nv_peer_mem: Fix SGls #75
Conversation
nv_peer_mem.c
Outdated
|
|
||
| if (page_table->page_size != NVIDIA_P2P_PAGE_SIZE_64KB) { | ||
| peer_err("nv_dma_map -- assumption of 64KB pages failed size_id=%u\n", | ||
| nv_mem_context->page_table->page_size); | ||
| return -EINVAL; | ||
| } | ||
|
|
||
| system_page_virt_start = nv_mem_context->addr & PAGE_MASK; | ||
| system_page_virt_end = (nv_mem_context->addr + nv_mem_context->size + PAGE_SIZE - 1) & PAGE_MASK; | ||
| system_page_offest = system_page_virt_start - nv_mem_context->page_virt_start; |
| sg->dma_length = nv_mem_context->page_size; | ||
| page_size = nv_mem_context->page_size; | ||
| if (i == nv_mem_context->npages - 1) | ||
| page_size = ((system_page_virt_end - system_page_virt_start + |
There was a problem hiding this comment.
This can be either system page size (4KB x86, 64KB POWER) or 64KB (GPU BAR1), right?
a comment would be useful
Also, that does not work if system page size > GPU_PAGE_SIZE, hypothetically. A BUG_ON/static assert would be useful.
There was a problem hiding this comment.
And don't forget on ppc64le, sometimes the system is using HUGE pages which is configurable (default 2MB on RHEL7/8, but should always be queried).
|
you are also changing nv_mem_get_page_size(), which is beyond the description of this patch. |
|
get_page_size() was a hack that over-rides the automatic page size detection, and for old kernels it makes certain requirements on the construction of the sgl that are incompatible with this. New kernels ignore the result of get_page_size() and use the automatic page size detection from the SGL. Returning a 4k PAGE_SIZE and a properly formed SGL is compatible both ways. |
|
For now, this change will not be merged as it fixes functionality but introduces a performance degradation |
Fix SGls to hold the correct data in both the first and the last entries. Signed-off-by: Artemy Kovalyov <artemyko@mellanox.com> Signed-off-by: Yishai Hadas <yishaih@mellanox.com>
Fix SGls to hold the correct data in both the first and the last entries.