 	u32 __iomem *cq_db;
 
+	u16 sq_tail;
 	u16 cq_head;
 	u8 cq_phase;
 
@@ -166,11 +168,17 @@ struct apple_nvme_iod {
 	struct scatterlist *sg;
 };
 
+struct apple_nvme_hw {
+	bool has_lsq_nvmmu;
+	u32 max_queue_depth;
+};
+
 struct apple_nvme {
 	struct device *dev;
 
 	void __iomem *mmio_coproc;
 	void __iomem *mmio_nvme;
+	const struct apple_nvme_hw *hw;
 
 	struct device **pd_dev;
 	struct device_link **pd_link;
@@ -215,10 +223,12 @@ static inline struct apple_nvme *queue_to_apple_nvme(struct apple_nvme_queue *q)
 
 static unsigned int apple_nvme_queue_depth(struct apple_nvme_queue *q)
 {
-	if (q->is_adminq)
+	struct apple_nvme *anv = queue_to_apple_nvme(q);
+
+	if (q->is_adminq && anv->hw->has_lsq_nvmmu)
 		return APPLE_NVME_AQ_DEPTH;
 
-	return APPLE_ANS_MAX_QUEUE_DEPTH;
+	return anv->hw->max_queue_depth;
 }
 
 static void apple_nvme_rtkit_crashed(void *cookie, const void *crashlog, size_t crashlog_size)
@@ -280,7 +290,28 @@ static void apple_nvmmu_inval(struct apple_nvme_queue *q, unsigned int tag)
 				     "NVMMU TCB invalidation failed\n");
 }
 
-static void apple_nvme_submit_cmd(struct apple_nvme_queue *q,
+static void apple_nvme_submit_cmd_t8015(struct apple_nvme_queue *q,
+				  struct nvme_command *cmd)
+{
+	struct apple_nvme *anv = queue_to_apple_nvme(q);
+
+	spin_lock_irq(&anv->lock);
+
+	if (q->is_adminq)
+		memcpy(&q->sqes[q->sq_tail], cmd, sizeof(*cmd));
+	else
+		memcpy((void *)q->sqes + (q->sq_tail << APPLE_NVME_IOSQES),
+			cmd, sizeof(*cmd));
+
+	if (++q->sq_tail == anv->hw->max_queue_depth)
+		q->sq_tail = 0;
+
+	writel(q->sq_tail, q->sq_db);
+	spin_unlock_irq(&anv->lock);
+}
+
+
+static void apple_nvme_submit_cmd_t8103(struct apple_nvme_queue *q,
 				  struct nvme_command *cmd)
 {
 	struct apple_nvme *anv = queue_to_apple_nvme(q);
@@ -590,7 +621,8 @@ static inline void apple_nvme_handle_cqe(struct apple_nvme_queue *q,
 	__u16 command_id = READ_ONCE(cqe->command_id);
 	struct request *req;
 
-	apple_nvmmu_inval(q, command_id);
+	if (anv->hw->has_lsq_nvmmu)
+		apple_nvmmu_inval(q, command_id);
 
 	req = nvme_find_rq(apple_nvme_queue_tagset(anv, q), command_id);
 	if (unlikely(!req)) {
@@ -685,7 +717,7 @@ static int apple_nvme_create_cq(struct apple_nvme *anv)
 	c.create_cq.opcode = nvme_admin_create_cq;
 	c.create_cq.prp1 = cpu_to_le64(anv->ioq.cq_dma_addr);
 	c.create_cq.cqid = cpu_to_le16(1);
-	c.create_cq.qsize = cpu_to_le16(APPLE_ANS_MAX_QUEUE_DEPTH - 1);
+	c.create_cq.qsize = cpu_to_le16(anv->hw->max_queue_depth - 1);
 	c.create_cq.cq_flags = cpu_to_le16(NVME_QUEUE_PHYS_CONTIG | NVME_CQ_IRQ_ENABLED);
 	c.create_cq.irq_vector = cpu_to_le16(0);
 
