NVMe: Re-introduce polling for completions
PATCH 1/1 NVMe: Re-introduce polling for completions
Multiple users have reported device initialization failure on some platforms due the driver not receiving expected interrupts. This is not the fault of any particular controller. This patch polls for completions on the admin queue in the watchdog timer, and starts the timer immediately after controller initialization completes, just before the first admin command is issued.
--- drivers/nvme/host/pci.c | 9 +++++++-- 1 file changed, 7 insertions(+), 2 deletions(-)
drivers/nvme/host/pci.c +7 -2
@@ -1354,6 +1354,7 @@ static void nvme_watchdog_timer(unsigned long data) {struct nvme_dev *dev = (struct nvme_dev *)data;u32 csts =readl (dev->bar +NVME_REG_CSTS ); +struct nvme_queue *admin_q = dev->queues[0 ];/* * Skip controllers currently under reset.
@@ -1369,6 +1370,10 @@ static void nvme_watchdog_timer(unsigned long data)return ; } +spin_lock_irq (&admin_q->q_lock); +nvme_process_cq (admin_q); +spin_unlock_irq (&admin_q->q_lock); +mod_timer (&dev->watchdog_timer,round_jiffies (jiffies +HZ )); }
@@ -1877,6 +1882,8 @@ static void nvme_reset_work(struct work_struct *work)if (result) goto out; +mod_timer (&dev->watchdog_timer,round_jiffies (jiffies +HZ )); + result =nvme_init_identify (&dev->ctrl);if (result) goto out;
@@ -1888,8 +1895,6 @@ static void nvme_reset_work(struct work_struct *work) dev->ctrl.event_limit =NVME_NR_AEN_COMMANDS ;queue_work (nvme_workq, &dev->async_work); -mod_timer (&dev->watchdog_timer,round_jiffies (jiffies +HZ )); -/* * Keep the controller around but remove all namespaces if we don't have * any working I/O queue.
Signature · 1 line
2.7.2
PATCH 1/1 NVMe: Re-introduce polling for completions
Quoted 9 lines · Keith Busch
Multiple users have reported device initialization failure on some platforms due the driver not receiving expected interrupts. This is not the fault of any particular controller. This patch polls for completions
on the admin queue in the watchdog timer, and starts the timer immediately
after controller initialization completes, just before the first admin command is issued.
Quoted 3 lines
Reported-by Tim Muhlemmer <muhlemmer at gmail.com>Reported-by Jon Derrick <jonathan.derrick at intel.com>Signed-off-by Keith Busch <keith.busch at intel.com>
PATCH 1/1 NVMe: Re-introduce polling for completions
In reply to Keith BuschQuoted 6 lines · Keith Busch
Multiple users have reported device initialization failure on some platforms due the driver not receiving expected interrupts. This is not the fault of any particular controller. This patch polls for completions on the admin queue in the watchdog timer, and starts the timer immediately after controller initialization completes, just before the first admin command is issued.
So how do any other PCIe device work given that almost no driver does unconditionaly polling?
drivers/nvme/host/pci.c static void nvme_watchdog_timer(unsigned long data) · line 1375
Quoted 3 lines
drivers/nvme/host/pci.c +3
-0+spin_lock_irq (&admin_q->q_lock); +nvme_process_cq (admin_q); +spin_unlock_irq (&admin_q->q_lock);
Add a comment on why this is done. Given that only the admin queue is polled it's going to be a really convoluted explanation which should be spelled out in detail.
PATCH 1/1 NVMe: Re-introduce polling for completions
Quoted 2 lines · Christoph Hellwig
So how do any other PCIe device work given that almost no driver does unconditionaly polling?
I honestly don't have a good answer to that, and I agree with you that this shouldn't be necessary ... but this is a harmless way to not frustrate people who purchase these devices.
I suspect other PCI device drivers either at least try to never use the legacy IRQ, or they've never been tested in platforms that break them.
Quoted 7 lines
drivers/nvme/host/pci.c static void nvme_watchdog_timer(unsigned long data) · line 1375
Quoted 3 lines
drivers/nvme/host/pci.c +3
-0+spin_lock_irq (&admin_q->q_lock); +nvme_process_cq (admin_q); +spin_unlock_irq (&admin_q->q_lock);Add a comment on why this is done. Given that only the admin queue is polled it's going to be a really convoluted explanation which should be spelled out in detail.
Sounds good.
PATCH 1/1 NVMe: Re-introduce polling for completions
Quoted 10 lines · Keith Busch
Quoted 2 lines · Christoph Hellwig
So how do any other PCIe device work given that almost no driver does unconditionaly polling?
I honestly don't have a good answer to that, and I agree with you that this shouldn't be necessary ... but this is a harmless way to not frustrate people who purchase these devices.
