Re: [PATCH v10 06/13] khugepaged: add mTHP support
From: Nico Pache <npache@redhat.com>
Date: 2025-09-08 22:30:22
Also in:
linux-doc, linux-mm, lkml
On Fri, Sep 5, 2025 at 4:15 AM Lorenzo Stoakes [off-list ref] wrote:
You've not responded to a ton of comments, I'm guesing I should assume in those cases you're acking the comments implicitly? :) Do let me know.
Yes! If they are obvious nits, then implicit ack. Sorry I wrote my response over a couple of days while making changes based on the review and trying to figure out a solution to the "creep" issue, and forgot to thoroughly review all the comments before sending this out. I'll make sure to go through it again.
On Tue, Sep 02, 2025 at 02:12:38PM -0600, Nico Pache wrote:quoted
On Wed, Aug 20, 2025 at 12:30 PM Lorenzo Stoakes [off-list ref] wrote:quoted
On Tue, Aug 19, 2025 at 07:41:58AM -0600, Nico Pache wrote:quoted
Introduce the ability for khugepaged to collapse to different mTHP sizes. While scanning PMD ranges for potential collapse candidates, keep track of pages in KHUGEPAGED_MIN_MTHP_ORDER chunks via a bitmap. Each bit represents a utilized region of order KHUGEPAGED_MIN_MTHP_ORDER ptes. If mTHPs are enabled we remove the restriction of max_ptes_none during the scan phase so we don't bailout early and miss potential mTHP candidates. A new function collapse_scan_bitmap is used to perform binary recursion on the bitmap and determine the best eligible order for the collapse. A stack struct is used instead of traditional recursion. max_ptes_none will be scaled by the attempted collapse order to determine how "full" an order must be before being considered for collapse. Once we determine what mTHP sizes fits best in that PMD range a collapse is attempted. A minimum collapse order of 2 is used as this is the lowest order supported by anon memory. For orders configured with "always", we perform greedy collapsing to that order without considering bit density. If a mTHP collapse is attempted, but contains swapped out, or shared pages, we don't perform the collapse. This is because adding new entries can lead to new none pages, and these may lead to constant promotion into a higher order (m)THP. A similar issue can occur with "max_ptes_none > HPAGE_PMD_NR/2" due to the fact that a collapse will introduce at least 2x the number of pages, and on a future scan will satisfy the promotion condition once again. For non-PMD collapse we must leave the anon VMA write locked until after we collapse the mTHP-- in the PMD case all the pages are isolated, but in the non-PMD case this is not true, and we must keep the lock to prevent changes to the VMA from occurring. Currently madv_collapse is not supported and will only attempt PMD collapse.Yes I think this has to remain the case unfortunately as we override sysfs-specified orders for MADV_COLLAPSE and there's no sensible way to determine what order we ought to be using.quoted
Signed-off-by: Nico Pache <npache@redhat.com>You've gone from small incremental changes to a huge one here... for the sake of reviewer sanity at least, any chance of breaking this up?I had this as two patches (one for the bitmap and one for implementing it), but I was asked to squash them :/Yeah it makes sense to combine those two, but maybe there's a better way of splitting things out.
I'll give it a shot but most of this patch belongs together in my eyes.
quoted
quoted
quoted
--- include/linux/khugepaged.h | 4 + mm/khugepaged.c | 236 +++++++++++++++++++++++++++++-------- 2 files changed, 188 insertions(+), 52 deletions(-)diff --git a/include/linux/khugepaged.h b/include/linux/khugepaged.h index eb1946a70cff..d12cdb9ef3ba 100644 --- a/include/linux/khugepaged.h +++ b/include/linux/khugepaged.h@@ -1,6 +1,10 @@ /* SPDX-License-Identifier: GPL-2.0 */ #ifndef _LINUX_KHUGEPAGED_H #define _LINUX_KHUGEPAGED_H +#define KHUGEPAGED_MIN_MTHP_ORDER 2I guess this makes sense as by definition 2 pages is least it could possibly be.Order, so 4 pages, 16kB mTHPRight, misread that! :) sorry. I guess then in fact there is no such thing as an order-1 mTHP? I did wonder how useful that'd be so makes more sense
Anon memory does not support order 1 THPs, im not sure if there are any plans do so either.
quoted
quoted
quoted
+#define KHUGEPAGED_MIN_MTHP_NR (1 << KHUGEPAGED_MIN_MTHP_ORDER)Surely KHUGEPAGED_MIN_NR_MTHP_PTES would be more meaningful?Sure!Thanks!quoted
quoted
quoted
+#define MAX_MTHP_BITMAP_SIZE (1 << (ilog2(MAX_PTRS_PER_PTE) - KHUGEPAGED_MIN_MTHP_ORDER))This is confusing - size of what?We need it like this due to ppc64 (and maybe others?), it used to be based on PMD_ORDER, but some arches fail to compile due to the PMD size only being known at boot time.That really sucks. Do please put this context in a comment though (see below for more discussion on this).
Ok, will do. although with my most recent changes we might just be able to get rid of this all together.
quoted
This compiles to 9 on arches that have 512 ptes. so 1 << (9 - 2) == 128What I'm saying is - what is this expressed in? There's no information on that here.
In my next implementation its probably going to be gone, but what it indicates is the bitmap size (128), where each bit equals a range of 4 pages. 128 *4 = 512.