@@ -713,7 +745,7 @@ static int apple_nvme_create_sq(struct apple_nvme *anv)
 	c.create_sq.opcode = nvme_admin_create_sq;
 	c.create_sq.prp1 = cpu_to_le64(anv->ioq.sq_dma_addr);
 	c.create_sq.sqid = cpu_to_le16(1);
-	c.create_sq.qsize = cpu_to_le16(APPLE_ANS_MAX_QUEUE_DEPTH - 1);
+	c.create_sq.qsize = cpu_to_le16(anv->hw->max_queue_depth - 1);
 	c.create_sq.sq_flags = cpu_to_le16(NVME_QUEUE_PHYS_CONTIG);
 	c.create_sq.cqid = cpu_to_le16(1);
 
@@ -765,7 +797,12 @@ static blk_status_t apple_nvme_queue_rq(struct blk_mq_hw_ctx *hctx,
 	}
 
 	nvme_start_request(req);
-	apple_nvme_submit_cmd(q, cmnd);
+
+	if (anv->hw->has_lsq_nvmmu)
+		apple_nvme_submit_cmd_t8103(q, cmnd);
+	else
+		apple_nvme_submit_cmd_t8015(q, cmnd);
+
 	return BLK_STS_OK;
 
 out_free_cmd:
@@ -970,11 +1007,13 @@ static const struct blk_mq_ops apple_nvme_mq_ops = {
 static void apple_nvme_init_queue(struct apple_nvme_queue *q)
 {
 	unsigned int depth = apple_nvme_queue_depth(q);
+	struct apple_nvme *anv = queue_to_apple_nvme(q);
 
 	q->cq_head = 0;
 	q->cq_phase = 1;
-	memset(q->tcbs, 0,
-	       APPLE_ANS_MAX_QUEUE_DEPTH * sizeof(struct apple_nvmmu_tcb));
+	if (anv->hw->has_lsq_nvmmu)
+		memset(q->tcbs, 0, anv->hw->max_queue_depth
+			* sizeof(struct apple_nvmmu_tcb));
 	memset(q->cqes, 0, depth * sizeof(struct nvme_completion));
 	WRITE_ONCE(q->enabled, true);
 	wmb(); /* ensure the first interrupt sees the initialization */
@@ -1069,49 +1108,54 @@ static void apple_nvme_reset_work(struct work_struct *work)
 
 	dma_set_max_seg_size(anv->dev, 0xffffffff);
 
-	/*
-	 * Enable NVMMU and linear submission queues.
-	 * While we could keep those disabled and pretend this is slightly
-	 * more common NVMe controller we'd still need some quirks (e.g.
-	 * sq entries will be 128 bytes) and Apple might drop support for
-	 * that mode in the future.
-	 */
-	writel(APPLE_ANS_LINEAR_SQ_EN,
-	       anv->mmio_nvme + APPLE_ANS_LINEAR_SQ_CTRL);
+	if (anv->hw->has_lsq_nvmmu) {
+		/*
+		 * Enable NVMMU and linear submission queues which is required
+		 * since T6000.
+		 */
+		writel(APPLE_ANS_LINEAR_SQ_EN,
+			anv->mmio_nvme + APPLE_ANS_LINEAR_SQ_CTRL);
 
-	/* Allow as many pending command as possible for both queues */
-	writel(APPLE_ANS_MAX_QUEUE_DEPTH | (APPLE_ANS_MAX_QUEUE_DEPTH << 16),
-	       anv->mmio_nvme + APPLE_ANS_MAX_PEND_CMDS_CTRL);
+		/* Allow as many pending command as possible for both queues */
+		writel(anv->hw->max_queue_depth | (anv->hw->max_queue_depth << 16),
+			anv->mmio_nvme + APPLE_ANS_MAX_PEND_CMDS_CTRL);
 
-	/* Setup the NVMMU for the maximum admin and IO queue depth */
-	writel(APPLE_ANS_MAX_QUEUE_DEPTH - 1,
-	       anv->mmio_nvme + APPLE_NVMMU_NUM_TCBS);
+		/* Setup the NVMMU for the maximum admin and IO queue depth */
+		writel(anv->hw->max_queue_depth - 1,
+			anv->mmio_nvme + APPLE_NVMMU_NUM_TCBS);
 
