On 7/10/23 14:31, Ming Lei wrote:
On Mon, Jul 10, 2023 at 02:01:18PM +0300, Sagi Grimberg wrote:
On 7/10/23 12:50, Ming Lei wrote:
On Mon, Jul 10, 2023 at 12:27:31PM +0300, Sagi Grimberg wrote:
I still want your patches for tcp/rdma that move the freeze.
If you are not planning to send them, I swear I will :)
Ming, can you please send the tcp/rdma patches that move the
freeze? As I said before, it addresses an existing issue with
requests unnecessarily blocked on a frozen queue instead of
failing over.
Any chance to fix the current issue in one easy(backportable) way[1] first?
There is, you suggested one. And I'm requesting you to send a patch for
it.
The patch is the one pointed by link [1], and it still can be applied on current
linus tree.
https://lore.kernel.org/linux-nvme/20230629064818.2070586-1-ming.lei@xxxxxxxxxx/
This is separate from what I am talking about.
All previous discussions on delay freeze[2] are generic, which apply on all
nvme drivers, not mention this error handling difference causes extra maintain
burden. I still suggest to convert all drivers in same way, and will work
along the approach[1] aiming for v6.6.
But we obviously hit a difference in expectations from different
drivers. In tcp/rdma there is currently an _existing_ bug, where
we freeze the queue on error recovery, and unfreeze only after we
reconnect. In the meantime, requests can be blocked on the frozen
request queue and not failover like they should.
In fabrics the delta between error recovery and reconnect can (and
often will be) minutes or more. Hence I request that we solve _this_
issue which is addressed by moving the freeze to the reconnect path.
I personally think that pci should do this as well, and at least
dual-ported multipath pci devices would prefer instant failover
than after a full reset cycle. But Keith disagrees and I am not going to
push for it.
Regardless of anything we do in pci, the tcp/rdma transport
freeze-blocking-failover _must_ be addressed.
It is one generic issue, freeze/unfreeze has to be paired strictly
for every driver.
For any nvme driver, the inbalance can happen when error handling
is involved, that is why I suggest to fix the issue in one generic
way.
Ming, you are ignoring what I'm saying. I don't care if the
freeze/unfreeze is 100% balanced or not (for the sake of this
discussion).
I'm talking about a _separate_ issue where a queue
is frozen for potentially many minutes blocking requests that
could otherwise failover.
So can you please submit a patch for each? Please phrase it as what
it is, a bug fix, so stable kernels can pick it up. And try to keep
it isolated to _only_ the freeze change so that it is easily
backportable.
The patch of "[PATCH V2] nvme: mark ctrl as DEAD if removing from error
recovery" can fix them all(include nvme tcp/fc's issue), and can be backported.
Ming, this is completely separate from what I'm talking about. This one
is addressing when the controller is removed, while I'm talking about
the error-recovery and failover, which is ages before the controller is
removed.
But as we discussed, we still want to call freeze/unfreeze in pair, and
I also suggest the following approach[2], which isn't good to backport:
1) moving freeze into reset
2) during resetting
- freeze NS queues
- unquiesce NS queues
- nvme_wait_freeze()
- update_nr_hw_queues
- unfreeze NS queues
3) meantime changes driver's ->queue_rq() in case that ctrl state is NVME_CTRL_CONNECTING,
- if the request is FS IO with data, re-submit all bios of this request, and free the request
- otherwise, fail the request
[2] https://lore.kernel.org/linux-block/5bddeeb5-39d2-7cec-70ac-e3c623a8fca6@xxxxxxxxxxx/T/#mfc96266b63eec3e4154f6843be72e5186a4055dc
Ming, please read again what my concern is. I'm talking about error recovery
freezing a queue, and unfreezing only after we reconnect,
blocking requests that should failover.
From my understanding, nothing is special for tcp/rdma compared with
nvme-pci.
But there is... The expectations are different.
All take two stage error recovery: teardown & [reset(nvme-pci) | reconnect(tcp/rdma)]
Queues are frozen during teardown, and unfreeze in reset or reconnect.
If the 2nd stage is failed or bypassed, queues could be left as frozen
& unquisced, and requests can't be handled, and io hang.
When tcp reconnect failed, nvme_delete_ctrl() is called for failing
requests & removing controller.
Then the patch of "nvme: mark ctrl as DEAD if removing from error recovery"
can avoid this issue by calling blk_mark_disk_dead() which can fail any
request pending in bio_queue_enter().
If that isn't tcp/rdma's issue, can you explain it in details?
Yes I can. The expectation in pci is that a reset lifetime will be a few
seconds. By lifetime I mean it starts and either succeeds or fails.
in fabrics the lifetime is many minutes. i.e. it starts when error
recovery kicks in, and either succeeds (reconnected) or fails (deleted
due to ctrl_loss_tmo). This can take a long period of time (if for
example the controller is down for maintenance/reboot).
Hence, while it may be acceptable that requests are blocked on
a frozen queue for the duration of a reset in pci, it is _not_
acceptable in fabrics. I/O should failover far sooner than that
and must _not_ be dependent on the success/failure or the reset.
Hence I requested that this is addressed specifically for fabrics
(tcp/rdma).
OK, I got your idea now, but which is basically not doable from current
nvme error recovery approach.
Even though starting freeze is moved to reconnect stage, queue is still
quiesced, then request is kept in block layer's internal queue, and can't
enter nvme fabric .queue_rq().
See error-recovery in nvme_[tcp|rdma]_error_recovery(), the queue is
unquiesced for fast-failover.
The patch you are pushing can't work for this requirement, also I don't
think you need that, because FS IO is actually queued to nvme mpath queue,
which won't be frozen at all.
But it will enter the ns request queue, and will block there. There is a
window where a *few* stray requests enter the ns request queue before
the current path is reset, those are the ones that will not failover
immediately (because they are blocked on a frozen queue).
However, you can improve nvme_ns_head_submit_bio() by selecting new path
if the underlying queue is in error recovery. Not dig into the current
code yet, I think it should have been done in this way already.
That is fine, and it happens to new requests already. The nvme-mpath
failover is not broken obviously. I am only talking about the few
requests that entered the ns queue when it was frozen and before the
nshead current_path was changed.
For those requests, we should fast failover as well, and we don't today.
