| Age | Commit message (Collapse) | Author |
|---|
|
The test of write combining was added before in mlx5_ib driver. It
opens UD QP and posts NOP WQEs, and uses BlueFlame doorbell. When
BlueFlame is used, WQEs get written directly to a PCI BAR of the
device (in addition to memory) so that the device handles them without
having to access memory.
In this test, the WQEs written in memory are different from the ones
written to the BlueFlame which request CQE update. By checking the
completion reports posted on CQ, we can know if BlueFlame succeeds or
not. The write combining must be supported if BlueFlame succeeds as
its register is written using write combining.
This patch reimplements the test in the same way, but using a pair of
SQ and CQ only. It is moved to mlx5_core as a general feature used by
both mlx5_core and mlx5_ib.
Besides, save write combine test result of the PCI function, so that
its thousands of child functions such as SF can query without paying
the time and resource penalty by itself. The test function is called
only after failing to get the cached result. With this enhancement,
all thousands of SFs of the PF attached to same driver no longer need
to perform WC check explicitly, which is already done in the system.
This saves several commands per SF, thereby speeds up SF creation and
also saves completion EQ creation.
Signed-off-by: Jianbo Liu <jianbol@nvidia.com>
Reviewed-by: Tariq Toukan <tariqt@nvidia.com>
Link: https://lore.kernel.org/r/4ff5a8cc4c5b5b0d98397baa45a5019bcdbf096e.1717409369.git.leon@kernel.org
Signed-off-by: Leon Romanovsky <leon@kernel.org>
|
|
mlx5 has a built in self-test at driver startup to evaluate if the
platform supports write combining to generate a 64 byte PCIe TLP or
not. This has proven necessary because a lot of common scenarios end up
with broken write combining (especially inside virtual machines) and there
is other way to learn this information.
This self test has been consistently failing on new ARM64 CPU
designs (specifically with NVIDIA Grace's implementation of Neoverse
V2). The C loop around writeq() generates some pretty terrible ARM64
assembly, but historically this has worked on a lot of existing ARM64 CPUs
till now.
We see it succeed about 1 time in 10,000 on the worst effected
systems. The CPU architects speculate that the load instructions
interspersed with the stores makes the WC buffers statistically flush too
often and thus the generation of large TLPs becomes infrequent. This makes
the boot up test unreliable in that it indicates no write-combining,
however userspace would be fine since it uses a ST4 instruction.
Further, S390 has similar issues where only the special zpci_memcpy_toio()
will actually generate large TLPs, and the open coded loop does not
trigger it at all.
Fix both ARM64 and S390 by switching to __iowrite64_copy() which now
provides architecture specific variants that have a high change of
generating a large TLP with write combining. x86 continues to use a
similar writeq loop in the generate __iowrite64_copy().
Fixes: 11f552e21755 ("IB/mlx5: Test write combining support")
Link: https://lore.kernel.org/r/6-v3-1893cd8b9369+1925-mlx5_arm_wc_jgg@nvidia.com
Tested-by: Niklas Schnelle <schnelle@linux.ibm.com>
Acked-by: Leon Romanovsky <leonro@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
|
|
rdma/ib_umem.h and rdma/ib_verbs.h are included by rdma/ib_umem_odp.h.
This patch removes the redundant entries.
Link: https://lore.kernel.org/r/20220823025131.862811-1-matsuda-daisuke@fujitsu.com
Signed-off-by: Daisuke Matsuda <matsuda-daisuke@fujitsu.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
|
|
This new function is defined only on ARM and serves to guarantee a barrier
in the WC operation. The barrier means that another run of this loop will
not combine with the stores this loop created.
On x86 this is happening implicitly because of the spin_unlock().
Link: https://lore.kernel.org/r/0-v1-c5dade92f363+11-mlx5_io_stop_wc_jgg@nvidia.com
Suggested-by: Pavel Shamis <Pavel.Shamis@arm.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
|
|
There is no need in include of module.h in the following files.
