Message ID | 1719478388-31917-1-git-send-email-yangge1116@126.com (mailing list archive) |
---|---|
State | New |
Headers | show |
Series | mm/gup: Use try_grab_page() instead of try_grab_folio() in gup slow path | expand |
On Thu, Jun 27, 2024 at 04:53:08PM +0800, yangge1116@126.com wrote: > From: yangge <yangge1116@126.com> > > If a large number of CMA memory are configured in system (for > example, the CMA memory accounts for 50% of the system memory), > starting a SEV virtual machine will fail. During starting the SEV > virtual machine, it will call pin_user_pages_fast(..., FOLL_LONGTERM, > ...) to pin memory. Normally if a page is present and in CMA area, > pin_user_pages_fast() will first call __get_user_pages_locked() to > pin the page in CMA area, and then call > check_and_migrate_movable_pages() to migrate the page from CMA area > to non-CMA area. But the current code calling __get_user_pages_locked() > will fail, because it call try_grab_folio() to pin page in gup slow > path. > > The commit 57edfcfd3419 ("mm/gup: accelerate thp gup even for "pages > != NULL"") uses try_grab_folio() in gup slow path, which seems to be > problematic because try_grap_folio() will check if the page can be > longterm pinned. This check may fail and cause __get_user_pages_lock() > to fail. However, these checks are not required in gup slow path, > seems we can use try_grab_page() instead of try_grab_folio(). In > addition, in the current code, try_grab_page() can only add 1 to the > page's refcount. We extend this function so that the page's refcount > can be increased according to the parameters passed in. > > The following log reveals it: > > [ 464.325306] WARNING: CPU: 13 PID: 6734 at mm/gup.c:1313 __get_user_pages+0x423/0x520 > [ 464.325464] CPU: 13 PID: 6734 Comm: qemu-kvm Kdump: loaded Not tainted 6.6.33+ #6 > [ 464.325477] RIP: 0010:__get_user_pages+0x423/0x520 > [ 464.325515] Call Trace: > [ 464.325520] <TASK> > [ 464.325523] ? __get_user_pages+0x423/0x520 > [ 464.325528] ? __warn+0x81/0x130 > [ 464.325536] ? __get_user_pages+0x423/0x520 > [ 464.325541] ? report_bug+0x171/0x1a0 > [ 464.325549] ? handle_bug+0x3c/0x70 > [ 464.325554] ? exc_invalid_op+0x17/0x70 > [ 464.325558] ? asm_exc_invalid_op+0x1a/0x20 > [ 464.325567] ? __get_user_pages+0x423/0x520 > [ 464.325575] __gup_longterm_locked+0x212/0x7a0 > [ 464.325583] internal_get_user_pages_fast+0xfb/0x190 > [ 464.325590] pin_user_pages_fast+0x47/0x60 > [ 464.325598] sev_pin_memory+0xca/0x170 [kvm_amd] > [ 464.325616] sev_mem_enc_register_region+0x81/0x130 [kvm_amd] > > Fixes: 57edfcfd3419 ("mm/gup: accelerate thp gup even for "pages != NULL"") > Cc: <stable@vger.kernel.org> > Signed-off-by: yangge <yangge1116@126.com> Thanks for the report and the fix proposed. This is unfortunate.. It's just that I worry this may not be enough, as thp slow gup isn't the only one using try_grab_folio(). There're also hugepd and memfd pinning (which just got queued, again). I suspect both of them can also hit a cma chunk here, and fail whenever they shouldn't have. The slight complexity resides in the hugepd path where it right now shares with fast-gup. So we may potentially need something similiar to what Yang used to introduce in this patch: https://lore.kernel.org/r/20240604234858.948986-2-yang@os.amperecomputing.com So as to identify whether the hugepd gup is slow or fast, and we should only let the fast gup fail on those. Let me also loop them in on the other relevant discussion. Thanks,
