On 7/28/2022 1:29 PM, Bart Van Assche wrote:
On 7/28/22 12:15, Asutosh Das (asd) wrote:
Hello John,
On 7/28/2022 12:10 PM, John Garry wrote:
On 19/07/2022 08:01, Can Guo wrote:
+
+ hba->nr_queues[HCTX_TYPE_DEFAULT] = num_possible_cpus();
+ hba->nr_queues[HCTX_TYPE_READ] = 0;
+ hba->nr_queues[HCTX_TYPE_POLL] = 1;
+
+ for (i = 0; i < HCTX_MAX_TYPES; i++)
+ host->nr_hw_queues += hba->nr_queues[i];
+
+ host->can_queue = hba->nutrs;
+ host->cmd_per_lun = hba->nutrs;
+
+ /* One more reserved for dev_cmd_queue */
+ hba->nr_hw_queues = host->nr_hw_queues + 1;
+
So this would mean that the host can accept .can_queue * .nr_hw_queues numbers of requests simultaneously - is that true?
That would mean that .can_queue * .nr_hw_queues numbers of request may be queued to the host.
Please can you elaborate if you see an issue.
Hi Asutosh,
The `host_tagset` flag has been introduced by John and Hannes some time ago. See also commit bdb01301f3ea ("scsi: Add host and host template flag 'host_tagset'"). This flag supports sharing tags across hardware queues and could be used to support UFSHCI 4.0 controllers that do not support the EXT_IID feature.
In order not to complicate the implementation further, I propose to fall back to the UFSHCI 3.0 compatibility mode for UFSHCI 4.0 controllers that do not support the EXT_IID feature.
To answer John's question: the maximum number of outstanding commands is 16 * hba->nutrs if EXT_IID is supported (EXT_IID is a four bits field). If the hardware queue index is encoded in the EXT_IID field, hba->nutrs is the number of commands per hardware queue.
Thanks,
Bart.
Thanks Bart, I wasn't aware of the background of John's Q. I will consider your proposal and get back.