Hi,
On 5/16/2023 1:21 PM, Fengnan Chang wrote:
Under legacy mode, if the number of continuous memsegs greater
than RTE_MAX_MEMSEG_PER_LIST, eal init will failed even though
another memseg list is empty, because only one memseg list used
to check in remap_needed_hugepages.
For example:
hugepage configure:
20480
13370
7110
startup log:
EAL: Detected memory type: socket_id:0 hugepage_sz:2097152
EAL: Detected memory type: socket_id:1 hugepage_sz:2097152
EAL: Creating 4 segment lists: n_segs:8192 socket_id:0 hugepage_sz:2097152
EAL: Creating 4 segment lists: n_segs:8192 socket_id:1 hugepage_sz:2097152
EAL: Requesting 13370 pages of size 2MB from socket 0
EAL: Requesting 7110 pages of size 2MB from socket 1
EAL: Attempting to map 14220M on socket 1
EAL: Allocated 14220M on socket 1
EAL: Attempting to map 26740M on socket 0
EAL: Could not find space for memseg. Please increase 32768 and/or 65536 in
configuration.
Unrelated, but this is probably a wrong message, this should've called
out the config options to change, not their values. Sounds like a log
message needs fixing somewhere...
EAL: Couldn't remap hugepage files into memseg lists
EAL: FATAL: Cannot init memory
EAL: Cannot init memory
Signed-off-by: Fengnan Chang <changfeng...@bytedance.com>
Signed-off-by: Lin Li <lilint...@bytedance.com>
---
lib/eal/linux/eal_memory.c | 2 ++
1 file changed, 2 insertions(+)
diff --git a/lib/eal/linux/eal_memory.c b/lib/eal/linux/eal_memory.c
index 60fc8cc6ca..36b9e78f5f 100644
--- a/lib/eal/linux/eal_memory.c
+++ b/lib/eal/linux/eal_memory.c
@@ -1001,6 +1001,8 @@ remap_needed_hugepages(struct hugepage_file *hugepages,
int n_pages)
if (cur->size == 0)
break;
+ if (cur_page - seg_start_page >= RTE_MAX_MEMSEG_PER_LIST)
+ new_memseg = 1;
I don't think this is quite right, because technically,
`RTE_MAX_MEMSEG_PER_LIST` is only applied to smaller page size segment
lists - larger page sizes segment lists will hit their limits earlier.
So, while this will work for 2MB pages, it won't work for page sizes
which segment list length is smaller than the maximum (such as 1GB pages).
I think this solution could be improved upon by trying to break up the
contiguous area instead. I suspect the core of the issue is not even the
fact that we're exceeding limits of one memseg list, but that we're
always attempting to map exactly N pages in `remap_hugepages`, which
results in us leaving large contiguous zones inside memseg lists unused
because we couldn't satisfy current allocation request and skipped to a
new memseg list.
For example, let's suppose we found a large contiguous area that
would've exceeded limits of current memseg list. Sooner or later, this
contiguous area will end, and we'll attempt to remap this virtual area
into a memseg list. Whenever that happens, we call into the remap code,
which will start with first segment, attempt to find exactly N number of
free spots, fail to do so, and skip to the next segment list.
Thus, sooner or later, if we get contiguous areas that are large enough,
we will not populate our memseg lists but instead skip through them, and
start with a new memseg list every time we need a large contiguous area.
We prioritize having a large contiguous area over using up all of our
memory map.
If, instead, we could break up the allocation - that is, use
`rte_fbarray_find_biggest_free()` instead of
`rte_fbarray_find_next_n_free()`, and keep doing it until we run out of
segment lists, we will achieve the same result your patch does, but have
it work for all page sizes, because now we would be targeting the actual
issue (under-utilization of memseg lists), not its symptoms (exceeding
segment list limits for large allocations).
This logic could either be inside `remap_hugepages`, or we could just
return number of pages mapped from `remap_hugepages`, and have the
calling code (`remap_needed_hugepages`) try again, this time with a
different start segment, reflecting how much pages we actually mapped.
IMO this would be easier to implement, as `remap_hugepages` is overly
complex as it is!
if (cur_page == 0)
new_memseg = 1;
else if (cur->socket_id != prev->socket_id)
--
Thanks,
Anatoly