So from what you say I can see it's the number of bits required in the bitmap, because we're being smart and only bothering to mark entries at a granularity of minimum mTHP size. OK so the idea is this is the number of PTE entries per entry in the bitmap.
Yeah it was an optimization that I was going for, but on second thought it's probably pretty useless and causes some inaccuracies in the max_ptes_none scaling and collapsing.
This is KEY missing context. We really need to spell things out here, the THP code is... confusing :) to put it politely, so we need to be especially careful to be mega clear.
Yeah its rather confusing, sorry, but I think my newer implementation that gets rid of this is actually much cleaner.
So please please PLEASE add a comment explaining all this. And clearly state that the unit here is bits.
ack, I'll be more clear with my newer implementation, and it should be easier to understand anyways.
quoted
quoted
If it's number of bits surely this should be ilog2(MAX_PTRS_PER_PTE) - KHUGEPAGED_MIN_MTHP_ORDER?This would only be 7? We need a 128 bit bitmapAgain missing context is you're storing bits per minimum mTHP as per above.quoted
quoted
This seems to be more so 'the maximum value that could contain the bits right? I think this is just wrong though, see below at DECLARE_BITMAP() stuff.quoted
+#define MTHP_BITMAP_SIZE (1 << (HPAGE_PMD_ORDER - KHUGEPAGED_MIN_MTHP_ORDER))Hard to know how this relates to MAX_MTHP_BITMAP_SIZE? I guess this is the current bitmap size indicating all that is possible, but if these are all #define's what is this accomplishing?One for compile time one for runtime. Kind of annoying but it was the easiest solution given the architecture limitations.OK, this context is fantastic + important for understanding, so let's put it in a comment :)
Or just get rid of it ;)
quoted
quoted
For all - please do not do (1 << xxx)! This can lead to sign-extension bugs at least in theory, use _BITUL(...), it's neater too.ack, thanks!OK sorry, based on David's feedback on this - just use 1UL << xxx here instead.
ack
(An aside that isn't really relevant now but - also I was wrong to suggest _BITUL() anyway BIT() is The Way (TM), for some reason I had it in my head that the former was better :P)quoted
quoted
NIT but the whitespace is all screwed up here. KHUGEPAGED_MIN_MTHP_ORDER and KHUGEPAGED_MIN_MTHP_NRquoted
#include <linux/mm.h>diff --git a/mm/khugepaged.c b/mm/khugepaged.c index 074101d03c9d..1ad7e00d3fd6 100644 --- a/mm/khugepaged.c +++ b/mm/khugepaged.c@@ -94,6 +94,11 @@ static DEFINE_READ_MOSTLY_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS); static struct kmem_cache *mm_slot_cache __ro_after_init; +struct scan_bit_state { + u8 order; + u16 offset; +}; + struct collapse_control { bool is_khugepaged;@@ -102,6 +107,18 @@ struct collapse_control { /* nodemask for allocation fallback */ nodemask_t alloc_nmask; + + /* + * bitmap used to collapse mTHP sizes. + * 1bit = order KHUGEPAGED_MIN_MTHP_ORDER mTHPI'm not sure what this '1bit = xxx' comment means?A single bit represents 1 << MIN_MTHP_ORDER (4) pages. Ill express that betterYeah again, this is the missing context that would have helped me feel a lot less confused here :Pquoted
quoted
quoted
+ */ + DECLARE_BITMAP(mthp_bitmap, MAX_MTHP_BITMAP_SIZE);Hmm this seems wrong.Should be a bitmap with 128 bits (for 4k page size). Not sure what's wrong here.quoted
DECLARE_BITMAP(..., val) is expessed as: #define DECLARE_BITMAP(name,bits) \ unsigned long name[BITS_TO_LONGS(bits)] So the 2nd param should be number of bits. But MAX_MTHP_BITMAP_SIZE is: (1 << (ilog2(MAX_PTRS_PER_PTE) - KHUGEPAGED_MIN_MTHP_ORDER)) So typically: (1 << (9 - 2)) = 128 And BITS_TO_LONGS is defined as: __KERNEL_DIV_ROUND_UP(nr, BITS_PER_TYPE(long)) So essentially this will be 128 / 8 on a 64-bit system so 16 bytes to store... 7 bits?I think you mean 64. 8 would be BYTES_PER_TYPEquoted
Unless I'm missing something here?Hmm, unless the DECLARE_BITMAP is being used incorrectly in multiple places, DECLARE_BITMAP(..., # of bits) is how this is intended to be used. I think it's an array of unsigned longs, so each part of the name[] is already 64 bits. hence the divide.Yup again this is due to the missing context above.quoted
quoted
quoted
+ DECLARE_BITMAP(mthp_bitmap_temp, MAX_MTHP_BITMAP_SIZE);Same comment as above obviously. But also this is kind of horrible, why are we putting a copy of this entire bitmap on the stack every time we declare a cc?The temp one is used as a scratch pad, Baolin also finds this useful in his file mTHP collapse useful for another use as well.Saying 'scratch pad' is really just saying 'temporary' using different words :) this isn't hugely helpful.
