Skip to contentSkip to conversation

nvme-pci: Pass the queue to SQ_SIZE/CQ_SIZE macros

PATCH Superseded Benjamin Herrenschmidt 3 participants 11 messages Jul 17, 2019 – Jul 18, 2019 NVMe

v2 of 4 Superseded Latest is v4 →

lore
Share & tools
lore
#0 lore raw

PATCH v2 1/3 nvme-pci: Pass the queue to SQ_SIZE/CQ_SIZE macros

This will make it easier to handle variable queue entry sizes later. No functional change.

Signed-off-by Benjamin Herrenschmidt <benh at kernel.crashing.org>
Reviewed-by Christoph Hellwig <hch at lst.de>
---
 drivers/nvme/host/pci.c | 30 +++++++++++++++---------------
 1 file changed, 15 insertions(+), 15 deletions(-)
drivers/nvme/host/pci.c +15 -15
index dd10cf78f2d3..8f006638452b 100644
@@ -28,8 +28,8 @@
 #include "trace.h"
 #include "nvme.h"
 
-#define SQ_SIZE(depth)		(depth * sizeof(struct nvme_command))
-#define CQ_SIZE(depth)		(depth * sizeof(struct nvme_completion))
+#define SQ_SIZE(q)	((q)->q_depth * sizeof(struct nvme_command))
+#define CQ_SIZE(q)	((q)->q_depth * sizeof(struct nvme_completion))
 
 #define SGES_PER_PAGE	(PAGE_SIZE / sizeof(struct nvme_sgl_desc))
 
@@ -1344,16 +1344,16 @@ static enum blk_eh_timer_return nvme_timeout(struct request *req, bool reserved)
 
 static void nvme_free_queue(struct nvme_queue *nvmeq)
 {
-	dma_free_coherent(nvmeq->dev->dev, CQ_SIZE(nvmeq->q_depth),
+	dma_free_coherent(nvmeq->dev->dev, CQ_SIZE(nvmeq),
 				(void *)nvmeq->cqes, nvmeq->cq_dma_addr);
 	if (!nvmeq->sq_cmds)
 		return;
 
 	if (test_and_clear_bit(NVMEQ_SQ_CMB, &nvmeq->flags)) {
 		pci_free_p2pmem(to_pci_dev(nvmeq->dev->dev),
-				nvmeq->sq_cmds, SQ_SIZE(nvmeq->q_depth));
+				nvmeq->sq_cmds, SQ_SIZE(nvmeq));
 	} else {
-		dma_free_coherent(nvmeq->dev->dev, SQ_SIZE(nvmeq->q_depth),
+		dma_free_coherent(nvmeq->dev->dev, SQ_SIZE(nvmeq),
 				nvmeq->sq_cmds, nvmeq->sq_dma_addr);
 	}
 }
@@ -1433,12 +1433,12 @@ static int nvme_cmb_qdepth(struct nvme_dev *dev, int nr_io_queues,
 }
 
 static int nvme_alloc_sq_cmds(struct nvme_dev *dev, struct nvme_queue *nvmeq,
-				int qid, int depth)
+				int qid)
 {
 	struct pci_dev *pdev = to_pci_dev(dev->dev);
 
 	if (qid && dev->cmb_use_sqes && (dev->cmbsz & NVME_CMBSZ_SQS)) {
-		nvmeq->sq_cmds = pci_alloc_p2pmem(pdev, SQ_SIZE(depth));
+		nvmeq->sq_cmds = pci_alloc_p2pmem(pdev, SQ_SIZE(nvmeq));
 		if (nvmeq->sq_cmds) {
 			nvmeq->sq_dma_addr = pci_p2pmem_virt_to_bus(pdev,
 							nvmeq->sq_cmds);
@@ -1447,11 +1447,11 @@ static int nvme_alloc_sq_cmds(struct nvme_dev *dev, struct nvme_queue *nvmeq,
 				return 0;
 			}
 
-			pci_free_p2pmem(pdev, nvmeq->sq_cmds, SQ_SIZE(depth));
+			pci_free_p2pmem(pdev, nvmeq->sq_cmds, SQ_SIZE(nvmeq));
 		}
 	}
 
