[ Upstream commit 271683bb2c ]
syzbot reported a uaf in page_pool_recycle_in_ring:
BUG: KASAN: slab-use-after-free in lock_release+0x151/0xa30 kernel/locking/lockdep.c:5862
Read of size 8 at addr ffff8880286045a0 by task syz.0.284/6943
CPU: 0 UID: 0 PID: 6943 Comm: syz.0.284 Not tainted 6.13.0-rc3-syzkaller-gdfa94ce54f41 #0
Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 09/13/2024
Call Trace:
<TASK>
__dump_stack lib/dump_stack.c:94 [inline]
dump_stack_lvl+0x241/0x360 lib/dump_stack.c:120
print_address_description mm/kasan/report.c:378 [inline]
print_report+0x169/0x550 mm/kasan/report.c:489
kasan_report+0x143/0x180 mm/kasan/report.c:602
lock_release+0x151/0xa30 kernel/locking/lockdep.c:5862
__raw_spin_unlock_bh include/linux/spinlock_api_smp.h:165 [inline]
_raw_spin_unlock_bh+0x1b/0x40 kernel/locking/spinlock.c:210
spin_unlock_bh include/linux/spinlock.h:396 [inline]
ptr_ring_produce_bh include/linux/ptr_ring.h:164 [inline]
page_pool_recycle_in_ring net/core/page_pool.c:707 [inline]
page_pool_put_unrefed_netmem+0x748/0xb00 net/core/page_pool.c:826
page_pool_put_netmem include/net/page_pool/helpers.h:323 [inline]
page_pool_put_full_netmem include/net/page_pool/helpers.h:353 [inline]
napi_pp_put_page+0x149/0x2b0 net/core/skbuff.c:1036
skb_pp_recycle net/core/skbuff.c:1047 [inline]
skb_free_head net/core/skbuff.c:1094 [inline]
skb_release_data+0x6c4/0x8a0 net/core/skbuff.c:1125
skb_release_all net/core/skbuff.c:1190 [inline]
__kfree_skb net/core/skbuff.c:1204 [inline]
sk_skb_reason_drop+0x1c9/0x380 net/core/skbuff.c:1242
kfree_skb_reason include/linux/skbuff.h:1263 [inline]
__skb_queue_purge_reason include/linux/skbuff.h:3343 [inline]
root cause is:
page_pool_recycle_in_ring
ptr_ring_produce
spin_lock(&r->producer_lock);
WRITE_ONCE(r->queue[r->producer++], ptr)
//recycle last page to pool
page_pool_release
page_pool_scrub
page_pool_empty_ring
ptr_ring_consume
page_pool_return_page //release all page
__page_pool_destroy
free_percpu(pool->recycle_stats);
free(pool) //free
spin_unlock(&r->producer_lock); //pool->ring uaf read
recycle_stat_inc(pool, ring);
page_pool can be free while page pool recycle the last page in ring.
Add producer-lock barrier to page_pool_release to prevent the page
pool from being free before all pages have been recycled.
recycle_stat_inc() is empty when CONFIG_PAGE_POOL_STATS is not
enabled, which will trigger Wempty-body build warning. Add definition
for pool stat macro to fix warning.
Suggested-by: Jakub Kicinski <kuba@kernel.org>
Link: https://lore.kernel.org/netdev/20250513083123.3514193-1-dongchenchen2@huawei.com
Fixes: ff7d6b27f8 ("page_pool: refurbish version of page_pool code")
Reported-by: syzbot+204a4382fcb3311f3858@syzkaller.appspotmail.com
Closes: https://syzkaller.appspot.com/bug?extid=204a4382fcb3311f3858
Signed-off-by: Dong Chenchen <dongchenchen2@huawei.com>
Reviewed-by: Toke Høiland-Jørgensen <toke@redhat.com>
Reviewed-by: Mina Almasry <almasrymina@google.com>
Link: https://patch.msgid.link/20250527114152.3119109-1-dongchenchen2@huawei.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit ee62ce7a1d ]
When enabling DMA mapping in page_pool, pages are kept DMA mapped until
they are released from the pool, to avoid the overhead of re-mapping the
pages every time they are used. This causes resource leaks and/or
crashes when there are pages still outstanding while the device is torn
down, because page_pool will attempt an unmap through a non-existent DMA
device on the subsequent page return.