On Thu, Jun 27, 2024 at 11:55 AM Peter Xu <peterx@redhat.com> wrote: > > On Thu, Jun 27, 2024 at 04:53:08PM +0800, yangge1116@126.com wrote: > > From: yangge <yangge1116@126.com> > > > > If a large number of CMA memory are configured in system (for > > example, the CMA memory accounts for 50% of the system memory), > > starting a SEV virtual machine will fail. During starting the SEV > > virtual machine, it will call pin_user_pages_fast(..., FOLL_LONGTERM, > > ...) to pin memory. Normally if a page is present and in CMA area, > > pin_user_pages_fast() will first call __get_user_pages_locked() to > > pin the page in CMA area, and then call > > check_and_migrate_movable_pages() to migrate the page from CMA area > > to non-CMA area. But the current code calling __get_user_pages_locked() > > will fail, because it call try_grab_folio() to pin page in gup slow > > path. > > > > The commit 57edfcfd3419 ("mm/gup: accelerate thp gup even for "pages > > != NULL"") uses try_grab_folio() in gup slow path, which seems to be > > problematic because try_grap_folio() will check if the page can be > > longterm pinned. This check may fail and cause __get_user_pages_lock() > > to fail. However, these checks are not required in gup slow path, > > seems we can use try_grab_page() instead of try_grab_folio(). In > > addition, in the current code, try_grab_page() can only add 1 to the > > page's refcount. We extend this function so that the page's refcount > > can be increased according to the parameters passed in. > > > > The following log reveals it: > > > > [ 464.325306] WARNING: CPU: 13 PID: 6734 at mm/gup.c:1313 __get_user_pages+0x423/0x520 > > [ 464.325464] CPU: 13 PID: 6734 Comm: qemu-kvm Kdump: loaded Not tainted 6.6.33+ #6 > > [ 464.325477] RIP: 0010:__get_user_pages+0x423/0x520 > > [ 464.325515] Call Trace: > > [ 464.325520] <TASK> > > [ 464.325523] ? __get_user_pages+0x423/0x520 > > [ 464.325528] ? __warn+0x81/0x130 > > [ 464.325536] ? __get_user_pages+0x423/0x520 > > [ 464.325541] ? report_bug+0x171/0x1a0 > > [ 464.325549] ? handle_bug+0x3c/0x70 > > [ 464.325554] ? exc_invalid_op+0x17/0x70 > > [ 464.325558] ? asm_exc_invalid_op+0x1a/0x20 > > [ 464.325567] ? __get_user_pages+0x423/0x520 > > [ 464.325575] __gup_longterm_locked+0x212/0x7a0 > > [ 464.325583] internal_get_user_pages_fast+0xfb/0x190 > > [ 464.325590] pin_user_pages_fast+0x47/0x60 > > [ 464.325598] sev_pin_memory+0xca/0x170 [kvm_amd] > > [ 464.325616] sev_mem_enc_register_region+0x81/0x130 [kvm_amd] > > > > Fixes: 57edfcfd3419 ("mm/gup: accelerate thp gup even for "pages != NULL"") > > Cc: <stable@vger.kernel.org> > > Signed-off-by: yangge <yangge1116@126.com> > > Thanks for the report and the fix proposed. This is unfortunate.. > > It's just that I worry this may not be enough, as thp slow gup isn't the > only one using try_grab_folio(). There're also hugepd and memfd pinning > (which just got queued, again). > > I suspect both of them can also hit a cma chunk here, and fail whenever > they shouldn't have. > > The slight complexity resides in the hugepd path where it right now shares > with fast-gup. So we may potentially need something similiar to what Yang > used to introduce in this patch: > > https://lore.kernel.org/r/20240604234858.948986-2-yang@os.amperecomputing.com > > So as to identify whether the hugepd gup is slow or fast, and we should > only let the fast gup fail on those. > > Let me also loop them in on the other relevant discussion. Thanks, Peter. I was actually typing the same thing... Yes, I agree my patch should be able to solve the problem. At the beginning I thought it is just a pure clean up patch, but it seems like it is more useful. I'm going to port my patch to the latest mm-unstable, then post it. > > Thanks, > > -- > Peter Xu >
I was complaining that switching from a folio to a page interface is a retro-step. But try_grab_page and try_grab_folio actually both take a strut page argument and do similar but different things. Yikes! > -int __must_check try_grab_page(struct page *page, unsigned int flags) > +int __must_check try_grab_page(struct page *page, int refs, unsigned int flags) This would now make it a try_grab_pages. Also please try to avoid the overly lone lines here and in the external declaration.
