On 6/28/23 5:11 PM, Michał Mirosław wrote:
Hi Muhammad,
I'd really like to reduce the number of conditionals in this code by a third. I believe it is possible if we agree to rework the API a bit (making it always a GET operation with optional WP) and relax some of the restrictions on the input values.
I have discussed this multiple times in revisions that why we need the GET, GET+WP and WP operations. We cannot remove only WP operation. Please consult previous revisions on why we need exclusive WP operation without GET.
Also please include *all* changes you make in a patch in the changelog
- I just noticed e.g. a `mmu_notifier` code that I haven't seen before
in the review.
mmu_notifier code was added v16. I'll try to write more longer specific changelog.
A general request from me would be to move all the page classification code to one place. That would mean replacing:
bitmap = PM_SCAN_FLAGS(is_written, false,
pmd_present(*pmd), is_swap_pmd(*pmd),
pmd_present(*pmd) && is_zero_pfn(pmd_pfn(*pmd)));
With:
category_flags = pmd_page_category_flags(*pmd);
... and adding a pmd_page_category_flags() and pte/hugepage versions together at the top. The idea here is to separate concerns: split classification of pages from the code acting on the classification results and be able to easily understand (and review for correctness) how corresponding classes are derived from PTE and PMD values.
This is just a cosmetic change. I don't think we should keep on doing cosmetic change when code is already readable and obvious. At this rate I would keep doing cosmetic changes on the argument that we are making is more easier to read. I've been working on this patch and particularly task_mmu.c file from last 1 year. The current code is in same flavor as the rest of the `task_mmu.c` file. All the functions above are using #ifdef THP.
I'd prefer the naming of the functions and variables to follow that it is a classification result, not a generic "bitmap" or "flags", so that it's harder to confuse them with page flags as used elsewhere in the kernel.
(inline review follows)
BTW, thanks for doing this!
On Wed, 28 Jun 2023 at 12:00, Muhammad Usama Anjum usama.anjum@collabora.com wrote: [...]
diff --git a/fs/proc/task_mmu.c b/fs/proc/task_mmu.c index 507cd4e59d07..be747fb89827 100644 --- a/fs/proc/task_mmu.c +++ b/fs/proc/task_mmu.c
[...]
+#define PM_SCAN_REQUIRE_UFFD (1ULL << 63)
+#define PM_SCAN_FOUND_MAX_PAGES (1) +#define PM_SCAN_END_WALK (-256)
[...]
+static int pagemap_scan_test_walk(unsigned long start, unsigned long end,
struct mm_walk *walk)
+{
struct pagemap_scan_private *p = walk->private;
struct vm_area_struct *vma = walk->vma;
if ((p->flags & PM_SCAN_REQUIRE_UFFD) && (!userfaultfd_wp_async(vma) ||
!userfaultfd_wp_use_markers(vma)))
return -EPERM;
p->vma_can_wp = userfaultfd_wp_async(vma) && userfaultfd_wp_use_markers(vma);
And then use that in classification functions and for skipping WP for pages not prepared for that. The PM_SCAN_REQUIRE_UFFD won't be needed then.
No we don't want to skip, but return error if a VMA isn't setup to use WP Async. Even after adding this new `vma_can_wp` variable. We'll have to check in every test_walk() if UFFD is initialized on the VMA for every VMA. At max, `vma_can_wp` can be setup to hold the true or false for PM_SCAN_REQUIRE_UFFD which makes it cosmetic change again.
if (vma->vm_flags & VM_PFNMAP)
return 1;
Why do we skip VM_PFNMAP vmas? This will skip over at least VDSO and VVAR pages. Those two are not that big a problem, but it should be at least documented what ranges are skipped and why.
We are following pagemap_read(). We don't want to expose any additional information which pagemap_read() doesn't show.
return 0;
+}
+static int pagemap_scan_output(unsigned long bitmap,
struct pagemap_scan_private *p,
unsigned long addr, unsigned int n_pages)
+{
struct page_region *cur_buf = &p->cur_buf;
if (!n_pages)
return -EINVAL;
How can this happen?
