Re: [PATCH 2/2] nvme-apple: Prevent tag collision across queues even if tag space is shared

From: Nick Chan

Date: Sat Jun 06 2026 - 11:51:19 EST




David Laight 於 2026/6/6 晚上10:29 寫道:
> On Sat, 06 Jun 2026 21:25:26 +0800
> Nick Chan <towinchenmi@xxxxxxxxx> wrote:
>
>> From: Yuriy Havrylyuk <yhavry@xxxxxxxxx>
>>
>> Apple NVMe controllers require tags of pending commands to not be shared
>> across admin and IO queues. However, on Apple A11 without linear SQ, it is
>> not possible for either queue to skip over some tags and must go from 0 to
>> the configured maximum before wrapping around.
>>
>> If a pending command tag is duplicated across queues, the firmware
>> crashes with: "duplicate tag error for tag N", with N being the tag.
>>
>> Instead of partitioning the tag space, which is not possible without
>> linear SQ, prevent tag collisions by keeping track of which tags are
>> currently in-flight across either queues, and return BLK_STS_RESOURCE to
>> temporaily block command submission when a collision would have occurred.
>
> I look at using the atomic64_xxx() functions rather than the bitmask ones.
> The for_each_bit_set() loop is then an atmomic64_andnot() call.

That does in fact simplify the loop code. However, using the atomic function
complicates the apple_nvme_reserve_tag_t8015() and
apple_nvme_release_tag_t8015() since those functions deal with set/clear a
a bit in terms of an integer (the tag).

Especially in the apple_nvme_reserve_tag_t8015() function the function body
is then

u64 tag_bit = BIT(nvme_tag_from_cid(cmd->common.command_id));

return !(atomic64_fetch_or(tag_bit, &anv->t8015_active_tags) & tag_bit);

The function would need to explictly convert the tag to a bit, and then
explictly extract the bit value after performing the atomic operation,
both which of could have been done by test_and_set_bit().

So I do not see any overall benefit for using atomic_xxx() functions.