When doing the "recursion", there is no easy way to read the bitmap between bits N0-N1, so we must manipulate it to extract that data; however, we still want to preserve the original state of the bitmap. The temp_bitmap is used to do these bit manipulations while not destroying the original state. I will see if there's a better way to read arbitrary ranges of the bitmap so we dont need to have two copies, but IIRC there is not-- I might have to expand the bitmap API to include this.
If we _need_ this we should give this a better name, and I also don't know why we need this on the stack for all collapse_control users. Can't you just have your 'scratch pad' local to the function that needs it or passed as a pointer?
If I cant find a solution to the problem above I will look into moving it to the local stack.
I'm sure it'd be useful to somebody to stick various temporary things in vm_area_struct for instance, but it'd be hugely egregious to do so...
The reason I didnt see this as a huge issue, is there is only one global CC that is reused for all khugepaged runs. It's a one time allocation and does not scale.
quoted
In general khugepaged always uses the same CC, so it doesn't not having to constantly allocate this.You're putting a LOT of data on the stack, and kernel stacks are very delicate, I'm quite concerned about this. BEFORE: struct collapse_control { bool is_khugepaged; /* 0 1 */ /* XXX 3 bytes hole, try to pack */ u32 node_load[64]; /* 4 256 */ /* XXX 4 bytes hole, try to pack */ /* --- cacheline 4 boundary (256 bytes) was 8 bytes ago --- */ nodemask_t alloc_nmask; /* 264 8 */ /* size: 272, cachelines: 5, members: 3 */ /* sum members: 265, holes: 2, sum holes: 7 */ /* last cacheline: 16 bytes */ }; 272 bytes. AFTER: struct collapse_control { bool is_khugepaged; /* 0 1 */ /* XXX 3 bytes hole, try to pack */ u32 node_load[64]; /* 4 256 */ /* XXX 4 bytes hole, try to pack */ /* --- cacheline 4 boundary (256 bytes) was 8 bytes ago --- */ nodemask_t alloc_nmask; /* 264 8 */ long unsigned int mthp_bitmap[2]; /* 272 16 */ long unsigned int mthp_bitmap_temp[2]; /* 288 16 */ struct scan_bit_state mthp_bitmap_stack[128]; /* 304 512 */ /* size: 816, cachelines: 13, members: 6 */ /* sum members: 809, holes: 2, sum holes: 7 */ /* last cacheline: 48 bytes */ }; 816 bytes. So you're roughly tripling the size of this struct and making this a thing in all callstacks that reference collapse_control. Kernel stack is a _super_ finite resource. I mean I can't really review your stack implementation thing until you've added a comment to help me understand what you are doing there (sorry but it's just too fiddly for me to first principles it), but I wonder if we truly need to be doing this? I wonder if we can just put this into somewhere allocated...quoted
quoted
quoted
+ struct scan_bit_state mthp_bitmap_stack[MAX_MTHP_BITMAP_SIZE]; +}; + +struct collapse_control khugepaged_collapse_control = { + .is_khugepaged = true, };Why are we moving this here?Because if not it doesn't compile.A reason as to why this is the case would be helpful :) But fair enough!quoted
quoted
quoted
/**@@ -854,10 +871,6 @@ static void khugepaged_alloc_sleep(void) remove_wait_queue(&khugepaged_wait, &wait); } -struct collapse_control khugepaged_collapse_control = { - .is_khugepaged = true, -}; - static bool collapse_scan_abort(int nid, struct collapse_control *cc) { int i;@@ -1136,17 +1149,19 @@ static int alloc_charge_folio(struct folio **foliop, struct mm_struct *mm, static int collapse_huge_page(struct mm_struct *mm, unsigned long address, int referenced, int unmapped, - struct collapse_control *cc) + struct collapse_control *cc, bool *mmap_locked, + unsigned int order, unsigned long offset) { LIST_HEAD(compound_pagelist); pmd_t *pmd, _pmd; - pte_t *pte; + pte_t *pte = NULL, mthp_pte; pgtable_t pgtable; struct folio *folio; spinlock_t *pmd_ptl, *pte_ptl; int result = SCAN_FAIL; struct vm_area_struct *vma; struct mmu_notifier_range range; + unsigned long _address = address + offset * PAGE_SIZE;This name is really horrible. please name it sensibly. It feels like address ought to be consistently the base of the THP or mTHP we wish to collapse, and if we need something PMD aligned for some reason we should rename _that_ to e.g. pmd_address. Orrr it could be mthp_address... Perhaps we could just figure that out here and pass only the address... aligning to PMD boundary shouldn't be hard/costly. But it may indicate we need further refactorisation so we don't need to paper over cracks + pass around a PMD address to do things when that may not be where the (m)THP range begins.Ok i'll rename them, but we still need to know the PMD address as we rely on it for a few key operations. Can we leave _address and rename address to pmd_address?No, we absolutely cannot leave _address as '_address', that's terrible and I'm just not going to live with that. I know it's trivial, but it's just such a horrid naming convention.