diff --git a/mm/gup.c b/mm/gup.c index 6ff9f95..bb58909 100644 --- a/mm/gup.c +++ b/mm/gup.c @@ -222,7 +222,7 @@ static void gup_put_folio(struct folio *folio, int refs, unsigned int flags) * -ENOMEM FOLL_GET or FOLL_PIN was set, but the page could not * be grabbed. */ -int __must_check try_grab_page(struct page *page, unsigned int flags) +int __must_check try_grab_page(struct page *page, int refs, unsigned int flags) { struct folio *folio = page_folio(page); @@ -233,7 +233,7 @@ int __must_check try_grab_page(struct page *page, unsigned int flags) return -EREMOTEIO; if (flags & FOLL_GET) - folio_ref_inc(folio); + folio_ref_add(folio, refs); else if (flags & FOLL_PIN) { /* * Don't take a pin on the zero page - it's not going anywhere @@ -248,13 +248,13 @@ int __must_check try_grab_page(struct page *page, unsigned int flags) * so that the page really is pinned. */ if (folio_test_large(folio)) { - folio_ref_add(folio, 1); - atomic_add(1, &folio->_pincount); + folio_ref_add(folio, refs); + atomic_add(refs, &folio->_pincount); } else { - folio_ref_add(folio, GUP_PIN_COUNTING_BIAS); + folio_ref_add(folio, refs * GUP_PIN_COUNTING_BIAS); } - node_stat_mod_folio(folio, NR_FOLL_PIN_ACQUIRED, 1); + node_stat_mod_folio(folio, NR_FOLL_PIN_ACQUIRED, refs); } return 0; @@ -729,7 +729,7 @@ static struct page *follow_huge_pud(struct vm_area_struct *vma, gup_must_unshare(vma, flags, page)) return ERR_PTR(-EMLINK); - ret = try_grab_page(page, flags); + ret = try_grab_page(page, 1, flags); if (ret) page = ERR_PTR(ret); else @@ -806,7 +806,7 @@ static struct page *follow_huge_pmd(struct vm_area_struct *vma, VM_BUG_ON_PAGE((flags & FOLL_PIN) && PageAnon(page) && !PageAnonExclusive(page), page); - ret = try_grab_page(page, flags); + ret = try_grab_page(page, 1, flags); if (ret) return ERR_PTR(ret); @@ -969,7 +969,7 @@ static struct page *follow_page_pte(struct vm_area_struct *vma, !PageAnonExclusive(page), page); /* try_grab_page() does nothing unless FOLL_GET or FOLL_PIN is set. */ - ret = try_grab_page(page, flags); + ret = try_grab_page(page, 1, flags); if (unlikely(ret)) { page = ERR_PTR(ret); goto out; @@ -1233,7 +1233,7 @@ static int get_gate_page(struct mm_struct *mm, unsigned long address, goto unmap; *page = pte_page(entry); } - ret = try_grab_page(*page, gup_flags); + ret = try_grab_page(*page, 1, gup_flags); if (unlikely(ret)) goto unmap; out: @@ -1636,22 +1636,20 @@ static long __get_user_pages(struct mm_struct *mm, * pages. */ if (page_increm > 1) { - struct folio *folio; /* * Since we already hold refcount on the * large folio, this should never fail. */ - folio = try_grab_folio(page, page_increm - 1, + ret = try_grab_page(page, page_increm - 1, foll_flags); - if (WARN_ON_ONCE(!folio)) { + if (WARN_ON_ONCE(ret)) { /* * Release the 1st page ref if the * folio is problematic, fail hard. */ gup_put_folio(page_folio(page), 1, foll_flags); - ret = -EFAULT; goto out; } } diff --git a/mm/huge_memory.c b/mm/huge_memory.c index 425374a..18604e4 100644 --- a/mm/huge_memory.c +++ b/mm/huge_memory.c @@ -1332,7 +1332,7 @@ struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr, if (!*pgmap) return ERR_PTR(-EFAULT); page = pfn_to_page(pfn); - ret = try_grab_page(page, flags); + ret = try_grab_page(page, 1, flags); if (ret) page = ERR_PTR(ret); diff --git a/mm/internal.h b/mm/internal.h index 2ea9a88..5305bbf 100644 --- a/mm/internal.h +++ b/mm/internal.h @@ -1227,7 +1227,7 @@ int migrate_device_coherent_page(struct page *page); * mm/gup.c */ struct folio *try_grab_folio(struct page *page, int refs, unsigned int flags); -int __must_check try_grab_page(struct page *page, unsigned int flags); +int __must_check try_grab_page(struct page *page, int refs, unsigned int flags); /* * mm/huge_memory.c