-	nvmeq->sq_cmds = dma_alloc_coherent(dev->dev, SQ_SIZE(depth),
+	nvmeq->sq_cmds = dma_alloc_coherent(dev->dev, SQ_SIZE(nvmeq),
 				&nvmeq->sq_dma_addr, GFP_KERNEL);
 	if (!nvmeq->sq_cmds)
 		return -ENOMEM;
@@ -1465,12 +1465,13 @@ static int nvme_alloc_queue(struct nvme_dev *dev, int qid, int depth)
 	if (dev->ctrl.queue_count > qid)
 		return 0;
 
-	nvmeq->cqes = dma_alloc_coherent(dev->dev, CQ_SIZE(depth),
+	nvmeq->q_depth = depth;
+	nvmeq->cqes = dma_alloc_coherent(dev->dev, CQ_SIZE(nvmeq),
 					 &nvmeq->cq_dma_addr, GFP_KERNEL);
 	if (!nvmeq->cqes)
 		goto free_nvmeq;
 
-	if (nvme_alloc_sq_cmds(dev, nvmeq, qid, depth))
+	if (nvme_alloc_sq_cmds(dev, nvmeq, qid))
 		goto free_cqdma;
 
 	nvmeq->dev = dev;
@@ -1479,15 +1480,14 @@ static int nvme_alloc_queue(struct nvme_dev *dev, int qid, int depth)
 	nvmeq->cq_head = 0;
 	nvmeq->cq_phase = 1;
 	nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
-	nvmeq->q_depth = depth;
 	nvmeq->qid = qid;
 	dev->ctrl.queue_count++;
 
 	return 0;
 
  free_cqdma:
-	dma_free_coherent(dev->dev, CQ_SIZE(depth), (void *)nvmeq->cqes,
-							nvmeq->cq_dma_addr);
+	dma_free_coherent(dev->dev, CQ_SIZE(nvmeq), (void *)nvmeq->cqes,
+			  nvmeq->cq_dma_addr);
  free_nvmeq:
 	return -ENOMEM;
 }
@@ -1515,7 +1515,7 @@ static void nvme_init_queue(struct nvme_queue *nvmeq, u16 qid)
 	nvmeq->cq_head = 0;
 	nvmeq->cq_phase = 1;
 	nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
-	memset((void *)nvmeq->cqes, 0, CQ_SIZE(nvmeq->q_depth));
+	memset((void *)nvmeq->cqes, 0, CQ_SIZE(nvmeq));
 	nvme_dbbuf_init(dev, nvmeq, qid);
 	dev->online_queues++;
 	wmb(); /* ensure the first interrupt sees the initialization */
Signature · 1 line
2.17.1
#1 lore raw

PATCH v2 2/3 nvme-pci: Add support for variable IO SQ element size

The size of a submission queue element should always be 6 (64 bytes) by spec.

However some controllers such as Apple's are not properly implementing the standard and require a different size.

This provides the ground work for the subsequent quirks for these controllers.

Signed-off-by Benjamin Herrenschmidt <benh at kernel.crashing.org>
---
 drivers/nvme/host/pci.c | 11 ++++++++---
 include/linux/nvme.h    |  1 +
 2 files changed, 9 insertions(+), 3 deletions(-)
drivers/nvme/host/pci.c +8 -3
index 8f006638452b..1637677afb78 100644
@@ -28,7 +28,7 @@
 #include "trace.h"
 #include "nvme.h"
 
-#define SQ_SIZE(q)	((q)->q_depth * sizeof(struct nvme_command))
+#define SQ_SIZE(q)	((q)->q_depth << (q)->sqes)
 #define CQ_SIZE(q)	((q)->q_depth * sizeof(struct nvme_completion))
 