Ok Ill rename it to something else!
quoted
quoted
quoted
VM_BUG_ON(address & ~HPAGE_PMD_MASK);@@ -1155,16 +1170,20 @@ static int collapse_huge_page(struct mm_struct *mm, unsigned long address, * The allocation can take potentially a long time if it involves * sync compaction, and we do not need to hold the mmap_lock during * that. We will recheck the vma after taking it again in write mode. + * If collapsing mTHPs we may have already released the read_lock. */ - mmap_read_unlock(mm); + if (*mmap_locked) { + mmap_read_unlock(mm); + *mmap_locked = false; + } - result = alloc_charge_folio(&folio, mm, cc, HPAGE_PMD_ORDER); + result = alloc_charge_folio(&folio, mm, cc, order); if (result != SCAN_SUCCEED) goto out_nolock; mmap_read_lock(mm); - result = hugepage_vma_revalidate(mm, address, true, &vma, cc, - BIT(HPAGE_PMD_ORDER)); + *mmap_locked = true; + result = hugepage_vma_revalidate(mm, address, true, &vma, cc, BIT(order));I mean this is kind of going back to previous commits, but it's really ugly to pass a BIT(xxx) here, is that really necessary? Can't we just pass in the order?Yes and no... currently we only ever pass the bit of the current order so we could get away with it, but to generalize it we want the ability to pass a bitmap of the available orders. Like in the case of future madvise_collapse support, we would need to pass a bitmap of possible orders.Can we just change that when we need it?
Sure, it was more so that I generalized it for madv_collapse and khugepaged, but then we dropped madv_collapse support, and I never undid the change.
'Future proofing' for an possible future implementation detail is just not how we should implement things. Right now we don't do that, so let's just pass the order. If in future we want to change it we can.
ack
quoted
quoted
It's also inconsistent with other calls like e.g. __collapse_huge_page_swapin() below which passes the order. Same goes obv. for all such invocations.quoted
if (result != SCAN_SUCCEED) { mmap_read_unlock(mm); goto out_nolock;@@ -1182,13 +1201,14 @@ static int collapse_huge_page(struct mm_struct *mm, unsigned long address, * released when it fails. So we jump out_nolock directly in * that case. Continuing to collapse causes inconsistency. */ - result = __collapse_huge_page_swapin(mm, vma, address, pmd, - referenced, HPAGE_PMD_ORDER); + result = __collapse_huge_page_swapin(mm, vma, _address, pmd, + referenced, order); if (result != SCAN_SUCCEED) goto out_nolock; } mmap_read_unlock(mm); + *mmap_locked = false; /* * Prevent all access to pagetables with the exception of * gup_fast later handled by the ptep_clear_flush and the VM@@ -1198,8 +1218,7 @@ static int collapse_huge_page(struct mm_struct *mm, unsigned long address, * mmap_lock. */ mmap_write_lock(mm); - result = hugepage_vma_revalidate(mm, address, true, &vma, cc, - BIT(HPAGE_PMD_ORDER)); + result = hugepage_vma_revalidate(mm, address, true, &vma, cc, BIT(order)); if (result != SCAN_SUCCEED) goto out_up_write; /* check if the pmd is still valid */@@ -1210,11 +1229,12 @@ static int collapse_huge_page(struct mm_struct *mm, unsigned long address, anon_vma_lock_write(vma->anon_vma); - mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, address, - address + HPAGE_PMD_SIZE); + mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, _address, + _address + (PAGE_SIZE << order));This _address is horrible. That really does have to change.quoted
mmu_notifier_invalidate_range_start(&range); pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */ +Odd whitespace...quoted
/* * This removes any huge TLB entry from the CPU so we won't allow * huge and small TLB entries for the same virtual address to@@ -1228,19 +1248,16 @@ static int collapse_huge_page(struct mm_struct *mm, unsigned long address, mmu_notifier_invalidate_range_end(&range); tlb_remove_table_sync_one(); - pte = pte_offset_map_lock(mm, &_pmd, address, &pte_ptl); + pte = pte_offset_map_lock(mm, &_pmd, _address, &pte_ptl);I see we already have a 'convention' of _ prefix on the pmd param, but two wrongs don't make a right...quoted
if (pte) { - result = __collapse_huge_page_isolate(vma, address, pte, cc, - &compound_pagelist, - HPAGE_PMD_ORDER); + result = __collapse_huge_page_isolate(vma, _address, pte, cc, + &compound_pagelist, order); spin_unlock(pte_ptl); } else { result = SCAN_PMD_NULL; } if (unlikely(result != SCAN_SUCCEED)) { - if (pte) - pte_unmap(pte);Why are we removing this?You're missing some of the comments, I'm guesing for most of the smaller stuff you're just implicitly acking them but this one was a question :)
Ah, for some of your reviews you comment, then later when you see another piece you say something like "ah ok this is why", for a lot of these cases I might have skipped replying because it seems that you figured it out. For this one we moved the pte_unmap further down, because if not, we could be prematurely unmapping a PTE, which is later referenced during the mTHP collapse, which was not the case when we only did PMD collapse.
quoted
quoted
quoted
spin_lock(pmd_ptl); BUG_ON(!pmd_none(*pmd)); /*@@ -1255,17 +1272,17 @@ static int collapse_huge_page(struct mm_struct *mm, unsigned long address, } /* - * All pages are isolated and locked so anon_vma rmap - * can't run anymore. + * For PMD collapse all pages are isolated and locked so anon_vma + * rmap can't run anymore */ - anon_vma_unlock_write(vma->anon_vma); + if (order == HPAGE_PMD_ORDER) + anon_vma_unlock_write(vma->anon_vma);Hmm this is introducing a horrible new way for things to go wrong. And there's now a whole host of terrible error paths that can go wrong very easily around rmap locks and yeah, no way we cannot do it this way. rmap locks are VERY sensitive and the ordering of the locking matters a great deal (see top of mm/rmap.c). So we have to be SO careful here. I suggest you simply have a boolean 'anon_vma_locked' or something like this, and get rid of these horrible additional code paths and the second order == HPAGE_PMD_ORDER check. We'll track whether or not the lock is held and thereby needs releasing that way instead.You've not responded to this suggestion re: refactoring here.