Best Regards,
Nick Chan

>
> -- David
>
>
>>
>> Cc: stable@xxxxxxxxxxxxxxx
>> Fixes: 04d8ecf37b5e ("nvme: apple: Add Apple A11 support")
>> Signed-off-by: Yuriy Havrylyuk <yhavry@xxxxxxxxx>
>> Co-developed-by: Nick Chan <towinchenmi@xxxxxxxxx>
>> Signed-off-by: Nick Chan <towinchenmi@xxxxxxxxx>
>> ---
>> drivers/nvme/host/apple.c | 65 +++++++++++++++++++++++++++++++++++++++++++++++
>> 1 file changed, 65 insertions(+)
>>
>> diff --git a/drivers/nvme/host/apple.c b/drivers/nvme/host/apple.c
>> index c1115e27a0d6..6354edf27225 100644
>> --- a/drivers/nvme/host/apple.c
>> +++ b/drivers/nvme/host/apple.c
>> @@ -203,6 +203,20 @@ struct apple_nvme {
>>
>> int irq;
>> spinlock_t lock;
>> +
>> + /*
>> + * Tags of pending commands must be unique across both Admin and IO
>> + * queue. However, on T8015, unlike T8103, without linear submission
>> + * queues, it is not possible for the either queue to skip some tags,
>> + * and both queues must go from 0 to their respective configured
>> + * maximum.
>> + *
>> + * Instead of reserving some tags for the admin queue, use a bitfield
>> + * to keep track of pending commands on either queue, and temporaily
>> + * block command submission by returning BLK_STS_RESOURCE until the
>> + * tag is freed on the other queue.
>> + */
>> + unsigned long t8015_active_tags;
>> };
>>
>> static_assert(sizeof(struct nvme_command) == 64);
>> @@ -290,6 +304,28 @@ static void apple_nvmmu_inval(struct apple_nvme_queue *q, unsigned int tag)
>> "NVMMU TCB invalidation failed\n");
>> }
>>
>> +static bool apple_nvme_reserve_tag_t8015(struct apple_nvme *anv,
>> + struct nvme_command *cmd)
>> +{
>> + u16 tag = nvme_tag_from_cid(cmd->common.command_id);
>> +
>> + if (WARN_ON_ONCE(tag >= BITS_PER_LONG))
>> + return false;
>> +
>> + return !test_and_set_bit(tag, &anv->t8015_active_tags);
>> +}
>> +
>> +static void apple_nvme_release_tag_t8015(struct apple_nvme *anv,
>> + __u16 command_id)
>> +{
>> + u16 tag = nvme_tag_from_cid(command_id);
>> +
>> + if (WARN_ON_ONCE(tag >= BITS_PER_LONG))
>> + return;
>> +
>> + clear_bit(tag, &anv->t8015_active_tags);
>> +}
>> +
>> static void apple_nvme_submit_cmd_t8015(struct apple_nvme_queue *q,
>> struct nvme_command *cmd)
>> {
>> @@ -652,6 +688,8 @@ static inline void apple_nvme_update_cq_head(struct apple_nvme_queue *q)
>> static bool apple_nvme_poll_cq(struct apple_nvme_queue *q,
>> struct io_comp_batch *iob)
>> {
>> + struct apple_nvme *anv = queue_to_apple_nvme(q);
>> + unsigned long completed_tags = 0;
>> bool found = false;
>>
>> while (apple_nvme_cqe_pending(q)) {
>> @@ -664,11 +702,26 @@ static bool apple_nvme_poll_cq(struct apple_nvme_queue *q,
>> dma_rmb();
>> apple_nvme_handle_cqe(q, iob, q->cq_head);
>> apple_nvme_update_cq_head(q);
>> +
>> + if (!anv->hw->has_lsq_nvmmu) {
>> + struct nvme_completion *cqe = &q->cqes[q->cq_head];
>> + u16 tag = nvme_tag_from_cid(READ_ONCE(cqe->command_id));
>> +
>> + if (!WARN_ON_ONCE(tag >= BITS_PER_LONG))
>> + __set_bit(tag, &completed_tags);
>> + }
>> }
>>
>> if (found)
>> writel(q->cq_head, q->cq_db);
>>
>> + if (!anv->hw->has_lsq_nvmmu && completed_tags) {
>> + unsigned long tag_bit;
>> +
>> + for_each_set_bit(tag_bit, &completed_tags, BITS_PER_LONG)
>> + clear_bit(tag_bit, &anv->t8015_active_tags);
>> + }
>> +
>> return found;
>> }
>>
>> @@ -790,6 +843,12 @@ static blk_status_t apple_nvme_queue_rq(struct blk_mq_hw_ctx *hctx,
>> if (ret)
>> return ret;
>>
>> + if (!anv->hw->has_lsq_nvmmu &&
>> + !apple_nvme_reserve_tag_t8015(anv, cmnd)) {
>> + ret = BLK_STS_RESOURCE;
>> + goto out_free_cmd;
>> + }
>> +
>> if (blk_rq_nr_phys_segments(req)) {
>> ret = apple_nvme_map_data(anv, req, cmnd);
>> if (ret)
>> @@ -806,6 +865,9 @@ static blk_status_t apple_nvme_queue_rq(struct blk_mq_hw_ctx *hctx,
>> return BLK_STS_OK;
>>
>> out_free_cmd:
>> + if (!anv->hw->has_lsq_nvmmu)
>> + apple_nvme_release_tag_t8015(anv, cmnd->common.command_id);
>> +
>> nvme_cleanup_cmd(req);
>> return ret;
>> }
>> @@ -1165,6 +1227,9 @@ static void apple_nvme_reset_work(struct work_struct *work)
>> if (ret)
>> goto out;
>>
>> + if (!anv->hw->has_lsq_nvmmu)
>> + WRITE_ONCE(anv->t8015_active_tags, 0);
>> +
>> dev_dbg(anv->dev, "Starting admin queue");
>> apple_nvme_init_queue(&anv->adminq);
>> nvme_unquiesce_admin_queue(&anv->ctrl);
>>
>