-	/*
-	 * This is probably a chicken bit: without it all commands where any PRP
-	 * is set to zero (including those that don't use that field) fail and
-	 * the co-processor complains about "completed with err BAD_CMD-" or
-	 * a "NULL_PRP_PTR_ERR" in the syslog
-	 */
-	writel(readl(anv->mmio_nvme + APPLE_ANS_UNKNOWN_CTRL) &
-		       ~APPLE_ANS_PRP_NULL_CHECK,
-	       anv->mmio_nvme + APPLE_ANS_UNKNOWN_CTRL);
+		/*
+		 * This is probably a chicken bit: without it all commands
+		 * where any PRP is set to zero (including those that don't use
+		 * that field) fail and the co-processor complains about
+		 * "completed with err BAD_CMD-" or a "NULL_PRP_PTR_ERR" in the
+		 * syslog
+		 */
+		writel(readl(anv->mmio_nvme + APPLE_ANS_UNKNOWN_CTRL) &
+			~APPLE_ANS_PRP_NULL_CHECK,
+			anv->mmio_nvme + APPLE_ANS_UNKNOWN_CTRL);
+	}
 
 	/* Setup the admin queue */
-	aqa = APPLE_NVME_AQ_DEPTH - 1;
+	if (anv->hw->has_lsq_nvmmu)
+		aqa = APPLE_NVME_AQ_DEPTH - 1;
+	else
+		aqa = anv->hw->max_queue_depth - 1;
 	aqa |= aqa << 16;
 	writel(aqa, anv->mmio_nvme + NVME_REG_AQA);
 	writeq(anv->adminq.sq_dma_addr, anv->mmio_nvme + NVME_REG_ASQ);
 	writeq(anv->adminq.cq_dma_addr, anv->mmio_nvme + NVME_REG_ACQ);
 
-	/* Setup NVMMU for both queues */
-	writeq(anv->adminq.tcb_dma_addr,
-	       anv->mmio_nvme + APPLE_NVMMU_ASQ_TCB_BASE);
-	writeq(anv->ioq.tcb_dma_addr,
-	       anv->mmio_nvme + APPLE_NVMMU_IOSQ_TCB_BASE);
+	if (anv->hw->has_lsq_nvmmu) {
+		/* Setup NVMMU for both queues */
+		writeq(anv->adminq.tcb_dma_addr,
+			anv->mmio_nvme + APPLE_NVMMU_ASQ_TCB_BASE);
+		writeq(anv->ioq.tcb_dma_addr,
+			anv->mmio_nvme + APPLE_NVMMU_IOSQ_TCB_BASE);
+	}
 
 	anv->ctrl.sqsize =
-		APPLE_ANS_MAX_QUEUE_DEPTH - 1; /* 0's based queue depth */
+		anv->hw->max_queue_depth - 1; /* 0's based queue depth */
 	anv->ctrl.cap = readq(anv->mmio_nvme + NVME_REG_CAP);
 
 	dev_dbg(anv->dev, "Enabling controller now");
@@ -1282,8 +1326,9 @@ static int apple_nvme_alloc_tagsets(struct apple_nvme *anv)
 	 * both queues. The admin queue gets the first APPLE_NVME_AQ_DEPTH which
 	 * must be marked as reserved in the IO queue.
 	 */
-	anv->tagset.reserved_tags = APPLE_NVME_AQ_DEPTH;
-	anv->tagset.queue_depth = APPLE_ANS_MAX_QUEUE_DEPTH - 1;
+	if (anv->hw->has_lsq_nvmmu)
+		anv->tagset.reserved_tags = APPLE_NVME_AQ_DEPTH;
+	anv->tagset.queue_depth = anv->hw->max_queue_depth - 1;
 	anv->tagset.timeout = NVME_IO_TIMEOUT;
 	anv->tagset.numa_node = NUMA_NO_NODE;
 	anv->tagset.cmd_size = sizeof(struct apple_nvme_iod);
@@ -1307,6 +1352,7 @@ static int apple_nvme_queue_alloc(struct apple_nvme *anv,
 				  struct apple_nvme_queue *q)
 {
 	unsigned int depth = apple_nvme_queue_depth(q);
+	size_t iosq_size;
 
 	q->cqes = dmam_alloc_coherent(anv->dev,
 				      depth * sizeof(struct nvme_completion),
@@ -1314,22 +1360,28 @@ static int apple_nvme_queue_alloc(struct apple_nvme *anv,
 	if (!q->cqes)
 		return -ENOMEM;
 