I think if anything this refactoring adds complexity... we simply postpone the unlock we normally do, to later if its mTHP collapse, because we havent fully isolated all the pages (like we do in the PMD case). We arent doing anything extremely different other than hold the already held lock for a little longer. We would still need these order == HPAGE_PMD_ORDER checks, but just be flipping a flag that says its locked, essentially adding more complexity for the same outcome.
I'm really not a fan of us arbitrarily messing with rmap locks like this, and we should very carefully keep track of whether we have/have not released them.
Its important to note, the code was already doing this, as I stated above, we simply hold it for longer because we haven't isolated all the pages.
Again rmap locking is a dangerous area, I've got personal experience of this (see top of mm/rmap.c for an indication of complexity here as well as https://kernel.org/doc/html/latest/mm/process_addrs.html).quoted
quoted
Also, and very importantly - are you 100% sure you can't possibly have a deadlock or issue beyond this point if you don't release the rmap lock?I double checked, this was added as a fix to an issue Hugh reported.quoted
The gap between these callers is rather small, and I see no way that it could skip the lock/unlock cycle.We're going to need more than this to be confident, you need to clearly justify why we won't encounter issues this way.
// WE ARE HOLDING THE LOCK ALREADY
if (order == HPAGE_PMD_ORDER) //ONLY RELEASE IF ITS PMD
anon_vma_unlock_write(vma->anon_vma);
result = __collapse_huge_page_copy(pte, folio, pmd, _pmd,
vma, _address, pte_ptl,
order, &compound_pagelist);
if (unlikely(result != SCAN_SUCCEED))
goto out_unlock_anon_vma; //LETS RELEASE THE LOCK IF WE FAIL
if (order == HPAGE_PMD_ORDER) {
...
// NO EXIT PATHS
} else { /* mTHP collapse */
...
// NO EXIT PATHS
}
folio = NULL;
result = SCAN_SUCCEED;
out_unlock_anon_vma:
if (order != HPAGE_PMD_ORDER)
anon_vma_unlock_write(vma->anon_vma); //RELEASE IF IT WASN'T
RELEASED EARLIER
Not sure if this is the verification you are looking for, but I really
dont see how this could go wrong, or how this is any different than
what the khugepaged code is already doing in a number of places.quoted
quoted
This is veeeery important, as there can be implicit assumptions around whether or not one can acquire these locks and you basically have to audit ALL code over which this lock is held. I'm speaking from hard experience here having bumped into this in various attempts at work relating to this stuff...quoted
result = __collapse_huge_page_copy(pte, folio, pmd, _pmd, - vma, address, pte_ptl, - &compound_pagelist, HPAGE_PMD_ORDER); - pte_unmap(pte); + vma, _address, pte_ptl, + &compound_pagelist, order); if (unlikely(result != SCAN_SUCCEED)) - goto out_up_write; + goto out_unlock_anon_vma;See above...quoted
/* * The smp_wmb() inside __folio_mark_uptodate() ensures the@@ -1273,33 +1290,115 @@ static int collapse_huge_page(struct mm_struct *mm, unsigned long address, * write. */ __folio_mark_uptodate(folio); - pgtable = pmd_pgtable(_pmd); - - _pmd = folio_mk_pmd(folio, vma->vm_page_prot); - _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma); - - spin_lock(pmd_ptl); - BUG_ON(!pmd_none(*pmd)); - folio_add_new_anon_rmap(folio, vma, address, RMAP_EXCLUSIVE); - folio_add_lru_vma(folio, vma); - pgtable_trans_huge_deposit(mm, pmd, pgtable); - set_pmd_at(mm, address, pmd, _pmd); - update_mmu_cache_pmd(vma, address, pmd); - deferred_split_folio(folio, false); - spin_unlock(pmd_ptl); + if (order == HPAGE_PMD_ORDER) { + pgtable = pmd_pgtable(_pmd); + _pmd = folio_mk_pmd(folio, vma->vm_page_prot); + _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma); + + spin_lock(pmd_ptl); + BUG_ON(!pmd_none(*pmd));I know you're refactoring this, but be good to change this to a WARN_ON_ONCE(), BUG_ON() is verboten unless it's absolutely definitely going to be a kernel nuclear event, so worth changing things up as we go.quoted
Yeah i keep seeing those warning in checkpatch, so Ill go ahead and edit it.Thanks!quoted
quoted
quoted
+ folio_add_new_anon_rmap(folio, vma, _address, RMAP_EXCLUSIVE); + folio_add_lru_vma(folio, vma); + pgtable_trans_huge_deposit(mm, pmd, pgtable); + set_pmd_at(mm, address, pmd, _pmd); + update_mmu_cache_pmd(vma, address, pmd); + deferred_split_folio(folio, false); + spin_unlock(pmd_ptl); + } else { /* mTHP collapse */ + mthp_pte = mk_pte(&folio->page, vma->vm_page_prot);I guess it's a rule that each THP or mTHP range spanned must span one and only one folio. Not sure &folio->page has a future though. Maybe better to use folio_page(folio, 0)?Ok sounds good I'll use that.Thanks!quoted
quoted
quoted