This was there to check validity of n_pages before proceeding. By doing static analysis, I can see that it isn't needed anymore and can be removed.
bitmap &= p->return_mask;
if (cur_buf->flags == bitmap &&
cur_buf->start + cur_buf->len * PAGE_SIZE == addr) {
BTW, maybe the ranges returned to the user could also use start .. end form, and then this would be simplified to `cur->flags == categories_to_report && cur->end == addr`.
cur_buf->len += n_pages;
p->found_pages += n_pages;
} else {
if (cur_buf->len) {
if (p->vec_buf_index >= p->vec_buf_len)
return PM_SCAN_END_WALK;
memcpy(&p->vec_buf[p->vec_buf_index], cur_buf,
sizeof(*p->vec_buf));
p->vec_buf_index++;
}
cur_buf->start = addr;
cur_buf->len = n_pages;
cur_buf->flags = bitmap;
p->found_pages += n_pages;
}
if (p->found_pages == p->max_pages)
return PM_SCAN_FOUND_MAX_PAGES;
Since we now return the address the walk ended at, what is the difference for PM_SCAN_END_WALK and PM_SCAN_FOUND_MAX_PAGES, and do we still need any of those instead of -ENOSPC or `n_pages != scan_output(...)` check?
Yes, we need two different return codes from here to judge if we need to wp the current range or not. When PM_SCAN_FOUND_MAX_PAGES is returned we need to wp the current range. But when END_WALK is returned we don't need to perform wp and return.
+static int pagemap_scan_pmd_entry(pmd_t *pmd, unsigned long start,
unsigned long end, struct mm_walk *walk)
+{
bool is_written, flush = false, interesting = true;
struct pagemap_scan_private *p = walk->private;
struct vm_area_struct *vma = walk->vma;
unsigned long bitmap, addr = end;
pte_t *pte, *orig_pte, ptent;
spinlock_t *ptl;
int ret = 0;
arch_enter_lazy_mmu_mode();
+#ifdef CONFIG_TRANSPARENT_HUGEPAGE
ptl = pmd_trans_huge_lock(pmd, vma);
if (ptl) {
unsigned long n_pages = (end - start)/PAGE_SIZE;
if (n_pages > p->max_pages - p->found_pages)
n_pages = p->max_pages - p->found_pages;
is_written = !is_pmd_uffd_wp(*pmd);
bitmap = PM_SCAN_FLAGS(is_written, false,
pmd_present(*pmd), is_swap_pmd(*pmd),
pmd_present(*pmd) && is_zero_pfn(pmd_pfn(*pmd)));
if (IS_PM_SCAN_GET(p->flags))
interesting = pagemap_scan_is_interesting_page(bitmap, p);
if (interesting) {
/*
* Break huge page into small pages if the WP operation
* need to be performed is on a portion of the huge page.
*/
if (is_written && IS_PM_SCAN_WP(p->flags) &&
n_pages < HPAGE_SIZE/PAGE_SIZE) {
spin_unlock(ptl);
split_huge_pmd(vma, pmd, start);
goto process_smaller_pages;
}
if (IS_PM_SCAN_GET(p->flags))
ret = pagemap_scan_output(bitmap, p, start, n_pages);
if (IS_PM_SCAN_WP(p->flags) && is_written && ret >= 0) {
make_uffd_wp_pmd(vma, start, pmd);
flush_tlb_range(vma, start, end);
}
}
spin_unlock(ptl);
arch_leave_lazy_mmu_mode();
return ret;
}
Could you split the THP code and use it here like:
enter_lazy_mmu();
ret = scan_thp(...); if (ret != -ENOTTY) { leave_lazy_mmu(); return ret; }
To avoid having #ifdef here, it can be moved to the scan_thp() only (either in the body or having a dummy `return -ENOTTY;` version in an #else block).
I'm following the flavour of code in this file as explained above. This isn't needed.
BTW, there's no `cond_resched()` in the THP case - is that intentional?
AFAIU, processing of THP here is fast as compared to 512 iterations over smaller pages. So cond_resched() isn't present here. But overall, I've not seen cond_resched() being used after THPs.
+process_smaller_pages: +#endif
orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, start, &ptl);
if (!pte) {
walk->action = ACTION_AGAIN;
return 0;
}
for (addr = start; addr < end && !ret; pte++, addr += PAGE_SIZE) {
ptent = ptep_get(pte);
is_written = !is_pte_uffd_wp(ptent);
bitmap = PM_SCAN_FLAGS(is_written, pagemap_scan_is_file(vma, ptent, addr),
pte_present(ptent), is_swap_pte(ptent),
pte_present(ptent) && is_zero_pfn(pte_pfn(ptent)));
if (IS_PM_SCAN_GET(p->flags)) {
interesting = pagemap_scan_is_interesting_page(bitmap, p);
If we consider GET as always done, this can be:
if (!is_interesting) continue;
And that would simplify the code greatly.
if (interesting)
ret = pagemap_scan_output(bitmap, p, addr, 1);
}
if (IS_PM_SCAN_WP(p->flags) && is_written && interesting &&
ret >= 0) {
make_uffd_wp_pte(vma, addr, pte);
flush = true;
}
}
if (flush)
flush_tlb_range(vma, start, addr);
Would optimizing the TLB flush range be beneficial here? If yes, the loop above could do:
flush_start = end; flush_end = end; ... for (...) { if (mark_wp) { make_wp(); if (flush_start > addr) flush_start = addr; flush_end = addr; } }
to reduce the flushed range.
