bus_dmamap_sync() ensures that memory that's prepared for PREWRITE can
be DMA'd immediately after it returns. The details differ, but this
mirrors atomic thread release semantics, at least for the buffers
synced.
For non-x86 platforms, bus_dmamap_sync() has the right syncing and
fences. So in the past, wmb() had been omitted for them.
For x86 platforms, the memory ordering is already strong enough to
ensure DMA to the device sees the current contents. As such, we don't
need the wmb() here. It translates to an sfence which is only needed
for writes to regions that have the write combining attribute set or
when some exotic opcodes are used. The nvme driver does neither of
these. Since bus_dmamap_sync() includes atomic_thread_fence_rel, we
can be assured any optimizer won't reorder the bus_dmamap_sync and the
bus_space_write operations. The wmb() was a vestiage of the pre-busdma
version initially committed to the tree.
Reviewed by: kib@, gallatin@, chuck@, mav@
Differential Revision: https://reviews.freebsd.org/D27448
- in nvme_qpair_process_completions() do dma sync before completion buffer
is used.
- in nvme_qpair_submit_tracker(), don't do explicit wmb() also for arm
and arm64. Bus_dmamap_sync() on these architectures is sufficient to ensure
that all CPU stores are visible to external (including DMA) observers.
- Allocate completion buffer as BUS_DMA_COHERENT. On not-DMA coherent systems,
buffers continuously owned (and accessed) by DMA must be allocated with this
flag. Note that BUS_DMA_COHERENT flag is no-op on DMA coherent systems
(or coherent buses in mixed systems).
MFC after: 4 weeks
Reviewed by: mav, imp
Differential Revision: https://reviews.freebsd.org/D27446
These swapping functions violate BUSDMA contract - we cannot write
to armed (by bus_dmamap_sync(PRE_..)) buffers. Remove them at least
from little endian machines until a better solution will be developed.
Reviewed by: imp
MFC after: 3 weeks
With 4KB page size the 2MB is the maximum we can address with one page PRP.
Going further would require chaining, that would add some more complexity.
On the other side, to reduce memory consumption, allocate the PRP memory
respecting maximum transfer size reported in the controller identify data.
Many of NVMe devices support much smaller values, starting from 128KB.
To do that we have to change the initialization sequence to pull the data
earlier, before setting up the I/O queue pairs. The admin queue pair is
still allocated for full MIN(maxphys, 2MB) size, but it is not a big deal,
since there is only one such queue with only 16 trackers.
Reviewed by: imp
MFC after: 2 weeks
Sponsored by: iXsystems, Inc.
Replace MAXPHYS by runtime variable maxphys. It is initialized from
MAXPHYS by default, but can be also adjusted with the tunable kern.maxphys.
Make b_pages[] array in struct buf flexible. Size b_pages[] for buffer
cache buffers exactly to atop(maxbcachebuf) (currently it is sized to
atop(MAXPHYS)), and b_pages[] for pbufs is sized to atop(maxphys) + 1.
The +1 for pbufs allow several pbuf consumers, among them vmapbuf(),
to use unaligned buffers still sized to maxphys, esp. when such
buffers come from userspace (*). Overall, we save significant amount
of otherwise wasted memory in b_pages[] for buffer cache buffers,
while bumping MAXPHYS to desired high value.
Eliminate all direct uses of the MAXPHYS constant in kernel and driver
sources, except a place which initialize maxphys. Some random (and
arguably weird) uses of MAXPHYS, e.g. in linuxolator, are converted
straight. Some drivers, which use MAXPHYS to size embeded structures,
get private MAXPHYS-like constant; their convertion is out of scope
for this work.
Changes to cam/, dev/ahci, dev/ata, dev/mpr, dev/mpt, dev/mvs,
dev/siis, where either submitted by, or based on changes by mav.
Suggested by: mav (*)
Reviewed by: imp, mav, imp, mckusick, scottl (intermediate versions)
Tested by: pho
Sponsored by: The FreeBSD Foundation
Differential revision: https://reviews.freebsd.org/D27225
Both the size (128 bytes) and ephemeral nature of allocations make it a great
fit for malloc.