+ mthp_pte = maybe_mkwrite(pte_mkdirty(mthp_pte), vma); + + spin_lock(pmd_ptl); + BUG_ON(!pmd_none(*pmd));having said the above, this is trictly introducing a new BUG_ON() which is a no-no, please make it a WARN_ON_ONCE().This one is more important, please do do this.quoted
quoted
quoted
+ folio_ref_add(folio, (1 << order) - 1);Again no 1 << x please.(as per David feedback elsewhere, 1UL << instead)
ack
quoted
quoted
Do we do something similar somewhere else for mthp ref counting? Can we share code somehow?quoted
Yeah but IIRC its only like 2 or 3 places that do something like this... most callers to folio_add_* do things in slightly different manners. Maybe something to look into for the future, but I think it will be difficult to generalize it.OK.quoted
quoted
quoted
+ folio_add_new_anon_rmap(folio, vma, _address, RMAP_EXCLUSIVE); + folio_add_lru_vma(folio, vma); + set_ptes(vma->vm_mm, _address, pte, mthp_pte, (1 << order));Please avoid 1 << order, and I think at this point since you reference it a bunch of times, just store a local var like nr_pages or sth?quoted
yeah not a bad idea!Thanks!quoted
quoted
quoted
+ update_mmu_cache_range(NULL, vma, _address, pte, (1 << order)); + + smp_wmb(); /* make pte visible before pmd */Can you give some detail as to why this will work here and why it is necessary?quoted
Other parts of the kernel do it when setting ptes before updating the PMD. I'm not sure if it's necessary, but better safe than sorry.Unfortunately this is a _totally_ unacceptable justification. We can't put in barriers based on 'better safe than sorry'. You need to analysis this and determine whether or not it's necessary. So the comment in pmd_install() seems to give an indication: if (likely(pmd_none(*pmd))) { /* Has another populated it ? */ mm_inc_nr_ptes(mm); /* * Ensure all pte setup (eg. pte page lock and page clearing) are * visible before the pte is made visible to other CPUs by being * put into page tables. * * The other side of the story is the pointer chasing in the page * table walking code (when walking the page table without locking; * ie. most of the time). Fortunately, these data accesses consist * of a chain of data-dependent loads, meaning most CPUs (alpha * being the notable exception) will already guarantee loads are * seen in-order. See the alpha page table accessors for the * smp_rmb() barriers in page table walking code. */ smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */ pmd_populate(mm, pmd, *pte); *pte = NULL; }
So this seems to indicate we do need it for some rare edge cases/arches, and that's why it exists in other callers too.
quoted
quoted
quoted
+ pmd_populate(mm, pmd, pmd_pgtable(_pmd));If we're updating PTE entriess why do we need to assign the PMD entry?quoted
We removed the PMD entry for GUP_fast reasons, then we reinstall the PMD entry after the mTHP is in place. Same as for PMD collapse.OK a comment to this effect would be useful.
Ok ill add a small comment. feel free to take a look at the failure handling around __collapse_huge_page_isolate to see another example of this. We remove the pmd (pmdp_collapse_flush), then if the isolation fails, we repopulate the original PMD. In the mTHP case, because we are not installing a new PMD, we still remove the PMD (for gup-fast reasons), modify the PTE entries within that PMD, then reinstall the original PMD. This prevent GUP-fast from accessing these PTEs/PMD mid change.
quoted
quoted
quoted
+ spin_unlock(pmd_ptl); + }This deeply, badly needs to be refactored into something that both shares code and separates out these two operations. This function is disgustingly long as it is, and that's not your fault, but let's try to make things better as we go.quoted
folio = NULL; result = SCAN_SUCCEED; +out_unlock_anon_vma: + if (order != HPAGE_PMD_ORDER) + anon_vma_unlock_write(vma->anon_vma);Obviously again as above, we need to simplify this and get rid of this whole bit.quoted
out_up_write: + if (pte) + pte_unmap(pte);OK I guess you moved this from above down here? Is this a valid place to do this?Yes if not we were potentially unmapping a pte early.quoted
quoted
mmap_write_unlock(mm); out_nolock: + *mmap_locked = false;This is kind of horrible, we now have pretty mad logic around who sets mmap_locked and where. Can we just do this at the call sites so we avoid that? I mean anything we do with this is hideous, but that'd be less confusing It hink.quoted
if (folio) folio_put(folio); trace_mm_collapse_huge_page(mm, result == SCAN_SUCCEED, result); return result; } +/* Recursive function to consume the bitmap */Err... please don't? Kernel stack is a seriously finite resource, we do not want recursion at all. But I'm not actually seeing any recursion here? Am I missing something?Yup this was before I realised it was an iterative implementation :) Though we are putting load on the stack anyway...