 #define SGES_PER_PAGE	(PAGE_SIZE / sizeof(struct nvme_sgl_desc))
@@ -100,6 +100,7 @@ struct nvme_dev {
 	unsigned io_queues[HCTX_MAX_TYPES];
 	unsigned int num_vecs;
 	int q_depth;
+	int io_sqes;
 	u32 db_stride;
 	void __iomem *bar;
 	unsigned long bar_mapped_size;
@@ -162,7 +163,7 @@ static inline struct nvme_dev *to_nvme_dev(struct nvme_ctrl *ctrl)
 struct nvme_queue {
 	struct nvme_dev *dev;
 	spinlock_t sq_lock;
-	struct nvme_command *sq_cmds;
+	void *sq_cmds;
 	 /* only used for poll queues: */
 	spinlock_t cq_poll_lock ____cacheline_aligned_in_smp;
 	volatile struct nvme_completion *cqes;
@@ -178,6 +179,7 @@ struct nvme_queue {
 	u16 last_cq_head;
 	u16 qid;
 	u8 cq_phase;
+	u8 sqes;
 	unsigned long flags;
 #define NVMEQ_ENABLED		0
 #define NVMEQ_SQ_CMB		1
@@ -488,7 +490,8 @@ static void nvme_submit_cmd(struct nvme_queue *nvmeq, struct nvme_command *cmd,
 			    bool write_sq)
 {
 	spin_lock(&nvmeq->sq_lock);
-	memcpy(&nvmeq->sq_cmds[nvmeq->sq_tail], cmd, sizeof(*cmd));
+	memcpy(nvmeq->sq_cmds + (nvmeq->sq_tail << nvmeq->sqes),
+	       cmd, sizeof(*cmd));
 	if (++nvmeq->sq_tail == nvmeq->q_depth)
 		nvmeq->sq_tail = 0;
 	nvme_write_sq_db(nvmeq, write_sq);
@@ -1465,6 +1468,7 @@ static int nvme_alloc_queue(struct nvme_dev *dev, int qid, int depth)
 	if (dev->ctrl.queue_count > qid)
 		return 0;
 
+	nvmeq->sqes = qid ? dev->io_sqes : NVME_NVM_ADMSQES;
 	nvmeq->q_depth = depth;
 	nvmeq->cqes = dma_alloc_coherent(dev->dev, CQ_SIZE(nvmeq),
 					 &nvmeq->cq_dma_addr, GFP_KERNEL);
@@ -2318,6 +2322,7 @@ static int nvme_pci_enable(struct nvme_dev *dev)
 				io_queue_depth);
 	dev->db_stride = 1 << NVME_CAP_STRIDE(dev->ctrl.cap);
 	dev->dbs = dev->bar + 4096;
+	dev->io_sqes = NVME_NVM_IOSQES;
 
 	/*
 	 * Temporary fix for the Apple controller found in the MacBook8,1 and
include/linux/nvme.h +1 -0
index 01aa6a6c241d..d5a4bc21f36b 100644
@@ -140,6 +140,7 @@ enum {
  * Submission and Completion Queue Entry Sizes for the NVM command set.
  * (In bytes and specified as a power of two (2^n)).
  */
+#define NVME_NVM_ADMSQES	6
 #define NVME_NVM_IOSQES		6
 #define NVME_NVM_IOCQES		4
 
Signature · 1 line
2.17.1
Reviewed-by on v2 2/3
#2 lore raw

Re: [PATCH v2 2/3] nvme-pci: Add support for variable IO SQ element size

Quoted 37 lines · Benjamin Herrenschmidt

The size of a submission queue element should always be 6 (64 bytes) by spec.

However some controllers such as Apple's are not properly implementing the standard and require a different size.

This provides the ground work for the subsequent quirks for these controllers.

Signed-off-by Benjamin Herrenschmidt <benh at kernel.crashing.org>
---
 drivers/nvme/host/pci.c | 11 ++++++++---
 include/linux/nvme.h    |  1 +
 2 files changed, 9 insertions(+), 3 deletions(-)
drivers/nvme/host/pci.c +2 -1
index 8f006638452b..1637677afb78 100644
@@ -28,7 +28,7 @@
 #include "trace.h"
 #include "nvme.h"
 
-#define SQ_SIZE(q)	((q)->q_depth * sizeof(struct nvme_command))
+#define SQ_SIZE(q)	((q)->q_depth << (q)->sqes)
 #define CQ_SIZE(q)	((q)->q_depth * sizeof(struct nvme_completion))
 
 #define SGES_PER_PAGE	(PAGE_SIZE / sizeof(struct nvme_sgl_desc))
@@ -100,6 +100,7 @@ struct nvme_dev {
 	unsigned io_queues[HCTX_MAX_TYPES];
 	unsigned int num_vecs;
 	int q_depth;
+	int io_sqes;
 	u32 db_stride;
 	void __iomem *bar;
 	unsigned long bar_mapped_size;

drivers/nvme/host/pci.c static inline struct nvme_dev *to_nvme_dev(struct nvme_ctrl *ctrl) · line 166

Quoted 5 lines
drivers/nvme/host/pci.c +1 -1
@@ -162,7 +163,7 @@ static inline struct nvme_dev *to_nvme_dev(struct nvme_ctrl *ctrl)
 struct nvme_queue {
 	struct nvme_dev *dev;
 	spinlock_t sq_lock;
-	struct nvme_command *sq_cmds;
+	void *sq_cmds;

It would be great if it can remain the existing data type for the SQEs... But I'm fine with this also.