Link: https://lore.kernel.org/r/3ab153e25c7ea59599022dc7fe3c409fcfe1aac1.1642960861.git.leonro@nvidia.com
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
|
|
Coverity reports "Potentially overflowing expression ..." warning, which
is correct thing to complain from the compiler point of view, but this is
not possible in the current code. Still, this is a small error as there
are some future situations that might need to use a 32 bit offset. Use ULL
so the calculation works up to 63.
Fixes: b045db62f6f6 ("RDMA/mlx5: Use ib_umem_find_best_pgoff() for SRQ")
Link: https://lore.kernel.org/r/20201125061704.6580-1-leon@kernel.org
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
|
|
Since devx uses the new rdma_for_each_block() to fill the PAS it can also
use ib_umem_find_best_pgsz().
However, the umem constructionin devx is complicated, the umem must still
respect all the HW limits such as page_offset_quantized and the IOVA
alignment.
Since we don't know what the user intends to use the umem for we have to
limit it to PAGE_SIZE.
There are users trying to mix umem's with mkeys so this makes them work
reliably, at least for an identity IOVA, by ensuring the IOVA matches the
selected page size.
Last user of mlx5_ib_get_buf_offset() so it can also be removed.
Fixes: aeae94579caf ("IB/mlx5: Add DEVX support for memory registration")
Link: https://lore.kernel.org/r/20201115114311.136250-7-leon@kernel.org
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
|
|
Delete custom logic in the QP in favor of more general variant.
Link: https://lore.kernel.org/r/20201115114311.136250-5-leon@kernel.org
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
|
|
SRQ uses a quantized and scaled page_offset, which is another variation of
ib_umem_find_best_pgsz(). Add mlx5_umem_find_best_quantized_pgoff() to
perform this calculation for each mailbox. A macro shows how the
calculation is directly connected to the mailbox format.
This new routine replaces the limited mlx5_ib_cont_pages() and
mlx5_ib_get_buf_offset() pairing which would reject valid configurations
rather than adjust the page_size to make it work.
In turn this is much more aggressive about choosing large page sizes for
these objects and when THP is enabled it will now often find a single page
solution.
Link: https://lore.kernel.org/r/20201115114311.136250-2-leon@kernel.org
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
|
|
Mixing these together is just a mess, make a dedicated version,
mlx5_ib_update_mr_pas(), which directly loads the whole MTT for a non-ODP
MR.
The split out version can trivially use a simple loop with
rdma_for_each_block() which allows using the core code to compute the MR
pages and avoids seeking in the SGL list after each chunk as the
__mlx5_ib_populate_pas() call required.
Significantly speeds loading large MTTs.
Link: https://lore.kernel.org/r/20201026132314.1336717-5-leon@kernel.org
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
|
|
This routine converts the umem SGL into a list of fixed pages for DMA,
which is exactly what rdma_umem_for_each_dma_block() is for, use the
common code directly.
Link: https://lore.kernel.org/r/20201026132314.1336717-2-leon@kernel.org
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
|
|
Most callers don't need this, and the few that do can get it as
ib_umem_num_pages(umem).
Link: https://lore.kernel.org/r/20201026131936.1335664-8-leon@kernel.org
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
|
|
This is the same as ib_umem_num_dma_blocks(umem, 1UL << page_shift), have
the callers compute it directly.
Link: https://lore.kernel.org/r/20201026131936.1335664-7-leon@kernel.org
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
|
|
Only alloc_mr_from_cache() needs order and can trivially compute it, so
lift it to the one call site and remove the NULL arguments.
Link: https://lore.kernel.org/r/20201026131936.1335664-6-leon@kernel.org
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
|
|
For the user MR path, instead of calling this after getting the umem, call
it as part of creating the struct mlx5_ib_mr and distill its output to a
single page_shift stored inside the mr.
This avoids passing around the tuple of its output. Based on the umem and
page_shift, the output arguments can be computed using:
count == ib_umem_num_pages(mr->umem)
shift == mr->page_shift
ncont == ib_umem_num_dma_blocks(mr->umem, 1 << mr->page_shift)
order == order_base_2(ncont)
And since mr->page_shift == umem_odp->page_shift then ncont ==
ib_umem_num_dma_blocks() == ib_umem_odp_num_pages() for ODP umems.
