On Wed, 2018-05-30 at 06:59 +0200, joro@8bytes.org wrote:
On Tue, May 29, 2018 at 04:10:24PM +0000, Kani, Toshi wrote:
Can you explain why you think allocating a page here is a major problem?
Because a larger allocation is more likely to fail. And if you fail the allocation, you also fail to free more pages, which _is_ a problem. So better avoid any allocations in code paths that are about freeing memory.
It only allocates a single page. If it fails to allocate a single page, this pud mapping request simply falls to pmd mappings, which is only as bad as your suggested plain revert always does for both pud and pmd mappings. Also, this func is called from ioremap() when a driver initializes its mapping. If the system does not have a single page to allocate, the driver has a much bigger issue to deal with than its request falling back to pmd mappings. Please also note that this func is not called from free-memory path.
If we just revert, please apply patch 1/3 first. This patch address the BUG_ON issue on PAE. This is a real issue that needs a fix ASAP.
It does not address the problem of dirty page-walk caches on x86-64.
This patch 3/3 fixes it. Hope my explanation above clarified.
The page-directory cache issue on x64, which is addressed by patch 3/3, is a theoretical issue that I could not hit by putting ioremap() calls into a loop for a whole day. Nobody hit this issue, either.
How do you know you didn't hit that issue? It might cause silent data corruption, which might not be easily detected.
If the test hit that issue, it would have caused a kernel page fault (freed & cleared pte page still unmapped, or this page reused and attribute data invalid) or MCE (pte page reused and phys addr invalid) when it accessed to ioremap'd address.
Causing silent data corruption requires that this freed & cleared pte page to be reused and re-initialized with a valid pte entry data. While this case is possible, the chance of my test only hitting this case without ever hitting much more likely cases of page fault or MCE is basically zero.
The simple revert patch Joerg posted a while ago causes pmd_free_pte_page() to fail on x64. This causes multiple pmd mappings to fall into pte mappings on my test systems. This can be seen as a degradation, and I am afraid that it is more harmful than good.
The plain revert just removes all the issues with the dirty TLB that the original patch introduced and prevents huge mappings from being established when there have been smaller mappings before. This is not ideal, but at least its is consistent and does not leak pages and leaves no dirty TLBs. So this is the easiest and most reliable fix for this stage in the release process.
If you think the page directory issue needs a fix ASAP, I believe taking patch 3/3 is much better option than the plain revert, which will introduce the fall-back issue I explained.
Thanks, -Toshi