It looks good to me.

Reviewed-by Minwoo Im <minwoo.im.dev at gmail.com>

Thanks,

#3 lore raw

Re: [PATCH v2 2/3] nvme-pci: Add support for variable IO SQ element size

Quoted 7 lines · Minwoo Im
drivers/nvme/host/pci.c +0 -1
-	struct nvme_command *sq_cmds;

drivers/nvme/host/pci.c static inline struct nvme_dev *to_nvme_dev(struct nvme_ctrl *ctrl) · line 166

Quoted 1 line
drivers/nvme/host/pci.c +1 -0
+	void *sq_cmds;

It would be great if it can remain the existing data type for the SQEs... But I'm fine with this also.

It looks good to me.

I changed it on purpose so we aren't tempted to index the array, since that's not always valid.

Cheers,
Ben.

#4 lore raw

PATCH 1/1 nvme-pci: Support shared tags across queues for Apple 2018 controllers

Another issue with the Apple T2 based 2018 controllers seem to be that they blow up (and shut the machine down) if there's a tag collision between the IO queue and the Admin queue.

This adds a quirk that offsets all the tags in the IO queue by 32 to avoid those collisions. It also limits the number of IO queues to 1 since the code wouldn't otherwise make sense (the device supports only one queue anyway but better safe than sorry).

The bug is typically triggered by tag collisions between SMART commands from smartd and IO commands, often at boot time.

Signed-off-by Benjamin Herrenschmidt <benh at kernel.crashing.org>

---

Note: This is the smallest way I found of doing this that keeps the impact self contained to pci.c. Feel free to suggest alternatives.

 drivers/nvme/host/nvme.h |  5 +++++
 drivers/nvme/host/pci.c  | 26 ++++++++++++++++++++------
 2 files changed, 25 insertions(+), 6 deletions(-)
drivers/nvme/host/nvme.h +5 -0
index 564b967058f4..eeb99e485898 100644
@@ -102,6 +102,11 @@ enum nvme_quirks {
 	 * Use non-standard 128 bytes SQEs.
 	 */
 	NVME_QUIRK_128_BYTES_SQES		= (1 << 11),
+
+	/*
+	 * Prevent tag overlap between queues
+	 */
+	NVME_QUIRK_SHARED_TAGS			= (1 << 12),
 };
 
 /*
drivers/nvme/host/pci.c +20 -6
index e399e59863c7..1055f19e57a4 100644
@@ -194,6 +194,7 @@ struct nvme_queue {
 	u16 cq_head;
 	u16 last_cq_head;
 	u16 qid;
+	u16 tag_offset;
 	u8 cq_phase;
 	u8 sqes;
 	unsigned long flags;
@@ -506,6 +507,7 @@ static void nvme_submit_cmd(struct nvme_queue *nvmeq, struct nvme_command *cmd,
 			    bool write_sq)
 {
 	spin_lock(&nvmeq->sq_lock);
+	cmd->common.command_id += nvmeq->tag_offset;
 	memcpy(nvmeq->sq_cmds + (nvmeq->sq_tail << nvmeq->sqes),
 	       cmd, sizeof(*cmd));
 	if (++nvmeq->sq_tail == nvmeq->q_depth)
@@ -967,9 +969,10 @@ static inline void nvme_ring_cq_doorbell(struct nvme_queue *nvmeq)
 static inline void nvme_handle_cqe(struct nvme_queue *nvmeq, u16 idx)
 {
 	volatile struct nvme_completion *cqe = &nvmeq->cqes[idx];
+	u16 ctag = cqe->command_id - nvmeq->tag_offset;
 	struct request *req;
 
-	if (unlikely(cqe->command_id >= nvmeq->q_depth)) {
+	if (unlikely(ctag >= nvmeq->q_depth)) {
 		dev_warn(nvmeq->dev->ctrl.device,
 			"invalid id %d completed on queue %d\n",
 			cqe->command_id, le16_to_cpu(cqe->sq_id));
@@ -982,14 +985,13 @@ static inline void nvme_handle_cqe(struct nvme_queue *nvmeq, u16 idx)
 	 * aborts.  We don't even bother to allocate a struct request
 	 * for them but rather special case them here.
 	 */
-	if (unlikely(nvmeq->qid == 0 &&
-			cqe->command_id >= NVME_AQ_BLK_MQ_DEPTH)) {
+	if (unlikely(nvmeq->qid == 0 && ctag >= NVME_AQ_BLK_MQ_DEPTH)) {
 		nvme_complete_async_event(&nvmeq->dev->ctrl,
 				cqe->status, &cqe->result);
 		return;
 	}
 