Just tried this. There is some bug or hidden restriction. It isn't working. Maybe we can look at it latter. Reducing the flush size may actually be beneficial.
pte_unmap_unlock(orig_pte, ptl);
arch_leave_lazy_mmu_mode();
cond_resched();
return ret;
+}
+#ifdef CONFIG_HUGETLB_PAGE +static int pagemap_scan_hugetlb_entry(pte_t *ptep, unsigned long hmask,
unsigned long start, unsigned long end,
struct mm_walk *walk)
+{
unsigned long n_pages = (end - start)/PAGE_SIZE;
struct pagemap_scan_private *p = walk->private;
struct vm_area_struct *vma = walk->vma;
bool is_written, interesting = true;
struct hstate *h = hstate_vma(vma);
unsigned long bitmap;
spinlock_t *ptl;
int ret = 0;
pte_t ptent;
if (IS_PM_SCAN_WP(p->flags) && n_pages < HPAGE_SIZE/PAGE_SIZE)
return -EINVAL;
Shouldn't this be checked after the `n_pages` is reduced? BTW, maybe check it only if the page `is_written` to avoid the conditional for all walks where WP is not used?
No, we (you and me) had reached to this current state after discussion. We are returning -EINVAL here for the user to understand that he had not passed the entire hugetlb page address range when specified WP op.
if (n_pages > p->max_pages - p->found_pages)
n_pages = p->max_pages - p->found_pages;
[...]
Proposing the relaxed API (and, while at it, less generic constant and field names) below.
+static int pagemap_scan_args_valid(struct pm_scan_arg *arg, unsigned long start,
unsigned long end, struct page_region __user *vec)
+{
/* Detect illegal size, flags, len and masks */
if (arg->size != sizeof(struct pm_scan_arg))
return -EINVAL;
if (!arg->flags)
return -EINVAL;
if (arg->flags & ~PM_SCAN_OPS)
return -EINVAL;
if (arg->flags & ~PM_SCAN_FLAGS) return -EINVAL;
We need if (!flags) condition for the case when flag is zero.
if (!(end - start))
return -EINVAL;
Remove. Nothing bad will happen when trying to scan an empty range.
We should return error. It is user's fault for not providing the correct range.
if ((arg->required_mask | arg->anyof_mask | arg->excluded_mask |
arg->return_mask) & ~PM_SCAN_BITS_ALL)
return -EINVAL;
if ((...) & ~PM_SCAN_PAGE_CATEGORIES) return -EINVAL;
if (!arg->required_mask && !arg->anyof_mask &&
!arg->excluded_mask)
return -EINVAL;
Remove. Will inspect all pages then.
if (!arg->return_mask)
return -EINVAL;
Remove. Will not return any ranges then. (But will WP, if requested.)
/* Validate memory range */
if (!IS_ALIGNED(start, PAGE_SIZE))
return -EINVAL;
if (!access_ok((void __user *)start, end - start))
return -EFAULT;
if (IS_PM_SCAN_GET(arg->flags)) {
Remove. Do GET always.
if (arg->vec_len == 0)
return -EINVAL;
Remove. Will end the walk at the first matching page (and return its address in `start`).
if (!vec)
return -EINVAL;
if (!access_ok((void __user *)vec,
arg->vec_len * sizeof(struct page_region)))
return -EFAULT;
Check only if vec_len != 0. BTW, return EFAULT for `!vec`.
}
if (IS_PM_SCAN_WP(arg->flags) && !IS_PM_SCAN_GET(arg->flags) &&
arg->max_pages)
return -EINVAL;
With return_mask == 0, arg->max_pages will be ignored anyway. We can just document that this limits the pages reported in the output vector. (And from that follows that if not returning anything, the value doesn't make a difference.)
I'm sorry. I don't see any value in channgig the interface again. The interface is based on requirement, not the simplification of the code.
[...]
p.max_pages = (arg.max_pages) ? arg.max_pages : ULONG_MAX;
p.found_pages = 0;
p.required_mask = arg.required_mask;
p.anyof_mask = arg.anyof_mask;
p.excluded_mask = arg.excluded_mask;
p.return_mask = arg.return_mask;
p.flags = arg.flags;
p.flags |= ((p.required_mask | p.anyof_mask | p.excluded_mask) &
PAGE_IS_WRITTEN) ? PM_SCAN_REQUIRE_UFFD : 0;
p.cur_buf.start = p.cur_buf.len = p.cur_buf.flags = 0;
p.vec_buf = NULL;
p.vec_buf_len = PAGEMAP_WALK_SIZE >> PAGE_SHIFT;
This needs to be less by 1 to account for the entry in p.cur_buf.
