On 25/03/21 4:14 am, Asutosh Das (asd) wrote: > On 3/23/2021 12:19 PM, Adrian Hunter wrote: >> On 23/03/21 5:13 pm, Asutosh Das (asd) wrote: >>> On 3/22/2021 11:12 PM, Adrian Hunter wrote: >>>> On 22/03/21 9:53 pm, Asutosh Das (asd) wrote: >>>>> On 3/19/2021 10:47 AM, Adrian Hunter wrote: >>>>>> On 19/03/21 2:35 am, Asutosh Das wrote: >>>>>>> During runtime-suspend of ufs host, the scsi devices are >>>>>>> already suspended and so are the queues associated with them. >>>>>>> But the ufs host sends SSU to wlun during its runtime-suspend. >>>>>>> During the process blk_queue_enter checks if the queue is not in >>>>>>> suspended state. If so, it waits for the queue to resume, and never >>>>>>> comes out of it. >>>>>>> The commit >>>>>>> (d55d15a33: scsi: block: Do not accept any requests while suspended) >>>>>>> adds the check if the queue is in suspended state in blk_queue_enter(). >>>>>>> >>>>>>> Call trace: >>>>>>> __switch_to+0x174/0x2c4 >>>>>>> __schedule+0x478/0x764 >>>>>>> schedule+0x9c/0xe0 >>>>>>> blk_queue_enter+0x158/0x228 >>>>>>> blk_mq_alloc_request+0x40/0xa4 >>>>>>> blk_get_request+0x2c/0x70 >>>>>>> __scsi_execute+0x60/0x1c4 >>>>>>> ufshcd_set_dev_pwr_mode+0x124/0x1e4 >>>>>>> ufshcd_suspend+0x208/0x83c >>>>>>> ufshcd_runtime_suspend+0x40/0x154 >>>>>>> ufshcd_pltfrm_runtime_suspend+0x14/0x20 >>>>>>> pm_generic_runtime_suspend+0x28/0x3c >>>>>>> __rpm_callback+0x80/0x2a4 >>>>>>> rpm_suspend+0x308/0x614 >>>>>>> rpm_idle+0x158/0x228 >>>>>>> pm_runtime_work+0x84/0xac >>>>>>> process_one_work+0x1f0/0x470 >>>>>>> worker_thread+0x26c/0x4c8 >>>>>>> kthread+0x13c/0x320 >>>>>>> ret_from_fork+0x10/0x18 >>>>>>> >>>>>>> Fix this by registering ufs device wlun as a scsi driver and >>>>>>> registering it for block runtime-pm. Also make this as a >>>>>>> supplier for all other luns. That way, this device wlun >>>>>>> suspends after all the consumers and resumes after >>>>>>> hba resumes. >>>>>>> >>>>>>> Co-developed-by: Can Guo <cang@xxxxxxxxxxxxxx> >>>>>>> Signed-off-by: Can Guo <cang@xxxxxxxxxxxxxx> >>>>>>> Signed-off-by: Asutosh Das <asutoshd@xxxxxxxxxxxxxx> >>>>>> >>>>>> I have some more comments that may help straighten things out. >>>>>> >>>>>> Also please look at ufs_debugfs_get_user_access() and >>>>>> ufs_debugfs_put_user_access() that now need to scsi_autopm_get/put_device >>>>>> sdev_ufs_device. >>>>>> >>>>>> It would also be good if you could re-base on linux-next. >>>>>> >>>>> >>>>> Hi Adrian >>>>> Thanks for the comments. >>>>> >>>>> I agree moving the code to wlun probe and other changes. >>>>> But it looks to me that it may not fully solve the issue. >>>>> >>>>> Please let me explain my understanding on this: >>>>> >>>>> (Please refer to the logs in v10) >>>>> scsi_autopm_*() are invoked on a sdev. >>>>> pm_runtime_get_suppliers()/rpm_put_suppliers() are on the supplier device. >>>>> >>>>> For the device wlun: >>>>> slave_configure(): >>>>> - doesn't set the rpm_autosuspend >>>>> - pm_runtime_getnoresume() >>>>> scsi_sysfs_add_sdev(): >>>>> - pm_runtime_forbid() >>>>> - scsi_autopm_get_device() >>>>> - device_add() >>>>> - ufshcd_wl_probe() >>>>> - scsi_autopm_put_device() >>>>> >>>>> For all other scsi devices: >>>>> slave_alloc(): >>>>> - ufshcd_setup_links() >>>>> Say all link_add: pm_runtime_put(&hba->sdev_ufs_device->sdev_gendev); >>>> >>>> With DL_FLAG_RPM_ACTIVE, links will 'get' not 'put' >>>> >>> I'm referring to the pm_runtime_put(sdev_ufs_device) after all the links are setup, that you suggested to add. >> >> Ok >> >>>>> slave_configure(): >>>>> - set rpm_autosuspend >>>>> scsi_sysfs_add_sdev(): >>>>> - scsi_autopm_get_device() >>>>> - device_add() -> schedules an async probe() >>>>> - scsi_autopm_put_device() - (1) >>>>> >>>>> Now the rpm_put_suppliers() can be invoked *after* pm_runtime_get_suppliers() of the async probe(), since both are running in different contexts. >>>> >>>> Only if the sd device suspends. >>>> >>> Correct. What'd stop the sd device from suspending? >>> We should be stopping the sd device from suspending here - imho. >> > > Hi Adrian, > Thanks for the comments. > >> You mean for performance reasons. That is something we can >> look at, but let's get it working first. >> > Not for performance reasons. I meant to say that this issue can be fixed if we stop the sd devices from suspending until the sd_probe() is completed. To me that looks like hiding the problem rather than fixing it. For example, Rafael's revert was a real issue that we uncovered. >From a maintenance point of view, hiding problems rather than fixing them, creates an unsustainable technical debt for the future. >>> >>>>> In that case, the usage_count of supplier would be decremented until rpm_active of this link becomes 1. >>>> >>>> Right, because the sd device suspended. >>>> >>>>> Now the