I suspect other PCI device drivers either at least try to never use the legacy IRQ, or they've never been tested in platforms that break them.
Before I resend with the requested code comments, I'd like to hear your opinion on an alternate solution (patch below).
Currently the driver uses legacy IRQ if available only until we know how many queues it can create. The legacy IRQ use is tripping up some h/w, but I've not heard such issues with MSI/MSI-x.
Would you prefer going straight to MSI-x? The driver does that when there is no INTx, but we wouldn't need to poll from the watchdog timer if we make MSI-x the default behavior,
---
drivers/nvme/host/pci.c +14 -8
@@ -1518,7 +1518,9 @@ static int nvme_setup_io_queues(struct nvme_dev *dev)* If we enable msix early due to not intx, disable it again before * setting up the full range we need. */ -if (!pdev->irq) +if (pdev->msi_enabled) +pci_disable_msi (pdev); +else if (pdev->msix_enabled)pci_disable_msix (pdev);for (i =0 ; i < nr_io_queues; i++)
@@ -1701,7 +1703,6 @@ static int nvme_pci_enable(struct nvme_dev *dev)if (pci_enable_device_mem (pdev))return result; - dev->entry[0 ].vector = pdev->irq;pci_set_master (pdev); if (dma_set_mask_and_coherent(dev->dev, DMA_BIT_MASK(64)) &&
@@ -1714,13 +1715,18 @@ static int nvme_pci_enable(struct nvme_dev *dev) }/* -* Some devices don't advertse INTx interrupts, pre-enable a single -* MSIX vec for setup. We'll adjust this later. +* Some devices and/or platforms don't advertise or work with INTx +* interrupts. Pre-enable a single MSIX or MSI vec for setup. We'll +* adjust this later. */ -if (!pdev->irq) { - result =pci_enable_msix (pdev, dev->entry,1 ); -if (result <0 ) - goto disable; +if (pci_enable_msix (pdev, dev->entry,1 )) { +pci_enable_msi (pdev); + dev->entry[0 ].vector = pdev->irq; + } + +if (!dev->entry[0 ].vector) { + result = -ENODEV ; + goto disable; } cap =lo_hi_readq (dev->bar +NVME_REG_CAP );
--
PATCH 1/1 NVMe: Re-introduce polling for completions
Going straight to MSI-x is a higher risk change since re-introducing polling restores code that was in place previously with no reported problems whereas changing the init path to something new has no test time on it.
Restoring the polling (with comments as requested) is a harmless way to regain general ecosystem compatibility.
Thanks,
Judy
-----Original Message-----
From: Linux-nvme [mailto:[email protected]] On Behalf Of Keith Busch
Sent: Tuesday, April 05, 2016 2:39 PM
To: Christoph Hellwig
Cc: Jens Axboe; linux-nvme at lists.infradead.org
Subject: Re: [PATCH] NVMe: Re-introduce polling for completions
Quoted 10 lines · Keith Busch
Quoted 2 lines · Christoph Hellwig
So how do any other PCIe device work given that almost no driver does unconditionaly polling?
I honestly don't have a good answer to that, and I agree with you that this shouldn't be necessary ... but this is a harmless way to not frustrate people who purchase these devices.
I suspect other PCI device drivers either at least try to never use the legacy IRQ, or they've never been tested in platforms that break them.
Before I resend with the requested code comments, I'd like to hear your opinion on an alternate solution (patch below).
Currently the driver uses legacy IRQ if available only until we know how many queues it can create. The legacy IRQ use is tripping up some h/w, but I've not heard such issues with MSI/MSI-x.
Would you prefer going straight to MSI-x? The driver does that when there is no INTx, but we wouldn't need to poll from the watchdog timer if we make MSI-x the default behavior,
---
drivers/nvme/host/pci.c +3 -2
@@ -1518,7 +1518,9 @@ static int nvme_setup_io_queues(struct nvme_dev *dev)* If we enable msix early due to not intx, disable it again before * setting up the full range we need. */ -if (!pdev->irq) +if (pdev->msi_enabled) +pci_disable_msi (pdev); +else if (pdev->msix_enabled)pci_disable_msix (pdev);for (i =0 ; i < nr_io_queues; i++)
@@ -1701,7 +1703,6 @@ static int nvme_pci_enable(struct nvme_dev *dev)if (pci_enable_device_mem (pdev))return result; - dev->entry[0 ].vector = pdev->irq;pci_set_master (pdev); if (dma_set_mask_and_coherent(dev->dev, DMA_BIT_MASK(64)) && @@ -1714,13 +1715,18 @@ static int nvme_pci_enable(struct nvme_dev *dev)
}
/*
- * Some devices don't advertse INTx interrupts, pre-enable a single
- * MSIX vec for setup. We'll adjust this later.