No, that would be bug.
/*
* Allocate smaller buffer to get output from inside the page walk
* functions and walk page range in PAGEMAP_WALK_SIZE size chunks. As
* we want to return output to user in compact form where no two
* consecutive regions should be continuous and have the same flags.
* So store the latest element in p.cur_buf between different walks and
* store the p.cur_buf at the end of the walk to the user buffer.
*/
if (IS_PM_SCAN_GET(p.flags)) {
if (p.vec_len != 0)
No, this is wrong.
p.vec_buf = kmalloc_array(p.vec_buf_len, sizeof(*p.vec_buf),
GFP_KERNEL);
if (!p.vec_buf)
return -ENOMEM;
}
if (IS_PM_SCAN_WP(p.flags)) {
mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA, 0,
mm, start, end);
mmu_notifier_invalidate_range_start(&range);
}
Please add a comment why this is needed.
This was added when we had left using change_pmd_range(). This is a notification that protection for the range has been changed.
walk_start = walk_end = start;
while (walk_end < end && !ret) {
if (IS_PM_SCAN_GET(p.flags)) {
if (p.vec_len)
p.vec_buf_index = 0;
/*
* All data is copied to cur_buf first. When more data
* is found, we push cur_buf to vec_buf and copy new
* data to cur_buf. Subtract 1 from length as the
* index of cur_buf isn't counted in length.
*/
empty_slots = arg.vec_len - vec_index;
p.vec_buf_len = min(p.vec_buf_len, empty_slots - 1);
}
ret = mmap_read_lock_killable(mm);
if (ret)
goto return_status;
This could be _interruptible() now as we can return early overwriting `start` so that userspace can detect that the walk was interrupted by a signal (start != end).
walk_end = min((walk_start + PAGEMAP_WALK_SIZE) & PAGEMAP_WALK_MASK, end);
ret = walk_page_range(mm, walk_start, walk_end,
&pagemap_scan_ops, &p);
mmap_read_unlock(mm);
if (ret && ret != PM_SCAN_FOUND_MAX_PAGES &&
ret != PM_SCAN_END_WALK)
goto return_status;
So now there is no difference in the two special error values.
Inside pagemap_scan_pmd_entry() the purpose is different.
walk_start = walk_end;
if (IS_PM_SCAN_GET(p.flags) && p.vec_buf_index) {
`if (p.vec_buf_index)` should be enough, as if not reporting the ranges, the index will never increase.
Agreed. Will update.
if (copy_to_user(&vec[vec_index], p.vec_buf,
p.vec_buf_index * sizeof(*p.vec_buf))) {
/*
* Return error even though the OP succeeded
*/
ret = -EFAULT;
goto return_status;
}
vec_index += p.vec_buf_index;
p.vec_buf_index = 0;
... so that there is no need to do that at every start of the loop iteration.
Okay.
}
}
if (p.cur_buf.len) {
if (copy_to_user(&vec[vec_index], &p.cur_buf, sizeof(p.cur_buf))) {
ret = -EFAULT;
goto return_status;
}
vec_index++;
}
ret = vec_index;
+return_status:
The label name looks too narrowing, considering what is being done here now. Maybe just 'out', as there is no other non-trivial exit path from the function?
Okay.
arg.start = (unsigned long)walk_end;
if (copy_to_user(&uarg->start, &arg.start, sizeof(arg.start)))
ret = -EFAULT;
if (IS_PM_SCAN_WP(p.flags))
mmu_notifier_invalidate_range_end(&range);
With removal of the OP_GET, there is only a single flag left. I don't think it is useful then to hide it behind a macro. It should be readable as `p.flags & PM_SCAN_DO_WP`.
[...]
+/* Bits are set in flags of the page_region and masks in pm_scan_args */ +#define PAGE_IS_WRITTEN (1 << 0) +#define PAGE_IS_FILE (1 << 1) +#define PAGE_IS_PRESENT (1 << 2) +#define PAGE_IS_SWAPPED (1 << 3) +#define PAGE_IS_PFNZERO (1 << 4)
Please add PAGE_IS_UFFD_WP_ENABLED (or a shorter name) - this would be populated from `p->vma_can_wp` in the `test_walk` implementation above.
Best Regards Michał Mirosław