To fix this, implement a simple tracking of outstanding DMA-mapped pages
in page pool using an xarray. This was first suggested by Mina[0], and
turns out to be fairly straight forward: We simply store pointers to
pages directly in the xarray with xa_alloc() when they are first DMA
mapped, and remove them from the array on unmap. Then, when a page pool
is torn down, it can simply walk the xarray and unmap all pages still
present there before returning, which also allows us to get rid of the
get/put_device() calls in page_pool. Using xa_cmpxchg(), no additional
synchronisation is needed, as a page will only ever be unmapped once.
To avoid having to walk the entire xarray on unmap to find the page
reference, we stash the ID assigned by xa_alloc() into the page
structure itself, using the upper bits of the pp_magic field. This
requires a couple of defines to avoid conflicting with the
POINTER_POISON_DELTA define, but this is all evaluated at compile-time,
so does not affect run-time performance. The bitmap calculations in this
patch gives the following number of bits for different architectures:
- 23 bits on 32-bit architectures
- 21 bits on PPC64 (because of the definition of ILLEGAL_POINTER_VALUE)
- 32 bits on other 64-bit architectures
Stashing a value into the unused bits of pp_magic does have the effect
that it can make the value stored there lie outside the unmappable
range (as governed by the mmap_min_addr sysctl), for architectures that
don't define ILLEGAL_POINTER_VALUE. This means that if one of the
pointers that is aliased to the pp_magic field (such as page->lru.next)
is dereferenced while the page is owned by page_pool, that could lead to
a dereference into userspace, which is a security concern. The risk of
this is mitigated by the fact that (a) we always clear pp_magic before
releasing a page from page_pool, and (b) this would need a
use-after-free bug for struct page, which can have many other risks
since page->lru.next is used as a generic list pointer in multiple
places in the kernel. As such, with this patch we take the position that
this risk is negligible in practice. For more discussion, see[1].
Since all the tracking added in this patch is performed on DMA
map/unmap, no additional code is needed in the fast path, meaning the
performance overhead of this tracking is negligible there. A
micro-benchmark shows that the total overhead of the tracking itself is
about 400 ns (39 cycles(tsc) 395.218 ns; sum for both map and unmap[2]).
Since this cost is only paid on DMA map and unmap, it seems like an
acceptable cost to fix the late unmap issue. Further optimisation can
narrow the cases where this cost is paid (for instance by eliding the
tracking when DMA map/unmap is a no-op).
The extra memory needed to track the pages is neatly encapsulated inside
xarray, which uses the 'struct xa_node' structure to track items. This
structure is 576 bytes long, with slots for 64 items, meaning that a
full node occurs only 9 bytes of overhead per slot it tracks (in
practice, it probably won't be this efficient, but in any case it should
be an acceptable overhead).
[0] https://lore.kernel.org/all/CAHS8izPg7B5DwKfSuzz-iOop_YRbk3Sd6Y4rX7KBG9DcVJcyWg@mail.gmail.com/
[1] https://lore.kernel.org/r/20250320023202.GA25514@openwall.com
[2] https://lore.kernel.org/r/ae07144c-9295-4c9d-a400-153bb689fe9e@huawei.com
Reported-by: Yonglong Liu <liuyonglong@huawei.com>
Closes: https://lore.kernel.org/r/8743264a-9700-4227-a556-5f931c720211@huawei.com
Fixes: ff7d6b27f8 ("page_pool: refurbish version of page_pool code")
Suggested-by: Mina Almasry <almasrymina@google.com>
Reviewed-by: Mina Almasry <almasrymina@google.com>
Reviewed-by: Jesper Dangaard Brouer <hawk@kernel.org>
Tested-by: Jesper Dangaard Brouer <hawk@kernel.org>
Tested-by: Qiuling Ren <qren@redhat.com>
Tested-by: Yuying Ma <yuma@redhat.com>
Tested-by: Yonglong Liu <liuyonglong@huawei.com>
Acked-by: Jesper Dangaard Brouer <hawk@kernel.org>
Signed-off-by: Toke Høiland-Jørgensen <toke@redhat.com>
Link: https://patch.msgid.link/20250409-page-pool-track-dma-v9-2-6a9ef2e0cba8@redhat.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit 43130d02ba ]
We noticed the kworker in page_pool_release_retry() was waken
up repeatedly and infinitely in production because of the
buggy driver causing the inflight less than 0 and warning
us in page_pool_inflight()[1].
Since the inflight value goes negative, it means we should
not expect the whole page_pool to get back to work normally.
This patch mitigates the adverse effect by not rescheduling
the kworker when detecting the inflight negative in
page_pool_release_retry().