-	req = blk_mq_tag_to_rq(*nvmeq->tags, cqe->command_id);
+	req = blk_mq_tag_to_rq(*nvmeq->tags, ctag);
 	trace_nvme_sq(req, cqe->sq_head, nvmeq->sq_tail);
 	nvme_end_request(req, cqe->status, cqe->result);
 }
@@ -1020,7 +1022,10 @@ static inline int nvme_process_cq(struct nvme_queue *nvmeq, u16 *start,
 
 	*start = nvmeq->cq_head;
 	while (nvme_cqe_pending(nvmeq)) {
-		if (tag == -1U || nvmeq->cqes[nvmeq->cq_head].command_id == tag)
+		u16 ctag = nvmeq->cqes[nvmeq->cq_head].command_id;
+
+		ctag -= nvmeq->tag_offset;
+		if (tag == -1U || ctag == tag)
 			found++;
 		nvme_update_cq_head(nvmeq);
 	}
@@ -1499,6 +1504,10 @@ static int nvme_alloc_queue(struct nvme_dev *dev, int qid, int depth)
 	nvmeq->qid = qid;
 	dev->ctrl.queue_count++;
 
+	if (qid && (dev->ctrl.quirks & NVME_QUIRK_SHARED_TAGS))
+		nvmeq->tag_offset = NVME_AQ_DEPTH;
+	else
+		nvmeq->tag_offset = 0;
 	return 0;
 
  free_cqdma:
@@ -2110,6 +2119,10 @@ static int nvme_setup_io_queues(struct nvme_dev *dev)
 	unsigned long size;
 
 	nr_io_queues = max_io_queues();
+
+	if (dev->ctrl.quirks & NVME_QUIRK_SHARED_TAGS)
+		nr_io_queues = 1;
+
 	result = nvme_set_queue_count(&dev->ctrl, &nr_io_queues);
 	if (result < 0)
 		return result;
@@ -2957,7 +2970,8 @@ static const struct pci_device_id nvme_id_table[] = {
 	{ PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2003) },
 	{ PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2005),
 		.driver_data = NVME_QUIRK_SINGLE_VECTOR |
-				NVME_QUIRK_128_BYTES_SQES },
+				NVME_QUIRK_128_BYTES_SQES |
+				NVME_QUIRK_SHARED_TAGS },
 	{ 0, }
 };
 MODULE_DEVICE_TABLE(pci, nvme_id_table);
#5 lore raw

Re: [PATCH] nvme-pci: Support shared tags across queues for Apple 2018 controllers

Quoted 12 lines · Benjamin Herrenschmidt

Another issue with the Apple T2 based 2018 controllers seem to be that they blow up (and shut the machine down) if there's a tag collision between the IO queue and the Admin queue.

This adds a quirk that offsets all the tags in the IO queue by 32 to avoid those collisions. It also limits the number of IO queues to 1 since the code wouldn't otherwise make sense (the device supports only one queue anyway but better safe than sorry).

The bug is typically triggered by tag collisions between SMART commands from smartd and IO commands, often at boot time.

Quoted 6 lines
Signed-off-by Benjamin Herrenschmidt <benh at kernel.crashing.org>

---

Note: This is the smallest way I found of doing this that keeps the impact self contained to pci.c. Feel free to suggest alternatives.

Looks like it's not enough ... the bug is a lot harder to hit but I still occasionally get a duplicate tag. I'm now wondering if it's unhappy about having tags bigger than q_depth... I wouldn't be surprised with anything here.

I'll try again with a reduce q_depth as well...

Ben.