Link: https://lore.kernel.org/r/20201026131936.1335664-5-leon@kernel.org
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
|
|
ib_umem_num_pages() should only be used by things working with the SGL in
CPU pages directly.
Drivers building DMA lists should use the new ib_num_dma_blocks() which
returns the number of blocks rdma_umem_for_each_block() will return.
To make this general for DMA drivers requires a different implementation.
Computing DMA block count based on umem->address only works if the
requested page size is < PAGE_SIZE and/or the IOVA == umem->address.
Instead the number of DMA pages should be computed in the IOVA address
space, not umem->address. Thus the IOVA has to be stored inside the umem
so it can be used for these calculations.
For now set it to umem->address by default and fix it up if
ib_umem_find_best_pgsz() was called. This allows drivers to be converted
to ib_umem_num_dma_blocks() safely.
Link: https://lore.kernel.org/r/6-v2-270386b7e60b+28f4-umem_1_jgg@nvidia.com
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
|
|
Now that we have direct and reliable detection of WC support by the
system, use is broadly. The only case we have to worry about is when the
WC autodetector cannot run.
For this fringe case generally assume that that WC is available, except in
the well defined case of no PAT support on x86 which is tested by calling
arch_can_pci_mmap_wc().
If WC is wrongly assumed to be available then it causes a small
performance hit on paths in userspace that are tuned to the assumption
that WC is available. There is no functional loss.
It is very unlikely that any platforms exist that lack WC and also care
about the micro optimization of WC in the fringe case where autodetection
does not work.
By removing the fairly bogus CONFIG tests this makes WC work broadly on
all arches and all platforms.
Link: https://lore.kernel.org/r/20200318100323.46659-1-leon@kernel.org
Signed-off-by: Yishai Hadas <yishaih@mellanox.com>
Reviewed-by: Michael Guralnik <michaelgur@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
|
|
Building MR translation table in the ODP case requires additional
flexibility, namely random access to DMA addresses. Make both direct and
indirect ODP MR use same code path, separated from the non-ODP MR code
path.
With the restructuring the correct page_shift is now used around
__mlx5_ib_populate_pas().
Fixes: d2183c6f1958 ("RDMA/umem: Move page_shift from ib_umem to ib_odp_umem")
Link: https://lore.kernel.org/r/20191222124649.52300-2-leon@kernel.org
Signed-off-by: Artemy Kovalyov <artemyko@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
|
|
Linux can run in all sorts of physical machines and VMs where write
combining may or may not be supported. Currently there is no way to
reliably tell if the system supports WC, or not. The driver uses WC to
optimize posting work to the HCA, and getting this wrong in either
direction can cause a significant performance loss.
Add a test in mlx5_ib initialization process to test whether
write-combining is supported on the machine. The test will run as part of
the enable_driver callback to ensure that the test runs after the device
is setup and can create and modify the QP needed, but runs before the
device is exposed to the users.
The test opens UD QP and posts NOP WQEs, the WQE written to the BlueFlame
is different from the WQE in memory, requesting CQE only on the BlueFlame
WQE. By checking whether we received a completion on one of these WQEs we
can know if BlueFlame succeeded and this write-combining must be
supported.
Change reporting of BlueFlame support to be dependent on write-combining
support instead of the FW's guess as to what the machine can do.
Link: https://lore.kernel.org/r/20191027062234.10993-1-leon@kernel.org
Signed-off-by: Michael Guralnik <michaelgur@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
|
|
This is the last creation API that is overloaded for both, there is very
little code sharing and a driver has to be specifically ready for a
umem_odp to be created to use the odp version.
Link: https://lore.kernel.org/r/20190819111710.18440-7-leon@kernel.org
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
|
|
When ODP is enabled with IB_ACCESS_HUGETLB then the required pages
should be calculated based on the extent of the MR, which is rounded
to the nearest huge page alignment.
Fixes: d2183c6f1958 ("RDMA/umem: Move page_shift from ib_umem to ib_odp_umem")
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Link: https://lore.kernel.org/r/20190815083834.9245-5-leon@kernel.org
Signed-off-by: Doug Ledford <dledford@redhat.com>
|
|
This value has always been set to PAGE_SHIFT in the core code, the only
thing that does differently was the ODP path. Move the value into the ODP
struct and still use it for ODP, but change all the non-ODP things to just
use PAGE_SHIFT/PAGE_SIZE/PAGE_MASK directly.