[1]
[Mon Feb 10 20:36:11 2025] ------------[ cut here ]------------
[Mon Feb 10 20:36:11 2025] Negative(-51446) inflight packet-pages
...
[Mon Feb 10 20:36:11 2025] Call Trace:
[Mon Feb 10 20:36:11 2025] page_pool_release_retry+0x23/0x70
[Mon Feb 10 20:36:11 2025] process_one_work+0x1b1/0x370
[Mon Feb 10 20:36:11 2025] worker_thread+0x37/0x3a0
[Mon Feb 10 20:36:11 2025] kthread+0x11a/0x140
[Mon Feb 10 20:36:11 2025] ? process_one_work+0x370/0x370
[Mon Feb 10 20:36:11 2025] ? __kthread_cancel_work+0x40/0x40
[Mon Feb 10 20:36:11 2025] ret_from_fork+0x35/0x40
[Mon Feb 10 20:36:11 2025] ---[ end trace ebffe800f33e7e34 ]---
Note: before this patch, the above calltrace would flood the
dmesg due to repeated reschedule of release_dw kworker.
Signed-off-by: Jason Xing <kerneljasonxing@gmail.com>
Reviewed-by: Mina Almasry <almasrymina@google.com>
Link: https://patch.msgid.link/20250214064250.85987-1-kerneljasonxing@gmail.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Signed-off-by: Sasha Levin <sashal@kernel.org>
Implement a memory provider that allocates dmabuf devmem in the form of
net_iov.
The provider receives a reference to the struct netdev_dmabuf_binding
via the pool->mp_priv pointer. The driver needs to set this pointer for
the provider in the net_iov.
The provider obtains a reference on the netdev_dmabuf_binding which
guarantees the binding and the underlying mapping remains alive until
the provider is destroyed.
Usage of PP_FLAG_DMA_MAP is required for this memory provide such that
the page_pool can provide the driver with the dma-addrs of the devmem.
Support for PP_FLAG_DMA_SYNC_DEV is omitted for simplicity & p.order !=
0.
Signed-off-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: Kaiyuan Zhang <kaiyuanz@google.com>
Signed-off-by: Mina Almasry <almasrymina@google.com>
Reviewed-by: Pavel Begunkov <asml.silence@gmail.com>
Reviewed-by: Jakub Kicinski <kuba@kernel.org>
Link: https://patch.msgid.link/20240910171458.219195-7-almasrymina@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Convert netmem to be a union of struct page and struct netmem. Overload
the LSB of struct netmem* to indicate that it's a net_iov, otherwise
it's a page.
Currently these entries in struct page are rented by the page_pool and
used exclusively by the net stack:
struct {
unsigned long pp_magic;
struct page_pool *pp;
unsigned long _pp_mapping_pad;
unsigned long dma_addr;
atomic_long_t pp_ref_count;
};
Mirror these (and only these) entries into struct net_iov and implement
netmem helpers that can access these common fields regardless of
whether the underlying type is page or net_iov.
Implement checks for net_iov in netmem helpers which delegate to mm
APIs, to ensure net_iov are never passed to the mm stack.
Signed-off-by: Mina Almasry <almasrymina@google.com>
Reviewed-by: Pavel Begunkov <asml.silence@gmail.com>
Acked-by: Jakub Kicinski <kuba@kernel.org>
Link: https://patch.msgid.link/20240910171458.219195-6-almasrymina@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Instead of doing __cacheline_group_begin() __aligned(), use the new
__cacheline_group_{begin,end}_aligned(), so that it will take care
of the group alignment itself.
Also replace open-coded `4 * sizeof(long)` in two places with
a definition.
Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>
Signed-off-by: Tony Nguyen <anthony.l.nguyen@intel.com>
Abstract the memory type from the page_pool so we can later add support
for new memory types. Convert the page_pool to use the new netmem type
abstraction, rather than use struct page directly.
As of this patch the netmem type is a no-op abstraction: it's always a
struct page underneath. All the page pool internals are converted to
use struct netmem instead of struct page, and the page pool now exports
2 APIs:
1. The existing struct page API.
2. The new struct netmem API.
Keeping the existing API is transitional; we do not want to refactor all
the current drivers using the page pool at once.
The netmem abstraction is currently a no-op. The page_pool uses
page_to_netmem() to convert allocated pages to netmem, and uses
netmem_to_page() to convert the netmem back to pages to pass to mm APIs,
Follow up patches to this series add non-paged netmem support to the
page_pool. This change is factored out on its own to limit the code
churn to this 1 patch, for ease of code review.