+ * Some devices and/or platforms don't advertise or work with INTx
+ * interrupts. Pre-enable a single MSIX or MSI vec for setup. We'll
+ * adjust this later.
*/
- if (!pdev->irq) {
- result = pci_enable_msix(pdev, dev->entry, 1);
- if (result < 0)
- goto disable;
+ if (pci_enable_msix(pdev, dev->entry, 1)) {
+ pci_enable_msi(pdev);
+ dev->entry[0].vector = pdev->irq;
+ }
+
+ if (!dev->entry[0].vector) {
+ result = -ENODEV;
+ goto disable;
}
cap = lo_hi_readq(dev->bar + NVME_REG_CAP);
--
_______________________________________________
Linux-nvme mailing list
Linux-nvme at lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-nvme
PATCH 1/1 NVMe: Re-introduce polling for completions
Quoted 1 line · Judy Brock-SSI
Going straight to MSI-x is a higher risk change since re-introducing polling restores code that was in place previously with no reported problems whereas changing the init path to something new has no test time on it.
Well, I wouldn't say "no" test time since we have a well tested path that goes straight to MSI-x in place for some time.
If we rely on polled completions during initialization, we potentially add up to 2 seconds for device initialization to complete (1 second polling interval * 2 commands using broken IRQ).
Maybe you want both?
PATCH 1/1 NVMe: Re-introduce polling for completions
Quoted 1 line
Maybe you want both?
I think that would be a good solution.
-----Original Message-----
From: Keith Busch [mailto:[email protected]]
Sent: Tuesday, April 05, 2016 3:30 PM
To: Judy Brock-SSI
Cc: Christoph Hellwig; Jens Axboe; linux-nvme at lists.infradead.org
Subject: Re: [PATCH] NVMe: Re-introduce polling for completions
Quoted 1 line · Judy Brock-SSI
Going straight to MSI-x is a higher risk change since re-introducing polling restores code that was in place previously with no reported problems whereas changing the init path to something new has no test time on it.
Well, I wouldn't say "no" test time since we have a well tested path that goes straight to MSI-x in place for some time.
If we rely on polled completions during initialization, we potentially add up to 2 seconds for device initialization to complete (1 second polling interval * 2 commands using broken IRQ).
Maybe you want both?
PATCH 1/1 NVMe: Re-introduce polling for completions
In reply to Judy Brock-SSIQuoted 1 line · Judy Brock-SSI
Restoring the polling (with comments as requested) is a harmless way to regain general ecosystem compatibility.
Please take your marketing talk elsewhere, there is no "ecosystem" to "synergize" in the linux kernel.
PATCH 1/1 NVMe: Re-introduce polling for completions
In reply to Keith BuschQuoted 10 lines · Keith Busch
Before I resend with the requested code comments, I'd like to hear your opinion on an alternate solution (patch below).
Currently the driver uses legacy IRQ if available only until we know how many queues it can create. The legacy IRQ use is tripping up some h/w, but I've not heard such issues with MSI/MSI-x.
Would you prefer going straight to MSI-x? The driver does that when there is no INTx, but we wouldn't need to poll from the watchdog timer if we make MSI-x the default behavior,
I would much prefer that variant over doing crazy non-standard polling.
PATCH 1/1 NVMe: Re-introduce polling for completions
In reply to Keith BuschQuoted 15 lines · Keith Busch
Quoted 10 lines · Keith Busch
Quoted 2 lines · Christoph Hellwig
So how do any other PCIe device work given that almost no driver does unconditionaly polling?
I honestly don't have a good answer to that, and I agree with you that this shouldn't be necessary ... but this is a harmless way to not frustrate people who purchase these devices.
I suspect other PCI device drivers either at least try to never use the legacy IRQ, or they've never been tested in platforms that break them.
Before I resend with the requested code comments, I'd like to hear your opinion on an alternate solution (patch below).
Quoted 7 lines
Currently the driver uses legacy IRQ if available only until we know how many queues it can create. The legacy IRQ use is tripping up some h/w, but I've not heard such issues with MSI/MSI-x.
Would you prefer going straight to MSI-x? The driver does that when there is no INTx, but we wouldn't need to poll from the watchdog timer if we make MSI-x the default behavior,
Can you resend this to Jens for 4.6 with a proper signoff?
PATCH 1/1 NVMe: Re-introduce polling for completions
Quoted 1 line · Christoph Hellwig
Can you resend this to Jens for 4.6 with a proper signoff?
Ooops just noticed you resent it and Jens picked it up. Feel free to ignore my comment :)