Reviewed-by: Shiraz Saleem <shiraz.saleem@intel.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
|
|
Convert various places to more readable code, which embeds
CONFIG_INFINIBAND_ON_DEMAND_PAGING into the code flow.
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
|
|
This no longer has any use, we can use container_of to get to the
umem_odp, and a simple flag to indicate if this is an odp MR. Remove the
few remaining references to it.
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
|
|
The patch simplifies mlx5_ib_cont_pages and fixes the following
issues in the original implementation:
First issues is related to alignment of the PFNs. After the check
base + p != PFN, the alignment of the PFN wasn't checked. So the PFN
sequence 0, 1, 1, 2 would result in a page_shift of 13 even though
the 3rd PFN is not 8KB aligned.
This wasn't actually a bug because it was supported by all the
existing mlx5 compatible device, but we don't want to require
this support in all future devices.
Another issue is because the inner loop didn't advance PFN so
the test "if (base + p != pfn)" always failed for SGE with
len > (1<<page_shift).
Fixes: e126ba97dba9 ("mlx5: Add driver for Mellanox Connect-IB adapters")
Signed-off-by: Ilya Lesokhin <ilyal@mellanox.com>
Reviewed-by: Eli Cohen <eli@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Doug Ledford <dledford@redhat.com>
|
|
Currenlty ODP supports only regular MMU pages.
Add ODP support for regions consisting of physically contiguous chunks
of arbitrary order (huge pages for instance) to improve performance.
Signed-off-by: Artemy Kovalyov <artemyko@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Doug Ledford <dledford@redhat.com>
|
|
Size of pages are held by struct ib_umem in page_size field.
It is better to store it as an exponent, because page size by nature
is always power-of-two and used as a factor, divisor or ilog2's argument.
The conversion of page_size to be page_shift allows to have portable
code and avoid following error while compiling on ARM:
ERROR: "__aeabi_uldivmod" [drivers/infiniband/core/ib_core.ko] undefined!
CC: Selvin Xavier <selvin.xavier@broadcom.com>
CC: Steve Wise <swise@chelsio.com>
CC: Lijun Ou <oulijun@huawei.com>
CC: Shiraz Saleem <shiraz.saleem@intel.com>
CC: Adit Ranadive <aditr@vmware.com>
CC: Dennis Dalessandro <dennis.dalessandro@intel.com>
CC: Ram Amrani <Ram.Amrani@Cavium.com>
Signed-off-by: Artemy Kovalyov <artemyko@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Acked-by: Ram Amrani <Ram.Amrani@cavium.com>
Acked-by: Shiraz Saleem <shiraz.saleem@intel.com>
Acked-by: Selvin Xavier <selvin.xavier@broadcom.com>
Acked-by: Selvin Xavier <selvin.xavier@broadcom.com>
Acked-by: Adit Ranadive <aditr@vmware.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
|
|
In this change we turn mlx5_ib_update_mtt() into generic
mlx5_ib_update_xlt() to perfrom HCA translation table modifiactions
supporting both atomic and process contexts and not limited by number
of modified entries.
Using this function we increase preallocated MRs up to 16GB.
Signed-off-by: Artemy Kovalyov <artemyko@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
The maximum page size in the mkey context is 2GB.
Until today, we didn't enforce this requirement in the code,
and therefore, if we got a page size larger than 2GB, we
have passed zeros in the log_page_shift instead of the actual value
and the registration failed.
This patch limits the driver to use compound pages of 2GB for mkeys.
Fixes: e126ba97dba9 ('mlx5: Add driver for Mellanox Connect-IB adapters')
Signed-off-by: Maor Gottlieb <maorg@mellanox.com>
Signed-off-by: Majd Dibbiny <majd@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Doug Ledford <dledford@redhat.com>
|
|
The 2nd parameter of 'find_first_bit' is the number of bits to search.
In this case, we are passing 'sizeof(tmp)' which is likely to be 4 or 8
because 'tmp' is an 'unsigned long'.