Quoted 124 lines
 drivers/nvme/host/nvme.h |  5 +++++
 drivers/nvme/host/pci.c  | 26 ++++++++++++++++++++------
 2 files changed, 25 insertions(+), 6 deletions(-)
drivers/nvme/host/nvme.h +5 -0
index 564b967058f4..eeb99e485898 100644
@@ -102,6 +102,11 @@ enum nvme_quirks {
 	 * Use non-standard 128 bytes SQEs.
 	 */
 	NVME_QUIRK_128_BYTES_SQES		= (1 << 11),
+
+	/*
+	 * Prevent tag overlap between queues
+	 */
+	NVME_QUIRK_SHARED_TAGS			= (1 << 12),
 };
 
 /*
drivers/nvme/host/pci.c +2 -0
index e399e59863c7..1055f19e57a4 100644
@@ -194,6 +194,7 @@ struct nvme_queue {
 	u16 cq_head;
 	u16 last_cq_head;
 	u16 qid;
+	u16 tag_offset;
 	u8 cq_phase;
 	u8 sqes;
 	unsigned long flags;
@@ -506,6 +507,7 @@ static void nvme_submit_cmd(struct nvme_queue
 *nvmeq, struct nvme_command *cmd,
 			    bool write_sq)
 {
 	spin_lock(&nvmeq->sq_lock);
+	cmd->common.command_id += nvmeq->tag_offset;
 	memcpy(nvmeq->sq_cmds + (nvmeq->sq_tail << nvmeq->sqes),
 	       cmd, sizeof(*cmd));
drivers/nvme/host/pci.c +2 -1
 	if (++nvmeq->sq_tail == nvmeq->q_depth)
@@ -967,9 +969,10 @@ static inline void nvme_ring_cq_doorbell(struct
 nvme_queue *nvmeq)
 static inline void nvme_handle_cqe(struct nvme_queue *nvmeq, u16
 idx)
 {
 	volatile struct nvme_completion *cqe = &nvmeq->cqes[idx];
+	u16 ctag = cqe->command_id - nvmeq->tag_offset;
 	struct request *req;
 
-	if (unlikely(cqe->command_id >= nvmeq->q_depth)) {
+	if (unlikely(ctag >= nvmeq->q_depth)) {
 		dev_warn(nvmeq->dev->ctrl.device,
drivers/nvme/host/pci.c +2 -3
 			"invalid id %d completed on queue %d\n",
 			cqe->command_id, le16_to_cpu(cqe->sq_id));
@@ -982,14 +985,13 @@ static inline void nvme_handle_cqe(struct
 nvme_queue *nvmeq, u16 idx)
 	 * aborts.  We don't even bother to allocate a struct request
 	 * for them but rather special case them here.
 	 */
-	if (unlikely(nvmeq->qid == 0 &&
-			cqe->command_id >= NVME_AQ_BLK_MQ_DEPTH)) {
+	if (unlikely(nvmeq->qid == 0 && ctag >= NVME_AQ_BLK_MQ_DEPTH))
 {
 		nvme_complete_async_event(&nvmeq->dev->ctrl,
 				cqe->status, &cqe->result);
 		return;
 	}
 
-	req = blk_mq_tag_to_rq(*nvmeq->tags, cqe->command_id);
+	req = blk_mq_tag_to_rq(*nvmeq->tags, ctag);
 	trace_nvme_sq(req, cqe->sq_head, nvmeq->sq_tail);
drivers/nvme/host/pci.c +4 -1
 	nvme_end_request(req, cqe->status, cqe->result);
 }
@@ -1020,7 +1022,10 @@ static inline int nvme_process_cq(struct
 nvme_queue *nvmeq, u16 *start,
 
 	*start = nvmeq->cq_head;
 	while (nvme_cqe_pending(nvmeq)) {
-		if (tag == -1U || nvmeq->cqes[nvmeq-
 >cq_head].command_id == tag)
+		u16 ctag = nvmeq->cqes[nvmeq->cq_head].command_id;
+
+		ctag -= nvmeq->tag_offset;
+		if (tag == -1U || ctag == tag)
 			found++;
drivers/nvme/host/pci.c +4 -0
 		nvme_update_cq_head(nvmeq);
 	}
@@ -1499,6 +1504,10 @@ static int nvme_alloc_queue(struct nvme_dev
 *dev, int qid, int depth)
 	nvmeq->qid = qid;
 	dev->ctrl.queue_count++;
 
+	if (qid && (dev->ctrl.quirks & NVME_QUIRK_SHARED_TAGS))
+		nvmeq->tag_offset = NVME_AQ_DEPTH;
+	else
+		nvmeq->tag_offset = 0;
 	return 0;
 

free_cqdma:

drivers/nvme/host/pci.c +4 -0
@@ -2110,6 +2119,10 @@ static int nvme_setup_io_queues(struct
 nvme_dev *dev)
 	unsigned long size;
 
 	nr_io_queues = max_io_queues();
+
+	if (dev->ctrl.quirks & NVME_QUIRK_SHARED_TAGS)
+		nr_io_queues = 1;
+
 	result = nvme_set_queue_count(&dev->ctrl, &nr_io_queues);
 	if (result < 0)
drivers/nvme/host/pci.c +2 -1
 		return result;
@@ -2957,7 +2970,8 @@ static const struct pci_device_id
 nvme_id_table[] = {
 	{ PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2003) },
 	{ PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2005),
 		.driver_data = NVME_QUIRK_SINGLE_VECTOR |
-				NVME_QUIRK_128_BYTES_SQES },
+				NVME_QUIRK_128_BYTES_SQES |
+				NVME_QUIRK_SHARED_TAGS },
 	{ 0, }
 };