Ill see if I can shrink this load!
quoted
quoted
quoted
+static int collapse_scan_bitmap(struct mm_struct *mm, unsigned long address, + int referenced, int unmapped, struct collapse_control *cc, + bool *mmap_locked, unsigned long enabled_orders)This is a complicated and confusing function, it requires a comment describing how it works.Ok will do!This is _VERY_ key - I need this I think to be able to sanely review this code.
ack, im also simplifying it a bit in the next iteration.
quoted
quoted
quoted
+{ + u8 order, next_order; + u16 offset, mid_offset; + int num_chunks; + int bits_set, threshold_bits; + int top = -1;Err why do we start at -1 then immediately increment it?You are correct, it was probably a leftover bit from my development phase. Seems I can just set it to 0 to begin with.Thanks!quoted
quoted
quoted
+ int collapsed = 0; + int ret; + struct scan_bit_state state; + bool is_pmd_only = (enabled_orders == (1 << HPAGE_PMD_ORDER));Extraneous outer parens.ackThanks!quoted
quoted
quoted
+ + cc->mthp_bitmap_stack[++top] = (struct scan_bit_state) + { HPAGE_PMD_ORDER - KHUGEPAGED_MIN_MTHP_ORDER, 0 };This is the same as cc->mthp_bitmap_stack[0] = ...; top = 1; No?quoted
no it would be bitmap_stack[0] = ... then top goes to -1 (at state =... ), and if we add more items (next_order) to the stack it would go top = 1 (adds one for each half of the split)OK, again going to need that comment to really grok all this... :)quoted
quoted
This is really horrible. Can we just have a helper function for this please?quoted
Seems kinda excessive for 4 lines and one caller.It's code that's very horrible to understand using a very unusual way of initialising a struct within the kernel.
ok, ill see if i can clean it up.
And having code be understandable when the compiler can inline for us does indeed justify this even in the instance of 4 lines and one caller :)
ack
quoted
quoted
Like: static int mthp_push_stack(struct collapse_control *cc, int index, u8 order, u16 offset) { struct scan_bit_state *state = &cc->mthp_bitmap_stack[index]; VM_WARN_ON(index >= MAX_MTHP_BITMAP_SIZE); state->order = order; state->offset = offset; return index + 1; }This would not work in its current state because its ++index in the current implementation. I would need to refactor, but the general idea still standsOK thanks. Something _like_ this would make things a great deal clearer.quoted
quoted
And can invoke via: top = mthp_push_stack(cc, top, order, offset); Or pass index as a pointer possibly also.quoted
+ + while (top >= 0) { + state = cc->mthp_bitmap_stack[top--];OK so this is the recursive bit... Oh man this function so needs a comment describing what it does, seriously. I think honestly for sake of my own sanity I'm going to hold off reviewing the rest of this until there's something describing the algorithm, in detail here, above the function.quoted
It's basically binary recursion with a stack structure, that checks regions of the bitmap in descending order (ie order 9, order 8, ...) if we go to the next order we add two items to the stack (left and right half). I will add a comment describing it at the top of the function.Right, to reiterate - this needs a _big_ comment. I'm sorry to split the review of this patch in two on this, because once you do that I'm going to inevitably do the deferred review on the algorithm, but I just feel this is the only sensible way I can determine whether what you intend to do here makes sense and is correctly implemented.
Its all good! This pushed me to come up with a better solution, by both removing the compression into 128 bits, and also come up with a solution to the creep (although its not the 0/511 that you and David were talking about).
Kinda need that 'what you intend to do' bit. So it should be an expansive and detailed comment explicitly explaining the algorithm and why you're doing it. Thanks!quoted
quoted
quoted
+ order = state.order + KHUGEPAGED_MIN_MTHP_ORDER; + offset = state.offset; + num_chunks = 1 << (state.order); + /* Skip mTHP orders that are not enabled */ + if (!test_bit(order, &enabled_orders)) + goto next_order; + + /* copy the relavant section to a new bitmap */ + bitmap_shift_right(cc->mthp_bitmap_temp, cc->mthp_bitmap, offset, + MTHP_BITMAP_SIZE); + + bits_set = bitmap_weight(cc->mthp_bitmap_temp, num_chunks); + threshold_bits = (HPAGE_PMD_NR - khugepaged_max_ptes_none - 1) + >> (HPAGE_PMD_ORDER - state.order); + + /* Check if the region is "almost full" based on the threshold */ + if (bits_set > threshold_bits || is_pmd_only + || test_bit(order, &huge_anon_orders_always)) { + ret = collapse_huge_page(mm, address, referenced, unmapped, + cc, mmap_locked, order, + offset * KHUGEPAGED_MIN_MTHP_NR); + if (ret == SCAN_SUCCEED) { + collapsed += (1 << order); + continue; + } + } + +next_order: + if (state.order > 0) { + next_order = state.order - 1; + mid_offset = offset + (num_chunks / 2); + cc->mthp_bitmap_stack[++top] = (struct scan_bit_state) + { next_order, mid_offset }; + cc->mthp_bitmap_stack[++top] = (struct scan_bit_state) + { next_order, offset }; + } + } + return collapsed; +} + static int collapse_scan_pmd(struct mm_struct *mm, struct vm_area_struct *vma, unsigned long address, bool *mmap_locked,@@ -1307,31 +1406,60 @@ static int collapse_scan_pmd(struct mm_struct *mm, { pmd_t *pmd; pte_t *pte, *_pte; + int i; int result = SCAN_FAIL, referenced = 0; int none_or_zero = 0, shared = 0; struct page *page = NULL; struct folio *folio = NULL; unsigned long _address; + unsigned long enabled_orders; spinlock_t *ptl; int node = NUMA_NO_NODE, unmapped = 0; + bool is_pmd_only; bool writable = false; - + int chunk_none_count = 0; + int scaled_none = khugepaged_max_ptes_none >> (HPAGE_PMD_ORDER - KHUGEPAGED_MIN_MTHP_ORDER); + unsigned long tva_flags = cc->is_khugepaged ? TVA_KHUGEPAGED : TVA_FORCED_COLLAPSE; VM_BUG_ON(address & ~HPAGE_PMD_MASK); result = find_pmd_or_thp_or_none(mm, address, &pmd); if (result != SCAN_SUCCEED) goto out; + bitmap_zero(cc->mthp_bitmap, MAX_MTHP_BITMAP_SIZE); + bitmap_zero(cc->mthp_bitmap_temp, MAX_MTHP_BITMAP_SIZE);Having this 'temp' thing on the stack for everyone is just horrid.As I mention above this serves a very good purpose, and is also expanded in another series by Baolin to serve another similar purpose too.Yeah, I'm not hugely convinced :) responded there. I can probably give more useful feedback on this once the big comment is added!