-	q->sqes = dmam_alloc_coherent(anv->dev,
-				      depth * sizeof(struct nvme_command),
+	if (anv->hw->has_lsq_nvmmu)
+		iosq_size = depth * sizeof(struct nvme_command);
+	else
+		iosq_size = depth << APPLE_NVME_IOSQES;
+
+	q->sqes = dmam_alloc_coherent(anv->dev, iosq_size,
 				      &q->sq_dma_addr, GFP_KERNEL);
 	if (!q->sqes)
 		return -ENOMEM;
 
-	/*
-	 * We need the maximum queue depth here because the NVMMU only has a
-	 * single depth configuration shared between both queues.
-	 */
-	q->tcbs = dmam_alloc_coherent(anv->dev,
-				      APPLE_ANS_MAX_QUEUE_DEPTH *
-					      sizeof(struct apple_nvmmu_tcb),
-				      &q->tcb_dma_addr, GFP_KERNEL);
-	if (!q->tcbs)
-		return -ENOMEM;
+	if (anv->hw->has_lsq_nvmmu) {
+		/*
+		 * We need the maximum queue depth here because the NVMMU only
+		 * has a single depth configuration shared between both queues.
+		 */
+		q->tcbs = dmam_alloc_coherent(anv->dev,
+			anv->hw->max_queue_depth *
+				sizeof(struct apple_nvmmu_tcb),
+			&q->tcb_dma_addr, GFP_KERNEL);
+		if (!q->tcbs)
+			return -ENOMEM;
+	}
 
 	/*
 	 * initialize phase to make sure the allocated and empty memory
@@ -1413,6 +1465,12 @@ static struct apple_nvme *apple_nvme_alloc(struct platform_device *pdev)
 	anv->adminq.is_adminq = true;
 	platform_set_drvdata(pdev, anv);
 
+	anv->hw = of_device_get_match_data(&pdev->dev);
+	if (!anv->hw) {
+		ret = -ENODEV;
+		goto put_dev;
+	}
+
 	ret = apple_nvme_attach_genpd(anv);
 	if (ret < 0) {
 		dev_err_probe(dev, ret, "Failed to attach power domains");
@@ -1444,10 +1502,17 @@ static struct apple_nvme *apple_nvme_alloc(struct platform_device *pdev)
 		goto put_dev;
 	}
 
-	anv->adminq.sq_db = anv->mmio_nvme + APPLE_ANS_LINEAR_ASQ_DB;
-	anv->adminq.cq_db = anv->mmio_nvme + APPLE_ANS_ACQ_DB;
-	anv->ioq.sq_db = anv->mmio_nvme + APPLE_ANS_LINEAR_IOSQ_DB;
-	anv->ioq.cq_db = anv->mmio_nvme + APPLE_ANS_IOCQ_DB;
+	if (anv->hw->has_lsq_nvmmu) {
+		anv->adminq.sq_db = anv->mmio_nvme + APPLE_ANS_LINEAR_ASQ_DB;
+		anv->adminq.cq_db = anv->mmio_nvme + APPLE_ANS_ACQ_DB;
+		anv->ioq.sq_db = anv->mmio_nvme + APPLE_ANS_LINEAR_IOSQ_DB;
+		anv->ioq.cq_db = anv->mmio_nvme + APPLE_ANS_IOCQ_DB;
+	} else {
+		anv->adminq.sq_db = anv->mmio_nvme + NVME_REG_DBS;
+		anv->adminq.cq_db = anv->mmio_nvme + APPLE_ANS_ACQ_DB;
+		anv->ioq.sq_db = anv->mmio_nvme + NVME_REG_DBS + 8;
+		anv->ioq.cq_db = anv->mmio_nvme + APPLE_ANS_IOCQ_DB;
+	}
 
 	anv->sart = devm_apple_sart_get(dev);
 	if (IS_ERR(anv->sart)) {
@@ -1625,8 +1690,19 @@ static int apple_nvme_suspend(struct device *dev)
 static DEFINE_SIMPLE_DEV_PM_OPS(apple_nvme_pm_ops, apple_nvme_suspend,
 				apple_nvme_resume);
 