MODULE_DEVICE_TABLE(pci, nvme_id_table);

#6 lore raw

PATCH v2 3/3 nvme-pci: Add support for Apple 2018+ models

In reply to Benjamin Herrenschmidt

Based on reverse engineering and original patch by

Paul Pawlowski <paul at mrarm.io>

This adds support for Apple weird implementation of NVME in their 2018 or later machines. It accounts for the twice-as-big SQ entries for the IO queues, and the fact that only interrupt vector 0 appears to function properly.

Signed-off-by Benjamin Herrenschmidt <benh at kernel.crashing.org>

# Conflicts:
# drivers/nvme/host/core.c

---
 drivers/nvme/host/nvme.h | 10 ++++++++++
 drivers/nvme/host/pci.c  | 21 ++++++++++++++++++++-
 2 files changed, 30 insertions(+), 1 deletion(-)
drivers/nvme/host/nvme.h +10 -0
index 716a876119c8..ced0e0a7e039 100644
@@ -92,6 +92,16 @@ enum nvme_quirks {
 	 * Broken Write Zeroes.
 	 */
 	NVME_QUIRK_DISABLE_WRITE_ZEROES		= (1 << 9),
+
+	/*
+	 * Use only one interrupt vector for all queues
+	 */
+	NVME_QUIRK_SINGLE_VECTOR		= (1 << 10),
+
+	/*
+	 * Use non-standard 128 bytes SQEs.
+	 */
+	NVME_QUIRK_128_BYTES_SQES		= (1 << 11),
 };
 
 /*
drivers/nvme/host/pci.c +20 -1
index 1637677afb78..7088971d4c42 100644
@@ -2081,6 +2081,13 @@ static int nvme_setup_irqs(struct nvme_dev *dev, unsigned int nr_io_queues)
 	dev->io_queues[HCTX_TYPE_DEFAULT] = 1;
 	dev->io_queues[HCTX_TYPE_READ] = 0;
 
+	/*
+	 * Some Apple controllers require all queues to use the
+	 * first vector.
+	 */
+	if (dev->ctrl.quirks & NVME_QUIRK_SINGLE_VECTOR)
+		irq_queues = 1;
+
 	return pci_alloc_irq_vectors_affinity(pdev, 1, irq_queues,
 			      PCI_IRQ_ALL_TYPES | PCI_IRQ_AFFINITY, &affd);
 }
@@ -2322,7 +2329,16 @@ static int nvme_pci_enable(struct nvme_dev *dev)
 				io_queue_depth);
 	dev->db_stride = 1 << NVME_CAP_STRIDE(dev->ctrl.cap);
 	dev->dbs = dev->bar + 4096;
-	dev->io_sqes = NVME_NVM_IOSQES;
+
+	/*
+	 * Some Apple controllers require a non-standard SQE size.
+	 * Interestingly they also seem to ignore the CC:IOSQES register
+	 * so we don't bother updating it here.
+	 */
+	if (dev->ctrl.quirks & NVME_QUIRK_128_BYTES_SQES)
+		dev->io_sqes = 7;
+	else
+		dev->io_sqes = NVME_NVM_IOSQES;
 
 	/*
 	 * Temporary fix for the Apple controller found in the MacBook8,1 and
@@ -3039,6 +3055,9 @@ static const struct pci_device_id nvme_id_table[] = {
 	{ PCI_DEVICE_CLASS(PCI_CLASS_STORAGE_EXPRESS, 0xffffff) },
 	{ PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2001) },
 	{ PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2003) },
+	{ PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2005),
+		.driver_data = NVME_QUIRK_SINGLE_VECTOR |
+				NVME_QUIRK_128_BYTES_SQES },
 	{ 0, }
 };
 MODULE_DEVICE_TABLE(pci, nvme_id_table);
Signature · 1 line
2.17.1
#7 lore raw

Re: [PATCH v2 3/3] nvme-pci: Add support for Apple 2018+ models

Quoted 2 lines

# Conflicts:
# drivers/nvme/host/core.c

I thought you were going to fix this up :)

But I can do that and this version of the series looks fine to me.