Ok! hopefully I find a solution so I can remove the temp one.
quoted
quoted
quoted
memset(cc->node_load, 0, sizeof(cc->node_load)); nodes_clear(cc->alloc_nmask); + + if (cc->is_khugepaged) + enabled_orders = thp_vma_allowable_orders(vma, vma->vm_flags, + tva_flags, THP_ORDERS_ALL_ANON); + else + enabled_orders = BIT(HPAGE_PMD_ORDER); + + is_pmd_only = (enabled_orders == (1 << HPAGE_PMD_ORDER));This is horrid, can we have a function broken out to do this please? In general if you keep open coding stuff, just write a static function for it, the compiler is smart enough to inline.ok, we do this is a few places so perhaps its the best approach.quoted
quoted
+ pte = pte_offset_map_lock(mm, pmd, address, &ptl); if (!pte) { result = SCAN_PMD_NULL; goto out; } - for (_address = address, _pte = pte; _pte < pte + HPAGE_PMD_NR; - _pte++, _address += PAGE_SIZE) { + for (i = 0; i < HPAGE_PMD_NR; i++) { + /* + * we are reading in KHUGEPAGED_MIN_MTHP_NR page chunks. if + * there are pages in this chunk keep track of it in the bitmap + * for mTHP collapsing. + */ + if (i % KHUGEPAGED_MIN_MTHP_NR == 0) { + if (i > 0 && chunk_none_count <= scaled_none) + bitmap_set(cc->mthp_bitmap, + (i - 1) / KHUGEPAGED_MIN_MTHP_NR, 1); + chunk_none_count = 0; + }This whole thing is really confusing and you are not explaining the algoritm here at all. This requires a comment, and really this bit should be separated out please.This used to be its own commit, but multiple people wanted it squashed... ugh. Which should we go with?quoted
quoted
+ + _pte = pte + i; + _address = address + i * PAGE_SIZE; pte_t pteval = ptep_get(_pte); if (is_swap_pte(pteval)) { ++unmapped;@@ -1354,10 +1482,11 @@ static int collapse_scan_pmd(struct mm_struct *mm, } } if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) { + ++chunk_none_count; ++none_or_zero; if (!userfaultfd_armed(vma) && - (!cc->is_khugepaged || - none_or_zero <= khugepaged_max_ptes_none)) { + (!cc->is_khugepaged || !is_pmd_only || + none_or_zero <= khugepaged_max_ptes_none)) { continue; } else { result = SCAN_EXCEED_NONE_PTE;@@ -1453,6 +1582,7 @@ static int collapse_scan_pmd(struct mm_struct *mm, address))) referenced++; } + if (!writable) { result = SCAN_PAGE_RO; } else if (cc->is_khugepaged &&@@ -1465,10 +1595,12 @@ static int collapse_scan_pmd(struct mm_struct *mm, out_unmap: pte_unmap_unlock(pte, ptl); if (result == SCAN_SUCCEED) { - result = collapse_huge_page(mm, address, referenced, - unmapped, cc); - /* collapse_huge_page will return with the mmap_lock released */ - *mmap_locked = false; + result = collapse_scan_bitmap(mm, address, referenced, unmapped, cc, + mmap_locked, enabled_orders); + if (result > 0) + result = SCAN_SUCCEED; + else + result = SCAN_FAIL;We're reusing result as both an enum value and as a storage for unmber colapsed PTE entries? Can we just use a new local variable? ThanksAgain you've skipped a ton of review comments here. You need to respond to everything.
Hmm we could, although it would serve no purpose in my eyes...
nr_collapsed = collapse_scan_bitmap()
if(nr_collapsed > 0)
result = SCAN_SUCCEED.
To me that is basically the same thing with extra steps. But maybe its
better for a code readiblity standpoint.quoted
quoted
quoted
} out: trace_mm_khugepaged_scan_pmd(mm, folio, writable, referenced, -- 2.50.1I will review the bitmap/chunk stuff in more detail once the algorithm is commented.ok thanks for the review.No problem! :)