It is likely that the number of bits of 'tmp' was expected here. So use
BITS_PER_LONG instead.
It has been spotted by the following coccinelle script:
@@
expression ret, x;
@@
* ret = \(find_first_bit \| find_first_zero_bit\) (x, sizeof(...));
Signed-off-by: Christophe JAILLET <christophe.jaillet@wanadoo.fr>
Acked-by: Majd Dibbiny <majd@mellanox.com>
Acked-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
|
|
Signed-off-by: Achiad Shochat <achiad@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
Signed-off-by: Eli Cohen <eli@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
The return type of find_first_bit() is architecture specific,
on ARM it is 'unsigned int', while the asm-generic code used
on x86 and a lot of other architectures returns 'unsigned long'.
When building the mlx5 driver on ARM, we get a warning about
this:
infiniband/hw/mlx5/mem.c: In function 'mlx5_ib_cont_pages':
infiniband/hw/mlx5/mem.c:84:143: warning: comparison of distinct pointer types lacks a cast
m = min(m, find_first_bit(&tmp, sizeof(tmp)));
This patch changes the driver to use min_t to make it behave
the same way on all architectures.
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Acked-by: Eli Cohen <eli@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
The new function allows updating the page tables of a memory region
after it was created. This can be used to handle page faults and page
invalidations.
Since mlx5_ib_update_mtt will need to work from within page invalidation,
so it must not block on memory allocation. It employs an atomic memory
allocation mechanism that is used as a fallback when kmalloc(GFP_ATOMIC) fails.
In order to reuse code from mlx5_ib_populate_pas, the patch splits
this function and add the needed parameters.
Signed-off-by: Haggai Eran <haggaie@mellanox.com>
Signed-off-by: Shachar Raindel <raindel@mellanox.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
|
|
This patch wraps together several changes needed for on-demand paging support
in the mlx5_ib_populate_pas function, and when registering memory regions.
* Instead of accepting a UMR bit telling the function to enable all
access flags, the function now accepts the access flags themselves.
* For on-demand paging memory regions, fill the memory tables from the
correct list, and enable/disable the access flags per-page according
to whether the page is present.
* A new bit is set to enable writing of access flags when using the
firmware create_mkey command.
* Disable contig pages when on-demand paging is enabled.
In addition the patch changes the UMR code to use PTR_ALIGN instead of
our own macro.
Signed-off-by: Haggai Eran <haggaie@mellanox.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
|
|
When dealing with umem objects, the driver assumed host page sizes
defined by PAGE_SHIFT. Modify the code to use arbitrary page shift
provided by umem->page_shift to support different page sizes.
Signed-off-by: Yishai Hadas <yishaih@mellanox.com>
Signed-off-by: Eli Cohen <eli@mellanox.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
|
|
There were many places where parameters which should be u8/u16 were
integer type.
Additionally, in 2 places, a check for a non-null pointer was added
before dereferencing the pointer (this is actually a bug fix).
Signed-off-by: Jack Morgenstein <jackm@dev.mellanox.co.il>
Signed-off-by: Eli Cohen <eli@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
This patch refactors the IB core umem code and vendor drivers to use a
linear (chained) SG table instead of chunk list. With this change the
relevant code becomes clearer—no need for nested loops to build and
use umem.
Signed-off-by: Shachar Raindel <raindel@mellanox.com>
Signed-off-by: Yishai Hadas <yishaih@mellanox.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
|
|
The driver is comprised of two kernel modules: mlx5_ib and mlx5_core.
This partitioning resembles what we have for mlx4, except that mlx5_ib
is the pci device driver and not mlx5_core.
mlx5_core is essentially a library that provides general functionality
that is intended to be used by other Mellanox devices that will be
introduced in the future. mlx5_ib has a similar role as any hardware
device under drivers/infiniband/hw.
Signed-off-by: Eli Cohen <eli@mellanox.com>
Signed-off-by: Jack Morgenstein <jackm@dev.mellanox.co.il>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[ Merge in coccinelle fixes from Fengguang Wu <fengguang.wu@intel.com>.
- Roland ]
Signed-off-by: Roland Dreier <roland@purestorage.com>
|