A dedicated zone unnecessarily avoids sharing buckets with 128-byte objects.
Reviewed by: imp
Differential Revision: https://reviews.freebsd.org/D27103
Remove code that supported pre-2011 kernels. CTLTYPE_S64 was defined
in rev 217616. All supported branches have it, so remove its compat
definition as OBE.
Instead, add arguments to vmapbuf. Since this argument is
always a pointer use a type of void * and cast to vm_offset_t in
vmapbuf. (In CheriBSD we've altered vm_fault_quick_hold_pages to
take a pointer and check its bounds.)
In no other situtation does b_data contain a user pointer and vmapbuf
replaces b_data with the actual mapping.
Suggested by: jhb
Reviewed by: imp, jhb
Obtained from: CheriBSD
MFC after: 1 week
Sponsored by: DARPA
Differential Revision: https://reviews.freebsd.org/D26784
This field was not in specs when the driver was written, but now there
are SSDs with the reported latency of 10s, where hardcoded value of 5s
seems to be not enough sometimes, causing shutdown timeout messages.
MFC after: 1 week
Sponsored by: iXsystems, Inc.
These functions were introduced before UMA started ensuring that freed
memory gets placed in domain-local caches. They no longer serve any
purpose since UMA now provides their functionality by default. Remove
them to simplyify the kernel memory allocator interfaces a bit.
Reviewed by: cem, kib
Sponsored by: The FreeBSD Foundation
Differential Revision: https://reviews.freebsd.org/D25937
Instead of panic after one second of polling, make the normal timeout
handler to activate, reset the controller and abort the outstanding
requests. If all of it won't happen within 10 seconds then something
in the driver is likely stuck bad and panic is the only way out.
In particular this fixed device hot unplug during execution of those
polled commands, allowing clean device detach instead of panic.
MFC after: 1 week
Sponsored by: iXsystems, Inc.
All callers are currently filtering bad nsid to this function,
however, we'll have undefined behavior if that's not true. Add the
KASSERT to prevent that.
Add the nvmeX device to the XPT_PATH_INQ nvme specific
information. while one could figure this out by looking up the
domain🚌slot:function, it's a lot easier to have the SIM set it
directly since the sim knows this.
r357614 added CTLFLAG_NEEDGIANT to make it easier to find nodes that are
still not MPSAFE (or already are but aren’t properly marked).
Use it in preparation for a general review of all nodes.
This is non-functional change that adds annotations to SYSCTL_NODE and
SYSCTL_PROC nodes using one of the soon-to-be-required flags.
Mark all obvious cases as MPSAFE. All entries that haven't been marked
as MPSAFE before are by default marked as NEEDGIANT
Approved by: kib (mentor, blanket)
Commented by: kib, gallatin, melifaro
Differential Revision: https://reviews.freebsd.org/D23718
BIO_READ and BIO_WRITE, we've handled this expanded syntax poorly in
drivers when the driver doesn't support a particular command. Do a
sweep and fix that.
Reported by: imp
SSD capacity in laptops is growing faster then RAM size, so my original
guess seems too low on second thought. Hopefully nobody will build large
array of those crappy SSDs.
MFC after: 2 weeks
X-MFC-with: 356474
This allows cheapest DRAM-less NVMe SSDs to use some of host RAM (about
1MB per 1GB on the devices I have) for its metadata cache, significantly
improving random I/O performance. Device reports minimal and preferable
size of the buffer. The code limits it to 1% of physical RAM by default.
If the buffer can not be allocated or below minimal size, the device will
just have to work without it.
MFC after: 2 weeks
Relnotes: yes
Sponsored by: iXsystems, Inc.
Within command completion processing the callback function may access
DMAed data buffer. Synchronize it before use, not after.
This allows to use NVMe disk on non-DMA coherent arm64 system.
