On 2022-11-16 16:21, Dean Luick wrote:
On 11/16/2022 9:15 AM, Robin Murphy wrote:
On 2022-11-16 14:40, Dean Luick wrote:
On 11/13/2022 10:35 AM, Christoph Hellwig wrote:
dma_alloc_coherent is an opaque allocator that only uses the GFP_ flags
for allocation context control. Don't pass GFP_USER which doesn't make
sense for a kernel DMA allocation or __GFP_COMP which makes no sense
for an allocation that can't in any way be converted to a page pointer.
Signed-off-by: Christoph Hellwig <hch@xxxxxx>
---
drivers/infiniband/hw/hfi1/init.c | 21 +++------------------
1 file changed, 3 insertions(+), 18 deletions(-)
diff --git a/drivers/infiniband/hw/hfi1/init.c b/drivers/infiniband/hw/hfi1/init.c
index 436372b314312..24c0f0d257fc9 100644
--- a/drivers/infiniband/hw/hfi1/init.c
+++ b/drivers/infiniband/hw/hfi1/init.c
@@ -1761,17 +1761,11 @@ int hfi1_create_rcvhdrq(struct hfi1_devdata *dd, struct hfi1_ctxtdata *rcd)
unsigned amt;
if (!rcd->rcvhdrq) {
- gfp_t gfp_flags;
-
amt = rcvhdrq_size(rcd);
- if (rcd->ctxt < dd->first_dyn_alloc_ctxt || rcd->is_vnic)
- gfp_flags = GFP_KERNEL;
- else
- gfp_flags = GFP_USER;
rcd->rcvhdrq = dma_alloc_coherent(&dd->pcidev->dev, amt,
&rcd->rcvhdrq_dma,
- gfp_flags | __GFP_COMP);
+ GFP_KERNEL);
A user context receive header queue may be mapped into user space. Is that not the use case for GFP_USER? The above conditional is what decides.
Why do you think GFP_USER should be removed here?
Coherent DMA buffers are allocated by a kernel driver or subsystem for the use of a device managed by that driver or subsystem, and thus they fundamentally belong to the kernel as proxy for the device. Any coherent DMA buffer may be mapped to userspace with the dma_mmap_*() interfaces, but they're never a "userspace allocation" in that sense.
My (seemingly dated) understanding is that GFP_USER is for kernel allocations that may be mapped into user space. The description of GFP_USER in gfp_types.h enforces my understanding. Is my uderstanding no longer correct? If not, then what is the point of GFP_USER? Is GFP_USER now mostly an artifact? Should its description be updated?
I think there's a subtle distinction between userspace memory which may
be shared with DMA, and DMA memory which may be shared with userspace.
My reading is that GFP_USER (vs. GFP_HIGHUSER or GFP_HIGHUSER_MOVABLE)
matters for the former case, where you might use alloc_pages(), but
intend to come back and do dma_map_single() at some point afterwards.
For dma_alloc_coherent(), it's clearly always a DMA allocation from the
off, so whether it might also end up mmap()ed into one or more processes
later doesn't really have much practical impact.
Presently, the difference between GFP_KERNEL and GFP_USER is __GFP_HARDWALL. This enforces cpuset allocation policy. If HARDWALL is not set, the allocator will back off to the nearest memory ancestor if needed. The back off seems like a reasonable general policy. I do have one concern that may be hypothetical: if GFP_KERNEL is used and a buffer is silently pushed out of the expected cpuset, this can lead to mysterious slowdowns.
Note that NUMA-aware dma_alloc_coherent() implementations do try to
place the buffer close to the device (since typically over its lifetime
it may be accessed by any CPU, but not any other device), therefore not
being bound by process restrictions is arguably right for that. If a
process is bound to a different node than a device it's trying to use
then *something's* going to have to eat the cross-node latency either
way. I suppose one could argue that GFP_USER might make sense when it's
known that the device is only ever going to use this buffer for work on
behalf of the current process, and thus it could be accounted to the
process instead of the kernel, but that seems incredibly niche and
likely to be far outweighed by people just using it wrong.
Thanks,
Robin.