Signed-off-by: Mina Almasry <almasrymina@google.com>
Reviewed-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://patch.msgid.link/20240628003253.1694510-6-almasrymina@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
56ef27e3 unexported page_pool_unlink_napi() and renamed it to
page_pool_disable_direct_recycling(). This is because there was no
in-tree user of page_pool_unlink_napi().
Since then Rx queue API and an implementation in bnxt got merged. In the
bnxt implementation, it broadly follows the following steps: allocate
new queue memory + page pool, stop old rx queue, swap, then destroy old
queue memory + page pool.
The existing NAPI instance is re-used so when the old page pool that is
no longer used but still linked to this shared NAPI instance is
destroyed, it will trigger warnings.
In my initial patches I unlinked a page pool from a NAPI instance
directly. Instead, export page_pool_disable_direct_recycling() and call
that instead to avoid having a driver touch a core struct.
Suggested-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: David Wei <dw@davidwei.uk>
Reviewed-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Having an OR in WARN_ON() makes me sad because it's impossible to tell
which condition is true when triggered.
Split a WARN_ON() with an OR in page_pool_disable_direct_recycling().
Signed-off-by: David Wei <dw@davidwei.uk>
Reviewed-by: Mina Almasry <almasrymina@google.com>
Reviewed-by: Somnath Kotur <somnath.kotur@broadcom.com>
Acked-by: Jesper Dangaard Brouer <hawk@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Pull dma-mapping updates from Christoph Hellwig:
- optimize DMA sync calls when they are no-ops (Alexander Lobakin)
- fix swiotlb padding for untrusted devices (Michael Kelley)
- add documentation for swiotb (Michael Kelley)
* tag 'dma-mapping-6.10-2024-05-20' of git://git.infradead.org/users/hch/dma-mapping:
dma: fix DMA sync for drivers not calling dma_set_mask*()
xsk: use generic DMA sync shortcut instead of a custom one
page_pool: check for DMA sync shortcut earlier
page_pool: don't use driver-set flags field directly
page_pool: make sure frag API fields don't span between cachelines
iommu/dma: avoid expensive indirect calls for sync operations
dma: avoid redundant calls for sync operations
dma: compile-out DMA sync op calls when not used
iommu/dma: fix zeroing of bounce buffer padding used by untrusted devices
swiotlb: remove alloc_size argument to swiotlb_tbl_map_single()
Documentation/core-api: add swiotlb documentation
We can save a couple more function calls in the Page Pool code if we
check for dma_need_sync() earlier, just when we test pp->p.dma_sync.
Move both these checks into an inline wrapper and call the PP wrapper
over the generic DMA sync function only when both are true.
You can't cache the result of dma_need_sync() in &page_pool, as it may
change anytime if an SWIOTLB buffer is allocated or mapped.
Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
page_pool::p is driver-defined params, copied directly from the
structure passed to page_pool_create(). The structure isn't meant
to be modified by the Page Pool core code and this even might look
confusing[0][1].
In order to be able to alter some flags, let's define our own, internal
fields the same way as the already existing one (::has_init_callback).
They are defined as bits in the driver-set params, leave them so here
as well, to not waste byte-per-bit or so. Almost 30 bits are still free
for future extensions.
We could've defined only new flags here or only the ones we may need
to alter, but checking some flags in one place while others in another
doesn't sound convenient or intuitive. ::flags passed by the driver can
now go to the "slow" PP params.
Suggested-by: Jakub Kicinski <kuba@kernel.org>
Link[0]: https://lore.kernel.org/netdev/20230703133207.4f0c54ce@kernel.org
Suggested-by: Alexander Duyck <alexanderduyck@fb.com>
Link[1]: https://lore.kernel.org/netdev/CAKgT0UfZCGnWgOH96E4GV3ZP6LLbROHM7SHE8NKwq+exX+Gk_Q@mail.gmail.com
Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
After commit 5027ec19f1 ("net: page_pool: split the page_pool_params
into fast and slow") that made &page_pool contain only "hot" params at
the start, cacheline boundary chops frag API fields group in the middle
again.
To not bother with this each time fast params get expanded or shrunk,
let's just align them to `4 * sizeof(long)`, the closest upper pow-2 to
their actual size (2 longs + 1 int). This ensures 16-byte alignment for
the 32-bit architectures and 32-byte alignment for the 64-bit ones,
excluding unnecessary false-sharing.