pm_runtime_get_suppliers() expects the link_active to be more than 1. >>>> >>>> Not sure what you mean here. pm_runtime_*put*_suppliers() won't do anything if the link count is 1. >>> I'm referring to the logs that I pasted before: >>> [ 6.941267][ T7] scsi 0:0:0:4: rpm_put_suppliers: [BEF] Supp (0:0:0:49488) usage_count: 4 rpm_active: 3 >>> >>> ------ T196 Context comes in while T7 is running ---------- >>> [ 6.941466][ T196] scsi 0:0:0:4: pm_runtime_get_suppliers: (0:0:0:49488): supp: usage_count: 5 rpm_active: 4 >>> -------------------------------------------------------------- >>> >>> [ 7.788397][ T7] scsi 0:0:0:4: rpm_put_suppliers: [AFT] Supp (0:0:0:49488) usage_count: 2 rpm_active: 1 >>> >>> I meant to say that, if the rpm_put_suppliers() is invoked after the pm_runtime_get_suppliers() as is seen above then the link_active may become 1 even *after* pm_runtime_get_suppliers() is invoked. >>> >>> I'm referring to the pm_runtime_get_suppliers() invoked from: >>> driver_probe_device() - say for, sd 0:0:0:x >>> |- pm_runtime_get_suppliers() - for sd 0:0:0:49488 >> >> I am hoping that was the problem that Rafael's revert dealt with. >> > I think the issue is in the sequence of events. > If rpm_put_suppliers() runs after pm_runtime_get_suppliers() this issue can occur. Then it would be a core PM issue not a UFS issue. > >>>> >>>>> Now then, there comes a time, that when sd_probe() schedules a suspend, the supplier usage_count becomes 0 and the link_active becomes 1. >>>>> And the supplier suspends before the consumer. >>>> >>>> sd probe first resumes the sd device which will resume the supplier. >>>> >>> Correct, but it'd again schedule a suspend (since autosuspend is enabled now) at the end of the sd_probe(). >>> Thereafter, pm_runtime_put_suppliers()(sd 0:0:0:49488) is invoked from driver_probe_device() which had actually invoked sd_probe(). >>> That'd make the link_active to 1 even when sd 0:0:0:x is active. >> >> If sd 0:0:0:x is active then rpm_get_suppliers() still has +1 rpm_active. pm_runtime_get_suppliers() also has +1 rpm_active. >> i.e. rpm_active is 3. If rpm_put_suppliers() is called, it means sd 0:0:0:x has really runtime suspended (not just waiting for autosuspend. Otherwise when the probe ends pm_runtime_put_suppliers() will drop rpm_active from 3 to 2. > In the good case it'd drop from 3 to 2. But in the bad case, I see that it drops to 1. That's when the supplier suspends before the consumer. > That would happen when rpm_put_suppliers() runs after the pm_runtime_get_suppliers() is completed and decrements the usage_count of supplier until link_active is 1. At that point yes, sd 0:0:0:x has really runtime-suspended. sd_probe() would resume it and schedule a suspend at the end of probe. > > IIUC, below is the sequence of events that can lead to this issue: > 1. sd 0:0:0:x schedules an async probe > 2. sd 0:0:0:x invokes scsi_autopm_put_device() > 3. async probe completes pm_runtime_get_suppliers() increments the rpm_active. > 4. suspend of sd 0:0:0:x is invoked and rpm_put_suppliers() is invoked which decrements the link_active (this was incremented in 3 above) > 5. sd_probe() is invoked which resumes it and schedules a suspend > 6. pm_runtime_put_suppliers() is invoked which decreases the link_active to 1 and supplier suspends before the consumer. > > So my solution was to stop sd 0:0:0:x from runtime suspending until the sd_probe() is done. I'll look into it, but maybe the following would help: diff --git a/drivers/base/power/runtime.c b/drivers/base/power/runtime.c index 18b82427d0cb..4f708b2a9359 100644 --- a/drivers/base/power/runtime.c +++ b/drivers/base/power/runtime.c @@ -312,6 +312,7 @@ static void rpm_put_suppliers(struct device *dev) list_for_each_entry_rcu(link, &dev->links.suppliers, c_node, device_links_read_lock_held()) { + link->supplier_preactivated = false; while (refcount_dec_not_one(&link->rpm_active)) pm_runtime_put(link->supplier); } > >> >> But it is a bit theoretical. Let's try it and see. >> >>> >>>>> >>>>> So I was wondering, what'd make sure that the pm_runtime_get_suppliers() from driver_probe_device() is invoked after scsi_autopm_put_device() (1) finishes the rpm_put_suppliers(). >>>>> >>>>> Am not sure if I'm missing something in this. >>>>> Do you think, the current changes alone can fix the above issue? >>>> >>>> Yes, but let's see. >>>> >>> Essentially, we should stop the sd device from runtime suspending until it's probe is done. Then allow the same. Does it make sense? >>> Please let me know what you think. >> >> I really think it would be good to try the changes that have been identified and see how it behaves. >> >> Then go from there. >> > Sure, I've pushed the changes v13 today. > I will test it after the changes are finalized. > Thank you! :-)