+static const struct apple_nvme_hw apple_nvme_t8015_hw = {
+	.has_lsq_nvmmu = false,
+	.max_queue_depth = 16,
+};
+
+static const struct apple_nvme_hw apple_nvme_t8103_hw = {
+	.has_lsq_nvmmu = true,
+	.max_queue_depth = 64,
+};
+
 static const struct of_device_id apple_nvme_of_match[] = {
-	{ .compatible = "apple,nvme-ans2" },
+	{ .compatible = "apple,t8015-nvme-ans2", .data = &apple_nvme_t8015_hw },
+	{ .compatible = "apple,nvme-ans2", .data = &apple_nvme_t8103_hw },
 	{},
 };
 MODULE_DEVICE_TABLE(of, apple_nvme_of_match);

-- 
2.50.1

On Thu, Aug 21, 2025 at 11:56:44PM +0800, Nick Chan wrote:
> +		writel(anv->hw->max_queue_depth | (anv->hw->max_queue_depth << 16),

One long line left here.

Otherwise this looks fine to me.

Do you want to merge this through the apple SOC tree?  If so:

Acked-by: Christoph Hellwig <hch@lst.de>

On 25.08.25 10:07, Christoph Hellwig wrote:
> On Thu, Aug 21, 2025 at 11:56:44PM +0800, Nick Chan wrote:
>> +		writel(anv->hw->max_queue_depth | (anv->hw->max_queue_depth << 16),
> 
> One long line left here.
> 
> Otherwise this looks fine to me.
> 
> Do you want to merge this through the apple SOC tree?  If so:
> 
> Acked-by: Christoph Hellwig <hch@lst.de>
> 
> 

I don't think that's necessary since there are no build time 
dependencies but if you want to I can take it through there.



Sven

On Wed, Aug 27, 2025 at 05:56:33PM +0200, Sven Peter wrote:
>> Do you want to merge this through the apple SOC tree?  If so:
>>
>> Acked-by: Christoph Hellwig <hch@lst.de>
>>
>>
>
> I don't think that's necessary since there are no build time dependencies 
> but if you want to I can take it through there.

Merging it through nvme sounds fine as well, I just through up
grabbing everything together would be easier.  I also noticed there's
another Apple hw enablement series that touches nvme, so I guess both
should go through the same tree?

On Tue, Sep 02, 2025 at 07:26:46AM +0200, Christoph Hellwig wrote:
> On Wed, Aug 27, 2025 at 05:56:33PM +0200, Sven Peter wrote:
> >> Do you want to merge this through the apple SOC tree?  If so:
> >>
> >> Acked-by: Christoph Hellwig <hch@lst.de>
> >>
> >>
> >
> > I don't think that's necessary since there are no build time dependencies 
> > but if you want to I can take it through there.
> 
> Merging it through nvme sounds fine as well, I just through up
> grabbing everything together would be easier.  I also noticed there's
> another Apple hw enablement series that touches nvme, so I guess both
> should go through the same tree?

yes, they should go through the same tree if they go in the same cycle.
They conflict in dt-bindings and possibly the driver. We should avoid
burden someone else with this conflict resolution.

Janne

On 03.09.25 11:53, Janne Grunau wrote:
> On Tue, Sep 02, 2025 at 07:26:46AM +0200, Christoph Hellwig wrote:
>> On Wed, Aug 27, 2025 at 05:56:33PM +0200, Sven Peter wrote:
>>>> Do you want to merge this through the apple SOC tree?  If so:
>>>>
>>>> Acked-by: Christoph Hellwig <hch@lst.de>
>>>>
>>>>
>>>
>>> I don't think that's necessary since there are no build time dependencies
>>> but if you want to I can take it through there.
>>
>> Merging it through nvme sounds fine as well, I just through up
>> grabbing everything together would be easier.  I also noticed there's
>> another Apple hw enablement series that touches nvme, so I guess both
>> should go through the same tree?
> 
> yes, they should go through the same tree if they go in the same cycle.
> They conflict in dt-bindings and possibly the driver. We should avoid
> burden someone else with this conflict resolution.

Good points, I'll pick up the latest version of this series on the 
weekend then and handle any conflicts in case the other one is ready for 
this cycle (which it probably will be) as well then.


Thanks,


Sven