::page_state_hold_cnt is used quite intensively on hotpath no matter if
frag API is used, so move it to the newly created hole in the first
cacheline.
Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
There are several functions taking pointers to data they don't modify.
This includes statistics fetching, page and page_pool parameters, etc.
Constify the pointers, so that call sites will be able to pass const
pointers as well.
No functional changes, no visible changes in functions sizes.
Reviewed-by: Ilias Apalodimas <ilias.apalodimas@linaro.org>
Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>
Signed-off-by: Tony Nguyen <anthony.l.nguyen@intel.com>
Now that the checks for direct recycling possibility live inside the
Page Pool core, reuse them when performing bulk recycling.
page_pool_put_page_bulk() can be called from process context as well,
page_pool_napi_local() takes care of this at the very beginning.
Under high .ndo_xdp_xmit() traffic load, the win is 2-3% Pps assuming
the sending driver uses xdp_return_frame_bulk() on Tx completion.
Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>
Link: https://lore.kernel.org/r/20240329165507.3240110-3-aleksander.lobakin@intel.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Since we have pool->p.napi (Jakub) and pool->cpuid (Lorenzo) to check
whether it's safe to use direct recycling, we can use both globally for
each page instead of relying solely on @allow_direct argument.
Let's assume that @allow_direct means "I'm sure it's local, don't waste
time rechecking this" and when it's false, try the mentioned params to
still recycle the page directly. If neither is true, we'll lose some
CPU cycles, but then it surely won't be hotpath. On the other hand,
paths where it's possible to use direct cache, but not possible to
safely set @allow_direct, will benefit from this move.
The whole propagation of @napi_safe through a dozen of skb freeing
functions can now go away, which saves us some stack space.
Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>
Link: https://lore.kernel.org/r/20240329165507.3240110-2-aleksander.lobakin@intel.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Now that direct recycling is performed basing on pool->cpuid when set,
memory leaks are possible:
1. A pool is destroyed.
2. Alloc cache is emptied (it's done only once).
3. pool->cpuid is still set.
4. napi_pp_put_page() does direct recycling basing on pool->cpuid.
5. Now alloc cache is not empty, but it won't ever be freed.
In order to avoid that, rewrite pool->cpuid to -1 when unlinking NAPI to
make sure no direct recycling will be possible after emptying the cache.
This involves a bit of overhead as pool->cpuid now must be accessed
via READ_ONCE() to avoid partial reads.
Rename page_pool_unlink_napi() -> page_pool_disable_direct_recycling()
to reflect what it actually does and unexport it.
Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>
Reviewed-by: Toke Høiland-Jørgensen <toke@redhat.com>
Link: https://lore.kernel.org/r/20240215113905.96817-1-aleksander.lobakin@intel.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Up to now, we were only subtracting from the number of used page fragments
to figure out when a page could be freed or recycled. A following patch
introduces support for multiple users referencing the same fragment. So
reduce the initial page fragments value to half to avoid overflowing.
Signed-off-by: Liang Chen <liangchen.linux@gmail.com>
Reviewed-by: Yunsheng Lin <linyunsheng@huawei.com>
Reviewed-by: Mina Almasry <almarsymina@google.com>
Reviewed-by: Ilias Apalodimas <ilias.apalodimas@linaro.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Releasing the DMA mapping will be useful for other types
of pages, so factor it out. Make sure compiler inlines it,
to avoid any regressions.
Signed-off-by: Mina Almasry <almasrymina@google.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Reviewed-by: Ilias Apalodimas <ilias.apalodimas@linaro.org>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Mute the periodic "stalled pool shutdown" warning if the page pool
is visible to user space. Rolling out a driver using page pools
to just a few hundred hosts at Meta surfaces applications which
fail to reap their broken sockets. Obviously it's best if the
applications are fixed, but we don't generally print warnings
for application resource leaks. Admins can now depend on the
netlink interface for getting page pool info to detect buggy
apps.
While at it throw in the ID of the pool into the message,
in rare cases (pools from destroyed netns) this will make
finding the pool with a debugger easier.
Reviewed-by: Eric Dumazet <edumazet@google.com>
Acked-by: Jesper Dangaard Brouer <hawk@kernel.org>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Report when page pool was destroyed. Together with the inflight
/ memory use reporting this can serve as a replacement for the
warning about leaked page pools we currently print to dmesg.