MFC after: 3 weeks
While there are subtle semantic differences between bool and boolean_t, none of
them matter in these cases. Prefer true/false when dealing with bool
type. Preserve a couple of TRUEs since they are passed into int args into CAM.
Preserve a couple of FALSEs when used for status.done, an int.
Differential Revision: https://reviews.freebsd.org/D20999
This trims the boot time a bit more for AWS and other platforms that have nvme
drives. There's no reason too do this inline. This has been in my tree a while,
but IIRC I talked to Jim Harris about this at one of our face to face meetings.
MFC After: 2 weeks
Don't needlessly pass around qpair pointers when the tracker knows what
qpair it's on. This will simplify code and make it easier to split
submission and completion queues in the future.
Signed-off-by: John Meneghini <johnm@netapp.com>
- For each queue pair precalculate CPU and domain it is bound to.
If queue pairs are not per-CPU, then use the domain of the device.
- Allocate most of queue pair memory from the domain it is bound to.
- Bind callouts to the same CPUs as queue pair to avoid migrations.
- Do not assign queue pairs to each SMT thread. It just wasted
resources and increased lock congestions.
- Remove fixed multiplier of CPUs per queue pair, spread them even.
This allows to use more queue pairs in some hardware configurations.
- If queue pair serves multiple CPUs, bind different NVMe devices to
different CPUs.
MFC after: 1 month
Sponsored by: iXsystems, Inc.
The NVMe standard (1.4) states
>>> 8.6 Doorbell Stride for Software Emulation
>>> The doorbell stride,...is useful in software emulation of an NVM
>>> Express controller. ... For hardware implementations of the NVM
>>> Express interface, the expected doorbell stride value is 0h.
However, hardware in the wild exists with a doorbell stride of 1
(meaning 8 byte separation). This change supports that hardware, as
well as software emulators as envisioned in Section 8.6. Since this is
the fast path, care has been taken to make this computation
efficient. The bit of math to compute an offset for each is replaced
by a memory load from cache of a pre-computed value.
MFC After: 3 days
Reviewed by: scottl@
Differential Revision: https://reviews.freebsd.org/D21514
When we suspend, we need to properly shutdown the NVME controller. The
controller may go into D3 state (or may have the power removed), and
to properly flush the metadata to non-volatile RAM, we must complete a
normal shutdown. This consists of deleting the I/O queues and setting
the shutodown bit. We have to do some extra stuff to make sure we
reset the software state of the queues as well.
On resume, we have to reset the card twice, for reasons described in
the attach funcion. Once we've done that, we can restart the card. If
any of this fails, we'll fail the NVMe card, just like we do when a
reset fails.
Set is_resetting for the duration of the suspend / resume. This keeps
the reset taskqueue from running a concurrent reset, and also is
needed to prevent any hw completions from queueing more I/O to the
card. Pass resetting flag to nvme_ctrlr_start. It doesn't need to get
that from the global state of the ctrlr. Wait for any pending reset to
finish. All queued I/O will get sent to the hardware as part of
nvme_ctrlr_start(), though the upper layers shouldn't send any
down. Disabling the qpairs is the other failsafe to ensure all I/O is
queued.
Rename nvme_ctrlr_destory_qpairs to nvme_ctrlr_delete_qpairs to avoid
confusion with all the other destroy functions. It just removes the
queues in hardware, while the other _destroy_ functions tear down
driver data structures.
Split parts of the hardware reset function up so that I can
do part of the reset in suspsend. Split out the software disabling
of the qpairs into nvme_ctrlr_disable_qpairs.
Finally, fix a couple of spelling errors in comments related to
this.
Relnotes: Yes
MFC After: 1 week
Reviewed by: scottl@ (prior version)
Differential Revision: https://reviews.freebsd.org/D21493
polling within a second. Panic if we don't. All the commands that use this
interface should typically complete within a few tens to hundreds of
microseconds. Panic rather than return ETIMEDOUT because if the command somehow
does later complete, it will randomly corrupt memory. Also, it helps to get a
traceback from where the unexpected failure happens, rather than an infinite
loop.