#8 lore raw

Re: [PATCH v2 3/3] nvme-pci: Add support for Apple 2018+ models

Quoted 4 lines · Christoph Hellwig
Quoted 2 lines

# Conflicts:
# drivers/nvme/host/core.c

I thought you were going to fix this up :)

Haha yeah I was ...

Quoted 1 line

But I can do that and this version of the series looks fine to me.

Thanks !

Cheers,
Ben.

Reviewed-by on v2 3/3
#9 lore raw

Re: [PATCH v2 3/3] nvme-pci: Add support for Apple 2018+ models

In reply to Benjamin Herrenschmidt
Quoted 58 lines · Benjamin Herrenschmidt

Based on reverse engineering and original patch by

Paul Pawlowski <paul at mrarm.io>

This adds support for Apple weird implementation of NVME in their 2018 or later machines. It accounts for the twice-as-big SQ entries for the IO queues, and the fact that only interrupt vector 0 appears to function properly.

Signed-off-by Benjamin Herrenschmidt <benh at kernel.crashing.org>

# Conflicts:
# drivers/nvme/host/core.c

---
 drivers/nvme/host/nvme.h | 10 ++++++++++
 drivers/nvme/host/pci.c  | 21 ++++++++++++++++++++-
 2 files changed, 30 insertions(+), 1 deletion(-)
drivers/nvme/host/nvme.h +10 -0
index 716a876119c8..ced0e0a7e039 100644
@@ -92,6 +92,16 @@ enum nvme_quirks {
 	 * Broken Write Zeroes.
 	 */
 	NVME_QUIRK_DISABLE_WRITE_ZEROES		= (1 << 9),
+
+	/*
+	 * Use only one interrupt vector for all queues
+	 */
+	NVME_QUIRK_SINGLE_VECTOR		= (1 << 10),
+
+	/*
+	 * Use non-standard 128 bytes SQEs.
+	 */
+	NVME_QUIRK_128_BYTES_SQES		= (1 << 11),
 };
 
 /*
drivers/nvme/host/pci.c +7 -0
index 1637677afb78..7088971d4c42 100644
@@ -2081,6 +2081,13 @@ static int nvme_setup_irqs(struct nvme_dev *dev, unsigned int nr_io_queues)
 	dev->io_queues[HCTX_TYPE_DEFAULT] = 1;
 	dev->io_queues[HCTX_TYPE_READ] = 0;
 
+	/*
+	 * Some Apple controllers require all queues to use the
+	 * first vector.
+	 */
+	if (dev->ctrl.quirks & NVME_QUIRK_SINGLE_VECTOR)
+		irq_queues = 1;
+
 	return pci_alloc_irq_vectors_affinity(pdev, 1, irq_queues,
 			      PCI_IRQ_ALL_TYPES | PCI_IRQ_AFFINITY, &affd);
 }

drivers/nvme/host/pci.c static int nvme_pci_enable(struct nvme_dev *dev) · line 2337

Quoted 10 lines
drivers/nvme/host/pci.c +6 -1
@@ -2322,7 +2329,16 @@ static int nvme_pci_enable(struct nvme_dev *dev)
 				io_queue_depth);
 	dev->db_stride = 1 << NVME_CAP_STRIDE(dev->ctrl.cap);
 	dev->dbs = dev->bar + 4096;
-	dev->io_sqes = NVME_NVM_IOSQES;
+
+	/*
+	 * Some Apple controllers require a non-standard SQE size.
+	 * Interestingly they also seem to ignore the CC:IOSQES register
+	 * so we don't bother updating it here.
+	 */

That is really interesting.

This also looks good to me.

Reviewed-by Minwoo Im <minwoo.im.dev at gmail.com>

Thanks,

Reviewed-by on v2 1/3
#10 lore raw
In reply to Benjamin Herrenschmidt
Quoted 5 lines · Benjamin Herrenschmidt

This will make it easier to handle variable queue entry sizes later. No functional change.

Signed-off-by Benjamin Herrenschmidt <benh at kernel.crashing.org>
Reviewed-by Christoph Hellwig <hch at lst.de>
Reviewed-by Minwoo Im <minwoo.im.dev at gmail.com>