Example output for a fake leaked page pool using some hacks
in netdevsim (one "live" pool, and one "leaked" on the same dev):
$ ./cli.py --no-schema --spec netlink/specs/netdev.yaml \
--dump page-pool-get
[{'id': 2, 'ifindex': 3},
{'id': 1, 'ifindex': 3, 'destroyed': 133, 'inflight': 1}]
Tested-by: Dragos Tatulea <dtatulea@nvidia.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Acked-by: Jesper Dangaard Brouer <hawk@kernel.org>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Advanced deployments need the ability to check memory use
of various system components. It makes it possible to make informed
decisions about memory allocation and to find regressions and leaks.
Report memory use of page pools. Report both number of references
and bytes held.
Acked-by: Jesper Dangaard Brouer <hawk@kernel.org>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
struct page_pool is rather performance critical and we use
16B of the first cache line to store 2 pointers used only
by test code. Future patches will add more informational
(non-fast path) attributes.
It's convenient for the user of the API to not have to worry
which fields are fast and which are slow path. Use struct
groups to split the params into the two categories internally.
Acked-by: Jesper Dangaard Brouer <hawk@kernel.org>
Reviewed-by: Mina Almasry <almasrymina@google.com>
Reviewed-by: Ilias Apalodimas <ilias.apalodimas@linaro.org>
Link: https://lore.kernel.org/r/20231121000048.789613-2-kuba@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Currently page pool supports the below use cases:
use case 1: allocate page without page splitting using
page_pool_alloc_pages() API if the driver knows
that the memory it need is always bigger than
half of the page allocated from page pool.
use case 2: allocate page frag with page splitting using
page_pool_alloc_frag() API if the driver knows
that the memory it need is always smaller than
or equal to the half of the page allocated from
page pool.
There is emerging use case [1] & [2] that is a mix of the
above two case: the driver doesn't know the size of memory it
need beforehand, so the driver may use something like below to
allocate memory with least memory utilization and performance
penalty:
if (size << 1 > max_size)
page = page_pool_alloc_pages();
else
page = page_pool_alloc_frag();
To avoid the driver doing something like above, add the
page_pool_alloc() API to support the above use case, and update
the true size of memory that is acctually allocated by updating
'*size' back to the driver in order to avoid exacerbating
truesize underestimate problem.
Rename page_pool_free() which is used in the destroy process to
__page_pool_destroy() to avoid confusion with the newly added
API.
1. https://lore.kernel.org/all/d3ae6bd3537fbce379382ac6a42f67e22f27ece2.1683896626.git.lorenzo@kernel.org/
2. https://lore.kernel.org/all/20230526054621.18371-3-liangchen.linux@gmail.com/
Signed-off-by: Yunsheng Lin <linyunsheng@huawei.com>
CC: Lorenzo Bianconi <lorenzo@kernel.org>
CC: Alexander Duyck <alexander.duyck@gmail.com>
CC: Liang Chen <liangchen.linux@gmail.com>
CC: Alexander Lobakin <aleksander.lobakin@intel.com>
Link: https://lore.kernel.org/r/20231020095952.11055-4-linyunsheng@huawei.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Currently when page_pool_create() is called with
PP_FLAG_PAGE_FRAG flag, page_pool_alloc_pages() is only
allowed to be called under the below constraints:
1. page_pool_fragment_page() need to be called to setup
page->pp_frag_count immediately.
2. page_pool_defrag_page() often need to be called to drain
the page->pp_frag_count when there is no more user will
be holding on to that page.
Those constraints exist in order to support a page to be
split into multi fragments.
And those constraints have some overhead because of the
cache line dirtying/bouncing and atomic update.
Those constraints are unavoidable for case when we need a
page to be split into more than one fragment, but there is
also case that we want to avoid the above constraints and
their overhead when a page can't be split as it can only
hold a fragment as requested by user, depending on different
use cases:
use case 1: allocate page without page splitting.
use case 2: allocate page with page splitting.
use case 3: allocate page with or without page splitting
depending on the fragment size.
Currently page pool only provide page_pool_alloc_pages() and
page_pool_alloc_frag() API to enable the 1 & 2 separately,
so we can not use a combination of 1 & 2 to enable 3, it is
not possible yet because of the per page_pool flag
PP_FLAG_PAGE_FRAG.
So in order to allow allocating unsplit page without the
overhead of split page while still allow allocating split
page we need to remove the per page_pool flag in
page_pool_is_last_frag(), as best as I can think of, it seems
there are two methods as below:
1. Add per page flag/bit to indicate a page is split or
not, which means we might need to update that flag/bit
everytime the page is recycled, dirtying the cache line
of 'struct page' for use case 1.
2. Unify the page->pp_frag_count handling for both split and
unsplit page by assuming all pages in the page pool is split
into a big fragment initially.
As page pool already supports use case 1 without dirtying the
cache line of 'struct page' whenever a page is recyclable, we
need to support the above use case 3 with minimal overhead,
especially not adding any noticeable overhead for use case 1,
and we are already doing an optimization by not updating
pp_frag_count in page_pool_defrag_page() for the last fragment
user, this patch chooses to unify the pp_frag_count handling
to support the above use case 3.
There is no noticeable performance degradation and some
justification for unifying the frag_count handling with this
patch applied using a micro-benchmark testing in [1].
1. https://lore.kernel.org/all/bf2591f8-7b3c-4480-bb2c-31dc9da1d6ac@huawei.com/
Signed-off-by: Yunsheng Lin <linyunsheng@huawei.com>
CC: Lorenzo Bianconi <lorenzo@kernel.org>
CC: Alexander Duyck <alexander.duyck@gmail.com>
CC: Liang Chen <liangchen.linux@gmail.com>
CC: Alexander Lobakin <aleksander.lobakin@intel.com>
Link: https://lore.kernel.org/r/20231020095952.11055-2-linyunsheng@huawei.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Currently, touching <net/page_pool/types.h> triggers a rebuild of more
than half of the kernel. That's because it's included in
<linux/skbuff.h>. And each new include to page_pool/types.h adds more
[useless] data for the toolchain to process per each source file from
that pile.
In commit 6a5bcd84e8 ("page_pool: Allow drivers to hint on SKB
recycling"), Matteo included it to be able to call a couple of functions
defined there. Then, in commit 57f05bc2ab ("page_pool: keep pp info as
long as page pool owns the page") one of the calls was removed, so only
one was left. It's the call to page_pool_return_skb_page() in
napi_frag_unref(). The function is external and doesn't have any
dependencies. Having very niche page_pool_types.h included only for that
looks like an overkill.
As %PP_SIGNATURE is not local to page_pool.c (was only in the
early submissions), nothing holds this function there. Teleport
page_pool_return_skb_page() to skbuff.c, just next to the main consumer,
skb_pp_recycle(), and rename it to napi_pp_put_page(), as it doesn't
work with skbs at all and the former name tells nothing. The #if guards
here are only to not compile and have it in the vmlinux when not needed
-- both call sites are already guarded.
Now, touching page_pool_types.h only triggers rebuilding of the drivers
using it and a couple of core networking files.
Suggested-by: Jakub Kicinski <kuba@kernel.org> # make skbuff.h less heavy
Suggested-by: Alexander Duyck <alexanderduyck@fb.com> # move to skbuff.c
Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>
Reviewed-by: Alexander Duyck <alexanderduyck@fb.com>
Link: https://lore.kernel.org/r/20230804180529.2483231-3-aleksander.lobakin@intel.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
All struct members of the driver-facing APIs are documented twice,
in the code and under Documentation. This is a bit tedious.
I also get the feeling that a lot of developers will read the header
when coding, rather than the doc. Bring the two a little closer
together by using kdoc for structs and functions.
Using kdoc also gives us links (mentioning a function or struct
in the text gets replaced by a link to its doc).
Reviewed-by: Randy Dunlap <rdunlap@infradead.org>
Tested-by: Randy Dunlap <rdunlap@infradead.org>
Acked-by: Jesper Dangaard Brouer <hawk@kernel.org>
Link: https://lore.kernel.org/r/20230802161821.3621985-3-kuba@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
page_pool_ring_[un]lock() use in_softirq() to decide which
spin lock variant to use, and when they are called in the
context with in_softirq() being false, spin_lock_bh() is
called in page_pool_ring_lock() while spin_unlock() is
called in page_pool_ring_unlock(), because spin_lock_bh()
has disabled the softirq in page_pool_ring_lock(), which
causes inconsistency for spin lock pair calling.
This patch fixes it by returning in_softirq state from
page_pool_producer_lock(), and use it to decide which
spin lock variant to use in page_pool_producer_unlock().
As pool->ring has both producer and consumer lock, so
rename it to page_pool_producer_[un]lock() to reflect
the actual usage. Also move them to page_pool.c as they
are only used there, and remove the 'inline' as the
compiler may have better idea to do inlining or not.
Fixes: 7886244736 ("net: page_pool: Add bulk support for ptr_ring")
Signed-off-by: Yunsheng Lin <linyunsheng@huawei.com>
Acked-by: Jesper Dangaard Brouer <brouer@redhat.com>
Acked-by: Ilias Apalodimas <ilias.apalodimas@linaro.org>
Link: https://lore.kernel.org/r/20230522031714.5089-1-linyunsheng@huawei.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Jesper points out that we must prevent recycling into cache
after page_pool_destroy() is called, because page_pool_destroy()
is not synchronized with recycling (some pages may still be
outstanding when destroy() gets called).
I assumed this will not happen because NAPI can't be scheduled
if its page pool is being destroyed. But I missed the fact that
NAPI may get reused. For instance when user changes ring configuration
driver may allocate a new page pool, stop NAPI, swap, start NAPI,
and then destroy the old pool. The NAPI is running so old page
pool will think it can recycle to the cache, but the consumer
at that point is the destroy() path, not NAPI.
To avoid extra synchronization let the drivers do "unlinking"
during the "swap" stage while NAPI is indeed disabled.
Fixes: 8c48eea3ad ("page_pool: allow caching from safely localized NAPI")
Reported-by: Jesper Dangaard Brouer <jbrouer@redhat.com>
Link: https://lore.kernel.org/all/e8df2654-6a5b-3c92-489d-2fe5e444135f@redhat.com/
Acked-by: Jesper Dangaard Brouer <brouer@redhat.com>
Link: https://lore.kernel.org/r/20230419182006.719923-1-kuba@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Commit c519fe9a4f ("bnxt: add dma mapping attributes") added
DMA_ATTR_WEAK_ORDERING to DMA attrs on bnxt. It has since spread
to a few more drivers (possibly as a copy'n'paste).
DMA_ATTR_WEAK_ORDERING only seems to matter on Sparc and PowerPC/cell,
the rarity of these platforms is likely why we never bothered adding
the attribute in the page pool, even though it should be safe to add.
To make the page pool migration in drivers which set this flag less
of a risk (of regressing the precious sparc database workloads or
whatever needed this) let's add DMA_ATTR_WEAK_ORDERING on all
page pool DMA mappings.
We could make this a driver opt-in but frankly I don't think it's
worth complicating the API. I can't think of a reason why device
accesses to packet memory would have to be ordered.
Acked-by: Ilias Apalodimas <ilias.apalodimas@linaro.org>
Acked-by: Somnath Kotur <somnath.kotur@broadcom.com>
Reviewed-by: Alexander Lobakin <aleksander.lobakin@intel.com>
Acked-by: Jesper Dangaard Brouer <brouer@redhat.com>
Link: https://lore.kernel.org/r/20230417152805.331865-1-kuba@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Recent patches to mlx5 mentioned a regression when moving from
driver local page pool to only using the generic page pool code.
Page pool has two recycling paths (1) direct one, which runs in
safe NAPI context (basically consumer context, so producing
can be lockless); and (2) via a ptr_ring, which takes a spin
lock because the freeing can happen from any CPU; producer
and consumer may run concurrently.
Since the page pool code was added, Eric introduced a revised version
of deferred skb freeing. TCP skbs are now usually returned to the CPU
which allocated them, and freed in softirq context. This places the
freeing (producing of pages back to the pool) enticingly close to
the allocation (consumer).
If we can prove that we're freeing in the same softirq context in which
the consumer NAPI will run - lockless use of the cache is perfectly fine,
no need for the lock.
Let drivers link the page pool to a NAPI instance. If the NAPI instance
is scheduled on the same CPU on which we're freeing - place the pages
in the direct cache.
With that and patched bnxt (XDP enabled to engage the page pool, sigh,
bnxt really needs page pool work :() I see a 2.6% perf boost with
a TCP stream test (app on a different physical core than softirq).
The CPU use of relevant functions decreases as expected:
page_pool_refill_alloc_cache 1.17% -> 0%
_raw_spin_lock 2.41% -> 0.98%
Only consider lockless path to be safe when NAPI is scheduled
- in practice this should cover majority if not all of steady state
workloads. It's usually the NAPI kicking in that causes the skb flush.
The main case we'll miss out on is when application runs on the same
CPU as NAPI. In that case we don't use the deferred skb free path.
Reviewed-by: Tariq Toukan <tariqt@nvidia.com>
Acked-by: Jesper Dangaard Brouer <brouer@redhat.com>
Tested-by: Dragos Tatulea <dtatulea@nvidia.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
